the university of queensland

THE UNIVERSITY OF QUEENSLAND AUSTRALIA

ESSAYS ON HOSPITAL EFFICIENCY MEASUREMENT

Kim-Huong Nguyen

A thesis submitted for the degree of Doctor of Philosophy at The University of Queensland in September 2010 School of Economics

Declaration by author This thesis is composed of my original work, and contains no material previously published or written by another person except where due reference has been made in the text. I have clearly stated the contribution by others to jointly-authored works that I have included in my thesis. I have clearly stated the contribution of others to my thesis as a whole, including statistical assistance, survey design, data analysis, significant technical procedures, professional editorial advice, and any other original research work used or reported in my thesis. The content of my thesis is the result of work I have carried out since the commencement of my research higher degree candidature and does not include a substantial part of work that has been submitted to qualify for the award of any other degree or diploma in any university or other tertiary institution. I have clearly stated which parts of my thesis, if any, have been submitted to qualify for another award. I acknowledge that an electronic copy of my thesis must be lodged with the University Library and, subject to the General Award Rules of The University of Queensland, immediately made available for research and study in accordance with the Copyright Act 1968. I acknowledge that copyright of all material contained in my thesis resides with the copyright holder(s) of that material.

Kim-Huong Nguyen

Statement of Contributions to Jointly Authored Works Contained in the Thesis

I hereby declare that the work contained within the results presented in this thesis are largely my own. Any contribution made to the research by colleagues, with whom I have worked at the University of Queensland or elsewhere, during my candidature, is fully acknowledged.

Statement of Contributions by Others to the Thesis as a Whole

No contributions by others.

Statement of Parts of the Thesis Submitted to Qualify for the Award of Another Degree

None.

ii

Publications

These published works are incorporated into the Thesis: 1. K.H. Nguyen & T. Coelli (2009). ‘Quantifying the effects of modelling choices on hospital efficiency measures: A meta-regression analysis’. CEPA Working paper series WP07/2009 and Conference Paper at the 1st Australian Workshop on Econometrics and Health Economics, Monash, 7th-9th April 2010. Incorporated as Chapter 2. 2. K.H. Nguyen & G. Lordan (2010). ‘Labour efficiency of Queensland public hospitals under the historical cost payment’. Submitting to the International Journal of Health Care Finance and Economics. Incorporated as Chapter 3. 3. K.H. Nguyen & D.S.P. Rao (2010). ‘The impact of financial autonomisation reform on the productive efficiency of public hospitals Vietnam’. Conference Paper at the 2nd Vietnam Economic Workshop, Australian National University, 2nd March 2010. Incorporated as Chapter 4. Additional Published Works by the Author Relevant to the Thesis but not Forming Part of it

None.

iii

Acknowledgment

It takes a long time to write a PhD thesis. For me it was (estimated) 13,390 hours 52 minutes 11 seconds of work. I would here like to express my thanks to those people who have spent their time supporting me and shared their knowledge to help me complete my thesis with the best possible result. I have been very fortunate with my supervisors. Professor Prasada Rao, Professor Tim Coelli and Dr. Grace Lordan were very good supervisors, who complemented each other wonderfully well. Prasada proposed a research topic and gave me the money that allowed me to hit the ground running, which is always a nice way to start. His perspectives and guidances were excellent. Tim, as my daily supervisor for two years, always brought a thorough eye and great enthusiasm. He also taught me the use of drawing pictures, and recently good alcohol in comprehending a difficult topics. Grace provided refreshing insights, critical evaluations and friendly lunches. I cannot thank them enough. For this research, data are essential. I’ve collected a lot of data. Many people helped with this, for which I would like to thank them wholeheartedly. Among those are Son Nghiem, Sarah Bales and Hoang Van Minh. I’m very grateful to many academic staff at the School of Economics, amongst whom are Chris O’Donnell and KK Tang, for their friendship, encouragement and insightful comments. Special thanks are due to Richard who has been a friend and mentor, a tremendous source of support throughout my graduate education. I have been fortunate to come across many funny and good friends, without whom life would be bleak. Big thanks to my PhD friends, Andrew, Eliana, Lebo, Leonora, Nick, and Sriram who are always funny, and more than anything, fantastically normal, for crisis supports. Thanks to all my friends for not always taking me too seriously. I also wish to thank the School of Economics for enjoyable free lunches, dinners, conferences, and above all a very friendly working environment. Warm thanks to my Australian family for their kindness and for never letting me feel that I am away from my parents: aunty Bridget and uncle Jervis, my defactor mother and fathers in-law Jane, David and Alan. I wish to express heartfelt gratitude to my boyfriend, Sam. Both of us have suffered from each other’s project iv

from the inception of our relationship, yet seen it grows and mature. I thank Sammy, most of all, for constantly reminding me that there are more important things in life than a PhD thesis, namely going to the beach and getting socially excited. It is a rare experience. I suggest everyone should try it (at least) once! I also thank Dori in Finding Nemo whose quote has become my mantra: “Just keep swimming, just keep swimming”. Perspective is, after all, everything. Finally, I devote my thesis to my loving parents - Nguyen Hung and Tran Thi Kim Loan, and my maternal grandfather - Tran Van Huan, who have always believed, since I was five years old, that one day I would be a “Tien Si” (PhD). I have done my part, now let’s see if the readers like it!

v

Abstract

The overall objective of this thesis is to contribute to the existing knowledge of efficiency measurements through the exploration of efficiency modelling issues and practical applications of different efficiency analysis frameworks. More specifically, this objective will be achieved through a collection of three essays, each of which deals with a distinctive topic. However, the umbrella theme uniting them is measuring productivity and efficiency of hospitals.

Quantifying the effects of modelling choices on hospital efficiency measures using meta-regression analysis Motivation and objective: This study recognises the sensitivity of efficiency estimates discussed in the literature and attempts to quantity the effects of modelling choices on estimated efficiencies using the meta-regression technique. Methodology and data: The meta-regression technique is used to quantify the degree to which modelling factors influence efficiency estimates. Modelling factors included in the regression are: sample size; dimension (number of variables); parametric versus non-parametric method; returns to scale (RTS) assumption; functional form; error distributional form; input versus output orientation; cost versus technical efficiency measure; and cross-sectional versus panel data. The data set is derived from 253 estimated models reported in 95 empirical analyses of hospital efficiency in the 22-year period from 1987 to 2008. Results and contributions: Sample size, dimension and RTS are found to have statistically significant effects at the 1% level. Sample size has a negative (and diminishing) effect on mean estimated efficiency; dimension has a positive (and diminishing) effect; while the imposition of constant returns to scale has a negative effect. This study is valuable contribution in that it quantifies the potential variation of efficiency estimates as a result of different model specifications and estimation approaches. The results can be used in improving the policy relevance of the empirical results produced by hospital efficiency studies.

vi

Measuring labour efficiency in Queensland public hospitals Motivation and objectives: This chapter attempts to measure efficiencies of public hospitals in Queensland with respect to medical labour usage for the period 1996-2003. It is motivated by the fact that the hospital sector is a labour intensive industry; therefore the improvements in productivity and efficiency are likely to come from the better use of human resources. Methodology and data: A labour input requirement frontier is estimated under the “true” random effect framework with the inclusion of selected control variables to capture the observed heterogeneity. Sensitivity analysis is performed on alternative model specification and estimation strategies. The data consists of 84 public hospitals in Queensland, and cover the period of 8 financial years. Results and contributions: The results suggest that inpatient outputs (represented by weighted episodes of surgical and medical care) greatly influence the labour requirement. The paper finds evidence to support the hypothesis that Queensland Health public hospitals are on average operating at a sub-optimal scale, and that teaching and principal referral hospitals tend to consume more labour resources. Labour requirements in city and town hospitals were also higher than shire and island ones. Hospitals located in the areas with large Indigenous population do not appear to require more medical labour resources. Apart from the policy implications of scale-inefficiencies and the evolution of labout efficiency during the studied period, the study contributes to the literature in being the first study examining labour efficiency in hospital sector under the latest generation of frontier model in a panel data framework.

Funding reforms and public hospital efficiency in Vietnam Motivation and objectives: Financing and provision of health services in Vietnam has been gradually privatised in the last two decades or so under the belief that a market-based allocation of health resources would deliver more efficient outcomes. However, market failure is a prominent characteristic of the health care market and there remain questions about the benefit of reforms, especially on efficiency improvement. This chapter attempts to evaluate the impact of financial autonomisation - one of the most recent reform initiatives that turns public service units into quasi-corporates - on productive efficiency of public hospitals in Vietnam between the two periods 1998-2000 (pre-reform) and 2005-2007 (post-reform). vii

Methodology and data: The meta-frontier framework is used to investigate the evolution of technical efficiencies and meta-technology ratios of different geographic and economic regions. The estimation is conducted using the Data Envelopment Analysis method on a dataset of 62 hospitals over 6 years, covering the two periods of pre-reform (1998-2000) and post-reform (2005-2007). Results and contributions: It is evident that the reform has created a more favourable operating environment for public hospitals in some regions of low poverty rates while the poorest regions appear to miss out on the reform benefits. Furthermore, a reduction in technical efficiency is observed overall, which implies that the gain from reform is not sufficient to outweigh the loss. Its contribution lies in the novel application of the recently developed meta-frontier framework to investigate effects of reform. This study also carries important policy implications for future reforms. It suggests that the reform has failed to improve productive efficiency of the public hospital sector while evidently sacrificing the objective of equity in health care.

viii

Keywords

health economics, econometrics, hospital, productivity analysis, efficiency measurement, stochastic frontier analysis, data envelopment analysis, queensland, vietnam.

ix

Australian and New Zealand Standard Research Classifications (ANZSRC)

140304 Panel Data Analysis 40% 140208 Health Economics 40% 140301 Cross-Sectional Analysis 20%

x

Contents

1 Introduction

1

1.1

Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1

1.2

Efficiency measurement for health care system improvement . . . . .

4

1.2.1

Policies and performance evaluation . . . . . . . . . . . . . . .

4

1.2.2

Efficiency measurement projects around the world . . . . . . .

7

1.2.3

Challenges in measuring efficiency . . . . . . . . . . . . . . . .

9

1.3

Objectives and organisation of the thesis . . . . . . . . . . . . . . . . 11

2 Quantifying the effects of modelling choices on hospital efficiency measures using meta-regression analysis 16 2.1

2.2

Modelling choices and efficiency estimates . . . . . . . . . . . . . . . 18 2.1.1

Variable choice and sample size . . . . . . . . . . . . . . . . . 18

2.1.2

Parametric versus non-parametric approaches . . . . . . . . . 24

2.1.3

Orientation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.1.4

Returns to scale and functional form . . . . . . . . . . . . . . 28

2.1.5

Summary of expected effects of modelling choice on mean efficiency estimates . . . . . . . . . . . . . . . . . . . . . . . . . 31

Methodology and data . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.2.1

Meta regression analysis . . . . . . . . . . . . . . . . . . . . . 33 xi

CONTENTS

2.2.2

Construction of the meta-dataset . . . . . . . . . . . . . . . . 35

2.2.3

Model specification . . . . . . . . . . . . . . . . . . . . . . . . 37

2.3

Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2.4

Implications for hospital efficiency studies . . . . . . . . . . . . . . . 49

2.5

Concluding remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53

3 Measuring labour efficiency in Queensland public hospitals 3.1

3.2

3.3

3.4

56

The public hospital system in Queensland . . . . . . . . . . . . . . . 57 3.1.1

Queensland Health system . . . . . . . . . . . . . . . . . . . . 57

3.1.2

Queensland public hospitals . . . . . . . . . . . . . . . . . . . 59

3.1.3

Funding models in Queensland public hospitals . . . . . . . . 60

3.1.4

Labour shortage and efficiency in Queensland Health public hospitals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 3.2.1

The input requirement function . . . . . . . . . . . . . . . . . 71

3.2.2

The “true” random effect model . . . . . . . . . . . . . . . . . 73

Data on Queensland public hospitals . . . . . . . . . . . . . . . . . . 76 3.3.1

Data collection . . . . . . . . . . . . . . . . . . . . . . . . . . 76

3.3.2

Choice of variables . . . . . . . . . . . . . . . . . . . . . . . . 79 3.3.2.1

The dependent variable . . . . . . . . . . . . . . . . 79

3.3.2.2

Output variables . . . . . . . . . . . . . . . . . . . . 80

3.3.2.3

Control variables . . . . . . . . . . . . . . . . . . . . 80

Empirical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82 3.4.1

Estimates of the “true” random effect model . . . . . . . . . . 82

3.4.2

Efficiency predictions . . . . . . . . . . . . . . . . . . . . . . . 88

xii

CONTENTS

3.5

Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

3.6


4 Funding reforms and public hospital efficiency in Vietnam 4.1

102

Vietnam’s health system and public hospitals . . . . . . . . . . . . . 104 4.1.1

Health sector financing . . . . . . . . . . . . . . . . . . . . . . 106

4.1.2

Administrative structure . . . . . . . . . . . . . . . . . . . . . 108

4.1.3

Reforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

4.1.4

Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

4.2

Research questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

4.3

The meta-frontier framework . . . . . . . . . . . . . . . . . . . . . . . 116

4.4

4.5

4.3.1

The meta-frontier . . . . . . . . . . . . . . . . . . . . . . . . . 118

4.3.2

Group frontiers . . . . . . . . . . . . . . . . . . . . . . . . . . 119

4.3.3

Technical efficiencies . . . . . . . . . . . . . . . . . . . . . . . 121

4.3.4

Meta-technology ratios . . . . . . . . . . . . . . . . . . . . . . 122

4.3.5

Evaluating the impact of reform using the meta-frontier framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 4.3.5.1

Efficiency improvement - the gain from reform . . . . 123

4.3.5.2

Efficiency changes by region - the distribution of the gain . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

Vietnamese public hospital data . . . . . . . . . . . . . . . . . . . . . 126 4.4.1

Input measures . . . . . . . . . . . . . . . . . . . . . . . . . . 128

4.4.2

Output measures . . . . . . . . . . . . . . . . . . . . . . . . . 129

Estimation method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 4.5.1

Data envelopment analysis . . . . . . . . . . . . . . . . . . . . 131

4.5.2

Estimation procedure for the meta-frontier and group frontiers 132 xiii

CONTENTS

4.6

4.7

4.5.2.1

Estimation of the meta-frontier . . . . . . . . . . . . 133

4.5.2.2

Estimation of the two reform frontiers . . . . . . . . 133

4.5.2.3

Estimation of regional (group) frontiers . . . . . . . . 134

Empirical results and discussion . . . . . . . . . . . . . . . . . . . . . 137 4.6.1

Meta technology ratios and the effect of reform . . . . . . . . 137

4.6.2

Sensitivity analysis using the bootstrap method . . . . . . . . 141


5 Conclusions

147

5.1

Thesis summary and contributions . . . . . . . . . . . . . . . . . . . 147

5.2

Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

A List of studies included in the meta-regression analysis

154

B The construction of cost weights for output aggregation for Vietnamese public hospitals 171 B.1 Aggregation issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171 B.2 Calculation of unit cost per service group . . . . . . . . . . . . . . . . 173 C The bootstrap results for estimated technical efficiency of Vietnamese public hospitals 177 D Review of provider payment methods (PPMs)

188

D.1 Classification of provider payment methods . . . . . . . . . . . . . . . 188 D.1.1 Fixed budget . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 D.1.1.1 Line item budget . . . . . . . . . . . . . . . . . . . . 191 D.1.1.2 Global budget . . . . . . . . . . . . . . . . . . . . . . 193 D.1.2 Floating budget

. . . . . . . . . . . . . . . . . . . . . . . . . 199 xiv

CONTENTS

D.1.2.1 Per diem . . . . . . . . . . . . . . . . . . . . . . . . 199 D.1.2.2 Per service (or scheduled fee) . . . . . . . . . . . . . 201 D.1.2.3 Episode of care (case-based payment) . . . . . . . . . 201 D.1.2.4 Fee for service . . . . . . . . . . . . . . . . . . . . . . 204 D.1.2.5 Performance based payment D.1.3 The mix approach

. . . . . . . . . . . . . 205

. . . . . . . . . . . . . . . . . . . . . . . . 207

D.2 Funding fooling mechanisms and provider payment methods . . . . . 209

xv

List of Figures

1.1

Total health expenditure of the world . . . . . . . . . . . . . . . . . .

2

1.2

The feedback process of performance measurement

6

2.1

Illustration of the omission and aggregation problems . . . . . . . . . 22

2.2

Illustration of increasing sample size

2.3

Illustration of parametric and non-parametric methods . . . . . . . . 25

2.4

Illustration of efficiencies produced by output versus input orientations 27

2.5

Illustration of efficiency estimates under CRS and VRS technologies . 29

2.6

Illustration of effect of functional form and returns to scale on efficiency estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

2.7

Distribution of studies by year and country . . . . . . . . . . . . . . . 38

2.8

Sample size and number of variables

2.9

Predicted mean efficiencies at the median . . . . . . . . . . . . . . . . 47

. . . . . . . . . .

. . . . . . . . . . . . . . . . . . 23

. . . . . . . . . . . . . . . . . . 40

2.10 Marginal effect of sample size and dimension on efficiency estimates . 48 3.1

Queensland Health Service Districts and their performance over time

3.2

Labour efficiency by hospital size . . . . . . . . . . . . . . . . . . . . 93

3.3

Ratio of estimated coefficients from different models with respect to the “true” random effect model . . . . . . . . . . . . . . . . . . . . . 96

xvi

89

LIST OF FIGURES

3.4

Distributions of predicted efficiencies by alternative models . . . . . . 97

3.5

Predicted labour efficiency by alternative models, by health district . 98

4.1

Child mortality rates in Vietnam and selected countries . . . . . . . . 105

4.2

Infant mortality rates by region . . . . . . . . . . . . . . . . . . . . . 105

4.3

Sources of total revenue in public hospitals (percent) . . . . . . . . . 107

4.4

Flows of health financing in Vietnam . . . . . . . . . . . . . . . . . . 108

4.5

Illustration of the meta-fronter and groups frontiers . . . . . . . . . . 121

4.6

Who benefits from the reform? . . . . . . . . . . . . . . . . . . . . . . 125

4.7

Comparison of before and after reform frontiers . . . . . . . . . . . . 126

4.8

Group definition, frontier specification, technical efficiencies and metatechnology ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

D.1 Classification of provider payment methods . . . . . . . . . . . . . . . 190 D.2 Efficiency incentives by line-item budget . . . . . . . . . . . . . . . . 192 D.3 Efficiency incentives by historical cost payment . . . . . . . . . . . . 196 D.4 Efficiency incentives by capitation method . . . . . . . . . . . . . . . 198 D.5 Efficiency incentives by per-diem reimbursement . . . . . . . . . . . . 200 D.6 Efficiency incentives by scheduled fee payment . . . . . . . . . . . . . 202 D.7 Efficiency incentives by case-based reimbursement . . . . . . . . . . . 204 D.8 Efficiency incentives by fee-for-service payment method . . . . . . . . 206 D.9 Efficiency incentives by pay-for-performance method . . . . . . . . . . 207 D.10 Matrix of provider payment methods and fund pooling mechanisms . 212

xvii

List of Tables

2.1

The expected effects of modelling choices on estimated mean efficiency 33

2.2

Variable names and definitions . . . . . . . . . . . . . . . . . . . . . . 42

2.3

Descriptive statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

2.4

Estimated results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2.5

Mean efficiencies (Linna et al., 2006) . . . . . . . . . . . . . . . . . . 52

2.6

Mean efficiencies (Grosskopf & Valdmanis, 1993) . . . . . . . . . . . . 53

2.7

Efficiency prediction and ranking . . . . . . . . . . . . . . . . . . . . 54

2.8

Efficiency predictions for developing and developed countries . . . . . 54

3.1

Summary of separations by public hospitals . . . . . . . . . . . . . . 60

3.2

Components of the old and new hospital funding models in Queensland public hospitals . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.3

Average salary (A$) of FTE staff in public hospitals, 2003-04 . . . . . 66

3.4

FTE public medical practitioners per 100,000 population . . . . . . . 67

3.5

Medical practitioners per 100,000 population (2004-05) . . . . . . . . 67

3.6

FTE nurses and other health professionals in 2003-04 . . . . . . . . . 68

3.7

Number of beds per 1,000 population in public hospitals, 2005-06 . . 69

3.8

Episodes of care in Queensland public and private hospitals . . . . . . 70

xviii

LIST OF TABLES

3.9

Summary of statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3.10 Parameter estimation for TRE . . . . . . . . . . . . . . . . . . . . . . 84 3.11 Estimates of labour elasticities . . . . . . . . . . . . . . . . . . . . . . 86 3.12 Predicted efficiency by Queensland Health Service Districts . . . . . . 90 3.13 Labour efficiency of teaching vs. non-teaching hospitals . . . . . . . . 91 3.14 Efficiency estimates by location of hospitals . . . . . . . . . . . . . . 92 3.15 Estimated coefficients by alternative models . . . . . . . . . . . . . . 95 3.16 Spearman’s correlation of predicted efficiencies . . . . . . . . . . . . . 98 4.1

Major health system reforms in Vietnam since the 1990s . . . . . . . 110

4.2

Health deflators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128

4.3

Summary of input statistics . . . . . . . . . . . . . . . . . . . . . . . 129

4.4

Summary of output statistics . . . . . . . . . . . . . . . . . . . . . . 129

4.5

Poverty rate in Vietnam by region . . . . . . . . . . . . . . . . . . . . 135

4.6

Output oriented technical efficiencies by poverty group . . . . . . . . 138

4.7

Meta-technology ratios by poverty group . . . . . . . . . . . . . . . . 139

4.8

Bootstrap results of selected hospitals (1998 and 2007) . . . . . . . . 143

4.9

Differences between original and corrected pre- and post-reform meta technology ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143

B.1 Cost weights for hospital outputs . . . . . . . . . . . . . . . . . . . . 175

xix

Glossary

A$ ACT AHCA

Australian Dollar, 57 Australian Capital Territory, 65 Australian Health Care Agreement, 57

CFM CRS

Casemix Funding Model, 61 Constant Returns to Scale, 26

DEA DGP DRG

Data Envelopment Analysis, 24 Data Generation Process, 138 Diagnosis Related Group, 7

F4S FTE

Fee For Service, 200 Full-time Equivalent, 65

GDP GP

Gross Domestic Product, 1 General Practitioners, 75

LOS

Length of Stay, 61

MO MOH MRA

Medical officer, 79 Ministry of Health, 105 Meta Regression Analysis, 33

NHCDC National Hospital Cost Data Collection, 61 NHS National Health Service, 8 NSW New South Wales, 65 NT Northern Territory, 65

xx

Glossary

OECD

Organisation of Economic Cooperation and Development, 1

P4P PHI PPM

Pay For Performance, 202 Private Health Insurance, 57 Provider Payment Method, 185

QH QLD

Queensland Health, 7 Queensland, 65

RAM

Resource Allocation Model, 61

SA SFA

South Australia, 65 Stochastic Frontier Analysis, 24

TAS TFE TRE

Tasmania, 65 “True” Fixed Effect Model, 73 “True” Random Effect Model, 73

UN US US$

United Nation, 101 United States of America, 36 US Dollar, 100

VIC VLSS VMO VND VRS

Victoria, 65 Vietnam Living Standard Survey, 110 Visiting medical officers, 75 Vietnamese Dong, 125 Variable Returns to Scale, 26

WA WHO

Western Australia, 65 World Health Organisation, 1

xxi

Chapter 1

Introduction

“Health care improvement starts from the ground up. It requires tenacious work to understand what does and what does not work in real life, and the engagement of countless providers and patients, institutions and communities.” (OECD, 2002).

1.1

Background

Health is a critical component of human well-being, a basis for a person’s ability to function and indeed, achieve anything at all. Though a precious endowment, health is often overlooked, until it deteriorates and professional help is desperately sought. Thus, the medical system occupies an essential position in society. It is not surprising then, according to the World Health Organisation (WHO), that most countries devote a large proportion of their available resources for producing health care services (as much as 8.7% of the world’s economic output in 20081 ). It is of some concern that health care costs have increased dramatically over the last few decades. Most countries have experienced increasing health expenditure both in real terms, and as a share of Gross Domestic Product (GDP). On average, while countries in the Organisation of Economic Cooperation and Development (OECD) spent around 7.7% of their GDP on health care in 1996, this share rose to around 9% ten years later. Although developing countries on average spend less than the OECD countries, in real terms as well as percentage of GDP, they also have experienced a similar trend. On average, per capita health expenditure has been growing at around 6% per annum over the period 1996-2006 (see Figure 1.1). 1

World Health Statistics Report (WHO, 2009)

1

1.1. BACKGROUND

Figure 1.1: Total health expenditure of the world

% GDP - OECD % GDP - World Per capita health expenditure (in $PPP) - OECD Per capita health expenditure (in $PPP) - World 9

1995.5

1997.5

1999.5

2001.5

2003.5

2005.5

3,500

8

2,500 2,000

7

$ PPP

Percentage of GDP

3,000

1,500 1,000

6

500 5

0

Source: WHO Statistics and OECD Health Statistics 2009.

What then, are the reasons for these significant increases in health care costs? Generally, one can identify three main causes. First, demand for health care is expanding much more rapidly than supply. Improved living standard yet a greying population (and associated chronic diseases), along with the alarming increase of diseases that relate to unhealthy lifestyle (such as obesity) have been identified as major forces pushing demand upward. In the meantime, health care supply has been struggling to keep up with demand because of resource constraints and severe regulatory restrictions on entry by providers2 . Second, the actual costs of producing health services are becoming ever more expensive due to advances in medical technology and rising wages. Rapid technological change is believed to explain a significant portion of the increase in real health care spending per capita that has occurred over the past decades. Dissemination of new technologies makes access to better services more feasible, but at the same time, encourages the development of excess capacity in different health facilities. Excess capacity can lead to overuse of these technologies, resulting in higher costs. Moreover, since health care is a labour-intensive industry, as societies become richer, 2

A detailed discussion on the impact of demand on health care costs can be found in Boutsolie (2010).

2

1.1. BACKGROUND

health services become increasingly expensive relative to those that can be produced in industry using less labour. Finally, inefficiency has been suggested as a source of cost inflation. This was succinctly summarised in Reid et al. (2005): “An estimated thirty to forty cents of every dollar spent on health care ... a half trillion dollars a year ... is spent on costs associated with: overuse, underuse, misuse, duplication, system failures ... and inefficiency.” There are many sources of inefficiency, including top-heavy medical bureaucracies, dysfunctional competition, uncoordinated delivery systems (such as insufficient attention to prevention and long-term results), lack of information on performance3 , and poorly designed financial incentive/reward mechanisms. In particular, providers lack incentives to conform to best practice and patients facing inappropriate price incentives to make prudent choices. Though far from ideal, the first and second causes of cost inflation may need to be simply accepted since we desire to improve quality and effectiveness of care as well as respond to higher demand. However, if we believe that a significant cause of rising expenditure is inefficient production of health services, effort should be made to ameliorate the situation4 . As concluded by the OECD (2004) report:“Ultimately, increasing efficiency may be the only way of reconciling rising demands for health care with financing constraints”. As part of a continuous effort by governments to eradicate inefficiency while maintaining equity and quality of care, review and evaluation of performance is important to find out what works and what does not. In order to do this, performance indicators have been developed to measure the success of a health care system at all levels in achieving its multiple objectives. The measurement of each objective will entail a PhD thesis in and of itself, however, this thesis focuses only the empirical aspects of efficiency assessment in the health care sector. More specifically, it will focus on the measurement of hospital efficiency. This segment often accounts for over one-third of total health expenditure in any country (Duckett, 2004), thus its performance has a significant impact on efficiency of the overall health care system. The rest of the introduction addresses the need for, and the recent developments and applications of efficiency measurement in the health care sector. It also summarises objectives and the structure of the thesis. 3

These do not necessary refer to X-inefficiency which is a common term used in the efficiency literature. It is noted that some inefficiency may be an acceptable trade-off with other policy goals, such as equity in accessing to care. 4

3

1.2. EFFICIENCY MEASUREMENT FOR HEALTH CARE SYSTEM IMPROVEMENT

1.2

Efficiency measurement for health care system improvement

An important question in the health policy design is how to improve efficiency while achieving other equally important objectives of health care. Improvement in the health care sector, however, does not usually happen automatically due to market failures and thus good performance cannot be taken for granted (WHO, 2000). The quest for efficiency improvement requires trial-and-error work that is carefully crafted, and subsequent evaluations in the light of different objectives to understand what does and what does not work in practice. Recent advances in information and communication technology and increasing demands for accountability by health care providers have highlighted the urgent need for robust performance measurement in the health care sector (Smith et al., 2008). The complexity of health care, along with some natural imperfections of this market, makes the scope of performance measurement enormous. It ranges from examining the state of a country’s health system to assessing activities of health service providers to reflecting the experiences of the individual users. Performance measurement also needs to reflect the multiple dimensionality of health care objectives, of which the three most commonly used are efficiency, equity, and quality of care5 .

1.2.1

Policies and performance evaluation

Governments around the world have tried different health care policies aimed at increasing efficiency, amongst other objectives, with varying degrees of success. Those policies include the introduction of structural reforms, and the management of supply and demand. Whichever approaches are to be employed, the concern is always to find a set of policies that provide proper incentives where individual players with private information and goals make decisions which are in accordance with the system’s overall objectives, or to monitor players where incentives are not properly aligned. 5

Effectiveness is also referred to as one of the major dimensions in performance measurement. It can be defined as “the degree of achieving desirable outcomes given the correct provision of evidence-based health care services”. However, efficiency in its broadest sense of “optimal use of available resources to yield maximum benefits or results” implies effectiveness. Health resources are needed to provide services that restore or improve health status of the patients. To achieve efficiency, the first and foremost condition therefore must be effectiveness.

4


Structural reforms deal with deregulation, decentralisation and pooling resources through taxation or health insurance. The movement towards tax-financed or social health insurance has came with a stronger emphasis on equity. Several approaches have been used to better replicate normal economic markets in health systems administrated publicly and funded through general taxation. They include decentralisation of management responsibility, i.e. giving public providers (especially hospitals) increased independence and managerial capacity, and separation of provision and purchasing functions within the public responsibilities. These methods are believed to increase flexibility for management at the micro-level and reduce administrative burden at the central government level, from which efficiency gain is realised. Demand-side control includes managed care and patient cost-sharing while supplyside relates to purchasing and payment arrangements, management of health technology, service mix and quality of care. The demand-side interventions can be effective administrative controls over heath service allocation, especially when managed care is implemented through disease management programs or gate-keeping system to specialist services. An acceptable cost-sharing scheme can reduce excessive demand and hence, reduce public finance burden while protecting access for the poor. They help lessen the impacts of some natural imperfections of the health care market such as externalities and moral hazard, thus helps increasing economic efficiency. Supply-side management is even a more powerful tool due to the demand’s dependency on supply of health care services. Efforts to manage health technologies and encourage competition amongst health care providers and insurers directly target expenditure while health care purchasing and payment arrangements try to align objectives of individual providers and those of the health system, especially quality of care and efficiency. It has been argued, for instance, that a case reimbursement scheme, either through scheduled-fee, fee-for-service or performance-based, could provide incentives for hospitals to provide quality services more efficiently. On the other hand, fixed budget methods such as global and line-item budgeting are often believed to be better at controlling cost although not necessarily improving efficiency and quality of services (detailed discussion can be found in Appendix D). The degree of success of these health policies and regulations, however, is usually uncertain. It largely depends on existing institutions that have developed historically, people’s beliefs and expectations, along with implementation of the new policies. It is then important to adopt an evidence-based approach, by which serious 5


evaluations follows the implementation to find out what works and what does not. The link between policy decisions, measurement and assessment/evaluation is shown in Figure 1.2. Data collection and measurement facilitates the analysis of some health care aspects. Interpretation of the analytical results in the light of objectives and environmental influences produces evaluations that feed into policy formulation and implementation. Changes in policies alter the behaviours of health care organisations and the system as a whole which produces new data and requires further measurement and evaluation. The process then continues. Figure 1.2: The feedback process of performance measurement

MEASUREMENT (DATA)

THE HEALTH CARE SYSTEM PREVENTIVE CARE PRIMARY CARE

EVALUATION (ANALYSIS)

TERTIARY CARE AGE CARE PHARMACEUTICAL MEDICAL EDUCATION

IMPLEMENTATION (ACTION)

Source: Adapted from Goddard et al. (1998).

This process emphasises the central role of measurement and assessment in policy design and implementation. It requires good monitoring systems as well as the development and utilisation of better data and accurate performance measures. In short, performance measurement and evaluation offers policy-makers opportunities for monitoring productivity and efficiency of providers, which is fundamental key to seek improvement. It provides feedback to clinical facilities and practitioners on their actions and how they compare with their peers. This is essential in the health care market where standard market-based indicators such as profitability and market share cannot be used to evaluate performance. Without measurement and evaluation, efficiency improvement cannot be systematic and is unlikely to be 6


sustained.

1.2.2

Efficiency measurement projects around the world

International organisations and countries around the world have researched, designed and implemented various schemes and indicators to measure their health systems’ performance. International organizations, especially the WHO and the OECD, also play a crucial role in pushing the efficiency measurement agenda. The Health System Performance Report by the WHO (2000) is a major step forward in methods for assessing and comparing national health system performance. Following this initiative, the OECD launched a Health Project, of which the main theme is performance measurement and improvement of health systems in the OECD countries. An outcome of this project is the Measuring Up Report that examines progress and challenges in the measurement and application of performance indicators (OECD, 2002). Since then various reports and studies have contributed to the methodological and technical aspects of measuring and comparing health system efficiency across countries (see for example: Hollingsworth & Wildman, 2003; Richardson et al., 2003; Greene, 2004; Lauer et al., 2004). The main challenge for cross-country analyses are the comparability of data and results. Multidimensionality of health systems requires more than one indicator for each objective that is to be measured. Comparison and ranking of countries by performance is, therefore, not straightforward, it is especially difficult to determine the relative importance of various objectives. This has led to development of a composite measure based on a single metric that can capture the concept of system performance. Such a measure is useful as it gives policy-makers some benchmarks, that is the best practice, to guide efficiency improvement. However, some methodological issues such as the development of a set of weights for individual objectives and indicators and the modelling of efficiency, including the treatment of exogenous influences, both observable and unobservable, on system performance, need further consideration. At the country level, assessment efforts focus on both measuring performance of the national health system and individual health service facilities. In some countries, this has became a routine management activity. For instance, in Norway, efficiency analyses were used as one argument for moving to financing providers

7


through the system of Diagnosis Related Group (DRG), that is DRG-based funding. This is currently used by some regional Health Enterprises to inform their internal resource allocation (Hollingsworth & Street, 2006). Australia has developed the National Health Performance Framework to evaluate the performance of its health care system. This framework covers nine dimensions, including efficiency (Committee, 2004). The state of Queensland has incorporated efficiency measurement in the Quality Hospital Report when evaluating hospital performance (QH, 2005). The New Zealand public health sector has used efficiency measurement to identify efficient expenditure levels to set prices for hospitals at the DRG level (Rouse & Swales, 2006). Another example is Finland introducing its Hospital Benchmarking project for the first time in 1997 to provide benchmarking information on hospital performance and productivity. In 2006, data from the project was integrated into the production of national statistics (Smith et al., 2008). In the UK, the National Health Service (NHS) has effectively used performance indicators in directly managing service performance of public provision of health services. All NHS health care organisations are assessed using about 40 performance indicators and the information is reported in an annual performance report. Those indicators cover dimensions of effectiveness and efficiency, as well as patient experience and resource capacity. They are chosen to reflect the wider range of services specified in the NHS Plan and include data from the patient and staff surveys, as well as clinical indicators6 . In the US, the Joint Commission on Accreditation of Health Care Organizations (JCAHO) has recently developed a set of core measures of performance, which it promotes across US hospitals. These indicators were agreed following a rigorous review of evidence, extensive industry consultation and pilot testing. The indicators deal with specific aspects of treatment of some relatively high volume conditions such as heart failure, community-acquired pneumonia and maternity services (Bankauskaite & Dargent, 2007). Country-level indicators usually suffer fewer problems of data compatibility due to their relatively unified data information management system. There is also little need for a composite measure at the national level as the information about the system performance with respect to individual objectives is necessary for decisionmakers to identify targets of future interventions/reforms. However, performance 6

Examples of NHS performance indicators can be seen at http://www.performance.doh. gov.uk/nhsperformanceindicators/hlpi2002/index.html, http://www.chi.nhs.uk/ratings/, http: //ratings2004.healthcarecommission.org.uk/ and http://ratings2005.healthcarecommission.org. uk/; accessed 24th December 2009.

8


assessment at the national level is not sufficient when it comes to understanding the sources of failure or under-achievement. The performance of a health care system is ultimately dependent on how well its institutions and facilities operate. They include hospitals, primary care facilities, preventive care programmes and the pharmaceutical industry. As pointed out by Smith et al. (2008) in the “Performance Measurement for Health System Improvement” report, productivity and efficiency is perhaps the most challenging measurement area of all, as it seeks to offer a comprehensive framework that links the resources used to the measures of quality and effectiveness. Therefore, this topic has attracted a large body of literature, of which a special emphasis has been placed on measuring efficiencies of health care institutions. The past two decades have seen an increasing volume of efficiency studies, driven partly by demand from decision makers and mostly by the greater availability of data, estimation techniques and software packages. In fact, performance assessment of hospitals, the largest segment of a health care system, has became a common and indeed, one of the major topics addressed in WHO Regional Office for Europe Health Systems Conference in June 2008 (Groene et al., 2008).

1.2.3

Challenges in measuring efficiency

Despite a fast growing body of literature that attracts both academics and decision-makers, performance measurement of the health care sector has not yet met its goal of guiding quality policy decisions for relevant actors within the health system. As noted by Hollingsworth & Street (2006), most analyses appear to be targeted at academic readers, and are rarely utilised by either policy-makers or health care institutions themselves. There are several reasons for this. The first challenge in measuring health care efficiency arises from the pursuit of multiple (and usually conflicting) objectives. For instance, due to resource constraints, a trade-off between efficiency and quality is usually inevitable. Likewise, there is a trade-off between equity and efficiency in health care. Fee-for-service financing schemes which are designed to create a market mechanism for the health sector have always failed to achieve equality. Though the health sector is not alone in this aspect, equity is arguably one of the highest priorities of any health care system because health is a major determinant of human welfare. Without an appropriate method to account for quality of care and equity priorities, one provider could look superior to others simply because it sacrifices these objectives for the so-called 9


efficiency through higher output volume. This problem is prominent in the health care efficiency literature, and so far, there has been little agreement on how it can be solved. A key element for reliable measurement is ensuring that data for performance measures are reported accurately and on a timely basis. This is the second issue in efficiency measurement. Health care services are complex and there is considerable diversity between providers in terms of the services they provide. Each case treated is actually different and the conventional definition of output cannot capture the heterogeneity of individual cases, in terms of severity and types of illness as well as quality of care. Providers of various scale, specialisation and location employ different mix of labour skills and technologies. Hence, there exists a significant scope for measurement errors of outputs and inputs (and indeed, they are always a heated topic of debate). Furthermore, many developing countries that are at an early stage of information system development suffer from data inconsistency due to the mixture of various data sources, ranging from administrative data, census to labor force surveys. This prevents efficiency comparisons across time, space and institution. Additionally, there is only a weak link between efficiency measurement and policy design and implementation. That is, many analyses fail to provide relevant and actionable information to guide improvement of health care provision practices. Many studies merely quantify the extent of inefficiency instead of identifying the nature, form and sources of inefficiency. Hence, they fail to answer questions of importance to decision makers: What are the sources of inefficiency and how can it be reduced? Efficiency results can also lack actionability because they are not transmitted in a way that facilitates understanding and appropriate action on the part of managers and policy makers. When used as a statistical black-box, efficiency measurement can do little to improve efficiency in real life. Hence, more attention should be paid to the presentation of performance measurement results, and how patients, providers, practitioners and the public interpret them and are influenced by them (Smith et al., 2008). Finally, reliability of efficiency measures is arguably the greatest challenge for it to become a part of the routine analytical tool of decision makers. Public agencies look for reliable guidance in formulating policies, especially when it comes to the search for the primary causes of inefficiency and improvement potentials. Health care providers expect the analyses to help reveal the factors that influence their performance so that appropriate adjustments can be made to achieve the best prac10

1.3. OBJECTIVES AND ORGANISATION OF THE THESIS

tice. However, many empirical studies have shown that estimated efficiency scores are sensitive to analytical techniques (see for example, Valdmanis, 1992; Grosskopf & Valdmanis, 1993; Magnussen, 1996; Parkin & Hollingsworth, 1997; Smith, 1997; Webster et al., 1998; Chirikos & Sear, 2000; Folland & Hofler, 2001; Jacobs, 2001; Hofmarcher et al., 2002; Gannon, 2005). Furthermore, the lack of an explicit recognition that production invariably takes place under conditions of uncertainty leads unsatisfactory measurement of efficiencies (O’Donnell et al., 2009). Therefore, measurement techniques are not yet able to offer policy makers reliable tools for assessing and regulating health care performance because they are not robust to both statistical noises and specification errors (Fried et al., 2008). There is a need for better understanding of the health service production process and its operational constraints, better definitions and measurements of health care outputs and inputs, and especially, more guidance for model construction and analytical techniques. These problems have prevented efficiency measurement (although well recognised as a potentially useful tool for health care planning and policy evaluation) from becoming a standard element of the analytical toolbox for decision makers. The field is young and further research is needed to realise the potential contributions of efficiency measurement in eradicating inefficiency in the health care sector. This thesis considers the reliability of efficiency results in empirical studies of the hospital sector and its relevance of to health care policies.

1.3

Objectives and organisation of the thesis

The overall objective of this thesis is to contribute to the existing knowledge of efficiency measurement in the health care sector through the exploration of efficiency modelling issues, practical applications of different efficiency analysis frameworks and their policy implications. These objectives will be achieved through a collection of three essays, each of which deals with a distinct topic. However, the overarching theme uniting them is the measurement of efficiency and productivity of hospitals. The hospital sector is chosen for its central role in a health care system. As noted by Duckett (2004), hospitals are key institutions, usually accounting for over onethird of total health expenditure. Their performance then has a significant impact on efficiency of the overall health care system. In fact, many governments and nongovernment organisations have initiated projects on hospital performance assessment to respond to concerns such as quality management, improving accountability of

11


hospitals and informing the public. More specifically, much of the thesis will be devoted to analysing efficiencies of the public hospital sectors in a developing (Vietnam) and a developed system (Queensland, Australia). The empirical studies are accompanied by a review of major health care payment methods, which acknowledges the central role of financing methods and payment policy on the performance of the health care sector in achieving the principal and usually conflicting objectives of equity, efficiency and quality of care. Empirical studies in hospital sector are important, however one should not ignore the importance of technical aspects of efficiency studies. The balance of this thesis therefore will be spent evaluating the impacts of various modelling choices on efficiency estimates. The thesis is contains three major chapters, each of which is self-contained. That is, relevant literature reviews, analytical methodology and data are described in the individual chapter, followed by findings and discussions. A summary of these chapters is as follows:

Chapter 2: Quantifying the effects of modelling choices on hospital efficiency measures using meta-regression analysis This chapter explores the variation of efficiency estimates under different modelling approaches. It recognises the sensitivity of efficiency estimates discussed in the literature and attempts to quantify the effects of modelling choices on estimated efficiencies using the meta-regression technique. The modelling variables discussed in this chapter include number of observations, number of variables used in the frontier model (including inputs, outputs, prices and control variables), parametric approach, assumptions on orientation, returns to scale, functional form and efficiency distribution. The chapter describes the expected effect of these factors (i.e. increasing or decreasing) on estimated efficiencies and uses the meta-regression technique to quantifies those effects. The data for this exercise is derived from 253 estimated models reported in 95 empirical analyses of hospital efficiency in the 22-year period from 1987 to 2008. A major contribution of this chapter is that it can be used to narrow the variation of efficiency estimates due to different sample sizes, variable choices and approaches, enhancing the reliability of results for policy formulation and across time and crosscountry/state/region comparisons. Furthermore, it is the first study applying to 12


meta-regression analysis to health care efficiency and it looks at a relatively large body of literature where the production units use a much more uniform technology. This chapter is accompanied by Appendix A that lists all the 95 empirical studies included in the meta-regression analysis. The list follows publication years and the alphabetical order. It also presents the study countries, estimation methods used, sample size and provides brief descriptions of inputs and outputs.

Chapter 3: Measuring labour efficiency in Queensland public hospitals This chapter investigates labour productivity and efficiency in the context of medical labour shortage in public hospitals in the State of Queensland, Australia. Labour productivity is a partial measure, as opposed to total productivity. However, it is a major determinant of the total productivity given that labour usually accounts for the largest share of inputs (between 60% and 80% during the period 2000-08, as indicated in the Ministerial Portfolio Statements by the Ministry for Health of the Queensland Government). This reflects the fact that the hospital sector is a labour intensive industry and therefore improvements in productivity and efficiency are likely to come from a better use of human resources. Queensland is chosen for the reason that medical labour shortage has became a critical issue during the period 1995-2007 when public hospitals in Queensland were reimbursed under a historical cost funding model. This problem is exacerbated due to high population growth and a fast aging population profiles. Therefore, an understanding of labour productivity and efficiency will be very useful for decision makers. A labour input requirement function is used to examine labour efficiency in public hospitals in Queensland. The sample contains 84 hospitals and covers the period 1996-2003. The estimation is performed under the “true” random effect framework proposed by Greene (2004, 2005b) with the inclusion of selected control variables to capture the observed heterogeneity. The study also conduct a sensitivity analysis on parameter and efficiency estimates through alternative model specifications.

Chapter 4: Funding reforms and public hospital efficiency in Vietnam This chapter presents the empirical results on the impact of financial autonomisation on productive efficiency of public hospitals in the context of a health sector in 13


transition - Vietnam. Vietnam is an ideal country for health sector studies for the reason that it has experienced various market-oriented reforms over the last decades to improve its productivity and economic growth. Financing and provision of health services have been gradually privatised under the belief that a market-based allocation of health resources would be more efficient in delivering outcomes. Financial autonomisation is one of the most recent reform initiatives that turns public service units into quasi-corporates. However, market failure is a prominent characteristic of the health care market and there remain questions about the benefit of reforms, especially in relation to efficiency improvement. The impact evaluation is performed on a sample of 62 public hospitals around the country, between the two periods 1998-2000 (pre-reform) and 2005-2007 (postreform). The hypothesis is that the financial autonomisation reform has increased the productive efficiency of public hospitals because it has given hospitals more autonomy in the management of resources, including labour, facility and equipment, which can increase internal efficiency. Furthermore, once faced with market-like incentives (of profitability), hospitals would strive to improve performance. However, the gain from reform (in the form of improved productivity and efficiency) might not be shared uniformly across regions due to different local conditions and rates of adoption. This study carries important policy implications for future reforms, especially in the context where it is evident that the reform outcomes have not been shared uniformly across regions and by different socio-economic groups. Another main contribution of this chapter lies in the novel application of the recently developed meta-frontier framework by Battese & Rao (2002) and O’Donnell et al. (2008) to investigate the effects of reform. Data Envelopment Analysis is used to estimate the efficiencies under the specified framework and the sensitivity analysis is performed using the bootstrapping technique. Chapter 4 is accompanied by two Appendices B and C. They contain information about the construction of output weights for the analysis and the full results of the sensitivity analysis conducted by bootstrapping.

Chapter 5: Conclusion This chapter recaps the motivations, objectives and achievements of the thesis. It emphasises the importance and relevance of efficiency measurement in offering

14


policy-makers an opportunity to achieve health system improvement. However, the complexity of health care services and the health care market makes the scope of performance measurement enormous. These errors usually come from data collection and analytical methodologies. Efficiency studies, therefore, need to pay special attention on analytical choices and whenever possible, sensitivity analysis and bias adjustments should be made to increase the reliability of results. This concluding chapter also summarises the main findings of the three papers, contributions, relevant policy recommendations, as well as short-comings of these studies. Overall, the three studies are quite unique and useful for various reasons, ranging from technical contributions (i.e. quantifying the effects of modelling choices on efficiency estimates) to empirical investigation of topical economic questions (i.e. labour efficiency in the context of historical cost funding and the impacts of marketoriented reforms on the productive efficiency of hospitals). It also lays out some research directions for the future, such as studying the effect of modelling choice on the variance of efficiency estimates, examining the total productivity of Queensland public hospitals(in contrast to labour productivity) in order to understand the degree of input substitution, shadow prices and price elasticities of demand for inputs, and investigating the impact of technological, scale and scope changes on the productivity of Vietnamese public hospitals. Finally, Appendix D contains a review of the incentive structures by various provider payment methods and their potential to achieve efficiency objectives. It is inspired by the fact that financing methods and payment policy are central to the performance of the sector. They accomplish far more than simply the transfer of resources due to their direct influence on supply behaviour. Because the hospital sector consumes the greatest share of health care resources, the way they are paid has a significant influence on the performance of the health care system as a whole. Many hospital payment methods and their combinations have come into existence to meet urgent policy needs or respond to various policy goals. Among these are equality of access and distribution of health services, recognition and reward for high quality services, and the promotion of efficient production.

15

Chapter 2

Quantifying the effects of modelling choices on hospital efficiency measures using meta-regression analysis

If efficiency analysis to be taken seriously, producer performance evaluation must be robust to both statistical noise and specification error. Fried et al. (2008) In recent years, applied academic research into health sector efficiency has expanded substantially. There are two major reasons for this growth: (i) an increasing demand for efficiency studies as an input into the decision making process and (ii) lower barriers of entry into this research field (Hollingsworth & Street, 2006). The demand for efficiency analyses is primarily due to a desire for better informed government policy decisions (e.g., assessing the effects of deregulation, mergers, and market structure on industry inefficiency) and also to help improve managerial performance (e.g., by identifying best and worst performers). Barriers to entry have fallen as a consequence of increased collection of computerised hospital data records and the wider availability of software packages that incorporate efficiency measurement methods (e.g., FRONTIER, LIMDEP and Stata for parametric models and DEAP, DEA-Solver and DEA-Frontier for non-parametric methods). Although the quality of efficiency analyses has significantly improved, controversy remains around the merits of different estimation strategies and methods, and their impact (both in direction and magnitude) on the efficiency estimates obtained. For policy-oriented studies that make use of efficiency estimates (such as those on health resource allocation), the reliability of results is a major concern. Health care providers expect the analysis to reveal the factors that influence their performance 16

so that appropriate adjustments can be made to achieve best practice. Public agencies and policy makers look for reliable guidance in formulating policies, especially when it comes to the search for the primary causes of inefficiency and improvement potentials. However, many empirical studies in the hospital efficiency literature have shown that the choice of methods and model specifications can greatly affect the estimated efficiency scores (see for example, Valdmanis, 1992; Grosskopf & Valdmanis, 1993; Magnussen, 1996; Parkin & Hollingsworth, 1997; Smith, 1997; Webster et al., 1998; Chirikos & Sear, 2000; Folland & Hofler, 2001; Jacobs, 2001; Hofmarcher et al., 2002; Gannon, 2005). There have been several systematic reviews of efficiency measurement in the health care sector such as those by Worthington (2000); Hollingsworth (2003); Worthington (2004); Erlandsen (2008); Hollingsworth & Peacock (2008); Hollingsworth (2008a); Rosko & Mutter (2008). While Erlandsen (2008)’s focus is at the macrolevel and on the possibility of comparing health care efficiency across countries, Rosko & Mutter (2008) provided a review of stochastic frontier applications on US hospitals, accompanied by an empirical application to demonstrate the process of making modelling choices. The more general reviews include those by Hollingsworth (2003); Worthington (2004) and Hollingsworth & Peacock (2008). They provide some statistical reviews on the growth of this research body and some discussion of the reliability of efficiency estimates, upon which relevant policy decisions were drawn. Although these reviews offer extensive overviews of the literature and some in-depth discussion on the reliability issue, it is noteworthy that none has attempted to quantify the degree to which methodological differences influence the diversity of results in this literature, using techniques such as meta-analysis. The aim of this chapter is two-fold: first, to provide an overview of the literature on hospital efficiency and relevant efficiency estimation methods and second, to examine the effect of modelling choices on efficiency estimates in the hospital efficiency literature. To this end, a discussion of various choices of estimation techniques, model specification and variables included in an efficiency analysis, is presented, followed by the empirical component, which consists of a statistical summary of the literature as well as a meta-regression analysis in order to identify the key factors that influence efficiency estimates.

17

2.1. MODELLING CHOICES AND EFFICIENCY ESTIMATES

2.1

Modelling choices and efficiency estimates

The influence of modelling choice on efficiency estimates is widely acknowledged in the efficiency literature. Although most studies do not have a choice in either the sample size or variables used due to data availability, decisions on analytical methods and model specifications, to a large extent, can be controlled to accommodate the research questions. Hence, there are good reasons for examining alternative model specifications and their results to ensure the reliability of the estimation. This is especially important for studies with a policy design focus, as other health economists have pointed out in earlier studies (for intance, Newhouse, 1994; Parkin & Hollingsworth, 1997; Folland & Hofler, 2001; Jacobs, 2001; Street & Jacobs, 2002; Chen, et al., 2005). If the estimates are to inform decision makers on funding or capacity utilisation, then hospitals incorrectly labelled as inefficient may receive less funding resources or may need to trim their production. If post evaluation of a health care policy on hospital behaviour is the issue of concern, a biased estimate of efficiency would be misleading to assess the true policy impacts.

2.1.1

Variable choice and sample size

The first major decision in modelling production technology relates to output and input choices. Inputs and outputs should be relevant and sufficient to capture the production process. In practice, problems with variable choice come under the form of imperfect measures of inputs and/or outputs, incorrect aggregation and variable omission1 . Although studies far too often do not have choice over quality of input and output data, it is worth emphasising that findings based on rudimentary measures of inputs and outputs should be interpreted with caution. Variables and aggregation in many situations are mainly attributed to different research questions or data availability, while in other cases are due to modelling choice. Its existence usually distorts findings. In the hospital efficiency literature, the generic problem is the variation in definitions and quality of input and output measures due to their multi-dimensionality. Ideally, the output of hospitals should be the incremental health improvement of 1

Inclusion of irrelevant variables is also another issue. However, in the hospital efficiency literature, it is far more often that a frontier model fails to capture all aspects of the health care service production than inclusion an extraneous variable, mainly because of data deficiency. Further more, it is suggested that exclusion of relevant variables is likely to be more damaging to frontier models than inclusion of irrelevant variables (Smith 1997).

18


the patient after receiving hospital treatment, which can theoretically be measured by the difference between health status before-treatment and that after-treatment. However, this output measure is generally unavailable and hospital efficiency studies have generally used activities as a proxy for outputs. These activities can relate to surgical procedures, inpatient episodes of care, emergency cases, and outpatient consultation sections. It is recognised that using activities to measure the performance of hospitals may not be problematic when there is good research evidence that activities are in fact leading to health improvements and/or there is no difference between organisations in activity implementation (i.e. effectiveness of treatment). When this is not the case, activity counts are less reliable as output measures of health care production (Jacobs, et al., 2006). The input side of hospital efficiency analysis is often considered less problematic. Hospital activities consume labour, medical and non-medical goods and capital (in the form of beds, infrastructure and medical equipment). These can be measured with high precision. Labour inputs usually come in categories of doctors (general practitioners and specialists), nurses, diagnosis and allied health professionals, carers and so on. There are also administrative and operational staff who take on the role of management and maintaining the capital stock. Similar to other service industries, labour accounts for a large part of the health service production. Non labour inputs such as medical goods, non-medical goods, materials and capital are usually measured by cost. Capital as an input in efficiency analysis, in principal, is defined as the capital consumed in the current period of analysis. However, measuring capital is challenging2 , partly because of the difficulty involved in first measuring the stock of existing infrastructure and equipment, and partly due to problems in attributing capital use to any particular period (Jacobs, et al., 2006). Imperfect (and sometimes non-existent) measures of inputs and outputs of hospital production means that often a study faces the problem of variable omission and/or aggregation bias. Common missing input variables are measures of capital stock and material inputs. Hence, the majority of studies utilise “number of beds” as the proxy for capital although this is far from ideal. On the output side, it is teaching and research variables that are often omitted. Bias created by an omitted variable is illustrated in Figure 2.1(a). It shows that the list of efficient and inefficient hospitals can alter significantly when one major variable is omitted. Assuming the production process involves two inputs X1 and X2 , in the first diagram, units A and C are identified as fully efficient (on the isoquants), while B is not. If X2 is 2

There is also a large body of literature dealing with capital measurement. It is, however, outside the scope of this paper.

19


omitted, mapping those units on X1 space produces quite different conclusions. Unit C becomes inefficient and the worst performer. Mean efficiency in this case would be much lower than in the case where X2 is not omitted. Benchmarking becomes even less reliable because the ranking has changed significantly. The other issue concerns aggregation of variables. Constraints on degrees of freedom and zero-value in some variables (not the same as missing data) usually lead to aggregation of variables. In most studies, the two main labour categories of doctors and nurses are produced by aggregating many sub-categories of very different skill levels, ranging from junior trainees to specialists or directors of nursing, with or without weights. Aggregation of administrative and domestic staff, or of allied health and health professional staff, is also common practice. On the output side, episodes and procedures in health care usually differ from one patient to the other, and hence the number of outputs can be considered a roughly the same as the number of patient treated. Aggregation is generally required to reduce the number of outputs. Since the development of casemix systems that take into account the differences in resources consumption for various types of treatments, studies have been using casemix information to aggregate outputs, often from more than several hundred output categories into one or two outputs. Many other analyses, most of which are early studies and studies using data from developing countries, use raw counts (or unweighted aggregation) of total number of inpatient and outpatient occasion of services. This can lead to biased results when particular health care units provide more or less complicated casemix services. Figure 2.1(b) provides an illustration of an input aggregation problem. The technology is represented by the convex isoquant. Linear aggregation of the two inputs is represented by the 45 degree straight line, where the two input variables (e.g., administrative and domestic staff) are allocated equal weights. Under a convex isoquant, A and D are identified as fully efficient, while a linear isoquant suggests three production units of B, C and D are inefficient. In this instance, the aggregation is likely to lead to an underestimation of the mean level of technical efficiency for these firms. While it is expected that inappropriate aggregation creates biases in efficiency measurement, this might be less problematic than missing variables as outputs and inputs are still captured (to some degree) in the production model3 . 3

Note that the omitted variable example in Figure 2.1(a) can also be viewed as a special case of an aggregation problem, where one of the weights is zero. One should also emphasise that the effects of aggregation can be reduced by selecting appropriate weights (e.g., wage levels) and/or by using non-linear aggregation methods, such as the Fisher index number formula in the place of a simple linear aggregation formula.

20


The question is then whether it is possible to predict the direction of impact on the average efficiency score by the inclusion or exclusion of a variable? Technically, the inclusion of another variable in the estimated model increases the dimension of the frontier. The illustrative examples in Figure 2.1 suggest that this may produce higher mean efficiency scores. The magnitude of this effect, however, depends on the omitted variable’s correlations with included variables. For instance, if the extra variable is an input and it is highly correlated to other input variables, omission of the variable is unlikely to significantly affect the results. On the other hand, if it is not strongly correlated then the impact on mean efficiencies can be notable. One example in the hospital efficiency literature is the study by Rosko & Chilingerian (1999). They added casemix variables to a basic translog function and found the basic translog case yielded lower efficiency scores compared to the one without casemix variables. In fact, the potential impact of dimensionality on efficiency scores was discussed by Nunamaker (1985) who found that variable set expansion, either through adding new variables or disaggregating existing variables, should produce an upward trend in mean efficiency scores. Other studies by Tauer (2001); Fare, et al. (2004); Barnum & Gleason (2005) also confirmed that aggregation of many outputs into fewer or one output introduces a downward bias on efficiency estimates, and the more outputs are aggregated, the greater the bias that may be expected. The opposite effect is generally observed for sample size. As illustrated in Figure 2.2, the increase of sample size will either push the production frontier up when new observations - points A - form part of the new frontier, as in Figure 2.2(a), or does not change the frontier at all when new observations - points C and D - lie entirely under the existing frontier, as in Figure 2.2(b). When the new observations form part of the new frontier, then the units that were once identified as efficient under the old frontier may now be identified as inefficient. When a new observation does not affect the position of the frontier (because it is either on or below the existing frontier) then it does not change the status of already identified efficient and inefficient units. Thus, on average, increasing the sample size is unlikely to result in an increase in mean efficiency scores4 . This observation is also recorded (Zhang & Bartels, 1998), who found the negative correlation between the estimated mean efficiency and the number of firms in the industry. When the sample is relatively small, the mean efficiency decreases quickly as number of observations increases. When sample sizes are large, the mean efficiency shows little change. Above a threshold, a mean efficiency seems to tend to be fairly constant. 4 Note that if the newly included data points are mostly quite efficient, but they do not shift the frontier, then it is possible that the mean efficiency level can increase, but this is less common.

21


Figure 2.1: Illustration of the omission and aggregation problems

X1

Isoquant

X 1A

A

X1B

B C

X1C

X2 (a) Omission of variables

X1 B A C D

Isoquant

Linear aggregation of inputs X1 and X2

X2 (b) Aggregation of variables

22


Figure 2.2: Illustration of increasing sample size

Y A

B

X (a) Added observations change the frontier’s position

Y

C

D

X (b) Added observations do not affect the frontier’s position

23


2.1.2

Parametric versus non-parametric approaches

The second main modelling choice relates to the decision between parametric and non-parametric approaches. While the non-parametric method involves mainly the use of linear programming techniques, the parametric method utilises production, cost, profit, revenue, and most recently, distance functions as alternative approaches to describe the production technology and estimating efficiency. Ideally, this choice should be based on the understanding of the production technology. It also depends on the analyst’s preference on the trade-off between some level of measurement error and the bias created by potentially incorrect parameterisation of the production technology. In production, measurement errors can come from the nature of the production process or during the sampling procedure. Data gathered from standardised manufacturing industries tend to have less measurement errors than those from multiple-output service industries. Measurement and sampling errors may have serious consequences in the non-parametric framework since no underlying error structure is specified. On the other hand, inappropriate choice of functional form in the parametric framework, both in the technology and the inefficiency distribution, confounds inefficiency with the effects of misspecification (Lovell, 1996). Several studies in the hospital efficiency literature have investigated the influence of estimation methods on efficiency predictions through testing for correlations between efficiency estimates (for examples, Linna & Hakkinen, 1998; Linna, 1998; Webster, et al., 1998; Linna & Hakkinen, 1999; Lopez-Casanovas & Saez, 1999; Chirikos & Sear, 2000; Jacobs, 2001; Gannon, 2005; Barbetta, et al., 2007). Even though the correlation between parametric-based and non-parametric-based efficiency estimates is generally quite high, they are usually lower than the correlation between efficiency scores produced by the same method but different model specifications (Jacobs, 2001). Gannon (2005) found lower efficiency scores when using the parametric method, suggesting that non-parametric efficiency measures (in this case data envelopment analysis - DEA) might not control for other factors such as the type of production process or other environmental factors. However, when it comes to determining the sources of inefficiency, both methods appear to lead to the same findings. Arguably, one could expect that the non-parametric approach yields lower efficiency scores compared to the stochastic frontier method. Since the former is deterministic, all deviations from the frontier are dubbed with inefficiency while the 24


latter allows for statistical/random noise. This effectively increases the efficiency scores predicted by the parametric method. However, this need not be the case since non-parametric methods can produce a frontier enveloping all the data points whilst stochastic frontier analysis (SFA) fits a frontier that may allow some data points to lie above it. Hence, it is not clear which method is more likely to produce higher mean efficiency scores. This issue is illustrated in Figure 2.3. Under the nonparametric frontier, production units A, C, E and F (holding up the frontier) are fully efficient. Under the parametric frontier, adjustment for the noise component can reveal different efficient units. In particular, noise adjusted C and E are C’ and and E’, respectively, and they are inefficient, while the reverse applies for D. Units A and F are still identified as fully efficient. Figure 2.3: Illustration of parametric and non-parametric methods

Y

Parametric frontier

E D’

F E’

C Nonparametric frontier

D C’ B

A A’

X

2.1.3

Orientation

The next choice in frontier modelling relates to orientation. Choice of output/input orientation is usually driven by the objective of production units under relevant production and management constraints. For instance, hospitals under an expenditure cap scheme tend to maximise output, while hospitals receiving reimbursement based on units of treatment appear to conserve cost. If maximising output (or outcome) is considered a relevant objective of a hospital, then an output orientation (output oriented DEA frontier or stochastic production frontier or output distance function) may be warranted. Alternatively, if the hospital is believed to 25


minimising inputs or cost, then stochastic cost frontier, input oriented DEA frontier or input distance function may be selected. In practice, the majority of parametric studies prefer a cost function because hospitals are multiple-output production units and cost function can accommodate multiple outputs. The underlying assumption of a cost function (and input orientation) is that of cost (input) minimising behaviour of hospitals. The assumption is defensible from the viewpoint of hospital managers who are constantly under the pressure of meeting a budget requirement. However, this assumption has received much criticism, especially from medical professionals who often argue that their objective is not minimising cost but improving lives through prevention and treatment of diseases. A number of authors argue that analysis and policy recommendations based on a one-sided cost angle, such as attempts to control expenditure or reward/punish on the basis of cost efficiency without accompanying incentives at the level of medical staff-patient relation will lead to bad medical practice, queues and resentment (e.g., Harris, 1977). The effect of orientation choice on efficiency estimates is illustrated in Figure 2.4. If the sample in the analysis contains mainly small and a few large hospitals, it is expected that most hospitals are operating in the increasing returns to scale region, and thus an input orientation approach would produce a higher efficiency level for small hospitals, and consequently, higher mean efficiency. The reverse applies to samples with mainly large hospitals. A sample with a balanced mix of hospital size is likely to generate similar mean efficiency score under either output or input orientations. It is noted that this issue only applies for the VRS frontier. In the CRS circumstance, output and input orientations produce identical technical efficiency (Coelli, et al., 2005). In the hospital efficiency literature, only a few studies apply both input and output oriented approaches to the same dataset since the hospital’s objective function usually needs to be specified in advance. Those that apply both approaches focus on the sensitivity of efficiency scores. Burgess & Wilson (1995, 1996) used both input and output oriented non-parametric approaches and found that the latter produced slightly higher efficiency scores5 . Webster et al. (1998) estimated both production and cost frontier for Australian private hospitals, and Chirikos & Sear (2000) calcu5

In Burgess & Wilson (1995), mean input oriented efficiency was 0.8395 and its output oriented counterpart was 0.8725. Their sample (of 1,480 hospitals in the US) contains mostly large hospitals, indicated by the average number of bed (weighted by scope of services) ranging from 1,800 to 7,000. They repeated this exercise with another larger sample of 2,246 large hospitals, and arrived at a similar result (Burgess & Wilson, 1996).

26


Figure 2.4: Illustration of efficiencies produced by output versus input orientations

CRS frontier

Y

VRS frontier

C

TE AO  TE AI A X (a) Hospitals in the IRS region

CRS frontier

Y

VRS frontier

C

B

TEBO  TEBI

X (b) Hospitals in the DRS region

27


lated efficiencies using output-oriented DEA and stochastic cost frontier and tested for efficiency correlations.

2.1.4

Returns to scale and functional form

Another modelling consideration involves selecting an appropriate model structure, including functional form, returns-to-scale and efficiency distribution. Selection of functional form and distributional assumption is applicable only to parametric methods while returns-to-scale is an issue under both parametric and nonparametric approaches. Constant returns to scale (CRS) can be imposed in DEA models by removing the weight constraint, and in parametric models by imposing coefficient restrictions. Returns to scale relates to whether production units are of the optimal size or not. This is one of the popular research questions in efficiency analysis. Some production technologies possess the property of constant returns to scale and the production size does not matter. Others (and the majority) do not. This brings to attention the question of how returns to scale should be modelled. CRS assumption is appropriate when all hospitals are operating at the optimal scale (i.e. productivity is scale dependent). However, imperfect competition, government regulations, valid social objectives, financial and labour constraints may cause the hospital to be not operating at the optimal scale (Coelli, et al., 2005). In this circumstance, if CRS is imposed on the model, efficiency estimates can be significantly biased. This bias is generally more serious than in the case where variable returns to scale (VRS) is assumed for a CRS technology. This is explained in Figure 2.5. The upper diagram shows a technology that would yield similar efficiency estimates under CRS and VRS as the distance difference from each data point to either CRS or VRS is very small. The lower digram is the opposite story, imposing CRS vastly underestimates efficiency. Moreover, Smith (1997) suggests that inappropriate use of the returns to scale assumption is particular damaging when the sample size is small. A variety of functional forms have been tried in hospital efficiency applications. These include linear, quadratic, cubic, Leontief, Cobb-Douglas and translog (with or without ad hoc restrictions on certain parameters), as well as their hybrids, i.e. inclusion of some variables to control for hospital heterogeneity. Among those, the Cobb-Douglas and translog functions are the most widely used. The translog function - a second order Taylor series expansion approximating some true but unknown generalised log function - has the flexibility advantage over its main rival, the Cobb28


Figure 2.5: Illustration of efficiency estimates under CRS and VRS technologies

CRS frontier

Y

VRS frontier

X (a) The VRS frontier is close to the CRS frontier

CRS frontier

Y

VRS frontier

X (b) The VRS and CRS frontiers are further apart

29


Douglas, for not assuming constant input elasticities and returns to scale for all hospitals by not restricting the squared terms and cross products to be zero. However, it consumes many more degrees of freedom6 , thus can only be handled well with large sample sizes. Some studies estimated both Cobb Douglas and translog functions and conducted statistical tests to choose the appropriate model (for examples, Chirikos, 1998a,b; Webster, et al., 1998; Lopez-Casanovas & Saez, 1999; Chirikos & Sear, 2000; Folland & Hofler, 2001; Rosko, 2001a). Some other studies employed different coefficient restriction strategies to mitigate the problem of multicollinearity and large degree of freedom caused by the translog form (for instance, Chirikos, 1998a,b; Carey, 2003, and more). Another decision on the model structure relates to the assumption regarding the efficiency distribution. The literature has reported half normal, truncated normal, exponential and gamma distributions, of which the first two are widely used, followed by the exponential distribution. Many studies use more than one distribution to compare efficiency scores or to test for appropriateness using likelihood ratio tests (Chirikos, 1998b; Linna & Hakkinen, 1998; Webster, et al., 1998; Fuiji & Ohta, 1999; Rosko, 1999; Yong & Harris, 1999; Fuiji, 2001; Street & Jacobs, 2002; Street, 2003). This exercise is straight forward for truncated and half normal distributions as the latter is a special case of the former. A common conclusion is that the estimated efficiencies obtained using different distributional assumptions are highly correlated, despite their variation in magnitude. Hence, it can be inferred that hospital ranking based on those efficiency estimates should be quite consistent. The various assumptions discussed above are expected to have different effects on predicted efficiency. While efficiency estimates appear to be robust when it comes to distributional assumption, they can be highly sensitive to functional form, including assumptions on returns to scale. A higher order and more flexible functional form is expected to fit the data more tightly, hence producing higher efficiency estimates; while the CRS assumption consistently generates lower efficiencies. This is illustrated in Figure 2.6. In the first diagram, production unit B is the only efficient hospital if the CRS assumption is imposed while units A, E, G and B are all efficient under VRS. As efficiency is measured as the distance to the frontier, the VRS model consistently predicts higher efficiencies than the CRS model. The second diagram illustrates different possible shapes of the frontier under various functional forms. Higher order functional form tends to fit the data more tightly, thus on average 6 If a translog function with m outputs, n input, and q control variables (all interaction terms between outputs and input included) is to be estimated then the number of slope coefficients to be estimated is h i (n+m)2 +3(n+m) +q . 2

30


producing higher efficiencies. Studies in the literature appear to be consistent with these predictions. Chirikos (1998b) reported that mean efficiency estimates were higher when he switched from Cobb-Douglas to translog model; irrespective of the assumption about the inefficiency term. Webster et al. (1998) obtained identical efficiencies using both functional forms for the production function, higher cost efficiencies when translog functions were used, irrespective of variable definitions (16% inefficiency for CobbDouglas and 4% for translog). Folland & Hofler (2001) reported that Cobb-Douglas and translog functions yield mean inefficiencies of 12.7% and 10.1%, respectively.

2.1.5

Summary of expected effects of modelling choice on mean efficiency estimates

The hospital efficiency literature has seen various applications of both parametric and non-parametric methods. Stochastic frontier regression, stochastic distance function and corrected ordinary least squares are representatives of the parametric approach. As for the non-parametric method, most studies employ DEA. Although parametric and non-parametric are different estimation strategies, they share a common limitation. That is, their results are generally sensitive to the underlying assumptions and the data used. Non-parametric methods like DEA are more sensitive to extreme data points whereas parametric methods like SFA or distance function produces efficiency estimates that vary by the functional form and distributional assumption imposed. As the choice of variables significantly influences efficiency estimates, comparisons across studies without taking into account these modelling factors should be taken with caution. Scattered in the efficiency literature are various discussions on individual issues such as the likely effect of sample size or dimension or functional form on computed efficiencies. However, it was not possible to identify any study where all such matters were put together, and the magnitude and direction of their impacts on efficiency are quantified. In Sections 2.2 and 2.3, the discussion is taken a step further by using the meta-regression method to analyse the effect of modelling choice on efficiency estimates. A summary of expected effects of methodological choice on estimated mean efficiency is shown in Table 2.1.

31


Figure 2.6: Illustration of effect of functional form and returns to scale on efficiency estimates

CRS frontier

Y

G VRS frontier

E B

F

D C

A X (a) RTS assumption under DEA

CRS frontier

Y

D NIRS frontier

E

VRS frontier

C B A X (b) RTS assumption under SFA

32

2.2. METHODOLOGY AND DATA

Table 2.1: The expected impacts of modelling choices on estimated mean efficiency Factors which push

Factor with ambiguous

Factors which push

mean efficiency

impact on mean

mean efficiency

upwards

efficiency

downwards

Number of variables

Orientation

Sample size

Pooled panel data Second order functional form

2.2 2.2.1

Parametric/Nonparametric

Constant returns to scale

Efficiency distribution

Methodology and data Meta regression analysis

Meta-analysis is a statistical method used to integrate the findings from a significantly large collection of empirical studies. It can help an analyst to investigate the relationship between a study’s features (research questions, analytical method etc), and its outcomes. Because it analyses the results from a group of studies, the problem of low statistical power in studies with small sample sizes is partly resolved, allowing more accurate data analysis and conclusions. It has been a useful tool in health-related research bodies that investigate the strength of relationship between variables, the relative impact of independent variables, both direction and size of the effect, and the overall effectiveness of interventions. The quality of a meta-analysis depends crucially on the quality of the systematic review of the relevant literature, on which it is based. A good meta-analysis usually aims for a complete (or relatively wide) coverage of relevant studies, detecting the presence of heterogeneity and employing sensitivity analysis to test the robustness of the main findings in those studies. In clinical research, meta-analysis is most often used to assess the clinical effectiveness of health care interventions by combining data from several randomised controlled trials (on new methods of treatment or different health care practices). In the pharmaceutical industry, it has been widely used to summarize the results of drug development programmes. It is recognised that the technique provides a useful means of summarizing the overall medical effectiveness of a drug application and of analysing less frequent outcomes in an overall safety evaluation. The attractiveness

33


of the meta-analysis approach in health-related research is largely due to the greater emphasis on evidence-based medicine and the need for reliable summaries of the vast volume of clinical research (Whitehead, 2002). In the field of economics, meta-analysis has been applied in both microeconomic and macroeconomic issues (see reviews by Brouwer et al., 1999; Florax et al., 2002). Usually known under the form of meta-regression analysis (MRA), these studies cover various topics, such as growth empirics and macroeconomic policies (de Mooij & Ederveen, 2001; Nijkamp & Poot, 2003; Abreu, et al., 2005; Doucouliagos & Paldam, 2005, 2006); the valuation of natural conservation and resources (Boyle, et al., 1994; Loomis & White, 1996; Brouwer, et al., 1999; Cavlovic, et al., 2000); the impact of public goods (Button & Rietveld, 2000; Croson & Marks, 2000), the labour market and wages (Card & Krueger, 1995; Doucouliagos, 1995, 1997; Fuller & Hester, 1998; Groot & van den Brink, 2000); and consumer behaviour (Espey, 1998; Espey & Thilmany, 2000; Espey & Kaufman, 2000; Dalhuisen, et al., 2001). Although vastly different in topics, the meta-regression analyses in these studies usually takes the form of a simple linear equation, in which the regressor set features characteristics of the primary studies, such as countries/regions, the types of data used, time frame of the analysis, relevant economic variables as well as analytical methods employed. This is used to examine the direction and size of the relationship between some macro or micro economic phenomena. By combining many small studies in a meta-regression, small but important effects that otherwise might not have been detected in a single study can be picked up and reduce the possibility of a type II error - where there seems to be no statistically significant relationship between variables, when in reality such a relationship exists (Pang & Song, 1999). Beside its strength, it is recognised that meta-analysis also has its own limitations. It might aggregate and generalise over the differences in primary research, especially when the literature coverage is not highly focused. It can also sometimes ignore qualitative variations between studies. This problem is usually overcome by extensive systematic reviews whereby lower quality studies are removed, and careful handling of qualitative variations through coding those features into the meta-data. Another concern over the quality of meta-analysis is publication bias. Valid conclusions might not be drawn from a meta-analysis if only significant findings are published (DeCoster, 2004). Last but not least, like any other quantitative analyses, the value and validity of the results of a meta-analysis are critically dependent upon the data available, i.e. the quality of the literature (Drummond, et al., 1997; Pang & Song, 1999).

34


Given its features, meta-analysis appears to be an ideal tool for examining the research question, i.e. the impact of methodological choices on hospital efficiency estimates. The analysis in this paper has the advantage of having a reasonably large number of research studies on a focused topic (hospital efficiency only); published in internationally recognised journals; and it unlikely to be unduly influenced by researcher biases because efficiency estimates do not necessarily require statistically significant findings. It should be noted that, to my knowledge, this is the first application of metaregression analysis to health care efficiency. However, this technique has been previously applied to efficiency studies in two other industries, namely agriculture and urban transport. Thiam et al. (2001) analysed 32 studies (51 models) in developing country crop farming (rice, maize, etc); Brons et al. (2005) analysed 33 studies in urban transport (buses, ferries, trams, metros, etc.), and Bravo-Ureta et al. (2007) analysed 167 studies (569 models) in agriculture (rice, wheat, vegetables, dairy, pigs and so on in many different countries). This study is a valuable contribution in that it is the first to study health care and that it looks at a relatively large data set (253 models from 95 studies) where the production units use a much more uniform technology (relative to these other studies that consider a much broader range of production activities).

2.2.2

Construction of the meta-dataset

The meta-dataset was constructed using a two-stage approach, a preliminary search, followed by a systematic review and key data entry. The preliminary search was conducted as follows. First, relevant studies were identified via the main economic research database (ECONLIT), web of science (WOS) and PubMed, in which keyword searches such as “efficiency”, “productivity”, “hospital”, “health care”, “health centre”, “data envelopment analysis”, “stochastic frontier”, “production frontier”, and “cost frontier” are used. Each relevant paper found via these three sources was then explored for references to other studies that might have been missed by the search or simply not covered in either ECONLIT or WOS or PubMed. These additional papers were then obtained from the respective journals or via standard web search engines (e.g. Google). This search resulted in more than 220 studies, covering the period of 1983-2008 (as of July, 2008). The majority are published papers in journals or chapters of books/reports. Some studies are working papers. Finally, efficiency studies on health care facilities other 35


than hospitals (such as physicians, hospital departments/wards, nursing homes, and health districts) were removed from the list. This exercise filtered out more than 120 papers and the three-step preliminary search process was completed. The second stage - systematic review - involves the critical appraisal of individual studies to identify the valid and applicable efficiency models. It is necessary because not all the models from the preliminary search papers could be included in the final meta-dataset. Either information with respect to model specifications and/or estimation techniques was unavailable or not clearly explained. The first step was to include only studies in English that were available as of July, 2008. Each paper was then carefully reviewed to determine its research questions, country/region in question, data years, analytical methods, model specifications, analytical results, validity and robustness of techniques, findings and policy implications. Then only the studies that supplied sufficient information on the model specification as well as estimated efficiency scores were included. Several hospital efficiency studies do not report estimated efficiencies because their main focus are factors that influence (in)efficiency level (see for examples, Hao & Pegels, 1994; Hollingsworth & Parkin, 1995; Morey, et al., 1995; Prior, 1996; Mobley & Magnussen, 1998; Gerdtham, et al., 1999b,a; Cremieux & Ouellette, 2001; Li & Rosenman, 2001; Brown, 2003). However, they account for less 10% of the preliminary search list. Many studies apply different approaches to the same hospital dataset (Linna & Hakkinen, 1998; Linna, 1998; Webster, et al., 1998; Lopez-Casanovas & Saez, 1999; Chirikos & Sear, 2000; Jacobs, 2001; Gannon, 2005; Barbetta, et al., 2007) or use different hospital data sets for comparison (Grosskopf & Valdmanis, 1993; Mobley & Magnussen, 1998; Dervaux, et al., 2004; Steinmann, et al., 2004; Linna, et al., 2006). The final meta-data set consists of 95 studies, from 1987-2008. Since many studies utilise different methods of estimation, and/or use more than one data sets, and/or apply several models to the same data sets, 253 estimated models (i.e. observations) can be extracted from these 95 studies. The meta-regression analysis is then performed on the meta-data set. Short-listed studies appeared in various types of journals (around 40 different journals). However, these sources can be grouped into five main categories. Studies appear in health economics journals, including Health Economics and Journal of Health Economics. Health and medical related journals have published a large number of hospital efficiency studies, of which thirty-six are included in the metaanalysis. This reflects the interest of medical and health professionals on efficiency issues. Hospital efficiency research also fits the publication criteria of various Eco36


nomics and Management journals, with the former published 22% and the latter 20% of the studies in concern. Six papers appear in other journals, including those with mathematical orientation and those of general interests. Four papers included in the meta-study are unpublished working papers. The complete list of the studies included in the meta-data set is presented in Appendix A. Of all the countries analysed, the US has the highest number of studies, accounting for 40%, followed by European OECD countries (close to 38%). Five studies focus on other non-US OECD countries (Japan, Taiwan and Australia). Research on hospital efficiency of other countries, mainly developing countries, accounts for 17% of all studies and most of them were published in recent years (from 2004-2007). The reason for such a distribution is the availability of data. Most OECD countries have well established information systems for health care management and data is generally made available to analysts. In developing countries, data deficiency, especially at the firm level, is often due to both less-developed information systems and lack of transparency, which are well documented. The distributions of included studies by year and by country can be seen in Figure 2.8(a) and 2.8(b).

2.2.3

Model specification

In the meta-regression, the dependent variable is the mean of technical efficiency score. Two thirds of the studies reported mean efficiency while the rest reported either group mean efficiencies or individual hospital efficiencies (Bitran & Josep, 1987; Grosskopf & Valdmanis, 1993; Lynch & Ozcan, 1994; Burgess & Wilson, 1995; Chang, 1998; O’Neil, 1998; Al-Shammari, 1999; Lopez-Casanovas & Saez, 1999; Sommersguter-Reichmann, 2000; Athanassopoulos & Gounaris, 2001; Osei, et al., 2005; Ramanathan, 2005; Renner, et al., 2005; Zere, et al., 2006; Arocena & GarciaPrado, 2007; Goncalves, et al., 2007; Hajialiafzali, et al., 2007; Masiye, 2007). The latter appeared in studies using small sample sizes (around 30 observations) and mean efficiency can be calculated by taking the average of reported efficiency scores. The former is a typical reporting style of studies that focused on comparing efficiencies of different hospital groups, such as by location, ownership type and/or year. Mean efficiency scores are then obtained by taking weighted average of groups’ mean efficiencies with the weights being the group sizes. Many studies reported technical, scale, allocative and cost efficiencies, of which cost and technical efficiency are somewhat comparable. Several non-parametric studies estimated cost efficiency using total cost as a single input variable, rather 37


Figure 2.7: Distribution of studies by year and country

1987

2

1992

1

1993

1

1994 1995

2 1

1996

1997

7 1

1998

9

1999

9

2000

10

2001

9

2002

3

2003

3

2004

8

2005

8

2006

9

2007 2008

8 1 (a) Distribution by year

European OECD 37.89%

United States 40%

NonOECD 16.84%

Other OECD 5.26%

(b) Distribution by country

38


than using input quantities and input prices in the standard manner. Although the total cost efficiency scores obtained are not strictly equivalent to those obtained using the standard method, they are identical when all hospitals face the same input price vector. It is assumed here that the primary studies using the total cost efficiency have made this assumption to ensure the comparability of cost efficiency estimates7 . In order to capture the difference between cost and technical efficiencies, a cost-efficiency dummy (COST − EF F ) was included in the meta-data. Exogenous variables included in the meta-regression were chosen based on approaches and model specifications in the primary studies. They include total number of variables included in the frontier model (including inputs, outputs and control variables), sample size, dummy variables to capture the type of data used (crosssection versus pooled panel data), analytical approaches (parametric versus nonparametric), orientation (input versus output), and model specifications (functional form and efficiency distributions). The first explanatory variable is the number of variables, which includes all inputs, outputs and control variables included in the model. Their squared terms and cross products were excluded because they represent the choice of functional form. Explanatory variables used to explain efficiency (in the one-stage or two-stage estimation approaches) were not included in the count because they do not alter the dimensions of the production space8 . Most studies incorporate around 6 to 9 input and output variables plus several control variables (apply only for parametric studies). Some notable exceptions include Bitran & Josep (1987); Jacobs (2001) with 15 output variables and Jacobs (2001) with 17, Maniadakis et al. (1999) using 8 input variables; Ferrier & Valdmanis (1996); Frech & Mobley (2000); Fuiji (2001) taking into account more than 10 environmental factors. Sample size is generally the number of individual hospitals included in the primary study. However, many studies estimate frontier models using panel data in a cross-sectional fashion, i.e. they pool the panel to construct one frontier, instead of estimating a separate frontier for each year. For those cases, sample size is the total number of observations, usually equal to number of individual hospitals multiplied by the number of years for balanced panel. Less than 20% of the studies applied frontier techniques on a sample of more than 500 hospitals (as shown in Figure 2.8). 7 This assumption might be more defensible than considering cost efficiency as technical efficiency. Considering this efficiency as technical efficiency means making a (very strong) assumption, that allocative efficiency is unity, i.e. all firms use optimal mix of inputs, which is unlikely to be the case. 8 In the one-stage estimation, explanatory variables can effect the numerical estimation of efficiency when they are included in the specification of the efficiency term. However, none of the included studies make use of this specification

39


A third of them are studies that use pooled panel instead of cross sectional data, including the study that has the largest sample, close to 4,800 observations (Deily, et al., 2000). It is expected that a pooled panel sample has less variation than a cross sectional sample. One hospital will be observed more than once, and thus variation from year to year is expected to be smaller than variation between different hospitals. This can potentially produce higher average efficiency scores. At the other end of the distribution, around 12% of the studies have a sample size of no more than 30 hospitals (Chang, 1998; O’Neil, 1998; Al-Shammari, 1999; SommersguterReichmann, 2000; Osei, et al., 2005; Ramanathan, 2005; Zere, et al., 2006; Arocena & Garcia-Prado, 2007; Goncalves, et al., 2007; Masiye, 2007; Kirigia, et al., 2008). Figure 2.8: Sample size and number of variables

Number of variables

25

20

15

10

5

0 0

1000

2000

3000

4000

5000

Sample size

Among the two main approaches, the parametric approach is used in approximately 25% of included studies. Twenty-six (around 45% of all parametric studies) use first order functional forms, including Cobb Douglas or some types of linear function. The rest use second order functional forms, including translog or translog with some coefficient restrictions, and cubic functions. They are used generally in studies with large samples, (average sample size of 1560), compared to those studies using Cobb-Douglas or linear form (average sample size of 507). This might be due to the fact that translog function consumes a large number of degrees of freedom and 40


hence requires a larger sample to achieve robust estimation. More than half of the parametric studies assume a half normal distribution for the efficiency term; 35% apply a truncated normal distribution and the rest use an exponential distribution. For papers using a non-parametric approach, the main choices involve orientation and returns to scale assumptions. The distribution of studies using constant and variable returns to scale is quite even, with the former accounting for 43% and the latter 57%. A large number of studies have chosen the input orientation based on the argument that hospitals (especially public hospitals) cannot choose their level of output, which depends on demand for health services. Hospitals then try to conserve inputs, which makes input (or cost) minimisation a reasonable assumption for DEA estimation. Recently, some countries have changed their financing arrangement for health service providers: instead of payments based on cost history or per diem, reimbursement for hospitals are based on output volume and sector average cost with a cap on the overall budget (i.e. global budget). The assumption of maximising output level, given the amount of health resources available, has been chosen in some studies to reflect this change. Ten variables are specified to capture the model options discussed above, of which eights are dummies. Apart from the two variables of sample size and number of observations, all other regressors are dummies that explain different methodological choices. The base case for the model is an output-oriented parametric cross-sectional model using a first order functional form with an efficiency term following a halfnormal distribution. Detailed descriptions of the variables are presented in Table 2.2. Table 2.3 contains some descriptive statistics. The average efficiency score from all studies is 84.1, with the highest being 98.9 and lowest 52. Interestingly, these come from the same study that uses different input and output variables and model specifications (Kibambe & Kocht, 2007). This is a striking example of how the choice of models and variables can significantly alter efficiency estimates, which leads one to question the degree to which policy should be influenced by this type of performance indicator. Amongst all reported cases, 82 estimated cost efficiency, of which a large proportion drew some conclusions about the possibility of cost saving or reimbursement for hospitals based on cost efficiency (for examples Puig-Junoy, 2000; Sahin & Ozcan, 2000; Fuiji, 2001; Giokas, 2001; Zere, et al., 2001; Kirigia, et al., 2004; Harrison & Ogniewski, 2005; Osei, et al., 2005; Renner, et al., 2005; Masiye, 2007; Lee, et al., 2008). The choice of functional form is driven by the possible impacts of the two contin-

41


Table 2.2: Variable names and definitions Variable name

Variable definition

EFF

Efficiency score

Reported average efficiency scores (on 0-100 scale)

COST-EFF

Cost efficiency

This dummy captures the difference between cost and technical efficiency. It takes value of 1 if the observations are cost efficiency, 0 if technical efficiency. This variable is designed to capture the effect, if any, of PANEL

Pooled panel

using pooled panel data instead of cross sectional data in

data

efficiency analysis. It takes value of 1 if pooled panel is used, 0 otherwise.

SIZE

Number of

Number of (hospital) observations included in the

observations

primary studies. Total number of output, input, input price, and control

DIMENSION

Number of

variables included in the frontier model. This does not

variables

include control variables used in the second stage of analysis and higher orders of variables used. Dummy variable to capture the method used in efficiency

NON-PARA

Method dummy

analysis. It takes value of 0 if parametric approach is chosen, and 1 if non-parametric approach.

INPUT-ORT

Orientation

This dummy takes value of 1 if input orientation is used

dummy

(including cost function), 0 otherwise. Returns to scale can be variable or constant returns to

CRS

Returns to scale

scale. It takes value of 1 if constant returns to scale, 0 otherwise.

2ND-ORDER TRUNCATED EXPONENTIAL

Functional form

This takes value of 1 for the second order functional form, 0 otherwise.

Truncated

This takes value of 1 if efficiency score is assumed

distribution

truncated normal distribution.

Exponential

This takes value of 1 if efficiency score is assumed

distribution

exponential distribution.

42


Table 2.3: Descriptive statistics Variable

Mean

Median

Std. Dev.

Min

Max

EFF

83.8289

86.1

9.4343

52

98.9

COST-EFF

0.3636

0

0.4820

0

1

PANEL

0.1383

0

0.3459

0

1

464.3715

131

809.1523

15

4739

DIMENSION

9.0514

9

3.8018

3

23

NON-PARA

0.7154

1

0.4521

0

1

INPUT-ORT

0.8735

1

0.3330

0

1

CRS

0.3557

0

0.4797

0

1

2ND-ORDER

0.1462

0

0.3541

0

1

TRUNCATED

0.0909

0

0.2881

0

1

EXPONENTIAL

0.0356

0

0.1856

0

1

SIZE

uous variables, dimension and sample size. Dimension is expected to have a positive impact on efficiency estimates while sample size is the opposite. Their effects are likely to be non-linear and diminishing when the dimension and the sample size increase. Two functional forms that appear to suit this expectation are quadratic and linear-log models. The specifications are as follows: The quadratic function: 1 EF F = α0 + α1 (COST − EF F ) + α2 (P AN EL) + α3 (SIZE) + α33 SIZE 2 2 1 2 + α4 (DIM EN SION ) + α44 DIM EN SION + α5 (N ON − P ARA) 2 + α6 (IN P U T − ORT ) + α7 (CRS) + α8 (2N D − ORDER) + α9 (T RU N CAT ED) + α10 (EXP ON EN T IAL) + ε.

(2.1)

The linear-log function: EF F = α0 + α1 (COST − EF F ) + α2 (P AN EL) + α3 ln (SIZE) + α4 ln (DIM EN SION ) + α5 (N ON − P ARA) + α6 (IN P U T − ORT ) + α7 (CRS) + α8 (2N D − ORDER) + α9 (T RU N CAT ED) + α10 (EXP ON EN T IAL) + ε.

(2.2)

In both cases, ε is the statistical noise, assumed to be identically and independently distributed, ε ∼ N [0, σ 2 ]. Arguably, the quadratic function might not be the ideal candidate. For dimension, it is required that the function has a non negative 43

2.3. RESULTS AND DISCUSSION

derivative throughout the domain. That is even when the dimension value becomes very large, one should not see the mean efficiency decline. Rather, it increases with a rate approaching zero. Unfortunately, the quadratic function with a maximum does not fulfil this requirement, as shown in the marginal effect of dimension on efficiency estimates: ∂EF F = α4 + α44 DIM EN SION. ∂DIM EN SION

(2.3)

In order to have the positive and diminishing marginal impact on efficiency estimates, the expected sign for α4 should be positive and α44 negative. As dimension increases, the marginal effect will eventually becomes negative. By symmetry, the same problem is encountered with sample size. The linear-log function does not have the same problem. The marginal effect of dimension on efficiency estimates is expressed as: 1 ∂EF F = α4 . ∂DIM EN SION DIM EN SION

(2.4)

And the marginal effect of sample size on efficiency is: ∂EF F 1 = α3 . ∂SIZE SIZE

(2.5)

When dimension increases, a positive α3 will ensure the marginal effect approaching zero but not turning negative. The opposite happens to size; a negative value of α4 allows the marginal effect of size on efficiency to approach zero from below as size increases.

2.3

Results and discussion

Both the quadratic and linear-log models are estimated using ordinary least squares regression. It is not necessary to use Tobit or limited dependent variable procedures, which are usually used when the dependent variable is bounded. There is no mean efficiency of 0 or 1 (or 100 in the percentage scale) in the meta-data and thus, Tobit estimates are exactly identical to its OLS counterparts. Table 2.4

44


presents results for the two estimated models. Most estimated coefficients have expected signs, although some are not significant at 10% level or better9 . Table 2.4: Estimated results Linear-log COST-EFF

Quadratic -0.693797

COST-EFF

(1.698731) PANEL

(1.721792)

3.965262**

PANEL

(1.756144) ln(SIZE)

-0.558482 2.575849 (1.773979)

-3.336676***

SIZE

-0.009742***

SIZE SQ

0.000004***

(0.5515723)

(0.002296) (0.000001)

ln(DIMENSION)

6.594401***

DIMENSION

(1.418766)

3.042650*** (0.653337)

DIMENSION SQ

-0.254391*** (0.061573)

NON-PARA

-2.735875

NON-PARA

(2.128559) INPUT-ORT

2.769407

CRS

-3.776708***

(2.161424) INPUT-ORT

(1.801119)

1.69552

2ND-ORDER TRUNCATED

2.604963 (2.575207)

4.462415

EXPONENTIAL

(3.222226) CONSTANT

0.207161 (2.402458)

2.717017 (2.396168)

EXPONENTIAL

-3.679944*** (1.216605)

(2.303395) TRUNCATED

2.897314 (1.820180)

CRS

(1.200323) 2ND-ORDER

-2.559747

5.439465* (3.276001)

86.6502***

CONSTANT

(3.885375)

71.490470*** (3.567707)

F-statistics

6.271594

F-statistics

4.846126

R-squared

0.205818

R-squared

0.195046

Adjusted R-squared

0.173000

Adjusted R-squared

0.154798

***, **, * denote significant at 1%, 5% and 10% level. The base case is a parametric, first order production frontier with the efficiency term assumed to follow a half normal distribution.

It is noted that neither of the linear-log nor the quadratic specification is nested within the other. This makes the direct comparison through coefficient tests (for 9 The interaction between size and dimension was considered in one of the original model but its coefficient is very small in magnitude and not statistically significant. Therefore, it was left out of the final models.

45


instance, the F-test or the maximum likelihood test) impossible. Therefore, the J-test is conducted to decide between the two models. This test is used for testing the specification of a non-linear regression model against the evidence provided by a non-nested alternative hypothesis (MacKinnon, et al., 1983). The procedure involves two steps: (i) estimate each model and save their predictions, (ii) each prediction is included as a regressor in the competing models. A significant coefficient of the prediction indicates the model, in which the prediction is included, is not correctly specified. If the prediction of model A is significant in model B, while the converse is insignificant then model A is preferred over model B, and vice versa. However, if the predictions of both models are either insignificant or significant in the other model, then neither of them is preferred. The t-ratio of the quadratic functions prediction in the linear-log models is 2.25 while that of the linear-logs function prediction in the quadratic model is 3.69. Both predictions are significant at the 5% level although if 1% level is used, the linear-log model is preferred. Between the two models, the linear-log also appears to fit the data better than the quadratic as indicated by R-squared and adjusted R-squared. Additionally, it suits the expectation of diminishing and asymptotic-to-zero marginal effects of dimension and sample size while the quadratic function does not have this property. Therefore, the linear-log is judged to be superior to the quadratic function. The discussion of results will be based on the linear log estimation. The estimated coefficient for SIZE, capturing the effect of sample size on mean efficiency, is negative while that for DIM EN SION , the variable that represents the influence of number of variables on efficiency, is positive. They are both significant at 1% level and in line with expectations. The negative sign of the coefficient for SIZE indicates that, everything else being equal, increasing the number of observations will yield a lower mean efficiency score. The marginal effect of SIZE is only -0.025, when evaluated at the sample median sample size of 131. However, at smaller sample sizes the marginal effect is larger. For example, a sample size of 30, yields a marginal effect of -0.111, suggesting that the addition of an extra 9 observations could lead to a reduction in mean efficiency of one percentage point. The effect of DIM EN SION on average efficiency score is more substantial. The marginal effect is 0.733, when evaluated at the sample median of 9 variables. However, as the number of variables decreases the marginal effect is larger. For example, a value of 3 yields a marginal effect of 2.198, suggesting that the addition of an extra variable could lead to an increase in mean efficiency of more than two percentage points. These larger effects at low values of SIZE and DIM EN SION 46


are evident in Figure 2.9, where predicted mean efficiencies are plotted for various values of these variables, and in Figure 2.10 where marginal effects are plotted. Figure 2.9: Predicted mean efficiencies at the median 100

95

95 Efficiency score

100

90 85

80 75

90

85 80 75 70

70 26

131

236

341

445

550

2

655

Sample size

5

8

11

14

17

20

23

Number of variables

(a) Predicted efficiencies w.r.t sample size

(b) Predicted efficiencies w.r.t dimension

As shown in Figure 2.10, as the number of variables included in the model increases, the average efficiency predictions drop quite quickly when the model size is fairly small. Inclusion of an extra variable into a model with more than 10 variables does not alter the average efficiency score very much. The opposite effect is observed for sample size: doubling the number of observations for a sample of more than 150 observations does not change efficiency estimates by very much. Zhang & Bartels (1998) also arrived at the similar conclusion on the sample size effect. They observed that when sample size was large, the mean technical efficiency shows little change and the mean efficiency tends to be constant after a threshold. Therefore, correcting for sample size has a major impact on the assessment of average efficiencies of an industry (Zhang & Bartels, 1998). The coefficient of the COST −EF F variable is expected to be negative, since cost efficiency is usually lower than technical efficiency, other things being equal, due to the fact that allocative inefficiency is also captured. A technically efficient hospital is not necessary allocatively efficient because it might use the wrong mix of inputs given their prices, thus its cost may be larger than it should be. The estimated coefficient is -0.69 and statistically insignificant, suggesting that this effect is small in these data. As expected, the coefficient for P AN EL variable returns a positive sign, suggesting the use of pooled panel tends to produce higher average efficiency scores, of around 4 percentage points. A possible explanation for this is that a hospital is 47


Figure 2.10: Marginal effect of sample size and dimension on efficiency estimates 10

120

240

360

480

3.5

600

0.00

3.0

-0.05

2.5

-0.10

2.0

-0.15

1.5

-0.20

1.0

-0.25

0.5

-0.30

0.0

-0.35

2

(a) Marginal effect of sample size

6

10

14

18

22

26

30

(b) Marginal effect of dimension

observed more than once in a pooled panel, and thus variation from year to year is expected to be smaller than variation between different hospitals when cross sectional is used. This can potentially produce higher average efficiency scores. From the discussion in Section 2.1.4, it is expected that, other things being equal, returns to scale (CRS) and functional form (2N D − ORDER) have predictable effects on efficiency score. Imposing CRS on the model tends to reduce efficiencies; while a higher order and more flexible functional form can predict higher efficiencies because it fits the data more tightly. However, the direction of the effect on efficiencies of the other model specification variables, such as approaches (N ON − P ARA), orientation (IN P U T − ORT ) and distribution assumption (T RU N CAT ED and EXP ON EN T IAL) are not unambiguous. Estimated coefficient for the variable CRS displays a negative and significant effect on mean efficiency score. The magnitude of the CRS coefficient implies that choosing a CRS technology instead of VRS will reduce the estimates of mean efficiency by around 4 percentage points. Whereas the expected effect of the CRS assumption on efficiency estimates is supported by the estimated model, the other results are less conclusive. The positive sign on the functional form (2N D −ORDER) coefficient is as expected. However, it is not statistically significant. This result might be the consequence of the inclusion of restricted translog functions into the “second order functional form” category. When a subset of the second order terms in the translog function are restricted to be

48

2.4. IMPLICATIONS FOR HOSPITAL EFFICIENCY STUDIES

zero, the flexibility advantage partly disappears, and hence it can behave in a similar manner to a first order function. Interestingly, the meta-analysis of farming industry by Bravo-Ureta et al. (2007) also found that the relationship between functional form and mean efficiency is inconclusive. The variable that captures the difference that non-parametric methods make on efficiency estimate compared to their parametric counterparts has a negative coefficient but it is also not statistically significant. This suggests that there is not enough evidence to say that studies using parametric method generally yield higher efficiency scores. The meta-analysis in urban public transport by Brons et al. (2005) also arrives to the same conclusion. It appears that the added flexibility of DEA and the noise component in SFA are cancelling each other out in the analysis. Similarly, the estimated coefficient for IN P U T −ORT variable displays a positive sign but is not statistically significant. This might indicate that samples used by the primary studies included in the meta-analysis are characterised by roughly the same number of hospitals in the increasing (small) and decreasing (large) returns to scale regions. The three distributions used in hospital efficiency studies are captured by two dummy variables: T RU N CAT ED and EXP ON EN T IAL; the half normal distribution is the base case. Their positive coefficients suggest that models using either truncated or exponential distributions, on average, yield higher efficiency score than those using the half normal distribution, and with the magnitude of 2.7 and 4.5 percentage points, respectively. This might be a consequence of the fact that the centre of mass of the exponential distribution is located near zero. However, neither of the two coefficients is statistically significant at the 10% level or better.

2.4

Implications for hospital efficiency studies

This section will demonstrate how the analysis and results of the meta regression can be used to guide model selection, to correct for potential biases due to sample size, variable choice and/or model specification, and to improve the comparison of hospital performance in different countries or states/regions. It can also be useful in comparing different study results in order to generalise the impact of various policy decisions on efficiency of the hospital industry. It is noted that comparing the performance of hospital industries in different countries does not imply the industry of higher mean efficiency is (absolutely) more efficient than the others that have 49


lower efficiency scores, except when hospitals of these countries are pooled as one sample in the analysis. It only indicates that within the hospital industry of the country, individual hospitals are, on average, closer to that country’s frontier (Zhang & Bartels, 1998). Firstly, the analysis and results suggest that the omission of inputs and outputs should be avoided as much as possible. This is not an easy task, especially when secondary sources are used. Data can have missing values for important variables or some required variables are not collected/constructed. Data deficiency can lead to small sample size, variable omission and/or aggregation, which usually leads to bias in efficiency estimated, as discussed in Section 2.1. Users of frontier methods, are therefore, can utilise some strategies to deal with those biases, for instance choosing between keeping the variables and dropping observations with missing values. The estimated coefficients of SIZE and DIM EN SION presented in Table 2.4 give an indication of the trade-off between dropping variables and removing observations: the bias on efficiency estimates can be much larger when a important variable (such as input or output) is missing. However, if many observations are removed, there comes a point where it is better off dropping the variable, especially when the value of dimension is quite large. A back-of-the-envelope calculation using the estimated coefficients reveals that dropping one from a set of ten variables has the same (negative) effect on estimated efficiency as removing 10 observations from a sample of 50, or 55 observations from a sample of 200. In the meantime, if the dimension is 5, losing one variable has the equivalent effect of removing half of a fifty-observation sample, or a-third of a two-hundred-observation sample. Second, the average efficiency estimates of a hospital industry can be adjusted post-estimation if the sample used is significantly smaller than the size of the industry, or some important outputs and inputs variables are not included in the estimated model. This adjustment, however, does not affect the ranking of hospitals because it is based on the average efficiency scores. The estimation can be made using the following formula: EF F adjusted = EF F estimated + α3 ln SIZE used − ln SIZE desired

+ α4 ln DIM EN SION used − ln DIM EN SION desired

(2.6)

in which, SIZE used and SIZE desired correspond to the actual sample size used and the “desired” number of observations, that is when all hospitals in the industry are included. By symmetry, this applies to the number of observations,

50


DIM EN SION used and DIM EN SION desired where DIM EN SION used is the number of variables actually included in the model while DIM EN SION desired represents the total number of inputs and outputs when no variable omission exist. Lastly, the estimated results can be used to compare hospital performance in different countries or states/regions. It is possible to use the coefficients presented in Table 2.4 to “correct” and compare the efficiency estimates of different studies (for instance, of various sample size and using different models). For instance, Linna et al. (2006), in their comparative study of Finnish and Norwegian hospitals, measured cost efficiencies by the non-parametric DEA approach. Two separate frontiers were estimated to predict within-country efficiencies. The studies applied two sets of costs, one adjusted for exchange rate differences and the other adjusted for input prices. Their argument is that cross-country differences in health care prices are not necessarily consistent with differences in general prices, hence input prices can be used to equalise the cost differences. The mean efficiencies indicates that the two hospital industries have almost equivalent levels of mean efficiency; both had mean VRS efficiencies of 92 and mean CRS efficiencies were 83 and 86 for Finnish and Norwegian hospitals, respectively. From the technical point of view, this comparison might not be totally convincing because the efficiency estimations were based on different sample sizes. The metaregression results suggest that countries with a larger sample of hospitals tend to have higher mean efficiency. The results in Table 2.4 can be used to adjust these estimates for the differences in modelling attributes. For instance, the base case at the median is chosen as the benchmark, and adjusted mean efficiencies are produced as follows:

EF F adjusted = EF F reported + α1 (COST − EF F ) + α2 (P AN EL) + α3 (ln SIZE − ln 131) + α4 (ln DIM EN SION − ln 9) + α5 (N ON − P ARA) + α6 (IN P U T − ORT ) + α7 (CRS) + α8 (2N D − ORDER) + α9 (T RU N CAT ED) + α10 (EXP ON EN T IAL) .

(2.7)

Here 131 and 9 are the values of size and dimension of the median; variables SIZE, DIM EN SION and all dummies takes their values from the reported model specification. For instance, Linna et al. (2006) uses a DEA input oriented model with five variables (four outputs and one input). Hence, DIM EN SION takes a value of 5, SIZE is 51 for Norway and 47 for Finland, COST −EF F , N ON −P ARA 51


and IN P U T − ORT dummies are 1, P AN EL, 2N D − ORDER, T RU N CAT ED and EXP ON EN T IAL are all 0. Table 5 presented the reported efficiencies from the study and their respective adjusted scores using the estimated coefficients of the linear-log model. The predicted efficiencies for both countries changed quite significantly. Overall, Finnish hospitals performed slightly better with respect to technical efficiency (90.88 vs. 90.61) but are less scale efficient than their Norwegian counterparts (85.94 vs. 89.21)10 . Table 2.5: Mean efficiencies (Linna et al., 2006) Reported efficiency

Predicted efficiency

Sample size

CRS

VRS

Scale effect

CRS

VRS

Scale effect

Norway

51

86.00

92.00

93.48

80.83

90.61

89.21

Finland

47

83.00

92.00

90.22

78.11

90.88

85.94

Taking another example in which the authors evaluated hospital performance using casemix adjusted outputs (Grosskopf & Valdmanis, 1993). This is one of the first papers in the hospital literature using casemix to take into account differences in severity and patients characteristics. The sample includes hospitals from the states of New York (49 hospitals) and California (59 hospitals). They found that New York hospitals, on average, are 6.9% more technically efficient than California hospitals but less scale efficient. After adjustment, the mean technical efficiencies of both New York and California hospitals are now slightly higher than the reported under both models while the effect of scale inefficiency is larger in both states. The adjusted result also implies that hospitals in New York are 7.5% more efficient than those in California (see Table 2.6). It is also possible to further look at the differences between estimated efficiencies from primary studies and the predicted efficiencies by the model through the change in rankings of different hospital sectors based on these two sets of efficiencies. Across the whole sample, close to 9% have changed the estimated efficiency by more than 10% while about 40% do not change their ranking substantially. This is reflected by a high Spearman’s correlation coefficient, around 93%. Table 2.7 presents the ranking of some low, medium and high efficiency estimates in various studies, associated with their new ranking based on the predicted efficiencies using the linear-log model. It is observed that the ranking has changed quite significantly for many observations. Some bottom performers show up in the middle range while the performance of some top-rated observations appears to be less impressive. However, it is noted that the significant change in ranking appear to happen with observations in the middle 10

The differences are not large in this case because the sample sizes are quite similar.

52

2.5. CONCLUDING REMARKS

Table 2.6: Mean efficiencies (Grosskopf & Valdmanis, 1993) New York (N=49) Adjusted output (8 variables) Reported efficiency

efficiency (predicted)

(7 variables)

Unadjusted output

(8 variables)

(7 variables)

CRS

VRS

CRS

VRS

CRS

VRS

CRS

93

86

92

88

87

85

86

86

92.47

efficiency (reported) Implied scale

Adjusted output

VRS Implied scale Predicted efficiency

California (N=59)

Unadjusted output

95.54

84.76 88.72

95.65 93.66

85.88

97.7 88.92

91.7

83.14 93.5

100 87.04

83.26 95.66

Note: The model with 8 variables includes 4 unadjusted outputs while the one with 7 variables includes 3 casemix adjusted outputs. Both of them have 4 input variables.

ranked group rather than the lowest and highest groups. For instance, observation of rank 89th by the reported efficiency has its adjusted efficiency of 98.83%; that is 12.5% different compared to the reported score. Similarly, one of the lower ranked observations (number 138) jumps into the top-fifty (number 37) with a predicted efficiency change of 11%. The comparison of developing and developed countries is presented in Table 2.8. The reported scores in the primary studies suggest that on average, hospitals in developing countries are much less efficient than those of the developed world, around 15.4% versus 9.5% inefficiency. It can be seen that the story changes with the adjusted efficiency predictions. The developing world is now not so far behind the developed countries, with less than one percent difference. It is hypothesised that this large change is primarily a consequence of developing country studies having access to data sets with sample sizes that are smaller relative to developed country studies11 .

2.5

Concluding remarks

This chapter provides a meta-regression analysis on the hospital efficiency literature, with the primary aim of explaining the influence of methodological choices on efficiency estimates. This contributes to the hospital efficiency literature by taking the systematic analysis of the literature a step further, by pooling all studies to11

However, as discussed earlier, it should be emphasised that comparisons of mean efficiencies across countries (or across any groups) can be misleading unless a single reference frontier is used.

53


Table 2.7: Efficiency prediction and ranking Old rank

Reported efficiency

Predicted efficiency

Difference

New rank

1

98.110

98.860

-0.750

3

2

97.800

98.757

-0.957

5

3

97.400

95.392

2.008

29

4

97.230

99.482

-2.252

2

5

96.650

99.868

-3.218

1

31

93.020

92.790

0.230

49

32

93.000

95.538

-2.538

26

33

93.000

91.652

1.348

57

34

92.990

90.732

2.258

65

35

92.700

93.076

-0.376

47

86

88.000

85.881

2.119

107

87

88.000

85.667

2.333

111

88

87.900

89.579

-1.679

72

89

87.890

98.825

-10.935

4

90

87.670

73.041

14.629

182

150

82.000

81.005

0.995

144

151

81.970

88.019

-6.049

86

152

81.830

86.557

-4.727

100

153

81.830

85.884

-4.054

106

154

81.710

79.216

2.494

156

237

60.000

54.074

5.926

237

238

58.100

53.052

5.048

238

239

56.800

44.299

12.501

241

240

54.000

48.512

5.488

239

241

52.000

47.121

4.879

240

Note: observations are ranked from most (rank = 1) to least efficient.

Table 2.8: Efficiency predictions for developing and developed countries Developing countries

Developed countries

Reported

84.66

90.42

Predicted

82.56

83.44

54


gether into a statistical analysis in order to examine the direction and magnitudes of the effects of modelling choices on mean efficiency scores. Various areas of future work are worthy of consideration. First, it might be of interest to examine the influence of modelling choices on the variance of efficiency estimates. It is obvious that the narrower the spread of the estimates, the more confidence is placed in the point estimates, in this case, mean efficiency. Second, an aspect of model specification that is not captured in this study is the range of output and input variables used in the primary studies. There are as many input and output definitions as the number of studies included, and accounting for their heterogeneity is a sizable challenge. Moreover, the model can be enriched by the inclusion of some control variables that reflect regions/countries or characteristics of the health care systems in the reported cases. Lastly, it might be useful to consider separating the number of variable measures into separate inputs, outputs and environmental variables; include dummies to capture differences from using DRG and/or quality-adjusted outputs, countries by living standard or health care system.

55

Chapter 3

Measuring labour efficiency in Queensland public hospitals

Health workforce performance is critical because it has an immediate impact on health service delivery and ultimately on population health. (WHO, 2006) Health care is a labour intensive industry and as a result health worker salaries and entitlements take up a large proportion of health care expenditure. According to the World Health Organisation, the world average proportion of government health expenditure paid to health workers is 42.2% (WHO, 2006). Americas and Eastern Mediterranean countries spend around 50% of their health expenditure on labour, followed by Western Pacific, including Australia, and Europe which pay 45% and 42.3% of their health budget, respectively, to health workers. It is argued that for productivity and efficiency enhancement to be achieved, better use of human resources is necessary. Against this background, it is remarkable that research on supply and utilisation of human resources has been overlooked in the health efficiency literature. To date, the main focus on hospital efficiency studies has been on either technical or cost efficiency. Many of these of studies utilise technical and cost efficiencies to shed light on policy issues such as how ownership and organisation structure induces efficiencies (see for examples, McKay, et al., 2002; Chang, et al., 2004; Dervaux, et al., 2004; Barbetta, et al., 2007; Lee, et al., 2008); the effects of financing and reimbursement on hospital behaviour and if particular mechanisms encourage efficiency seeking practices (Sommersguter-Reichmann, 2000; Biorn, et al., 2003; Liu & Mills, 2005; Aletras, et al., 2007); the efficiency enhancement pressure from external factors 56

3.1. THE PUBLIC HOSPITAL SYSTEM IN QUEENSLAND

such as market structure and competition forces (Puig-Junoy, 2000; Rosko, 2001a,b; Carey, 2003; Grosskopf, et al., 2004; Bates, et al., 2006; Ferrari, 2006a). Labour efficiency can be examined using an input orientated stochastic frontier. While it is theoretically established in the literature, applications of this approach are less than a handful, at best in the general efficiency literature (Kumbhakar & Tsionas, 2006), with only one example being identified in the primary care literature (Lordan, 2009). This method is conceptually similar to a cost frontier, in that they are both input-oriented, focusing on the conservation of resources used to produce the required outputs. However, unlike the cost frontier, it does so without any economic optimisation implied, instead technical optimisation is the focus. This approach is appropriate when labour cost make up a large percentage of total cost of the organisations. Furthermore, when the labour supply is constrained and the organisations do not have great control over the number of employed while facing fixed capital resources in the short run, the scope for efficiency improvement does not come from allocative efficiency but technical efficiency enhancement, especially from better utilisation of labour resources. The main purpose of this chapter is to analyse the labour efficiency of the public hospitals in the state of Queensland, Australia for the period 1996-2003. This period is characterised by hospitals funding based on a historical cost reimbursement model, which has since been replaced by a prospective casemix funding method in 2007. During this period, labour shortage, especially shortage in public hospitals, also gradually built up to become an economic and political issue. At the same time, Queensland has experienced rapidly population expansion, declining share of public hospitals and fast-growing private hospital sector.

3.1 3.1.1

The public hospital system in Queensland Queensland Health system

Queensland has the most widely geographically distributed health system in Australia owing to its geographically dispersed population. More than 48% of its population resides outside major cities, of which a half lives in rural and remote areas (Davies, 2005). Since 2006, public health services in Queensland have been delivered through 21 health service districts after the consolidation of the previous

57


37 districts1 . Each district is responsible for the management and delivery of health services to its communities. All health districts include hospital and community based clinical services such as oral health, child and youth health, community health, Aboriginal and Torres Strait Islander health, women’s health, mental health, home and community care services. As a part of the Australian public health system, Queensland Health is characterised by universal access to public hospitals and medical care. Indeed, it was the first state in Australia to introduce free and universal public hospital treatment following the nationalisation of the public hospitals in 1944. Funding for public health care services in Queensland is shared between the Commonwealth, the State and municipal governments, and the non-government sector. The amount of Commonwealth funding is governed by the Australian Health Care Agreement (AHCA) which includes both general and specific purpose grants. General grants are provided to Queensland public hospitals while specific purpose grants relate to initiatives such as quality and safety, mental health and palliative care, and are subject to a variety of conditions and performance criteria. The primary objective of the agreement is to secure access for the community to public hospital services based on the principle of freedom of choice to receive free-of-charge public health care services to ensure equitable access. Under the new AHCA 20032008, Queensland is granted A$8.021 billion, including A$7.7 billion for the general components. Another source of finance comes from private health insurance (PHI). Queensland, as well as other Australian States have experienced increasing levels of PHI coverage under the influence of an AHCA’s provision on adjusting funding to reflect movements of private health insurance. However, Queensland continues to have a low level of PHI coverage compared to the national average. For instance, over the period 2001-2007, around 40% of people in Queensland were covered by PHI for hospital care, compared to the national average of 45% (DHA, 2007a, page 9). By expenditure category, employee expenses account for approximate 60% of Queensland health’s total expenditure while supplies and services represent 22%. By service type, hospital expenditure (inpatient and outpatient services) is the largest expenditure component and accounts for around 64% of the total Queensland Health’s budget (QH, 2005). 1

The year 2008 saw another administration shake-up with 21 health districts merged into 15.

58


3.1.2

Queensland public hospitals

Public hospitals in Queensland are organised on a zonal self-sufficiency model, by role delineation and by network (Surrao, et al., 2002). There are a small number of large teaching hospitals (more than 500 beds), supplemented by medium sized base hospitals in regional centres and numerous small hospitals in remote and very remote locations. Together, the State has a network of 178 public hospitals with a total number of licensed beds of 10,106 and 277 primary and community health centres, and a workforce comprising of 43,785 full time equivalents, the third largest in Australia (ABS, 2007, table 2.2). Hospital size generally reflects the size of the population they serve as well as the complexity of procedures and treatment that is provided. Expenditure on public hospitals accounts for approximately half of Queensland’s public funding for health service delivery. Since 1998-99, there has been a substantial increase in recurrent expenditure on public hospitals in Queensland in order to match the rate of funding growth made by the Australian Government. In 200203, the Queensland Government funded 46% of the A$3 billion recurrent costs for public hospitals. Of the remaining 54%, the Australian Government’s contribution was roughly the same, and the rest (around 3.8%) came from non-government sources. Recurrent expenditure per person in public hospital services for Queensland increased from A$408 in 1998-99 to A$614 in 2005-06 but still lower than the national average (A$501 and A$665, respectively). The cost of treating public hospital patients include paying staff and fee for visiting medical practitioners, buying and operating medical technologies, buying daily medical supplies and goods and providing support services such as meals, cleaning, security and computer systems (DHA, 2007b). Queensland Health’s 178 public facilities and the Mater public hospital networks have been delivering an increasing amount of both inpatient and outpatient services. In the financial year 2005-06, Queensland Health recorded 750,317 admissions, with 49% of the same day. Around 90% of total admissions were provided without charge, with the balance to private and compensable patients. In the same year, 9.15 million outpatient occasions of service were provided. They include accident and emergencies (around 1.3 million), outpatient services (including specialist medical services, allied health services and ancillary, accounting for 33% of all non-admitted occasions of services), and other services. Outpatient services are provided free of charge for public patients. Over the ten-year period, from 1995-96 to 2004-05, outpatient

59


service grew by 38% while inpatient service grew by only 16% (details in Table 3.1). This can be explained by the fact that many diseases can now be treated under outpatient setting. The number of patient days have been fairly constant and slightly decreased, around 2% over 10 years. As the episodes of care increased, the decrease in patient days could be due to increasing efficiency in acute care. Table 3.1: Summary of separations by public hospitals Accrued

Episodes of

Total outpatient

% public

patient days

care

occasions of service

patients

1995-96

2,614,268

631,717

6,359,321

88.6

1996-97

2,556,605

646,425

6,743,027

89.3

1997-98

2,544,481

683,898

6,903,948

90.3

1998-99

2,539,383

707,227

7,212,528

91.5

1999-00

2,460,254

706,530

7,249,798

92.1

2000-01

2,447,346

687,952

8,414,029

92.4

2001-02

2,469,698

694,264

8,639,993

92.5

2002-03

2,468,230

701,753

8,722,659

92.4

2003-04

2,506,819

720,673

8,648,059

92.0

2004-05

2,559,589

733,231

8,758,590

92.6

Source: Queensland Health monthly activity collection (QH, 2006)

3.1.3

Funding models in Queensland public hospitals

Until 2007, Queensland public hospitals have been funded on a historical basis2 (Surrao et al., 2002; QH, 2008). The annual budget is calculated using a twoelement Hospital Funding Model3 . The fixed grant component does not depend on throughputs and aims at covering hospital overhead costs. The variable component is dependent on output targets and average price, which in turns depends on hospital groupings. The service volumes per annum (target) were capped at the district (multi-hospital) level. This model experienced eleven iterations with slight modifications from year to year since first introduced in 1995. The historical cost funding model is essentially a global budget modality, in which annual budgets are based on the previous period. Adjustments are usually 2

Although casemix has been used in Queensland Health, it is not for the funding purpose, but more as a management and information tool to allow benchmarking, to encourage performance improvement and ultimately, to achieve “fair, accessible, comprehensive and cost-effective public hospital system” (Surrao et al., 2002). 3 Outlier inpatient cases, outpatient occasion of services and emergency visits are funded through the (scheduled) fee for service avenue.

60


made for demonstrated and predicted changes in demand for health services, effects of inflation and possible new investment in equipment and personnel4 . In the Queensland Health setting, this method is reflected through target (quantity) and price setting norms for the variable component. At the District level, the (acute weighted separation) targets were negotiated for each Health Service District based on forecasts of hospital activities (Surrao et al., 2002; QH, 2000). These targets act as global budgets at the district level. They are then distributed across the hospitals to derive hospital acute weighted separation targets. This target is then multiplied by the average weighted separation price (cost weight) to determine the hospitals acute inpatient budget. The average weighted separation prices were based on the standard prices for different DRGs derived from National Public Hospital Weights and the hospital groups. Hospitals are divided in five groups with the base payment price increasing with size. This is because Queensland assumes diseconomies of scale and thus, for the same DRG, prices are higher in larger hospitals than in smaller hospitals (Duckett, 1998). Prices were also adjusted for allocated overhead, medical pathology and critical care components. Further adjustments were made to account for outliers, if any (see description in QH, 1998, 1999, 2000). Outliers and exceptional case payments are applied for long and extra-long stay cases or some special cases with high cost. The main purpose of a global budget scheme is to control the aggregate expenditure and sometimes, the allocation of funds across hospitals in the catchment area (or regions of the country). It usually leaves hospitals with considerable amount of flexibility on how its budget is spent within each period. Feldman & Lobo (1997) argued that because hospitals exercise this control, the central authority can absolve itself from the responsibility of “micro-managing” the production of health care services. The job is turned over to professionals who are better qualified to determine the optimal quality and quantity of care to patients. This can theoretically encourage the development of changes to service delivery patterns, and improve efficiency, good clinical practice, and quality. Furthermore, this method is believed to have advantages of cost containment, funding certainty, cheaper administration, and improved co-ordination in service planning (Dredge, 2004). In practice, efficiency can be compromised due to the deterioration of quality of care, budget rationing, and locked in pattern of services (allocative inefficiency). If quality standards, and requirements for regular collection and assessment of infor4

Detailed description of this funding method can be found in Section D.1.1.2 of Appendix D.

61


mation on quality, is not explicitly built in the payment scheme, this method will provide little incentive for quality. There is also a risk that becoming more cost efficient (i.e. saving resources) will lead to a lower budget for the coming years. This can discourage providers from striving for more efficient ways to deliver services, especially when the budget allocation is fairly generous. When hospitals face inadequate budget, historical budgetting can lead to the under-treatment of admitted patients, reduction in the number of admissions or admission of the less complicated casemix, i.e. cream skimming (detailed discussion can be found in Section D.1.1.2 of Appendix D). All of these appear to transpire the case of Queensland Health hospitals, as shortage of funding and inadequate monitoring of performance have been been highlighted issues in the past years (see discussion in Davies, 2005). The historical budget also tends to create an institutional inertia that locks in existing patterns of resource use. Davies (2005) identifies two major problems in Queensland. First, if the original budget was not fixed sufficiently to provide an adequate coverage of different service ranges, history-based budgeting practice is unlikely to create a break-through. Second, if the original budget was based on the then need of the community, subsequent budgets would fail to take into account changes in demographic structure and population size. This is especially serious in a State that has experienced fast population growth like Queensland. It is reported that the percentage change in weighted population from 1998-99 to 2003-04 in Queensland is 14.3%, the highest in Australia, and about 4% above the national average (Davies, 2005). In 2007, Queensland Health introduced a new funding model, based on population and regional needs and casemix. It aims at further decentralising decisionmaking and increasing funding certainty for health service providers, in order to promote equity, accountability, transparency, stability and predictability. It covers at least 90% of state-wide public hospital activity (outputs) and applying to over 30 of the top facilities in Queensland (QH, 2008). The new funding scheme is also a two tier model (see the description of components of the old and new models in Table 3.2). The Resource Allocation Model (RAM) informs the allocation of Queensland Health’s budget between its Area Health Services, based on the health needs of their respective populations. It takes into account the population size, age distribution and other factors affecting relative need for health services. This design aims at achieving an equitable distribution of funding. This setting is fairly similar to the old funding model.

62


The Casemix Funding Model (CFM) is an output-based allocation formula where a facility is being funded based on its individual activities. District and facility budgets are determined by the number and clinical acuity of treated patients through the use of resource homogeneous grouping of clinical services for like patients - the Australian Refined Diagnosis Related Groups (AR-DRGs) system. There is a single set of DRG cost weights and a single base price applied across the State. This is where the major change lies, compared to the old funding model where prices vary by hospital groups. The Queensland Health’s prices, cost weights and parameters for payment are derived from the National Hospital Cost Data Collection (NHCDC) cost weight relativities, which in turn are based the use of casemix groups established across patients and hospitals all over Australia. The fundamental assumption behind casemix grouping is that the resource required for a case is determined by the demographic, diagnostic, and treatment profile. The rates reflect the historical costs of both individual hospitals and the entire network of hospitals. Therefore casemix based payment creates incentives for hospitals to control unit cost through reducing the inpatient length of stay (LOS) and increasing efficiency while maintaining quality of care5 . The targets (estimated activities) are negotiated between Queensland Health and the districts. A district is provided with a casemix budget which includes their approved activity targets for each financial year. Where there is significant deviation of actual activity from the approved planned or budgeted activity, the district is required to demonstrate or explain why such a significant variance has occurred6 . This feature is similar to the old funding model as the activity target acts as the volume cap at the district level. However, its casemix-based characteristic suggests a mix funding scheme which aims at controlling total costs (i.e. global budget) and promoting productivity and efficiency (i.e. casemix based). This is an improvement from the old mechanism. The Casemix Funding Technical Document (QH, 2008) claims that the new funding model contains economic initiatives and incentives to reward and encourage more efficient patient treatment, good quality care and equitable funding. As a casebased payment scheme, it can also affect service patterns by encouraging providers to favour treatments under lower cost settings while maintaining quality of care. 5

For detailed discussion of this funding model, see Section D.1.2 in Appendix D. It is described in The Casemix Funding Technical Document (QH, 2008) that districts should not expect that funding will be made available where districts have exceeded their activity targets. 6

63


However, as the new system has only been implemented in the last year and sufficient data are not yet available, it is not possible to investigate the evolution of productivity and efficiency under old versus new funding models. On the other hand, the changing context in Queensland Health still provides an opportunity to examine how well the system has performed under the old funding system, especially to address the issues of health workforce shortage. Therefore, this is the focus in the remaining discussion of this chapter. Table 3.2: Components of the old and new hospital funding model in Queensland public hospitals Old model* Elements

New CFM system**

Price/Classification

Elements

Price/Classification

Variable components Acute Inpatient Sub and non-acute patient (General wards)

AR-DRG and hospital

Acute inpatients

groupings

(General wards)

National Health Data Dictionary (NHDD) episode (per diem price)

Sub and non-acute patient (Designated

Critical care (Designated

Clinical Services

ward)

Capability Framework

Sub and non acute

AN-SNAP V1

patients (SNAP)

units)

AR-DRG

Designated Psychiatric

Mental Health Unit

Mental health

Unit

grouping

(acute/designated ward)

AN-SNAP V1 Mental health ward/unit grouping - CFM ward designation Queensland Health

Ambulantory

NHCDC and hospital

Outpatients

groupings

outpatient clinic classification system (based on NHCDC)

Emergency Department Home Dialysis

Triage category/discharge disposition Four models

Fixed component (the elements vary from year to year) Clinical education and

Teaching grant

research Emergency department

Research grant

fixed component

High cost outlier grant Infrastructure grant Other special grants Source: * Hospital funding model for Queensland Public hospitals (QH, 1998, 1999, 2000) ** Casemix Funding in Queensland (QH, 2008)

64


3.1.4

Labour shortage and efficiency in Queensland Health public hospitals

The Health System Review Report (QH, 2005) maintains that Queensland hospitals operate very efficiently compared to other Australian States and Territories. Queensland has: (i) spent 11% less per casemix adjusted separation (A$2,929 compared to A$3,293); (ii) achieved a similar (slightly lower) number of public hospital beds per 1000 people; (iii) maintained lower relative lengths of stay in hospital (0.94 compared to 0.99, controlling for the average complexity of cases)7 ; and (iv) achieved a similar rate of same day admissions (49%) (QH, 2005). However, those indicators only provide a partial representation of efficiency and can be misleading for several reasons. Firstly, here efficiency was not assessed in the context of demand. It is evident that service provision from public hospitals has not been keeping pace with demand. Over the period 2000-01 and 2003-04 hospital admissions in Queensland have grown by 3.8% while population growth has been 4.7%. The decentralised nature of Queensland’s population, along with a fast growth rate, necessitates some duplication of health service infrastructure and dilution of the medical workforce across the State. Therefore, there needs to be greater investment (for maintenance and new technological infrastructure) for the same outcome in a less decentralised setting (or less outcomes if investment stays the same) (Davies, 2005). It is also pointed out that some rural and regional districts do not have adequate infrastructure or capacity to fully support their expected role including clinical governance, training and support services (QH, 2005). This results in Queensland expenditure per separation on public hospitals being below the national average8 . This does not always mean efficiency gain. This can be the evidence that quality is compromised so less resources were consumed. One can also argue that it is due to an under-funded hospital system - a possible consequence of the historical funding model. Secondly, the key driver for the lower cost per casemix adjusted separation is lower expenditure on nursing, allied health and medical services (staff numbers and average salaries) (QH, 2005). As seen in Table 3.3, compared to the Australian av7

These two indicators suggest that Queensland Health public hospitals have the lowest patient days per 1,000 population, 690 days, around 120 days lower than the national average. 8 The 2003 Productivity Commission report records that in 2000-01, Queensland recorded the lowest government real recurrent expenditure per person on public hospitals (in 1999-00 dollars) at A$660 per person, well below the national average of A$776 per person, a gap of A$116 per person. This trend has continued. For the 2003-04, Queensland again recorded the lowest government real recurrent expenditure per person on public hospitals (in 2001-02 dollars) at A$712 per person, well below the national average of A$895 per person, a gap of A$183 per person Davies (2005).

65


erage, Queensland Health pays around 6% less in average salaries for public hospital staff. Queensland’s salary for medical doctors (excluding visiting medical doctors) is the second lowest in Australia, close to 13% lower than the national average. The salary for nurses is also lower than the national average, only higher than Tasmania. The same picture is seen for all other labour categories, except for diagnostic and allied health professionals. Given the differences in income, labour mobility allows medical workers, especially doctors and nurses, to migrate to other States or overseas where higher income is offered. Moreover, it has been suggested that the difference between private and public salaries, together with working environment and lifestyle choices (Davies, 2005), makes it increasingly difficult for public hospitals to attract nurses and medical officers, especially specialists (DHA, 2008). Hence, the lower relative wage as a cost saving device in the public hospitals is not always an encouraging efficiency indicator. It can well be one of the attribute to the shortage of medical workers in Queensland. Table 3.3: Average salary (A$) of full-time equivalent (FTE) staff in public hospitals, 2003-04 Staffing

Salaried

category

medical

Nurses

Other

Diagnostic

care staff

officers

NSW

116,880

65,284

...

Admin

Domestic

and

and

and

allied

clerical

other

health

staff

staff

53,769

50,366

36,914

Average

61,481

VIC

133,174

62,315

...

43,356

44,404

42,645

60,756

QLD

105,388

57,422

38,273

59,419

42,084

38,665

56,719

WA

138,997

61,407

39,944

54,823

45,361

39,348

61,417

SA

107,378

57,546

...

50,328

42,546

34,923

56,307

TAS

102,624

56,202

...

59,505

40,708

47,853

56,742

ACT

130,376

64,828

52,350

62,147

54,642

41,831

65,003

NT

130,376

24,828

52,350

62,147

54,642

41,831

65,003

Average

120,627

61,969

39,134

50,515

46,280

39,995

60,083

Source: Australian Hospital Statistics 2002-03 (ABS, 2003)

Tables 3.4 and 3.5 provide further evidence of the health worker shortage picture in Queensland. Despite the increasing trend of the number of full time equivalent medical practitioners per 100,000 population (around 12% over the whole period), Queensland is still below the national average and only ahead of Tasmania. It can also be seen from Table 3.4 that in 2004 the number of employed medical staff per 100,000 population in Queensland is much lower than the national average and indeed the lowest of all States. It has the lowest number of doctors per 100,000 population in remote areas, and second lowest in the number of doctors per 100,000 66


population in very remote areas. Additionally, Queensland hospitals have tended to recruit more and more doctors from the international market in recent years9 (Davies, 2005). Table 3.4: FTE public medical practitioners per 100,000 population 2001-02

2003-04

2005-06

% change

NSW

99

100

115

16.16

VIC

94

109

114

21.28

QLD

91

94

102

12.09

WA

90

96

105

16.67

SA

113

109

128

13.27

TAS

73

77

97

32.88

ACT

91

97

121

32.97

NT

128

123

147

14.84

96

101

112

16.67

Australia

Source: Medical Labour Force 2005 (AIHW, 2008)

Table 3.5: Medical practitioners per 100,000 population in 2004-05 Major

Inner

Outer

city

regional

regional

Remote

Very

NSW

354

205

112

98

92

VIC

352

179

141

220

...

QLD

275

147

145

54

83

WA

285

107

128

145

121

SA

389

124

123

97

132

TAS

...

364

134

58

...

ACT

393

...

...

...

...

NT

...

...

285

306

56

Australia

338

187

140

126

91

remote

Source: Medical Labour Force 2004 (AIHW, 2006)

Similar to doctors, there is also a shortage of nurses in Queensland (see for example Davies, 2005; QH, 2005). The Queensland Health Systems Review (QH, 2005) report informs that Queensland Health has experienced a critical shortage of nurses. It has the lowest number of nurses per capita of any state in Australia (except Tasmania), 15% below the national average. The same story can be seen for the group of diagnostics and allied health professionals; Queensland Health has 40% less other health professional staff than the national average. Apart from medical worker shortage, the expenditure data also indicates that 9 Those overseas trained doctors tend to be drawn now from developing countries rather than the Australia’s traditional markets of United Kingdom, Ireland and South Africa. This trend is more pronounced in the regional health care facilities than it is in urban and tertiary hospitals (Davies, 2005).

67


Table 3.6: FTE nurses and other health professionals in 2003-04 Nurses Number

Other health professionals

Per 1,000

Number

Per 1,000

NSW

31,865

4.8

10,005

1.5

VIC

24,028

4.9

10,784

2.2

QLD

14,661

3.9

3,231

0.9

WA

8,158

5.3

2,230

1.5

SA

7,813

4.0

1,965

1.0

TAS

1,806

3.8

349

0.7

ACT

1,479

7.4

349

1.8

NT Australia

941

2.9

261

0.8

90,751

4.6

29,174

1.5

Source: Queensland Health System Review (QH, 2005, page 209)

the hospital administration costs per weighted separation in Queensland are approximately 15.6% (or A$66) lower than the national average. It is also reported in QH (2005) that there were insufficient numbers of basic administrative support staff in hospitals resulting in clinicians being diverted inappropriately to administrative tasks10 . The Commission Inquiry report also pointed out that the recruitment of extra medical officers (e.g. specialists) is sometimes not accompanied with provision of funding for supporting administration, nursing and allied health staff (Davies, 2005). One could argue that this is more an issue of allocative efficiency (i.e. choice of sub-optimal staff mix). However, when there is lack of administrative support, the work-flow is affected and fewer cases can be seen/treated for a given number of FTE medical staff, which implies lower productivity and technical efficiency. This is one of the areas where efficiency improvements can be made, especially when it is identified that the Corporate Office has taken on a range of administration functions that should not be performed by a central office and would be better performed closer to health service delivery (QH, 2005). Secondly, a slightly lower number of beds per 1,000 population, doubled with lower lengths of stay, can be seen as a sign of infrastructure shortage, rather than efficiency improvement. The number of available beds in Queensland public hospitals decreased from 10,164 in 1995-96 to 9,340 in 2004-05 (QH, 2006). The number of bed per 1,000 population in 1998-99 was 3.2, higher than national average (indeed second highest, only less than Northern Territory) but fell to 2.5 by 2005-06, lower 10 Queensland Health spends $9.40 (82%) more per person overall on health administration than other states. The higher administration costs most likely reflect Queensland Health’s more centralised structure with costs being recorded corporately rather than locally and some inconsistencies in how administrative staff are defined in different states (QH, 2005)

68


than the national average (of 2.7) (DHA, 2007b), as shown in Table 3.7. This trend is consistent with the argument that funding has failed to catch up with health care need due to the historical budget practice. Furthermore, lower relative lengths of stay in hospital can be a sign of the use of early discharge as a management tool rather than the hospital’s ability to increase throughputs. This is a relevant argument in the context of bed and medical staff shortage and long waiting lines for acute treatments like in Queensland. Since the information about overall hospital quality of care as well as patient satisfaction is not available, it is impossible to decide if the lower length of stay in Queensland Health is the result of efficiency improvement efforts or premature discharge11 . However, one should not ignore the possibility that cost saving might be achieved at the expense of service quality; therefore cost information alone cannot be used as an indicator of efficiency improvement. Table 3.7: Number of bed per 1,000 population in public hospitals, 2005-06

NSW

Major

Inner

Outer

Remote

cities

regional

regional

2.7

3.2

3.7

6.6

Very

Total

remote 6.1

2.9

VIC

2.4

2.6

2.7

2.4

...

2.4

QLD

2.4

2.1

3.1

4.1

8.6

2.5

WA

2.4

1.2

4.1

4.2

3.4

2.5

SA

2.8

2.3

5.1

7.7

7.4

3.2

TAS

...

2.9

2.2

2.2

3.6

2.7

ACT

2.2

0.0

...

...

...

2.2

NT Australia

...

...

2.7

5.2

1.0

2.8

2.5

2.6

3.3

5.0

4.7

2.7

Source: Australian Hospital Statistics 2006-2007 (ABS, 2007)

Another example of public hospitals failing to support health demand can be seen through the expanding private health sector in Queensland. Services provided in Queensland’s private hospitals have been growing at a faster rate compared to that of public hospitals. This trend is reflected through a significant expansion of private free-standing day facilities. Queensland now has 51 of these facilities compared with 4 in 1991-92. Additionally, public patient admissions per 1,000 population (and hospital admissions per 1000 population) has decreased and below the national average in 2003-04 (180 compared to 199 in 1998-99) (Allen, 2006, page 11

It was not possible to obtain data on service quality (such as average waiting time for emergency services and scheduled surgical procedures, percentage of errors in the handling of cases etc) for the studied period. Hence, the potential trade-off between quality and cost is not examined in the study. If the quality difference between hospitals presents, it is expected to see that hospitals that cut cost through quality reduction will look more efficient than those that did not.

69

3.2. METHODOLOGY

14). This has resulted in the public hospital share of overall hospital activity12 falling from 61% of all hospital activity in Queensland in 1999-00 to 51% in 2005-06, the lowest level in Australia. This is partly due to the incentive that the Commonwealth Government introduced to encourage the participation on private health insurance. However, there is evidence suggesting that private health insurance is only a part of the story. There were 466 patient days per 1,000 population in Queensland private hospitals, higher than those of any other States, and around 35% higher than the national average. At the same time, the patient-days-per-1,000-population in public hospitals is of the opposing situation. Allen (2006) suggests that this is the consequence of more patients choosing to be treated in private hospitals, rather than more patients using public facilities. In fact, Queensland has a substantially larger proportion of private patients opting to use private hospitals compared to the national average. At the same time, the proportion of private patients in Queensland public hospitals is only two-thirds of the national average (18 per 1000 population compared to 27) (Allen, 2006, pg 15). Table 3.8: Episodes of care in Queensland public and private hospitals Year

Public

% share

Private

hospitals

% share

Total

hospitals

1999-00

706,530

60.96%

452,506

39.04%

1,159,036

2000-01

687,952

56.66%

526,313

43.34%

1,214,265

2001-02

694,264

53.93%

593,116

46.07%

1,287,380

2002-03

701,753

53.82%

602,166

46.18%

1,303,919

2003-04

720,673

57.16%

540,048

42.84%

1,260,721

Source: Queensland Health System Review (QH, 2005)

The workforce shortage is and will continue to be a major challenge for Queensland Health, especially in the context of ageing population with costly chronic diseases, global health workforce shortage and an environment of significant competition between jurisdictions for clinical staff. Therefore, the continuous improvement of productivity and efficiency of the medical labour force is one of the necessary condition to satisfy increasing demand in the labour shortage situation.

3.2

Methodology

A hospital can be approximately represented as a production unit that transforms labour, capital and material inputs into medical services for the purpose of 12

Here, hospital activity is measured by number of separations per 1,000 population

70

3.2. METHODOLOGY

improving the health status of the patients. In the hospital context, capital (such as building, equipment, and number of beds) is usually fixed in the short term. Material consumption, on the other hand, is highly driven by output volume but fairly fixed for a particular case. The hospital can choose the labour mix to maximise its outputs (the number of cases treated or number of patient days) or alternatively, to minimise the amount of labour input usage, thus cost. Therefore, efficient use of the labour force is essential to improve total productivity and reduce overall inefficiency in hospitals. This section explains why the input requirement function is appropriate to examine labour efficiency in Section 3.2.1 and discusses the estimation strategy in Section 3.2.2.

3.2.1

The input requirement function

To investigate the labour efficiency of the hospital workforce, an input requirement function is considered. As an input-oriented method, it is conceptually similar to a cost frontier, in that they both focus on the conservation of resources used to produce the required outputs. This is intuitive in the case of public hospitals that take demand as given, in the sense that patients cannot be turned away if the treatments are available13 . Moreover, the capped budget mechanism in Queensland Health public hospitals does not give incentive to increase outputs. The primary objective becomes resource containment. Therefore an output oriented approach is not intuitive while the input orientation can address this optimisation objective. Unlike the cost function, the labour requirement function does not requires any input price information as its focus is on technical optimisation rather than economic optimisation. This is an advantage as the market output prices are neither observed (public patients pay zero effective price in Queensland Health public hospitals) nor reflect the true opportunity costs of the services provided (charges for private patients in public hospitals are regulated by schedulled fees). A hospital is deemed inefficient if it uses more labour to produce a given output mix than would an otherwise efficient hospital, where capital inputs are more or less fixed. This approach has been used by Kumbhakar & Hjalmarsson (1995) who estimate labour inefficiency for a panel of Swedish social insurance offices. The authors argue that this approach is appropriate because labour costs make up about 80 percent of total cost of Swedish social insurers. This is also true in the Queensland public hospital setting, as labour is the biggest cost driver in these organisations, 13

The fact that treatment is rationed using waiting lists is another issue.

71

3.2. METHODOLOGY

accounting for more than 60% of operational expenditure and arguably the most scarce (see Section 3.1.4). Like the cost frontier, the input requirement frontier accommodates multiple outputs, which is useful in the hospital setting where a natural aggregator for hospital outputs is not. As a parametric approach, it takes into account noise in the production process. While it is necessary to choose an appropriate functional form and distribution for the efficiency term in this setting this is not cause for overt concern. Considering the former, attention surrounding the issue of choosing a functional form has dampened in recent years (Greene, 2008). This is due to the existence of the translog functional form. This is a second order approximation of an unknown functional form and therefore is highly flexible. In fact, the translog can approximate most other functional forms (Intriligator, 1978) and is most useful in cases where the researcher does not have any a priori reason to choose one functional form over another. It is for this reason that the translog functional form (see for examples, Zuckerman, et al., 1994; Chirikos, 1998a; Chirikos & Sear, 2000; Folland & Hofler, 2001; Rosko, 2001a; Rosenman & Friesner, 2004) or an alternative model that is nested in this form (Chirikos, 1998a; Puig-Junoy & Ortun, 2004) are the most common in the SFA literature. Considering the choice of distribution for the efficiency component, while, theoretically it is expected that choosing one distribution over another will yield very different efficiency estimates (Jacobs, et al., 2006), given that each distribution imposes a different skew on the efficiencies, findings in Chapter 2 and evidence from the literature suggests the contrary. For example, Hollingsworth (2003), using WHO (2000) data, conclude that there is good comparability between the inefficiency estimates emanating from both the half-normal and truncated- normal distributions. In particular, the authors find that correlations between the two distributions are close to one. Kumbhakar & Lovell (2000) examined the differences in inefficiency values across all four distribution types. The data considered were of 124 U.S. electric utilities (these data were originally gathered as part of a study by Christensen & Greene (1976) and this dataset has been used in many subsequent papers as it has a well defined production process). The authors concluded that choosing one distribution over another should not greatly affect the overall efficiency results. Other studies also arrive to a similar conclusion (see for example, Fuiji & Ohta, 1999; Rosko, 1999, 2001a,b; Rosko & Proenca, 2005; Zuckerman, et al., 1994; Jacobs, et al., 2006). Therefore, for functional form and error distribution choice the evidence in the literature implies that emanating results should not be sensitive to

72

3.2. METHODOLOGY

such choices.

3.2.2

The “true” random effect model

The labour requirement function can be estimated under the stochastic frontier framework. A typical SFA model is described in Equation 3.1. The dependent variable is aggregated labour (Lit ) while Yit represent output variables (k = 1, ..., K). The composite error term consists of the inefficiency component uit and the stochastic term vit . By construction, uit represents the proportion by which labour usage exceeds the minimum labour required to produce the output mix. It is assumed to follow a particular distribution. The most frequently used are the half normal, exponential and truncated normal from below at zero (Murillo-Zamorano, 2004). Although been no general agreement on which distribution should be chosen over the other, many studies have examined the correlations between efficiency scores generated by different distributional assumptions. The conclusions have been consistent that these are highly correlated and thus it can be inferred that hospital ranking based on those efficiency estimates should be quite consistent (see for examples Chirikos, 1998b; Linna & Hakkinen, 1998; Webster et al., 1998; Fuiji & Ohta, 1999; Rosko, 1999; Yong & Harris, 1999; Fuiji, 2001; Street & Jacobs, 2002; Street, 2003). For this reason, half normal distribution will be used in this analysis. The stochastic term vit is the usual noise term and represents random events not under control of hospitals such as unforeseen equipment failure, luck and measurement error in the dependent variable. Hence, it is not unreasonable to assume that uit and vit are uncorrelated given that uit is within the hospitals control whereas vit is not. It is also necessary to assume zero correlations between the error terms and the independent variables since the services offered by the hospital are exogenous. ln Lit =α0 +

K X

K

αk ln Yitk +

k=1

K

1 XX αkl ln Yitk ln Yitl + uit + vit 2 k=1 l=1

uit ∼ N + 0, σu2 vit ∼ N 0, σv2

(3.1)

It is necessary to incorporate hospital level heterogeneity into Equation 3.1 given that attributes such as staff per patient, location, teaching status and size may impact on how a hospital is organised and how it provides its services, as well as the service mix produced. A common approach is to introduce them into the 73

3.2. METHODOLOGY

deterministic part of Equation 3.1 (as shown in Equation 3.2, in which Zit are relevant control variables). The predicted inefficiencies of individual hospitals now vary with both outputs and these attribute variables. ln Lit = α0 + +

K X k=1 J X

K

αk ln Yitk +

K

1 XX αkl ln Yitk ln Yitl 2 k=1 l=1

(3.2)

βj ln Zitj + uit + vit

j=1

The inclusion of attribute variables can only remove observed heterogeneity, not latent heterogeneity. Unobserved heterogeneity reflects variables that are not quantifiable or simply missing in the model. If latent heterogeneity is not taken into account, the estimated inefficiency picks up these effects. Inefficiency is, thus, confounded with hospital heterogeneity and in many cases, can significantly shuffle the performance ranking. This issue is discussed in Greene (2003a). The paper analyses the WHO data under various models and concludes that models that make no distinction between technical inefficiency and individual (country) heterogeneity and those that do (the preferred specification in the paper) brought a substantial change in the estimated results and ranking. Therefore it is pertinent that the unobserved hospital heterogeneity is distinguished from the inefficiency component (Greene, 2005a). Attempts to deal with this issue have been made since the 1980s. Pitt & Lee (1981) and Schmidt & Sickles (1984) proposed to use fixed-effects or random-effects models in the panel data frontier estimation setting. However, by definition these models dub hospital specific heterogeneity as efficiency. Moreover, the treatment of the hospital specific effects as inefficiencies forces the inefficiencies to be constant over time. Some other authors have extended the model to include time-varying inefficiency (such as Cornwell, et al., 1990; Kumbhakar, 1990; Lee & Schmidt, 1993; Battese & Coelli, 1992). In these models, some flexible functions of time are included in the inefficiency component to allow a time varying effect. However, the latent effects are still dubbed as inefficiency and when these are correlated with the explanatory variables, the random-effects estimators are affected by heterogeneity bias (Farsi, et al., 2003). Recently, the proposal to accommodate time-varying inefficiency and to distinguish unobserved heterogeneity and inefficiency by Greene (2004, 2005a) takes the form of “true” fixed (TFE) and “true” random effect (TRE) models. Equation 3.3

74

3.2. METHODOLOGY

and Equation 3.4 specify these two models in the labour requirement setting: K

K X

K

1 XX ln Lit = αkl ln Yitk ln Yitl + αi + vit + uit αk ln Yitk + 2 k=1 l=1 k=1 2 + uit ∼ N 0, σu vit ∼ N 0, σv2

(3.3)

and K X

K

K

1 XX ln Lit = αk ln Yitk + αkl ln Yitk ln Yitl + α0 + wi + uit + vit 2 k=1 k=1 l=1 uit = |Uit | with Uit ∼ N 0, σu2 wi ∼ N 0, σw2 vit ∼ N 0, σv2

(3.4)

Both models retain the assumptions of labour cost minimisation as well as the error term construction as discussed in Equation 3.1. Now αi (in Equation 3.3) and wi (in Equation 3.4) represent hospital specific effects which measure unobserved heterogeneity. While the TFE model can be estimated by maximum likelihood estimation where maximisation of the full log likelihood function is achieved by “brute force”, the TRE model can be estimated using simulated maximum likelihood by integrating out wi (see Greene, 2005a, for details). The main difference between these two models is the specification of hospital heterogeneity. The random effect specification requires an additional assumption that the time-invariant, hospital specific random term wit is uncorrelated with all regressors in the model while the fixed effect allows the heterogeneous term to be correlated with the included variables. Like the usual fixed effect model, the TFE model cannot accommodate timeinvariant regressors. In this case all the heterogeneity, even observed, rather than only omitted time-invariant factors as in the case of the TRE, is captured in the fixed effects. Many observed characteristics of hospitals, such as location or teaching status, are constant over the study period but likely to influence the hospital performance. Their exclusion forces the inefficiency estimators to pick up hospital heterogeneity that is not related to inefficiency (Farsi, et al., 2005). Moreover, as pointed out by Greene (2005a), the presence of the individual effects may create

75

3.3. DATA ON QUEENSLAND PUBLIC HOSPITALS

an incidental parameter problem, where estimates of the individual effects αit may be inconsistent and affect efficiency estimates in short panel. Therefore, the TRE model, with the inclusion of some control variables, is considered here14 . Estimation of the parameters for TRE is achieved by maximum simulated likelihood. Inefficiency estimates (uit ) are extracted after the random effects (wi ) is integrated out of using pseudo random draws from a distribution. In this case Halton draws are used. The Halton draws are based on an “intelligent” set of values for the simulation and proved to dramatically speed up the process of maximization by simulation. A small number of Halton draws is as effective as or more so than a large number of pseudorandom draws using a random number generator (Greene, 2003b). Efficiencies, uit , are computed using the Jondrow, Lovell, Materov, and Schmidt (JLMS) estimator (Equation 3.5), after integrating out the random effects (Greene, 2003a). λσ φ [it (wi )λ/σ] −it (wi )λ E [uit |it (wi )] = + (3.5) 1 + λ2 σ Φ [−it (wi )λ/σ] Where σ 2 = σu2 + σv2 ; λ = σu /σv ; φ(.) is the probability density function of the standard normal distribution and Φ(.) is the cumulative distribution function of the standard normal distribution.

3.3 3.3.1

Data on Queensland public hospitals Data collection

The data was collected from Queensland Health through several databases, including the InfoBank and Casemix unit for the period from 1994-95 to 2004-05. The InfoBank dataset contains two cost variables, seven inputs (of which six are labour inputs), two outputs, number of bed days and occupancy rate. The two cost variables are labour and non-labour costs, recorded at constant price level. The only non-labour input is the number of beds, which is usually used as a rough indicator of the stock of capital and hospital size. 14 The individual effect in this model absorb all unobserved heterogeneity such as scale effects due to different size and labour standby.

76


Six labour categories include medical officers, nurses, allied health, administrative staff and operation (domestic) staff, measured in full-time equivalent (FTE). Medical officers include both staff MOs and visiting medical officers (VMOs). All VMOs are specialists while staff MOs is made up of two main groups, general practitioners (GPs) and specialists. Large hospitals employ predominantly specialists and only a few GPs while small rural hospitals tend to have mostly GPs. In Queensland large hospitals usually involve teaching, with trainees and resident medical officers usually making up approximately half the medical staff. The nurse group consists of nurses at all skill levels, including registered, enrolled, clinical and assistant nurses. The allied health segment is composed of two main groups: professional and technical officers. The former consists of non nursing and non-medical health care professionals, such as therapists, dieticians, optometrists, podiatrists, psychologists, social workers, pharmacists, medical laboratory scientists, physical scientists and radiographers. The latter are technicians who work as medical laboratory and renal dialysis technicians, electroencephalograph recordists and so on. The administrative group supports the hospital administrative infrastructure in areas such as general administration, accounting, personnel and admission/discharge. It is made up of clerks, receptionists, administrators, IT personnel and non-clinical managers. There are several levels of administrative officers with distinctive skill and job complexity. The operational segment contains staff working in laundry, kitchen, wards, hospital grounds and maintenance areas, and a small number of semi-technical staff (such anaesthetic assistants, sterilisation unit operators, therapist assistant, pathology attendants, etc). The two outputs available are raw counts of number of inpatient episodes of care and number of outpatient occasion of services. Outpatient occasions of service usually take the form of simple examination and/or treatment during short visits or consultation sessions. Since the variation of resource consumption between different outpatient cases is quite small compared to that of inpatients QH (2008), the raw count of outpatient visits have been widely accepted in the hospital efficiency literature. The raw count of inpatient cases, on the other hand, is not satisfactory. Resource consumption between two episodes of inpatient care can be vastly different. Therefore, casemix information should be incorporated into the analysis to distinguish between hospitals that treat complicated cases with those who admit mostly simple ones. This requires the use of a second source of information, the Casemix dataset. The Casemix dataset contains information about number of inpatient cases,

77


grouped by DRGs). Since there are more than 600 DRGs, there are far too many output variables to be included in the model. Therefore, DRG cost weights are used to aggregate these DRGs into a composite output. The cost weight for each DRG is an index designed to capture the resource consumed to treat a condition classified in that particular DRG. The weight data is then collected from the series of NHCDC conducted during the study period. The aggregated output is, therefore, adjusted for casemix differences. This convenience comes at the cost of losing information on the relative importance of different outputs to health labour utilisation. Hence, a compromise position was adopted where all DRGs were aggregated into two main groups of treatment: those involving surgeries and those with only medical interventions. The information on surgical and medical classification was obtained through AR-DRGs classification. The two datasets were merged and screened for consistency and validity. All observations with missing data in inputs or outputs were dropped. In addition, outpatient centres and clinics were removed because they have different organisational structures, staffing and management practice compared to larger hospitals. Clinics and outpatient centres offer simple inpatients services or only outpatient treatment while hospitals are well equipped with specialised departments and treat more complicated casemix. Dropping these observations ensures homogeneity. Output and input information for the first year were missing in all observations. Labour data for the last year was inconsistent with previous years. Therefore, the first year and the last year from the final dataset are excluded. During this period, Royal Brisbane and Royal Women hospitals were consolidated into one Royal Brisbane and Women’s, starting 2002-03. Observations from 7 years of the two individual hospitals (from 1994-2002) were combined under an “artificial” hospital so that it could match Royal Brisbane and Women’s of 2002-03 and 20030415 The final dataset is a balanced panel of 84 hospitals from 21 districts16 , over eight years (from 1995-96 to 2002-03) with six labour inputs, one capital input, three outputs and two cost variables. 15 It is noted that the merge of two hospitals might imply the different structure between the “artificial” hospital and the actual unified one. This information has been verified with medical professionals at Queensland Health. They suggested that the consolidation would not change the working flow and nature of the two hospitals very much. This is also shown through a very similar level of output and input consumptions pre- and post-consolidation. 16 Although the 21-district system did not exist from 1996 to 2003 but this location division need not affect the analysis. As far as we know, there was no change of health district divisions during the studied periods. The consolidation of 37 districts into 21 happened in 2005 (and again in 2008 to 15 districts). This happened particularly in the urban areas with the formation of the “Northside” and “Southside” Districts

78


The final step in the data preparation involved aggregation of the six labour categories into one labour variable using wages data. Wage information was obtained from the database of the Australian Bureau of Statistics. However, given the available data it was necessary to assume that wages are uniform across hospitals, regardless of location and hospital type. This assumption is acceptable in the case of Queensland public hospitals because the salary package of a defined medical position, by both scale and category, is determined centralised by Queensland Health with slight variation of incentive bonuses and awards17 . Therefore, this figure purely reflects weighting across job types. Extra information on individual hospitals is obtained through various sources; one of them is the Queensland Health website. It is possible to identify teaching hospitals, categories, and their location (through Health Service District and Hospital Profile sections). There are twenty-one teaching hospitals in the final sample. Most of them are large or medium-sized, and are also the principal referral hospitals in their districts or regions. Nine hospitals are classified in category A - large and specialised hospitals, fifteen in category B - medium-sized, and the rest are smaller hospitals. The majority of small hospitals concentrate in health service districts of Central Queensland, Central West, Mt Isa, South West, Toowoomba and Darling Downs, Townsville and West Moreton South Burnett. With respect to location, hospitals belong to city, town, shire or islands. Although shires do not necessarily refer to being rural or remote, especially for those that are close to the Brisbane city, the majority of shires locate in rural areas. Eight health service districts cover areas where the Indigenous population is higher than Queensland average.

3.3.2 3.3.2.1

Choice of variables The dependent variable

The independent variable of the input requirement function is aggregated labour (Lit ). As discussed in Section 3.3.1, this variables is constructed as a linear aggregation of six labour categories with their respective wages as weights. The weights reflect relative importance of job types. The six labour categories are medical officers, nurses, allied health, administrative staff and operation/domestic staff. The wage for each labour category is uniform across hospitals, regardless of location and hospital type. 17

For details, see the Queensland Health Certified Agreement with medical officers and nurses in http: //www.health.qld.gov.au/hrpolicies/

79


3.3.2.2

Output variables

Three output variables are selected for the analysis. They are number of outpatient occasions of service (OU P T ), weighted surgical inpatient (W EP SS ) and weighted medical inpatient (W EP SM ) episodes of care. Coefficients of these variables are expected to be positive. That is, more labour is required to provide larger amount of services. The size of individual coefficient reflects its relative importance in labour requirement. OU P T is a raw count number of outpatient occasions of service. It includes non-admitted examination and/or treatment, visits and consultation sessions in the hospital. Using the raw count of outpatient visits have been widely accepted in the hospital efficiency literature because the variation of resource consumption between different outpatient cases is quite small. Raw count of number of inpatient cases are usually undesirable. Resource consumptions by episodes of inpatient cares for two disease types can be vastly different. Hospitals that treat more complicated casemixes will be identified as less inefficient compared to those that treat simple cases (especially when the sample contains large general hospitals and small hospitals). The availability of DRG cost weight information for individual DRG enable the construction of aggregate inpatient outputs that better reflect the casemix produced by individual hospitals. As discussed in Section 3.3.1, weighted surgical inpatient (W EP SS ) and weighted medical inpatient (W EP SM ) episodes of care are constructed from the DRG data to capture inpatient services.

3.3.2.3

Control variables

Among many hospital characteristics and environmental variables, the following are included in the analysis18 : medical officers (MO) per inpatient19 , hospital categories, teaching and referral status, location, occupancy rate and Indigenous population. Note that capital and material inputs are not included in the analysis due to the focus on labour efficiency, as contrast to the total factor productivity, technical and efficiencies. 18

See Table 3.9 for variable definitions This variable takes the form of a dummy instead of the original ratio. Both specification suggests the same conclusion but the dummy specification is used for interpretation convenience. 19

80


MO-per-patient is a dummy variable that captures the difference between hospitals that use more than the average amount of medical officers per patient and those that do not. This dummy was created by dividing the amount of FTE medical officers by the number of inpatient cases. This dummy is expected to have a negative coefficient as hospitals with lower MO per patient naturally uses less labour resources. Hospital categories, teaching and referral status, and location are also dummy variables that potentially influence hospital performance. Large and more specialised hospitals usually employ more staff. It is expected that larger hospitals would require more labours and thus, positive coefficients for the category dummies. Likewise, the teaching dummy is expected to return a positive coefficient as teaching hospitals require extra staff hour to attend residents and medical students. The coefficients for referral, city and town dummies might take a positive value as referral and city/town hospitals usually receive a high and sometimes unpredicted volume of inpatients which requires them to have better extra labour resource to attend emergency situations. The occupancy rate, calculated as the ratio of “occupied bed-days” on “total bed-days” in a hospital, measures the standing capacity of hospital and its ability to cope with uncertain demand. By construction, this variable is determined by total bed-days available, average length of stay and number of inpatients. The sign of the occupancy rate variable is less definite. This variable has been discussed in the hospital literature as indicator of quality of care (see for example, Keegan, 2008; Sprivulis, et al., 2006; Cunningham, et al., 2006) although there remains some concerns over its use. One can argue that the lower the occupancy rate, the more standing capacity the hospital has, and thus waiting time for a patient to get inpatient service will be shorter. It may also imply that there will be more resources devoted for patient treatment and better quality of care. The alternative argument is also persuasive. Some large (and busy) hospitals still achieve the same quality of care while operating at higher occupancy rates because they enjoy economies of scale (Smet, 2007). Indeed, high occupancy rate might be an indicator of higher productivity and higher volume of output produced. These two opposing arguments lead to disagreement on the implication of high occupancy rate on resource utilisation. A high occupancy rate implies more output and better quality thus more resource consumption, and on the other hand, a high occupancy rate can be a consequence of low reserve capacity, which implies lower resource utilisation (Smet, 2007). Lastly, a dummy is included to indicate a relatively high Indigenous population. 81

3.4. EMPIRICAL RESULTS

The health disadvantage among Aboriginal and Torres Strait Islander people is well documented. According to the QH (2007) report, there remains a great difference in general health status between Aboriginal and Torres Strait Islander people and non-Indigenous people although there have been a few small gains in some areas in recent years. The age-adjusted death rate of Aboriginal and Torres Strait Islander people in Queensland is estimated to be more than three times greater than that for the total population. The Aboriginal and Torres Strait Islander infant mortality rate is two and half times greater than the total population rate. Hence, it is reasonable to expect that hospitals in areas with larger aboriginal populations will receive, on average, higher flows of patients. Summaries of all variable statistics are provided in Table 3.9

3.4

Empirical results

The data are mean corrected prior to estimation. The advantage of doing this is that the estimated coefficients of the first order terms of outputs can be interpreted as labour elasticity at the mean, and thus economies of scale can be observed directly. The TRE model is estimated using NLOGIT4 and 500 Halton draws are used.

3.4.1

Estimates of the “true” random effect model

The coefficient estimates and their associated standard errors for the TRE model are reported in Table 3.10. All coefficients of the first order output terms are positive (as expected, given that increased labour (ceteris paribus) is required to produce more outputs), and statistically significant at the 5% level. As noted before, these coefficients can be interpreted as labour elasticities because the data is mean corrected. The coefficients imply that a 1% increase in number of medical inpatient cases (number of surgical inpatient cases) requires an additional 0.463% (0.209%) of labour. Compared to outpatient services, the coefficients of medical (W EP SM ) and surgical inpatients (W EP SS ) are two of the most influential variables on labour requirement. This reflects the fact that inpatient care is a major part of acute hospitals, and is labour resource intensive.

82


Table 3.9: Summary of statistics Variable

Definition

Aggregated

Labour aggregation using

labour

wages as weights

OUTP WEPS S WEPS M

Outpatient occasions of services Weighted episodes surgical care Weighted episodes of medical care

MO-per-

1 if MO-per-inpatient is less

inpatient

than average

Category A Category B Teaching Referral City Town

1 if a hospital is large and specialised 1 if a hospital is medium sized 1 if a hospital is classified as teaching 1 if a hospital is a main regional referral 1 if a hospital locates in the city 1 if a hospital locates in town

Occupancy

Total bed days divided by

rate

number of beds

Mean

Std. Dev.

Min

Max

336.482

688.042

15.027

4,691.146

83,410.430

133,854.100

1,291.000

101,1976.000

3,095.526

7,162.305

0.470

42,060.620

5,138.749

8,466.130

93.600

51,536.260

0.82292

0.38202

0

1

0.10714

0.30953

0

1

0.17857

0.38328

0

1

0.25000

0.43334

0

1

0.39286

0.48875

0

1

0.29762

0.45755

0

1

0.04762

0.21312

0

1

0.57118

0.20812

0.08691

1.14089

0.46429

0.49909

0

1

1 if a hospital locates in the Indigenous

area with the Indigenous

population

population higher than Queensland average

83


Table 3.10: Parameter estimation for TRE Coefficients

Std. Err

Output variables ln(OU T P )

0.12729***

0.01853

ln(W EP SS )

0.20869***

0.01101

ln(W EP SM )

0.46310***

0.01890

[ln(OU T P )]2

0.07565***

0.02280

[ln(W EP SS )]2

0.02574***

0.00456

0.07968**

0.03606

[ln(W EP SM

)]2

ln(OU T P ) × ln(W EP SS )

-0.00078

0.00724

ln(OU P T ) × ln(W EP SM )

-0.05060*

0.02702

0.01868*

0.00989

ln(W EP SS ) × ln(W EP SM ) Control variables MO-per-inpatient

-0.10615***

0.01377

Category A

0.41820***

0.03397

Category B

0.25017***

0.02522

Teaching

0.26478***

0.02038

Referral

0.08852***

0.01051

City

0.18952***

0.01574

Town

0.27209***

0.01962

Occupancy rate

-0.01106

0.02736

Indigenous population

-0.01227

0.00888

Variance parameters for the compound error Mean for random parameter (constant) Scale

parameter20

-0.72285***

0.02778

0.19183***

0.00482

Sigma (σ)

0.14679***

0.00463

Lambda (λ)

1.71086***

0.18802

Log-likelihood

418.0112

Sample size

672

***, **, * denote significant at 1%, 5% and 10% level.

84


The sum of output coefficients is also the indicator of scale economies: the first order output coefficients summing to a value smaller than unity indicates the presence of increasing returns to scale, to unity indicates constant returns to scale and to a value greater than one indicates the presence of decreasing returns to scale. Thus, the estimation suggests there is evidence of increasing returns to scale at the mean (coefficients summing to 0.799). The average labour elasticities with respect to the three outputs at every data points are also computed. Those numbers represent the relative importance of three outputs to labour utilisation as well as the change of economies of scale over time. As shown in Table 3.11, returns to scale did not change substantially and the relative importance of the three outputs was fairly constant although a slight decreasing trend is observed over time21 . Contributions of each output in scale economics are quite the distinctive. While outpatient services gain increasingly significant impact on labour requirement (close to 12% increase), surgical inpatient service’s contribution experienced overall decline (more than 6.5% decrease). Medical inpatient service stayed relatively constant. This reflects a trend in medical practices in the past years: shifting from surgical to medical interventions and from medical inpatient to outpatient treatments for less complicated cases. Cost-saving objectives, technological developments and the introduction of new drugs together are the main drivers of those practices. The hypothesis of increasing returns to scale can be tested formally using a likelihood ratio test. The log likelihood value of the unconstrained model is much larger than that of the model with constant returns to scale (CRS) constraint, and the calculated likelihood ratio is 30.93. The hypothesis of CRS is rejected at 1% level of significance. These estimations suggest the existence of unexploited scale economies. This is consistent with the Queensland situation: the public hospital system spreads itself thinly to handle the majority of acute care separations and to account for most regional and remote areas with very small populations. PC (2009) noted that Queensland has one of the highest concentrations of small-scale hospitals in comparison to other States: over 80% of hospitals have no more than 50 beds. However, they still need to provide a reasonable (to full) range of specialist inpatient, outpatient, emergency and diagnostic services at all times. Hospital size has implications for resource efficiency. Compared to large hospitals, small hospitals are less likely to be able to 21 Note that the average elasticities are smaller than the estimated elasticities at the mean (the first order coefficients of outputs). This is due to the left skewed distribution of public hospitals. That is there are many small hospitals in Queensland. Estimation at the median will probably give a closer number to the average elasticities. However, this does not change the conclusion that increasing returns to scale exist in this sector

85


take advantage of economies of scale or reallocate their resources when work-flows vary, thereby usually operating relatively less efficient (PC, 2009). Fast population growth and the expansion of small towns around Queensland might contribute to changes in scale efficiency in the future22 . Table 3.11: Estimates of labour elasticities Elasticities of

OU T P

W EP SS

W EP SM

SUM

1996

0.1032

0.1185

0.3776

0.5993

1997

0.1038

0.1235

0.3829

0.6102

1998

0.1049

0.1231

0.3824

0.6104

1999

0.1052

0.1222

0.3822

0.6096

2000

0.1149

0.1175

0.3719

0.6042

2001

0.1162

0.1147

0.3688

0.5998

2002

0.1175

0.1132

0.3661

0.5968

2003

0.1155

0.1107

0.3662

0.5924

Although the results indicate that on average Queensland Health public hospitals operate at the sub-optimal scale, it can be argued that scale efficiencies should not be an issue in health care facilities. It is unreasonable to recommend closing small public hospitals or merging them to create bigger and more “scale optimal” hospitals. Small hospitals are designed to serve low population areas in order to ensure equity of access; their importance to communities cannot be underestimated. However, they are viable operations for the private sector because small size and rural location mean those hospitals are often relatively expensive to operate and commercially less profitable. Therefore, following a strict efficiency maximising policy would lead to the closing or downsizing public facilities in rural and remote areas, and thus deprive people living in there from getting equal access to health care services. Turning to other hospital characteristics, the MO-per-inpatient dummy variable is negative, as expected and significant at 1% level. This implies that hospitals having higher-than-average ratios, on average, tend to use more labour than those having smaller-than-average ratios. The difference in labour requirement is approximately 11.2%. The large and significant coefficients for the category dummies suggest that labour utilisation varies by size. Large and specialised hospitals require 51.9% more 22

When hospitals expand to cope with larger population, they usually enjoy economics of scope as well as scale economies due to the high degree of complementarity of health services (facility, equipment and medical skills are in many case interchangeable between emergency, outpatient and inpatient services). However, economies of scope is not the focus here because economies of scale is more relevant to the equity of access implication

86


labour resources than smaller hospitals for the same volume of outputs. This number is 28.4% for medium sized hospitals. This can be explained by the fact that larger hospitals usually keep a certain volume of resources permanently on standby to serve unpredicted demand. Without the standby capacity, the need to divert resources to emergencies can interrupt and constrain the delivery of other services, especially surgical. Additionally, larger hospitals usually provide a wider-range of complementary services such as catering and laundry. The estimated results also indicate that on average, teaching hospitals require around 30.3% more labour resources than the non-teaching ones. This is intuitive given that in teaching hospitals, medical staff need to attend trainees and medical students and thus, more working hours are required. This finding is consistent with other studies that investigated the effect of teaching hospitals on cost. For instance, Butler (1995); Cameron (1985); Kane et al. (2005); Sloan et al. (1983); Frick et al. (1985) found that the university hospitals are more costly than their non-teaching counterparts. However, it is noted by Pardes (1997) that teaching and non-teaching hospitals produce different outputs, (i.e. non teaching hospitals do not produce the output of research and trained students) which make comparison difficult. Using output-cost ratios without removing the effect of teaching activities introduces a downward bias on efficiency of teaching hospitals. In this study, we include the teaching dummy to capture this effect. Being a principal referral hospital means using more labour, as the coefficient of the referral dummy is positive and statistically significant at 1% level. This confirms the expectation that a referral hospital usually has a higher reserve capacity to deal with extra unforeseen demand, and thus uses more labour. In Queensland, a referral hospital is usually the largest one in the service area, capable of dealing with complex cases, equipped with more modern technologies and specialised staff. In order to service patients referred/transferred from other smaller regional hospitals, extra spare capacity is usually required. According to our estimation, a referral hospital on average will need 9.3% more labour compared to a non-referral one. Estimated coefficients of the two location dummies, city and town, are positive as expected, and significant. This suggests that city and town hospitals tend to require more labour, compared to hospitals locating in shire and islands. The difference in their magnitudes is an interesting observation: town hospitals need 10.6% more labour than city hospitals, other things being equal. This indicates that town hospitals cannot smooth out their resource utilisation as well their city counterparts. This can be explained by the difference in population densities between town and 87


cities and hospitals’ strategies to cope with unpredicted demands. In large urban areas, it is much easier for a hospital to refer patients to a nearby hospital if it does not have spare capacity while in small urban areas like towns, the closest referral or equivalent would be rather far away. Occupancy is insignificant which indicates that it is indeed has opposing effects on resource requirements. High occupancy rates implies more output produced and thus more input consumption. On the other hand, it can be the consequence of low reserve capacity, which suggests less resources are needed. The last dummy variable - Indigenous population - is not statistically different from zero, suggesting that hospitals in the area with a higher than average Indigenous population do not require more labour resources. This observation is counter to expectation but may be explained by the clinical ice-berg associated with the Indigenous population propensity to present to formal care.

3.4.2

Efficiency predictions

The means of predicted labour efficiency for each health service district over eight years are presented in Table 3.12. On average, labour efficiency improvement over the studied period is 4.18% with a noticeable improvement in the first three years, reaching the highest in 1998 at 92.8%, and then slightly decreases over the rest of the period. This efficiency score implies that hospitals in Queensland, on average, could use 7.2% less amount of labour to deliver the same amount of activities. Across districts, efficiencies appeared to fluctuate from year to year, with some changes of large magnitudes, from positive 31.5% (Torres Straight) to negative 7.4% (Southside district). It is quite noticeable that the performance of most districts in the North Queensland (with the exception of the Torres Strait) is worse than those in South East Queensland (the exception is Sunshine Coast and Southside districts) (see Figure 3.1). Seven districts experienced an overall downward trend of efficiency; these are Cairns and Hinterland, Cape York, Mackay and Mt Isa in the North Queensland, Southside and the Sunshine Coast of South East Queensland, and Central West district. It is noted that Mt Isa and Sunshine Coast districts - the two top performers in 1996 - have fallen behind other districts in 2003, lower than the Queensland average efficiency. Two districts that have made significant progress in productivity and efficiency improvement were Torres Strait (31.5%) and Fraser Coast (23.03%), becoming the most efficient from the two least efficient districts over the study period. 88


Figure 3.1: Queensland Health Service Districts and their performance over time

Districts have experienced deterioration (or almost zero growth) of labour efficiency: most districts in North Queensland except for Townsville; Sunshine Coast and Cooloola, and South Side of South East Queensland

Most health service districts in South East Queensland, except for the Southside and Sunshine Coast, have experienced efficiency growth. Torres Strait is the only district in North Queensland that achieved efficiency improvement.

89


Among 21 health service districts, 8 have larger than the Queensland average Indigenous populations. Mean estimated efficiency for these suggest that they are slightly less efficient than hospitals locating in areas with lower than average Indigenous population (efficiency estimates of 0.628 vs. 0.635). However, the Wilcoxon rank-sum (Mann-Whitney) test suggests that there is no significant difference in predicted efficiency between the two hospital groups (z-statistics equals -0.551). Table 3.12: Predicted efficiency by Queensland Health Service Districts Districts

1996

1997

1998

1999

2000

2001

2002

2003

C&H

0.9112

0.9179

0.9410

0.9089

0.8847

0.8817

0.8999

0.8733

Cape York

0.8883

0.8992

0.9241

0.9457

0.9218

0.9471

0.8785

0.8717

Central Qld

0.8778

0.9236

0.9390

0.9226

0.9047

0.9081

0.8976

0.9099

Central West

0.9023

0.9096

0.9328

0.9210

0.9316

0.9308

0.9216

0.8833

Fraser Coast

0.8075

0.7738

0.8545

0.8573

0.9010

0.9410

0.9665

0.9684

Gold Coast

0.8933

0.8831

0.8652

0.9401

0.9362

0.9477

0.9387

0.9304

Mackay

0.9004

0.9008

0.9400

0.9052

0.9122

0.9359

0.9075

0.8942

Mater Public

0.8954

0.9126

0.9286

0.9205

0.8909

0.9005

0.9236

0.9205

Mt Isa

0.9303

0.9366

0.9566

0.9202

0.9147

0.8789

0.8786

0.8874

Northside

0.8836

0.9267

0.9105

0.9219

0.8968

0.9161

0.9412

0.9294

PA

0.8169

0.8904

0.9444

0.9438

0.8799

0.9011

0.9441

0.9423

RBW

0.8554

0.8774

0.9461

0.9488

0.9196

0.9049

0.9365

0.9156

RCH

0.9035

0.9038

0.9432

0.9031

0.9274

0.9031

0.9368

0.9175

South West

0.9194

0.9351

0.9231

0.8897

0.9185

0.8638

0.9149

0.9257

Southside

0.8975

0.9354

0.9360

0.8843

0.8910

0.8675

0.8524

0.8312

SC&C

0.9211

0.9226

0.9594

0.9310

0.9152

0.8788

0.8616

0.8575

T&DD

0.8120

0.8579

0.9232

0.9038

0.9196

0.9237

0.9281

0.9393

Torres Strait

0.7212

0.8074

0.9345

0.9483

0.9301

0.9078

0.9192

0.9484

Townsville

0.8916

0.9284

0.9440

0.9212

0.9286

0.8919

0.8800

0.8927

WMSB

0.8257

0.8875

0.9050

0.9349

0.9429

0.9069

0.9249

0.9267

Wide Bay

0.8852

0.9122

0.9349

0.9198

0.9125

0.8928

0.9267

0.9404

Average

0.8733

0.8972

0.9279

0.9187

0.9133

0.9062

0.9133

0.9098

C&H: Cairns & Hinterland; PA: Princess Alexandra RBW: Royal Brisbane & Women’s; RCH: Royal Children’s SC&C: Sunshine Coast & Cooloola; T&DD: Toowoomba & Darling Downs WMSB: West Moreton South Burnett

Over the eight-year period, teaching hospitals appear to perform better compared to the non-teaching ones (see Table 3.13). They have done exceptionally well in the last two years. There are a few possible explanations for this observation. Teaching hospitals usually have better skilled doctors and nurses, who are eligible to teach medical students and mentor junior practitioners. Better skills help them to increase 90


the probability of making appropriate diagnoses and treatments, thus raising the overall effectiveness of hospital services that contribute to higher efficiency. On the other hand, this result can be entirely driven by the fact that students and interns who take up some of diagnostic work in the teaching hospitals are not included in the labour measure; thus teaching hospitals appear to look more efficient with the use of medical staff. Table 3.13: Labour efficiency of teaching vs. non-teaching hospitals 1996

1997

1998

1999

2000

2001

2002

2003

Non-teaching

0.8743

0.9034

0.9304

0.9128

0.9155

0.9038

0.8988

0.9014

Teaching

0.8820

0.9079

0.9300

0.9204

0.9083

0.9026

0.9333

0.9212

As shown in Table 3.14, city hospitals are the most efficient, while hospitals that locate in towns and shires seem to perform equally well, and on average, are better than hospitals on islands. This seems to follow the common expectation/findings (among which is the most recent research report by the Productivity Commission (PC, 2009)) that urban hospitals (city and town) are more efficient than rural hospitals (shire and islands). Firstly, urban hospitals are usually large as they serve high population density areas, thus having high volume of patients. Furthermore, operation at a large scale is directly linked to better equipment and more specialised departments. Hence, in addition to patients in their service areas, they also serve patients who are referred to them from other smaller hospitals, which further increases their demand for health workers. This allows them to exploit the economies of scale and scope - the two important sources of high productivity. Albeit having the lowest mean efficiency in the first two years, the performance of island hospitals has improved significantly (overall increase of 31.5%). This trend is also observed in town hospitals although at a lower degree (around 18.8% up). From 1999 to 2003, town and island hospitals slightly outperformed their counterparts in cities and shires. There was little improvement seen in shire hospitals, while city hospitals raised their efficiency by 3.6% over the course of eight years. This may explain why the Wilcoxon rank-sum (Mann-Whitney) tests do not reject the hypotheses that there is no significant difference in predicted efficiencies between hospitals in different locations. On average, large and specialised hospitals achieved the highest efficiency. It is observed that large and specialised as well as medium hospitals have improved their efficiencies over time while the smaller hospitals have worsen. Although smaller hospitals are expected to achieve lower productivity compared to the larger ones 91


Table 3.14: Efficiency estimates by location of hospitals 1996

1997

1998

1999

2000

2001

2002

2003

City

0.8802

0.8992

0.9269

0.9196

0.9130

0.8984

0.9168

0.9116

Town

0.7853

0.8422

0.9192

0.9199

0.9223

0.9229

0.9159

0.9330

Shire

0.8870

0.9154

0.9325

0.9108

0.9127

0.9043

0.9019

0.9003

Island

0.7212

0.8074

0.9345

0.9483

0.9301

0.9078

0.9192

0.9484

because they do not have economies of scale and scope, efficiency deterioration does not necessarily relate to size or location, especially when the smallest hospital group did not share the same experience. They experienced growth, as shown in Figure 3.2 although to a lesser extent compared to the larger sized group. It is rather difficult to speculate any causes. The best guess comes from two major forces: population growth and aging in Queensland, and the shortage of funding, and thus medical staff in small hospitals. The former creates more health services demand for hospitals, which effectively increasing throughputs. Standing resources in the smallest hospitals were employed more often, resulting in to higher labour efficiently. This did not happen in slightly larger (but still small) hospitals because their resources might have always been just sufficient to serve lower demand. Funding shortage and less attractive working environment could result in less experienced and qualified medical staff, thus lower labour productivity. It would be beneficial for decision makers in Queensland Health to investigate the possible causes of these trends to avoid adverse factors and enhance the productivity-augmented practices. In addition, they could aim to address this situation in the new funding model.

92

3.5. SENSITIVITY ANALYSIS

Figure 3.2: Labour efficiency by hospital size

Efficiency estimates

0.95 0.93 0.91 0.89

0.87 0.85 1996

3.5

1997

1998

1999

2000

2001

Large&Specialised

Medium

Smaller

Smallest

2002

2003

Sensitivity analysis

It is noted that transparent and correct policy conclusions cannot be made if one is not aware of the risks of using one model specification over the others. Estimated results, both coefficients and efficiencies, are obviously driven by modelling choices as shown in Chapter 2. Here, the sensitivity analysis considers the effects on the coefficients and efficiency estimates when the labour requirement function is estimated using three alternative models. This exercise demonstrates the robustness (or the lack of which) of coefficient and efficiency estimates under different estimation strategies. The three alternative estimation strategies considered here are the usual random effect (RE), a pooled panel frontier (POOL) and a time-varying model (BC). RE is the usual random effect model with the individual specific effects assumed to follow a normal distribution. It is estimated using maximum likelihood and the individual inefficiencies are estimated by the conditional expectation of the inefficiency term (see Equation 3.5). A limitation of this model is that the inefficiency estimates are time-invariant. However, one can argue that eight years is not a too long panel, hence efficiency variation over time might be minimal. It is expected that the efficiency estimates by RE are lower than those by TRE. The TRE picks up latent heterogeneity through the specification of wi while heterogeneity is dubbed 93


with inefficiency in the RE. POOL treats the data as a series of sectional sub-samples pooled together. This specification allows inefficiency estimates to vary over time, as shown in Equation 3.2. However, its obvious disadvantage is that the panel structure is ignored because each observation is treated as individual hospitals. It also assumes away the hospitals’ heterogeneity, which usually lead to overestimation of inefficiency. The last model is the time-varying model BC (Battese & Coelli, 1992) where inefficiency is defined as uit = f (t)Ui with Ui follows a half normal distribution, and the time function is specified as f (t) = exp [−η(t − T )]. The exponential specification implies that technical efficiency must either increase at a decreasing rate (η < 0), decrease at an increasing rate (η > 0) or remain constant (η = 0). This assumption might be too restrictive if the actual efficiency fluctuates over time. These models are selected for several reasons. Firstly, it is useful to compare the estimates of TRE and the others because the main difference lies in the treatment of the unobserved heterogeneity. While the other models dubs it with inefficiency, the TRE separates it from any inefficiency effect. Secondly, POOL and TRE are quite similar in the inefficiency term structure when allowing it to vary over time. However, the former ignores the panel structure of the data and thus, is less “efficient” than TRE because some cross section characteristics will be ignored. Finally, although BC takes into account the panel structure of the data and allows inefficiencies to vary, its time function can be too restrictive compared to the inefficiency specification by TRE. The estimated coefficients of four models are reported in Table 3.15. Overall, the labour elasticities of the three outputs are significant and their magnitudes are fairly constant across models. The evidence of increasing returns to scale is observed across all models. Most squared and cross terms of outputs are also significantly different from 0 at 10% level or better. Compared to other models, the POOL produces larger coefficient estimates for the OU T P and W EP SS and lower coefficient estimates for W EP SM . This difference is likely to come from ignoring the panel structure of the data set by the POOL. Although this does not create a different order of relative importance of the outputs, their magnitudes suggests that the two types of inpatient occasions of services, medical and surgical, equally influence the labour requirement. Estimates from the other three models imply that medical inpatient services are relatively more important than surgical ones. This is plausible because medial inpatient services take up 94


the largest share of hospital outputs and cover much larger range of diseases and thus labour skills. The RE, however, appears to over-estimate the effect of medical inpatient services, reflected through the largest W EP SM coefficients. Based on the obvious distinction between this model and the other two, the TRE and BC, it can be inferred that the over-estimation may come from the time-invariant structure of the efficiency term. Table 3.15: Estimated coefficients by alternative models

Constant

TRE

RE

POOL

BC

-0.72285***

-0.15243***

-0.23571***

-1.00908***

ln(OU T P )

0.12729***

0.12136***

0.18306***

0.10535***

ln(W EP SS )

0.20869***

0.20303***

0.34139***

0.19424***

ln(W EP SM )

0.46310***

0.51220***

0.35939***

0.46293***

[ln(OU T P )]2

0.07565***

0.07625*

0.15934***

0.08543**

0.02574***

0.02728***

0.04851***

0.02461***

0.07968**

2

[ln(W EP SS )]

2

[ln(W EP SM )]

0.12300*

0.16454***

0.07851

ln(OU T P ) × ln(W EP SS )

-0.00078

0.00116

0.02686**

-0.00402

ln(OU P T ) × ln(W EP SM )

-0.05060*

-0.06654

-0.15882***

-0.05302

0.01868*

0.01019

ln(W EP SS ) × ln(W EP SM ) MO-per-inpatient

-0.10615***

-0.10510***

-0.01833 -0.15611***

0.02026 -0.10566***

Category A

0.41820***

0.22096

0.10752*

0.71490**

Category B

0.25017***

0.01224

0.03721

0.55489**

Teaching

0.26478***

0.16318

0.15482***

0.23328

Referral

0.08852***

0.05314

0.07190***

0.04642

City

0.18952***

0.19317*

0.05408*

0.19106

0.18702***

0.25896***

Town

0.27209***

0.25351**

Occupancy rate

-0.01106

-0.02771

-0.04654

0.00751

Indigenous population

-0.01227

0.02119

0.00398

-0.01287

Log-likelihood

418.011

396.814

183.684

422.362

Lambda

1.71086***

3.62965***

0.68175***

8.07300

Sigma

0.14697***

0.39297***

0.20614***

1.37263

***, **, * denote significant at 1%, 5% and 10% level.

As for the control variables, despite the differences in magnitude, all coefficients have their expected sign23 . The comparison of the estimated coefficients for the control variables24 is illustrated in Figure 3.3. As TRE is the benchmark model, the ratio of estimates coefficients were produced by dividing coefficients of RE, POOL and BC to the respective ones of TRE. The results show that the coefficients of 23

The coefficient of the constant term is not of interest in this case as it has a different interpretation in the case of the TRE. 24 Occupancy rate and Indigenous population are excluded in this figure as they are not statistically significant in any model.

95


Figure 3.3: Ratio of estimated coefficients from different models with respect to the “true” random effect model 25 20 15 Category B

10 Category A

5 0

Referral MO per patient

Teaching

City

Town

-5

RE

POOL

BC

the category dummies (especially category B) and city location are significantly influenced by model specifications, reflected by much larger than unity ratios, while the other coefficient estimates appear to be quite robust across models. It is noticed from Figure 3.4 that the RE model produces much lower efficiencies compared to the TRE that distinguish hospital unobserved heterogeneity from inefficiencies. The usual random effect specification is expected to produce lower efficiencies as it pools together the “persistent inefficiency” and “time-varying inefficiency”, which is captured in the composite error term in the TRE model. In both the POOL and BC models, hospital latent heterogeneity is assumed to equal zero. It is predicted that ignoring the unobserved specific effects will over-state the inefficiency estimates as this heterogeneity is subsumed into the inefficiencies giving favour to hospitals with favourable environmental conditions, among other things. It is also observed that the POOL produces very similar mean efficiency values to the benchmark model while the BC predicts a much lower mean efficiency. It is also noted that the removal of hospital individual effects (in the TRE model) increases the discriminating power and reveals a more obvious efficiency trend of Queensland Health public hospitals (compared to the POOL model). Figure 3.5 presents the predicted efficiencies grouped by health service districts. Alternative models produce quite different ranking of district performance. The 96


0

0.0

2

0.5

4

1.0

6

8

1.5

10

2.0

12

Figure 3.4: Distributions of predicted efficiencies by alternative models

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.6

0.8

1.0

0.6

0.8

1.0

(b) RE

0

0.0

5

1.0

10

2.0

15

3.0

(a) TRE

0.4

0.0

0.2

0.4

0.6

0.8

1.0

0.0

(c) POOL

0.2

0.4

(d) BC

pooled and true random effect models produce a quite similar ranking, while the other two models give a slightly different picture. An explanation for the ranking difference is the way each model treats efficiency and individual heterogeneity, as described above. A closer look at the predicted efficiencies through the Spearman rank correlation confirms the variation of efficiency predictions across models. Table 3.16 summarises the correlations between the various sets of efficiency predictions for individual hospitals. Efficiency estimates from the RE model exhibits very low correlations with those by other models. The POOL model produces efficiencies that are weakly correlated with those of the benchmark model and the BC. This can be explained by its specification: the pooled model shares the “efficiency term” specification with the bench mark model while omitting the hospital heterogeneous effect just like the BC model.

97


Figure 3.5: Predicted labour efficiency by alternative models, by health district

C&H Wide Bay WMSB Townsville Torres Strait T&DD

Cape York Central Qld

1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00

Central West Fraser Coast Gold Coast

SC&C

Mackay

Southside

Mater Public

South West

Mt Isa RCH

RBW

TRE

PA

RE

Northside

POOL

BC

Table 3.16: Spearman’s correlation of predicted efficiencies TRE TRE

1.00000

RE

-0.06640

RE

POOL

BC

1.00000

(0.08530) POOL BC

0.54780

-0.58820

(0.00000)

(0.00000)

1.00000

0.04240

-0.62750

0.59140

(0.27200)

(0.00000)

(0.00000)

Note: two-sided p-value in parentheses

98

1.00000


3.6

Concluding remarks

This chapter has used a stochastic labour requirement frontier to examine the labour efficiency of Queensland Health public hospitals from 1996 to 2003. This function is estimated using the true random effect framework developed by Greene (2004, 2005a) with the inclusion of selected control variables to capture the observed hospital heterogeneity. This framework is chosen because it can accommodate timevarying efficiency, time-invariant control variables and latent heterogeneity all together. This makes it superior to other alternatives such as the usual random and fixed effect models and other time-varying models such as the BC, as well as its twin sister - the “true fixed effect”. Sensitivity analysis was performed on alternative model specifications, including the usual random effect, a pooled panel frontier, and the time-varying model following the specification of (Battese & Coelli, 1992). The results indicate that individual efficiency estimates are rather sensitive to different estimation strategies, especially between the models with hospital heterogeneity specifications and those without. Hence, it is important to bear that there are risks in choosing one model over others. On the other hand, the comparison of the estimated coefficients for labour elasticities of outputs suggests that the general specification of selected variables is quire robust. The magnitudes of estimated labour elasticities of outputs are fairly constant across models. The analysis brings some implications for Queensland policy makers and future studies. First, the results suggest that despite overall improvement of labour efficiency over the period 1996-2003, most improvement came from the first three years. This trend is observed consistently across hospital categories, either by location or size. Furthermore, amongst 21 health service districts, some experienced slight deterioration in their performance while others, especially the Fraser Coast and Torres Strait districts showed considerable growth of efficiencies. A closer examination of the medical labour performance by location reveals that island and town hospitals achieved significant labour efficiency gains over the study period. By 2003, they outperformed city and shire hospitals. Another noticeable movement is the continuous deterioration of labour efficiency in medium-small-sized hospitals over the eight year periods while the opposing trend is observed by the rest. Understanding the causes of these movements would certainly benefit future planning and management strategies by promoting efficiency augmented practices while eliminating adverse factors. This is important because the improvement of labour productivity and effi-

99


ciency is necessary for overall efficiency improvement. It is therefore worthwhile for Queensland Health to further investigate these trends. Second, although the possibility of closing down and scaling up small hospitals in rural and remote areas may not be an option due to equity concerns, scale efficiency has implications for future residential development and (industrial) project evaluations. This is especially true in the case of Queensland, where population and thus health resources are spread thinly across the State, and the cost of inefficiency is this significant. This comes from not only scale inefficiency but also distance and service mix. Hospitals in remote areas are likely to incur a higher cost of transporting hospital supplies as well as greater difficulty attracting staff, which may necessitate higher wages. Some hospitals in remote areas have an added responsibility to provide primary health and aged care services, which would otherwise not be provided in their areas (PC, 2009). Opening a new residential town, usually far away from everywhere else means a new (and inefficiently small) hospital or an outpatient clinic is needed. This requires reallocation of medical staff from other places Therefore, the social cost of scale inefficiency ought to be one of the arguments against the opening of new residential towns in rural and remote areas25 . Third, different provider payment methods are believed to have various effects on efficiency seeking behaviours26 . Queensland Health has recently replaced the historical cost funding formula with a casemix funding model, aiming at efficiency and quality improvement. This change would undoubtedly influence behaviours of medical staff and their performance. The historical cost payment is usually discussed in the context of lacking efficiency and quality incentives due to its irresponsiveness to demand changes and budget rationing to keep the tight (and usually insufficient) capped budget. During the study period, the historical cost funding method was implemented in Queensland Health public hospitals while resident medical staffs receive a fixed salary payment. The new casemix funding model is believed to positively affect productivity by encouraging medical staff to favour treatments under lower cost settings while maintaining quality of care (QH, 2008). It would therefore be beneficial to investigate the behaviour changes of medical staff and the evolution of labour efficiency before and after the reform. Like many other studies, this analysis fails to address the relationship between quality of care and efficiency. It was not possible to obtain data on service qual25

It is noted that there may be more powerful arguments in favour of opening new towns, for instance to accommodate new industrial projects. They are beyond the scope of this thesis. 26 As summarised in Appendix D

100


ity (such as average waiting time for emergency services and scheduled surgical procedures, percentage of errors in the handling of cases etc) for the study period although there is evident that Queensland Health does have a quality control system. Therefore, the effect on the results due to quality differences among hospitals were not explored. This potential produces a negative bias toward hospitals that provide better quality of care since higher quality tends to go hand in hand with more resources consumed per case, through longer length of stay and more medical hour attendance27 . Another drawback of this paper relates to labour measures due to the lack of more detailed separation of staff skill mix within each labour category. Although the sensitivity analysis on the dependent variable of labour demonstrates that the results are robust to different labour measures, the risk should still not be underestimated. Staffing is generally not a stable variable, with wide variability in levels and ratios across settings. Hospitals with higher skilled health workers might be able to produce more outputs per FTE unit of staff, thus being more efficient. A significant staffing inequality between urban and rural/remote hospitals is usually observed because it is very difficult to recruit a sufficient number of high skill doctors for some regions of Queensland, either because it is quite remote, or due to the lack of the professional and “technologically dynamic” environment that only major cities would accommodate. If higher and lower skilled staff are grouped into one health worker class, and there is no adjustment for the skill differences, this might introduce ambiguity in the results, more specifically understate efficiency of hospitals with lower skilled health staff. This might be the reason for our observation of the slight differences between predicted efficiencies between teaching and non-teaching hospitals, city and town hospitals and shire and island hospitals.

27

It has been discussed in Rosko & Mutter (2008) that the lack of quality adjustment is an important issue in efficiency analysis, that however sometimes is overrated

101

Chapter 4

Funding reforms and public hospital efficiency in Vietnam

“Health reform is not a purely technical exercise; it is as much about capacity, responsibilities and accountability.” Lieberman & Wagstaff (2009). Vietnam is classified as a low-income country with a GDP per capita of US$724 (in 2006) and a population of 84 million (data for 2006 by World Development Indicators). According to the World Bank, over 15% of the population are classified as poor, of which around 93% live in rural areas (calculated from WB, 2008b, page 5). Recent studies on the health system and public hospitals of Vietnam have praised Vietnam as a successful story: its performance is comparable to many countries of a much higher income rank (see for example Adams, 2005; WB, 2007, 2008a). Most of the vital health indicators are better than would be expected for a country at its development level, and some indicators are even comparable to those of much wealthier countries (Adams, 2005). The health sector has also been successful in providing preventive health services and in controlling key communicable diseases (UNDP, 2003). These outcomes result from an extensive health care delivery network, an increasing number of qualified health workers, and expanding national public health programs. However, it is believed that while the system has been doing better than expected, challenging issues regarding equity and efficiency remain. There have been many studies looking at the issue of health care equity in Vietnam (for examples, Segall et al., 2002; Sepehri et al., 2003; Wagstaff & van Doorslaer, 2003; Thorson & Johansson, 2004; Luong et al., 2007; Chaudhuri & Roy, 2008; Dao et al., 2008). These outnumber the studies on efficiency. The exhaustive literature 102

search resulted in only two studies on hospitals efficiency in Vietnam by Nguyen & Giang (2007) and Uslu & Pham (2008). Both of them have employed a DEA two-stage framework, that is technical efficiency computed by DEA followed by a regression analysis. However, the former looks at technical efficiencies of non-state providers while the latter makes use of a richer dataset (101 hospitals over 9 yeas, 1998-2006) of public hospitals to investigate the effects of regulatory changes on hospital productivity and efficiency. Uslu & Pham (2008) reported an average annual productivity growth of 1.9% over the studied period. Importantly, user fee and financial autonomy are found to have increased technical efficiency although the improvement is not uniform across regions and hospital types. The problem with this two-stage approach is that it suffers from the “separability problem” of input-output production space and the environmental variable space described in Simar & Wilson (2008). That is a typical two-stage analysis requires an implicit (but not stated) assumption that the environmental variables do not influence the shape or boundary of the technology set (for the first stage of DEA estimation). This is a rather restrictive assumption, especially when one has the knowledge that environmental factors play a role in defining the technology frontier. Also inspired by the question of reform impact, the study presented in this chapter uses an alternative approach to this question. The meta-frontier framework neither requires the specification of environmental measures, which in many cases are not available, nor does it suffer from separability problem discussed above. It captures the “environmental effects” in the measurement of “technology gap” and benchmarks a producer’s performance in the context of its operating environment, i.e. the environment-specific (or group) frontier. More specifically, this chapter focuses on analysing the effects on technical efficiency of the financial autonomisation reform (the introduction of Decrees 10/2002 and 43/2006) in the current institutional arrangement. This reform requires that public providers become more self-sufficient financially. Providers are encouraged to earn more income from patients and are allowed to use extra revenues to pay staff and invest in equipment and facility. It is claimed that the reform had created stronger economic incentives for expanding management scope, increasing in productivity of medical staff and promoting efficiency of health service provision (Nguyen et al., 2008). These aspects are discussed in detailed in the next sections.

103

4.1. VIETNAM’S HEALTH SYSTEM AND PUBLIC HOSPITALS

4.1

Vietnam’s health system and public hospitals

Despite being one of the poorest countries in Asia, Vietnam’s overall state of health has been in much better shape than what is seen in countries at the same level of development. Data from the WHO shows that overall Vietnam spent 6.6% of GDP on health in 2006 (WHO, 2008), comparable to other low and middleincome countries. Its annual total health expenditure per capita was around US$46, of which 32.3% (or US$15) was from public sources, while per capita income was estimated at US$724. Vietnam’s healthy life expectancy (HALE) at birth in 2006 was 72 years, not very different from life expectancies in wealthier countries such as Brazil, Malaysia, and Hungary1 . Between 2000 and 2005, the life expectancy (of males and females together) improved from 67.8 years to 71.3 attributable to a decline in the infant mortality rate, from 36.7 per 1,000 to 17.8 and the mortality rate of children under 5 years, from 42.0 per 1,000 to 27.5. During the same period, the maternal mortality rate also fell from 85 per 100,000 to 80, the fraction of children with low weight at birth from 7.3% to 5.1%, malnutrition among children under five years from 33.8% to 25.2%. This places Vietnam in a relatively similar position to many countries with considerably higher per capita income2 (WB, 2007). However, these positive outcomes have not been shared equally across regions and by different socio-economic groups. The gap in child survival prospects between the poor and the rich has widened: the smallest reductions in both infant and under-five mortality rates were observed in the poorest quintile during the period 1997-2000. Improvement in child vaccination rates was not seen in the poorest groups while the richest quintile experienced a 55% increase in its immunisation rate. Regional differences are also alarming. Infant mortality has fallen faster in richer southern regions and the Red River Delta while remaining high in the Northern Mountains and Central Highlands (WB, 2007). 1 As indicated in the UN report of World Population Prospects (UN, 2006) and the World Health Statistics 2008 (WHO, 2008). 2 There is no doubt that this progress is also possible due to rapid income growth and better community infrastructure. However, Vietnam still remains a good performer in terms of infant and under-five child mortality even after adjusted for economic growth (WB, 2007).

104


Figure 4.1: Child mortality rates in Vietnam and selected countries

(a) Infant mortality rate (per 1,000)

(b) Under-five mortality rate (per 1,000)

Source: Vietnam Development Report 2009: Social protection (WB, 2007).

Figure 4.2: Infant mortality rates by region

Source: Vietnam Development Report 2009: Social protection (WB, 2007).

105


4.1.1

Health sector financing

For most of its modern history, Vietnam has relied on the public system to finance and deliver health care services to its population. Government-owned and administered health care providers receive state subsidies to provide services for the population, especially primary care and inpatient services, and some community care. However, the country has seen a decreasing role of government in both funding and provision of health care services in the last twenty years. Since about the same time that the country adopted market-oriented reforms (early 1990s), financing and provision of health services have been gradually privatised (WB, 2007). Private payment, including hospital fees and drug expenditure, has become the dominant source of health financing. By 2005, public sources of funding accounted for less than 25% of the total health expenditure. As a percentage of total general government spending, expenditure in health has actually decreased. More than 70% of Vietnam’s total health expenditure mobilized from private payments has been mainly in the form of household direct out of-pocket payments paid at the point of services3 . This is high compared to other other low and middle-income countries4 . Privatisation of service provision has happened at a much slower pace. In 2007, there were over 1,000 government run versus only around 50 privately owned and operated hospitals, totally offering around 144,000 active beds. Private providers only account for 4% of inpatient services and 11% of preventive health care although it plays a greater role in the provision of out-patient services, where it handles 60% of visits (WB, 2007). The financial flows of the Vietnam health system are described in Figure 4.4. As for public hospitals and health centres, revenue comes from the state subsidies, social health insurance and fee for service revenues collected from patients, and sometimes foreign aid. Foreign aid (or Official Development Assistance) is a temporary financial resource facilitating the long-term policies aimed at improving economic and social access to health care services. The state subsidies, usually based on bed norm, aim at covering some of the fixed costs of public hospitals (including salaries, utilities and essential running costs) to provide basic care at the community levels (immunisation, family planning, malaria control and so on) and inpatient care for the poor and the 3

Private health insurance is still practically non-existing in Vietnam due to a very small number of insurance companies offering this service. 4 In fact, Vietnam was identified as having the highest incidence of catastrophic health care spending in the world, together with Brazil (WB, 2007).

106


Figure 4.3: Sources of total revenue in public hospitals (percent)

1.2

23.2

1.1 30.7

1.6

4.0

6.3

34.7

27.0

24.9

8.8

31.0

7.2 10.1 11.9

17.0

5.1

5.9

3.8

6.1

32.8

32.1

35.7

34.8

16.2

14.3

7.1

7.3

7.0

38.8

35.4

33.0

21.4

23.0

28.0

32.7

34.3

32.0

14.8 12.5

13.2

20.3

68.4 58.1

51.8

52.0

54.0

47.7

48.9

45.8

46.2

38.8

1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006

State budget

Social health insurance

User fees

Others

Source: Data section in official website of Ministry of Health, Vietnam, at http://www.moh.gov.vn/homebyt/vn/portal/InfoList.jsp?area=58&cat=1450.

eligibles. Although subsidies remain an important income source of public providers, its contribution to the health system as a whole has been declining (WB, 2007). The balance of the budget comes from user-fee payments and health insurance reimbursement. Health insurance reimbursement accounts for only a modest proportion of the revenue, and usually comes from public health insurance, which is a mandatory contribution covering only public servants, formal sector employees and “people of merit”5 . Public health insurance organisation, the Vietnam Social Security (VSS), plays a very limited role as an informed purchaser of health services as it acts largely as passive payer of bills (WB, 2007). In 2006, the national health insurance schemes covered about 42% of the population, including the scheme for the poor, but financially only accounts for approximately 13% of total health expenditure. 5

People of merit include mothers, widows and orphans of veterans, army invalids, and the elderly aged 90 and over.

107


Since 1989, hospitals in Vietnam have been authorized to collect fees for services and drugs provided to the patients if they do not have social health insurance or exemption status. Patient payments finance a large part of inpatient services, most of the cost of drugs and outpatient care (WB, 2007). The charged fee is a mixture of per-diem and per-service specified in a fee schedule which was introduced in 1995 and revised recently in 2005. Figure 4.4: Flows of health financing in Vietnam

Financial resources

Fund pooling

Foreign aid

State budget for health

Allocation and management Ministry of Health Provincial health department

Purchase of services Public health providers

Ministries Businesses Employers

Individuals Households

Private health providers

Social health insurance fund Central/ Provincal social insurance

Private health insurance

Pharmacies

Source: Adapted from Ladinsky et al. (2000), page 85 and Nguyen et al. (2008), page 25.

4.1.2

Administrative structure

Vietnam’s government hospitals and other facilities are classified under different functional and administrative levels. The Ministry of Health (MOH) is the highest authority in the health sector, followed by the Provincial, District and Communal People’s Committees. One important aspect of this hierarchy is the fact that MOH 108


sets prices for health care services provided in all public facilities through the fee schedule. It is also involved directly in the training of the health workforce, and the production of pharmaceutical products. The smallest and most basic units in the system are health centres at the commune level, each of which serves around 7,000 to 12,000 people. They usually have around 3 to 5 health workers, usually a medical doctor, an assistant doctor, a midwife and a nurse). They are responsible for primary care, including preventive care, normal delivery, provision of drugs, family planning, and overall health promotion in the community. The next level of administration in the hierarchy is the district level which consists of inner commune clinics, district hospitals or polyclinics. They provide basic treatment for common diseases, emergency care and limited inpatient services. On average, the almost 600 district level hospitals have fewer than 80 beds, and vary widely in the technical sophistication and quality of services. Another level up, the 64 provinces of Vietnam have 320 general and specialised provincial hospitals. They treat diseases that are beyond the capability of the district hospitals. Province hospitals (including large city hospitals) are administered by the Provincial Health Department. Ranging from 300-500 beds, they are significantly larger than district hospitals and provide curative and outpatient services for both local and regional populations. At the central level, there are around 20 specialised and general hospitals under MOH management providing highly specialised treatments with advanced techniques. These are the largest and most technically up-to-date facilities with an average of over 500 beds. Almost all are located in the largest cities of Vietnam (Hanoi, Ho Chi Minh City, Da Nang and Hue).

4.1.3

Reforms

Vietnam’s Doi Moi economic reform, which began in the late 1980s, has had considerable impact on the health sector, mainly through the legalisation of private medical practices, the commercialisation of pharmaceutical industry and the introduction of user-pay fees in 1995. Since then, there have been several major reforms, notably the introduction of pro-poor programs which waive fees completely or partly for the poor and some target groups (during the 1990s and revised in 2003 as Decision 139), increased hospital financial autonomisation (in 2002 and 2006), 109


the revision of the schedule fees and expansion of social health insurance coverage (in 2005). Details of these reforms are presented in Table 4.1. One of the most recent and influential initiatives has been the financial autonomisation which takes the form of Decrees 10/2002 and 43/2006. This policy requires that public service providers attain financial self-sufficiency. Health service providers are encouraged to earn more income from patients and to use these extra revenues to pay higher salaries to staff and invest in modern medical equipment. While Decree 10/2002 empowered public service providers with financial autonomy through allowance of more flexible salary schemes (i.e. additional bonus and salary allowed), development of welfare and reserve funds, and investments in other organisations, it did not grant autonomy in staffing (i.e. staff recruitment, promotion and firing) and in organisational restructuring (i.e. establish, dissolve and reorganise wards and departments). Decree 46/2006 was introduced to address this “shortcoming”. According to the World Bank, these autonomisation reforms give public service providers more autonomy and rely on market or “market-like” incentives to improve performance (Nguyen et al., 2008). Table 4.1: Major health system reforms in Vietnam since the 1990s

Years

1993

1994

1995

Reforms National health insurance program Payment for commune health centre workers

Schedule fee

Major initiatives National health insurance program covers initially civil servants, formal sector workers and “people of merit” on a mandatory basis. Central government resumed the responsibility for paying the salaries of commune health centres workers who were previously paid by their communes. Schedule for user fees for consultation and physical examinations, inpatient days, technical services and lab test was introduced. The aim was to ensure a broad common set of user fees rather than allowing facilities to charge whatever the local market would bear.

110


Years

Reform initiatives

1990s

Fee waiver for target groups

1996 (and 2002)

State Budget Law

2003

Decision 139/2003

2002

Hospital autonomy (Decree 10/2002)

2005

Social health insurance

Main points Some target groups received fee waivers. This policy aimed at reducing out of pocket spending on health care for the poor. However, it has usually been only weakly targeted the poor and is relatively ineffective at reducing out of pocket spending, not least because most of the out of pocket spending was on drugs not fees. Government responsibilities in health were increasingly decentralised, with local governments being given increasing rights in the planning and execution of spending decisions. The new rules gave local governments more freedom in setting priorities in health and between health and other sectors, and in the way provinces allocated resources across districts and communes. Extend insurance coverage to the poor and other disadvantaged groups through the use of tax financed enrollment, with central government picking up the bulk of the cost but provincial governments responsible for some co-financing, identification of beneficiaries and implementation. Increased autonomisation of public health service providers was gradually introduced from 2002 onwards. The Decree aims at aligning the incentives of providers (hospital staff) with the overall performance of the health facility. Hospitals are given greater control over their spending and to a lesser degree over pay and employment, user charges for non-basic services, and investment. Benefit package became more generous with co-payment scrapped; all formal sector workers were required to enroll; the insurer could contract with private providers.

111


Years

Reform initiatives

Main points

Following Decree 10/2002, Decree 43/2006 allows further autonomy. It stipulated that the first 25% of net revenues are set aside for facility upgrading and hospital staff gets an outright share (in terms of Hospital additional income) of the remaining 75%. There are autonomy no caps on the additional income. It also grants power 2006 (Decree to the hospital director to raise net revenue (through 43/2006) cutting cost and raising revenue). The hospital director now has full control over manpower, i.e. hiring, firing, promotion, assignments and subordinate units (create, merge, dismantle, etc). Source: Adapted from Lieberman & Wagstaff (2009) and Adams (2005)

4.1.4

Challenges

Reforms have transformed Vietnamese health system and brought about great benefits. However, there remain questions about efficiency improvement, equity of access and utilisation, financial protection for the poor and near-poor, quality of care and the sustainability of the system. For the purpose of the study, only issues relating to equity and efficiency are discussed here. Although stated as an important goal of the health system, equity is becoming more and more difficult to sustain under the existing financing and provider payment arrangement. Despite various policy efforts to subsidies the poor, there is growing evidence that poor people cannot afford health care services any more (MOH, 2002) since the introduction of the hospital user fees in early 1990s and the liberalisation of economic policies that has encouraged private medical practice and the free trade of medicines and drugs. The 1998 Vietnam Living Standard Survey (VLSS) reveals that the cost of a single service contact with a public hospital takes up about 22% of annual nonfood consumption expenditures for a person in the bottom two quintiles. This cost is greater if the hospital contact involves admission. For a low and middle-

112


income person, a single admission to a public hospital accounts for about 60% of annual non-food consumption (Sepehri et al., 2006). According to the 2008 Vietnam Development Report (WB, 2007) and Dao et al. (2008), of 34 million poor people without enough money to pay for health services, around 18 million are classified as not “poor enough” to be exempted. The other poor are eligible for social insurance and concessions but the actual out-of-pocket payment per hospital contact, including unofficial payments, remain a large proportion of their income. In 1998, average out of pocket expenditure per hospital contact by the insured was about 44% of that of the uninsured (Sepehri et al., 2006), which is close to 30% of annual non-food consumption for a low and middle-income person. Inequality in utilisation, both quantity and quality of care, is the direct consequence of various reform initiatives, including user-fee policy and financial autonomisation. Apart from services provided free under several national programmes, virtually all health care services and goods provided at public health care facilities attract user fees (MOH, 2002). The consequence of it is that a high proportion, around 66% of people with health problems (the majority of them are poor or near poor) delay seeking care from a suitable health professional or self medicate or buy drugs over the counter without a doctor’s prescription (MOH, 2002). When the delay is not possible, hospitalisation usually pushes them into the “poverty trap”. The paper by Sepehri et al. (2006) on social health insurance found a wide variation in the use of hospital outpatient services across income quintiles, ranging from as low as 0.47 contact per capita for those in the lowest income quintile and as high as 1.1 contacts for those in the highest income quintile. The study by Dao et al. (2008) found that user fees have a negative impact on service utilisation by increasing inequality and catastrophic health expenditure. The poor had a service utilisation rate of 30% compared to 40% the rest of the population. The average number of health service contacts for the poorest is 8.3 per year, compared to 10.4 for the richest. In principle, the poor are exempted from fees or entitled to have a “health insurance card for the poor” under Decision 139 (2002). In practice, insurance coverage benefits are quite limited, leaving poor and near-poor patients with a relatively large financial burden. For catastrophic illness that requires hospital admission, the indirect and informal out-of-pocket expenses (medical and non-medical) are unaffordable for the majority (Sepehri et al., 2006). Therefore, the poor are more likely to use polyclinics. Moreover, the public budget is often insufficient to give insurance and exemption because there are so many poor people that need health care. Hospitals that treat a large proportion of exempted (and poor) patients are

113


always in budget distress. This leads to the practice of discriminating between exempted patients and fee-paying patients: the latter receive lower quality services (WB, 2007). This practice includes, but is not limited to, long waiting times, low quality drugs, shared beds and pre-mature discharges. A response to this practice has been increases in voluntary and involuntary unofficial payments. It is reported that unofficial payments can be higher than the user fees and account for two-thirds of total out-of-pocket payments (Dao et al., 2008). The second issue relates to allocative efficiency: the bias toward inpatient services and a more expensive care style in particular. The main driver of this bias is the existing provider payment system which follows a mix of regulated per-diem and fee for service for hospital charges and bed-norm for state subsidies. The practice of bed-based budgeting has been a major reason for inpatient care bias. State subsidies are based on the hospital’s official number of beds, which gives hospitals a strong incentive to use their beds, or at least report them as being used, in order to protect their budgets. This practice also encourages hospitals to jealously guard their bed stock and put patients in beds even when they could be treated at a lower level or on an outpatient basis (WB, 2007). This is reflected through the imbalance between the number of inpatient admissions and outpatient visits. The inpatient admission rate in Vietnam is not much different from that of other countries despite relatively low bed stock, while its outpatient visit rate appears to be much lower. Furthermore, it is also reflected in the higher occupancy rates reported in the MOH data: the average exceeding 100% since 2000, especially for higher level hospitals. It is likely that many inpatient admissions could have been avoided, at least in part, through high quality outpatient care. The fee-for-service model in general provides an incentive to increase volume, even if this means delivering unnecessary services. It also gives providers an incentive to focus on interventions that are more profitable. The regulated fee schedule in Vietnam was introduced in 1995 and the interventions covered are relatively lowtech. Since then, it has been updated with new and relatively high-tech services at their current prices. Meanwhile, the existing (old) service prices have not been updated and lag behind the actual cost of production. This price bias give incentives for hospitals to drift toward case types requiring tests and other interventions that are priced favourably, i.e. more high-tech and high-cost style of care, even when the clinical effectiveness of those interventions is questionable. This incentive becomes more intense under the recent financial autonomisation reform6 . 6

WB (2007) reports that the implementation of Decree 10/2002 together with fee for service arrange-

114


Financial autonomisation reform was introduced under Decree 10/2002 and further enhanced in Decree 43/2006. The two Decrees confer upon hospitals increased discretion over financial operations, management of human resources, organisation of services, and choices of services offered. They are allowed to earn net income, distribute it amongst their staff and invest it. They can (and most have) establish wards for fee-paying-only patients, providing better quality of care than in regular wards. In essence, hospitals can operate in a corporate fashion, playing a much greater, more direct, and formally sanctioned role in shaping the costs, qualities and distributions of services at the point of delivery. This greater decision-making power is believed to be an important advance of the two Decrees and ought to provide more flexibility to the hospital sector in order to improve productivity and scope of services. However, there is a risk that this flexibility can be exploited in the pursuit of profit. The scope to set prices within the scheduled fee bands leads to higher prices in areas where the public can afford the higher prices or where there is little competition, threatening the poor and middle income groups with unaffordable health care. This incentive scheme also has inevitably made larger facilities, with greater net revenue, and thus more generous bonus packages, more attractive to medical workers. The effect has been to encourage doctors and other medical staff to try to work in such facilities, creating imbalances in the system between rural areas and urban areas. Furthermore, as a result of their financial autonomy, those hospitals can explore ways to attract patients, including patients with only mild diseases, resulting in more patients bypassing the commune and district health facilities, especially when quality medical workers all move up to larger facilities. This creates the imbalance of service utilisation at different levels: overcrowding faced in provincial and central hospitals while district and commune providers act mostly as expensive pharmacies or preventive care centres. Lastly, the scope for providers to induce demand for unnecessary care is high because they are paid on a fee-for-service basis and are given a strong incentive to generate revenues. This drives up the cost of health care while resulting in little improvement of health outcomes. ment has led to 47% increase in out of pocket payment by the third year of implementation (2005). It is difficult to know if the increase is justified in medical terms but there is a risk of supply-induced demand in autonomous hospitals.

115

4.2. RESEARCH QUESTIONS

4.2

Research questions

The Joint Annual Health Review produced by the MOH and the Health Partnership Group in 2008 has claimed that Decrees 10/2002 and 43/2006 were to bring benefits to the health sector. They have created stronger economic incentives for medical staff to increase productivity, expand the scope of management and promote efficiency in the provision of health services (Nguyen et al., 2008). The report has provided some initial evaluations of Decree 10/2002 (page 86-90 of the Report). None of them however directly relate to or suggest evidence of efficiency and productivity improvement - one of the ultimate goals of this reform. The purpose of this study is two-fold. First, it seeks to evaluate whether or not financial autonomisation reform (mostly the implementation of Decree 10/2002)7 has induced productivity growth and efficiency improvement. Second, the study attempts to understand the distribution of reform gains (if any) across different regions of the country in order to examine if hospitals at all levels and regions benefit equally from the reform. The findings are expected to have important policy implications. The financial autonomisation reform, by design, has explicitly sacrificed the equity objective. If it is successful with respect to the efficiency objective, the study will give an indication of the trade-off magnitude. If no evidence of efficiency improvement is found, this generally implies a failure of this policy with respect to both equity and efficiency targets. In this case, there should be further reforms that are more carefully crafted to fix the problem. Furthermore, understanding who gains/loses from these reforms will provide opportunities for further investigation and better targeted policies for hospitals of different levels and/or regions. The reform impact evaluation is performed using the meta-frontier analysis framework proposed by Battese, et al. (2004) and O’Donnell, et al. (2008).

4.3

The meta-frontier framework

The meta-frontier framework for measuring efficiency and the technology gap is an extension of the usual frontier framework. It is a useful tool of analysis as it 7

Decree 43/2006 was not introduced and implemented until late 2006 and thus the effect, if any will not be felt in 2006 but in and after 2007. The impact, if any, would mainly come from Decree 10/2002. This is worth noting because Decree 10/2002 does not give complete autonomy over staffing decisions (hiring and firing) and organisational structure.

116

4.3. THE META-FRONTIER FRAMEWORK

recognises that not all production units are “free” to adopt/ use the existing production technology. Socio-economic conditions, regulations, or geographic location can serve as barriers that prevent the units choosing the optimum mix of inputs and/or to produce the best output combination, and/or to use the most productive technology available. For instance, if a government regulates the maximum number of beds that hospitals can have or maintains a fixed budget for labour (especially under the provider payment method that is based on bed norms8 ), then hospitals cannot choose to expand their production capacity. Alternatively, due to the limited labour market in the area of operation, hospitals cannot employ the labour required to produce some particular outputs like complex surgeries. In measuring and interpreting efficiency levels, ignoring these factors may be unfair to producers that face restrictive/ less favourable operating conditions. In the efficiency literature, a common way to deal with this is to incorporate such factors of restriction into the model to capture the variations of efficiency estimates. This has been done in both one-stage (usually stochastic frontier applications) and two-stage procedure (both stochastic frontier and data envelopment analysis applications). The one-stage SFA approach includes environmental variables in the efficiency component of the composite error term (see for example Fuiji & Ohta, 1999; Rosko, 1999; Rosko & Chillingerian, 1999; Deily et al., 2000; Frech & Mobley, 2000; Fuiji, 2001; Li & Rosenman, 2001; Rosko, 2001a; Street & Jacobs, 2002; Carey, 2003; Rosko & Proenca, 2005; Smet, 2007). In the two-stage approach, efficiency scores are estimated in the first stage (either by SFA or DEA), and then regressed on environmental variables (some examples include Lynch & Ozcan, 1994; Ferrier & Valdmanis, 1996; Chang, 1998; Dalmau-Matarrodona & Puig-Junoy, 1998; Linna & Hakkinen, 1999; Rosko & Chilingerian, 1999; Folland & Hofler, 2001; Zere et al., 2001; McKay et al., 2002; Biorn et al., 2003; Sari, 2003; Bilodeau et al., 2004; Grosskopf et al., 2004; Chen et al., 2005; Bates et al., 2006; Barbetta et al., 2007; Uslu & Pham, 2008) The single and two-stage approaches are both feasible when some measures of the environmental factors are available. In practice, many of these may be missing due to data deficiency or simply difficult to measure. There is another problem with the two-stage procedure, as pointed out by Simar & Wilson (2008): the implicit (but not usually stated) assumption that the environmental variables do not influence the shape or boundary of the technology set. That means there is a separability between the input-output production space and the environmental variable space. This is a 8

Further discussion about this method can be found in subsection D.1.1.2 in AppendixD

117


restrictive assumption, especially when one has the knowledge that environmental factors do play a role in defining the technology frontier. The meta-frontier framework, in contrast, neither requires the specification of environmental measures nor restricts interaction between the input-output production space and the environmental variable space. It captures the effects of such factors in the measurement of “technology gap” and benchmarks producer performance in the context of their operating environments, i.e. the environment-specific (or group) frontier9 . For these very reasons, the meta-frontier is chosen to address the research question of this study: the impact of reform on technical efficiency. The next section describes the analytical framework laid out by Battese, et al. (2004) and O’Donnell et al. (2008) that explains the concept of the meta-fronter, group-frontiers, and meta-technology ratios in the context of hospital industry. The meta-frontier is the frontier that envelopes all the group frontiers. The difference between each group frontier and the meta-frontier is defined as the meta-technology gap. The meta-technology ratio then is derived from technical efficiency estimates of individual hospitals with respect to the meta-frontier and the respective group frontiers.

4.3.1

The meta-frontier

Hospitals are production units that use multiple inputs such as labour, equipment, drugs and medical goods to produce multiple outputs such as diagnostic tests and procedures, emergency services, outpatient visits and inpatient care. The meta-technology for the hospital industry is therefore defined over a multiple-input multiple-output space. The meta-technology set contains all technologically feasible combinations of inputs (X) and outputs (Y ) under the available technology (T ). The technology T represents the existing stage of knowledge available in the industry that makes the transformation of X into Y possible. Mathematically, the technology, the output 9

The inevitable trade-off is that it is not possible to fully separate the marginal effect of individual environmental factor if more than one present. A strategy to deal with this could be grouping of producers by different environmental factors and repeating the meta-frontier estimation for each type of grouping. This requires a large sample with relatively even distribution across groups.

118


set (P (X)) and the input set (L(Y )) are defined as: T = {(X, Y ) : X ≥ 0; Y ≥ 0; X can produce Y }

(4.1)

P (X) = {Y : (X, Y ) ∈ T }

(4.2)

L (Y ) = {X : (X, Y ) ∈ T } .

(4.3)

The boundary of the meta-technology set is referred to as the meta-frontier. These definitions are identical to the usual concepts of the technology set and frontier (as discussed in Fare & Primont, 1995; Coelli et al., 2005). For the purpose of measuring efficiency, technology can be conveniently and equivalently expressed through the input or output distance functions, depending on the relevant objective function of the production units. The incentive structure of the Vietnam’s health system that encourages public hospitals to be self-sufficient financially through increasing revenue (as analysed in Section 4.1) indicates an output maximising behaviour of hospitals. Thus, the output distance function is discussed and used here. The output distance defines the maximum amount by which a hospital can radically expand its output vector, given the input vector. It is given by: Y ∈ P (X) . (4.4) DO (X, Y ) = inf θ > 0 : θ θ The output distance function possesses standard properties (non-increasing, quasiconcave in outputs, nondecreasing, concave and linear homogeneous in inputs) if the output set P (X) satisfied the standard regularity conditions (nonempty, bounded, closed and weak disposability of outputs). A hospital i using inputs Xi to produce outputs Yi is considered technically efficient if and only if DO (Xi , Yi ) = 1, i.e. it is impossible to increase outputs without using more inputs.

4.3.2

Group frontiers

The group-specific technologies are subsets of the meta-technology. The existence of group-specific technologies is due to constraints in the operating environment of the hospitals. Resources, regulations, socio-economic and geographic conditions may

119


prevent hospitals from having access to all feasible parts of the technology set. In other words, not all input-output combinations defined by the meta-technology set are accessible to some hospitals. Different environment characteristics allow the division of hospitals into groups. The group technology set is then defined as all input-output combinations available for the hospitals in the group. Formally, the k th group technology set is written as: T k = {(X, Y ) : X ≥ 0; Y ≥ 0; X can be used by hospitals in group k to produce Y } . (4.5) Since T k is a sub-set of the “unrestricted” meta-technology set T , all feasible input-output combinations in T k belongs to T and all the group technology sets together fill up the meta-technology set. This can be illustrated in Figure 4.5 where there are three group frontiers enveloped by the meta frontier FM . Hospital B using input XB to produce YB belongs to Group 2 and is on the frontier F2 . A hospital in Group 1 is not able to choose B’s input-output combination due to the constraints that Group 1 faced (for instance, cannot reduce the number of medical staff because they are state employees). Likewise, A is not feasible for any hospital in either Group 2 or 3 even when input expansion is possible. This may be due to a more productive technology that Group 1 can get access to, but is not available for the other two groups. It is noted, however, that there are infeasible input-outcome combinations for all the groups frontiers: those in the areas between the frontier FM and the dotted line. This implies that the empirical meta-frontier need not be convex if there exists no hospitals, both theoretically and practically, free from constraints to reach this area. To complete the discussion, output set and output distance function of hospitals in group k are defined as: P k (X) = Y : (X, Y ) ∈ T k Y k k DO (X, Y ) = inf θ > 0 : ∈ P (X) . θ θ

(4.6) (4.7)

k The output distance function DO (X, Y ) can take value no less than the metaoutput distance function DO (X, Y ). This is because the meta-frontier envelops all group frontiers and thus the maximum amount by which output of a hospital in group k can radially expand cannot exceed the maximum amount of feasible radial expansion under the meta-technology. The group output distance functions and

120


Figure 4.5: Illustration of the meta-fronter and groups frontiers

Y

Meta-frontier FM

YMA

Group frontier F1

YA

A

YMB

Group frontier F2

YB B

Group frontier F3

XA

XB

X

their corresponding output sets also follow regular properties (similar to those of their meta counterparts).

4.3.3

Technical efficiencies

Following the output distance function’s definition, a hospital is considered technically efficient if and only if DO (X, Y ) = 1. This means it is not technically feasible to expand outputs without using more inputs. If the distance function of a hospital takes the value 0.85, it implies that the output vector Y is 85% of the maximum output that could be produced using the input vector X. Or equivalently, this hospital is 85% efficient. Formally, the technical efficiencies of a hospital i measured against the metafrontier (i.e. the existing stage of knowledge in the industry) and the group frontier (representing the existing stage of knowledge and the physical, socio-economic con-

121


ditions that characterise group k) are defined as: T EiM (X, Y ) = DO,i (X, Y ) k (X, Y ) . T Eik (X, Y ) = DO,i

(4.8) (4.9)

Referring to Figure 4.5, hospital B using input XB to produce YB belongs to Group 2 and is on the frontier F2 . It is technically efficient given its group frontier F2 , that is T EB2 (X, Y ) = 1. However, if being measured against the meta-frontier, hospital B is technical inefficient because a higher output level (YBM ) is feasible with input XB . The short-fall in outputs is not caused by hospital B not doing the right thing, but because it is prevented from reaching the full potential - the meta technology. Hospital A, on the other hand, is not technical efficient with respect to its group frontier F1 . Therefore, its technical inefficiency with respect to the meta-frontier T EAM (X, Y ) contains two inefficiency components: (i) not doing the best it can given the group technology and (ii) the short-fall of frontier F1 from the metafrontier FM . It is useful to distinguish these two sources as producers should not be punished for operating in a less favourable environment if they are doing the best they can. Furthermore, measuring the gap between group and the meta frontiers gives an indication of the “cost” of environmental constraints.

4.3.4

Meta-technology ratios

The technology gap between the group frontiers and the meta frontier can be measured by the meta-technology ratio, which is formally defined as:

M T Rk (X, Y ) =

DO,i (X, Y ) T E M (X, Y ) = . k T E k (X, Y ) DO,i (X, Y )

(4.10)

Assume that hospital A uses input vector XA to produce output vector YA and achieves 90% efficiency with respect to its group k frontier; its technical efficiency score measured against the meta-frontier is 81%. The meta-technology ratio is calculated as 0.81 , or 90%. This means that the maximum output that could be 0.90 produced by the hospital in group k using the input vector XA is 90% of the output that is feasible using the meta-technology.

122


Rearranging Equation 4.13 to obtain T E M (X, Y ) = T E k (X, Y ) × M T Rk (X, Y ) .

(4.11)

This implies that technical efficiency with respect to the meta-frontier can be decomposed into two components: technical efficiency measured with reference to group k frontier and the meta-technology ratio for group k. The decomposition in Equation 4.11 is important because it allows policy makers to understand the source of inefficiencies, either internal to hospitals (i.e. technical efficiency with respect to its group frontier) or external such as regulations, labour market, demand or burden of disease. Such knowledge is useful in devising policies to improve hospital performance: to design better incentive schemes for internal improvement and/or to improve the operating environment for hospitals.

4.3.5

Evaluating the impact of reform using the meta-frontier framework

The application of the meta-frontier framework can address two questions: whether reform creates a more favourable operating environment for public hospitals, leading to better performance, i.e. an expansion of the group production possibility set and more efficient behaviour; and whether hospitals of all levels and regions benefit equally(if any) from the reform. The first question concerns the gain from reform and the second the distribution of the reform benefit across hospitals.

4.3.5.1

Efficiency improvement - the gain from reform

Consider the meta-frontier constructed with data on hospitals in both pre- and post reform periods. It represents the unconstrained production possibility set available under both conditions. Equation 4.11 suggests that the post-reform technology is more progressive than the pre-reform technology. It also implies that a higher post-reform T E M can be the result of improvements in either M T Rk or T E k or both, given the growth rate of one dominates the falling rate of the other. Improvement of both T E k and M T Rk means hospitals become better at what they are doing with respect to their own group frontier and the gap between the group’s technology and what is feasible is narrowed.

123


Although this is obviously the ultimate objective of reform, achieving one of the two can also be considered an improvement (i.e. positive gain) under some circumstances. Reform can lead to the expansion of the group technology feasibility set (i.e. M T Rk growth) and thus improvement of technical efficiencies. This does not necessarily implies that all hospitals experience rising T E k because some might be slower in the catch up process. Another situation arises where reform induces efficiency behaviours rather than positively alters the operating environment. It is expected to see an increase in T E k but little or no change in M T Rk . However, a positive change in T E M due to T E k rising faster than M T Rk falling might not be considered an “improvement”. A lower post-reform meta-technology ratio (on average) implies the reform creates more constraints on hospitals, which is represented by the contraction of the production possibility set. Therefore, even when reform induces efficiency behaviours (expressed by higher T E k ), the potential for future improvement is limited.

4.3.5.2

Efficiency changes by region - the distribution of the gain

Now, consider the distribution question of reform benefits. Figure 4.6 illustrates an extreme situation to highlight the two potential sources of change. After the reform, all hospitals stay the same except for N , moving from N pre to N post . This indicates the expansion of the technology possibility after the reform for Group 2 and the hospital industry as a whole. Technical efficiencies of hospitals in Group 2 in relation to their group frontier, on average, decrease as the group frontier expands. The Group 1’s post-reform frontier stays the same and thus, technical efficiencies of all hospitals in Group 1 (A, B and C) in relative to their group frontier are constant. As the meta-frontier is pushed upward after reform, the distance from Group 1 frontier to the post-reform frontier is further from that to the pre-reform meta-frontier. This implies that on average, technical efficiencies of hospitals in Group 1 in relation to the meta-technology (of the hospital industry) decrease and their M T Rs will be lower. This analysis suggests that hospitals in Group 1 do not enjoy the benefits of reform as their production possibility has not expanded to reach the post-reform frontier. Hospitals in Group 2, except for N , have not yet achieved efficiency gain from the reform but they would be able to, because some environmental constraints have been removed which results in the upward shift of the post-reform frontier.

124


Figure 4.6: Who benefits from the reform?

Y Npost

Post-reform Frontier

Post-reform frontier of Group-2

Npre

Pre-reform Frontier

P C A

B

Group-1 frontier (no change)

Q Pre-reform frontier of Group-2

X

This can be generalised in the less extreme cases, as illustrated in Figure 4.7, where the pre- and post-reform group frontiers shift and hospitals change their output-input choices between pre- and post-reform periods. Reform creates more difficulties for hospitals in Group 1, which can be seen as most hospitals (A, B and C) have moved inward from the meta-frontier and the post-reform group frontier (the dotted curve) is sitting below the pre-reform frontier. In contrast, most hospitals in Group 2 (such as N , P , and Q) have improved as they move toward the metafrontier under the expansion of their post-reform technology possibility set. The reform clearly brings about improvement to these hospitals. This is reflected by higher meta-technology ratios. Efficiency measures of these hospitals, in relation to the meta-technology, increases as a consequence. These two illustrations suggest that it is essential to decompose technical efficiency in order to understand the source of change. If a hospital experiences an increase in TE in relation to the meta-frontier (better T E M ), it can be the result of its getting better at what it has been doing (increased T E k ), the group frontier shrinks due to extra environmental constraints and distortions (lower meta-frontier 125

4.4. VIETNAMESE PUBLIC HOSPITAL DATA

Figure 4.7: Comparison of before and after reform frontiers

Y

Npost Meta-frontier (pre and post-reform)

P

post

Post-reform frontier of Group-2

Qpost C

pre

Bpre

C

A

Q

pre

Npre

Spost

Bpost Post-reform frontier of Group-1

Ppre

post

Spre

pre

Pre-reform frontier of Group-2

Apost Pre-reform frontier of Group-1

X

ratio M T Rk ) or the combination of both. Policies then can be examined to improve the operating environment for hospitals or to provide better incentive schemes for internal improvement.

4.4

Vietnamese public hospital data

The data was extracted from the Hospital Inventory database managed by the Vietnam’s Ministry of Health. Hospital inventory report is a compulsory annual survey conducted at the end of each financial year (between October and December). It is designed to serve as a management tool for MOH: monitoring the utilisation of assets and labour, and the provision of services in terms of both quantity and quality; and evaluating and ranking performance of individual hospitals. The survey contains two major parts: part A collects information about hospital activities (service provision, expenditure, revenue and medical labour force), and part B is reserved for evaluation of infrastructure, labour and professional activities. The reporting structure has not changed significantly in the last ten years of implementation, which makes it easier to compare data collected through these surveys over 126


time. Apart from records on activities, revenue and expenditure, the hospital inventory report also identifies various characteristics of the individual hospital such as hospital class, type of government unit and location. Hospital class is a form of accreditation, determined by the Ministry of Health based on the scale and scope of services and quality of care. There are four classes of hospitals, class 1 (large and complex general hospitals) to 4 (small sized hospitals, usually at the district level). As discussed in Section 4.1, all public health providers in Vietnam fall into different administrative levels: commune, province, district and central. These variables, especially location, prove to be useful for the comparison purposed as they allow categorisation of hospitals by class, government unit and location. The two separate datasets were extracted for the purpose of the study: prereform (from 1998 to 2000) and post-reform (2005-2007) periods10 . The first contains more than 400 hospitals and polyclinics of various sizes while the second does not cover polyclinics. This is, however, not an issue because the research questions in this study are not applicable to polyclinics. They have not been targets of the autonomisation reform, at least in the first stage under Decree 10/2002 while general central, provincial and some large district hospitals are. Furthermore, these hospitals account for a large share of health service provision and generally determine the performance of the public hospital sector as a whole. Lastly, hospitals and polyclinics are very different providers. Hospitals, providers of acute and tertiary care, treat complex cases and offer surgical and high-tech tests and procedures for inpatients and outpatients. In contrast, polyclinics act mainly as providers of outpatient care, some preventive services (such as immunisation) and drugs. The sample with only general and tertiary hospitals is more homogeneous, which is required to ensure meaningful benchmarking in DEA. The final dataset consist of a six-year balanced panel of 62 hospitals, representing all eight administrative regions in Vietnam, covering two periods before (1998-2000) and after/during the process of financial autonomisation (2005-2007). 10

It would be ideal to have complete data for the studied hospitals in the period 2001-2004, especially for the year 2001, the last year before the reform was introduced under Decree 10/2002. However, due to data deficiency, it is not possible to have a balanced panel for all those years.

127


4.4.1

Input measures

Medical resources in hospitals are usually divided into (i) labour (e.g. doctors, nurses, allied and other health professionals, technicians, administrative and domestic staff), (ii) capital (e.g. building, medical equipment and machines, beds etc), and (iii) medical goods and supplies (such as drugs, bloods, chemicals and other medical consumables). The hospital inventory survey collects the hospital input information in the form of expenditure, number of available beds and number of medical staff. All expenditure variables are measured in current price Vietnamese Dong (VND)11 , which are converted to constant price (2004) using annual health deflators. The deflators are constructed using the total health expenditures measured in current price and constant price (2004), extracted from the National Accounts data of the General Office of Statistics, Vietnam12 . The year 2004 is used because the output cost weights are derived from 2004 hospital cost data (see details in Section 4.4.2 and Appendix B). The deflators, as well as the output cost weights are applied to all hospitals across regions. Table 4.2: Health deflators Year

Bil. VND (current price)

Bil. VND (1994 price)

Deflator (2004 price)

1995

3,642

3,009

0.5838

1996

4,007

3,220

0.6002

1997

4,381

3,348

0.6312

1998

4,979

3,566

0.6735

1999

5,401

3,707

0.7028

2000

5,999

3,946

0.7333

2001

6,417

4,151

0.7457

2002

7,057

4,464

0.7625

2003

8,865

4,853

0.8811

2004

10,851

5,234

1.0000

2005

12,412

5,640

1.0615

2006

14,093

6,082

1.1177

2007

16,160

6,572

1.1861

2008

18,592

7,117

1.2601

Source: General Office of Statistics, Vietnam

In the final dataset, four variables are retained: labour cost, number of beds, recurrent and medical goods expenditures because data on the number of staff is incomplete and so are some expenditure categories. The summary of input variables 11 12

Expenditure are in million Dong. The data was accessed on 14/10/2009 at http://www.gso.gov.vn/default_en.aspx?tabid=491.

128


are presented in Table 4.3. Table 4.3: Summary of input statistics Variable

Mean

Std. Dev.

Min

Max

Beds

340.00

250.11

17.00

1,083.00

12,771.90

25,165.85

149.20

208,111.70

5,536.92

6,740.60

86.65

43,004.11

22,374.80

40,797.85

40.26

295,904.70

Labour expenditure Recurrent expenditure Medical good expenditure

4.4.2

Output measures

There is fairly detailed data on the output side. There are eleven main output categories, some of which are broken down into sub-categories by user type (such as children, the poor and the insured). Inpatient services are captured by the number of inpatient days and number of inpatient cases. Outpatient data consists of the number of consultations and the number of scheduled outpatient visits. Surgical operations, birth deliveries, three diagnostic tests (blood, bio-chemical and microorganisms) and four diagnostic procedures (x-ray, ultra sound, endoscopy and CT MRI) make up the eleven output variables. Table 4.4 presents the statistics of these output variables. Table 4.4: Summary of output statistics Variable

Mean

Std. Dev.

Min

Max

Inpatient days

135,514.60

105,643.10

3,790.00

534,466.00

Outpatient visits

180,990.80

192,634.00

6,673.00

1,520,403.00

Surgical operations

4,546.96

6,145.31

0.00

37,946.00

Deliveries

2,079.79

2,701.39

0.00

29,691.00

Biochemical tests

158,101.10

237,536.30

0.00

1,430,563.00

Blood tests

228,401.60

359,019.80

742.00

2,701,106.00

Micro-organism tests

37,963.37

73,712.36

0.00

462,247.00

X-ray

31,163.27

40,227.30

0.00

313,300.00

1,114.42

2,592.46

0.00

21,876.00

Ultra sound

15,144.50

18,206.37

0.00

122,834.00

Endoscopy

1,629.94

3,262.49

0.00

29,572.00

CT-scan

All data on output categories are expressed in raw counts of cases/ tests/ procedures. There is no casemix information since such a system has not yet been implemented in Vietnamese public hospitals. This is a weakness of the dataset and a potential source of bias. For instance, hospitals that treat more complicated 129


casemixes may be identified as less inefficient compared to those that treat simple cases. Complicated casemixes often require more resources, including medical labour time, drugs and modern equipment, which makes the input count per complex case much higher than that per simple case. This is especially true when the sample is a mix of large and smaller-sized hospitals. A large number of outputs and a relatively small sample size (372 observations in total) makes it essential to aggregate of some outputs to preserve degrees of freedom. The organisation and provider payment mechanism in Vietnam suggests three output categories: inpatient days, outpatient visits, diagnostic and treatment procedures. The distinction between inpatient and outpatient services is obvious: they consume different amounts and types of health resources. The separation of diagnostic/treatment procedures from those two stems from the activity classification and provider payment system (per-diem and fee for services). Payment for inpatients (per-diem) only covers the hotelling service provided during a hospital stay (including some nursing and medical supplies). Procedures or tests performed on the patients (such as X-rays or surgical operations) which are usually the main part of a treated case, are paid under the fee-for-service settings. That is each service is paid separately. The same arrangement applies for outpatients: patients receive examination and consultation from medical staff, for which hospitals get a per-case payment; any extra procedures and tests required are recorded and charged separately from the outpatient visit/consultation. The aggregation of the nine diagnostic/treatment procedures and tests is challenging due to data deficiency. Unweighted aggregation is not desirable because a surgical procedure requires more medical resources, especially labour, than an X-ray or a blood test. The hospital inventory does not collect information on shares of expenditure or revenue for these output categories. To address this problem, information from another dataset is utilised13 . This dataset contains information of hospital outputs and inputs by departments. The number of department varies from one hospital to another but cover all eleven output categories specified above (that is surgeries, inpatients, outpatients and nine diagnostic/treatment procedures). Inputs of respective departments include labour expenditure and detailed lists of equipment used, from which the cost weights of individual output category are derived. These cost weights are then used to aggregate the nine output categories into one output labelled “diagnostic/treatment procedure”. Details of the construction of the output weights are presented in Appendix B. 13

We thank Sarah Bales who generously gave up her time to extract this dataset and provide insights and suggestions on hospital costing methods.

130

4.5. ESTIMATION METHOD

The inclusion of tests and procedures is an improvement compared to the recent study of public hospitals in Vietnam by Uslu & Pham (2008). This study includes only the number of inpatient days, outpatient visits and operations as outputs. This practice ignores an important role that tests and procedures plays in diagnosis and treatment for both inpatient and outpatient cases. In fact, they take up a sizable part of the hospital output. The omission of these variables can bias estimation of efficiency as discussed in Section 2.1.1 of Chapter 2.

4.5

Estimation method

As noted in Section 4.3, the impact of regional differences on technical efficiencies of hospitals can be investigated using the meta-frontier framework. This process can be performed using the data envelopment analysis approach.

4.5.1

Data envelopment analysis

DEA is a linear programming approach to the construction of the production frontier. The concept was originally proposed by Farrell (1957) and further developed by Charnes et al. (1978) in a constant return to scale input-oriented DEA model. Subsequent papers introduced modifications including variable returns to scale and output-orientated DEA (Fare et al., 1983; Banker et al., 1984). Since then, there have been various advancements in DEA methodologies and applications, including measurement of allocative efficiency when price information is available (Ferrier & Lovell, 1990; Fare et al., 1985), non-oriented measure of efficiencies which reflects the potential for improvement in desired input and output directions (see for instance, Fare & Lovell, 1978; Fare et al., 1985, 1997), restrictions on input/output weights by assurance region (Thompson et al., 1986, 1990; Roll et al., 1991; PedrajaChaparro et al., 1997) and cone-ratio (Charnes et al., 1989, 1990; Brokett et al., 1997). DEA has been primarily used in the public, service and not-for-profit sectors in which managerial objectives are unclear and market price information is either missing or unreliable (Lovell, 2006). This method is especially useful in health care analysis because there is usually lack of price data associated with outputs, and sometimes inputs. Furthermore, it can handle multiple inputs and outputs without imposing a functional form. It also allows different measurement units

131


for both inputs and outputs. For instance, one output can be units of lives saved while another can be measured in dollars without requiring any priori trade-off between the two. This partly explains why the health care efficiency literature has been dominated by applications using data envelopment analysis (see reviews such as Hollingsworth, 2008b; Rosko & Mutter, 2008, for details). Nonetheless, DEA does not take into account noise (both measurement errors and stochastic factors). Recently, there have been some attempts to develop the statistical properties of DEA estimators and introduce some stochastic flavours into the framework in order to improve the reliability of its estimates (see for instance Sengupta, 1990; Ferrier & Hirschberg, 1997; Kneip et al., 1998; Simar & Wilson, 1998, 1999a,b; Sengupta, 2000; Huang & Li, 2001; Cazals et al., 2002). An output oriented DEA model will be used for the estimation. It seeks to maximum proportional increase of outputs that is feasible while holding inputs fixed. This results in a frontier enveloping all the data points and technical efficiencies are measured as the amount of radial expansion of outputs that can be achieved in order for the hospital to be on the frontier. The rationale for an output oriented model is that this reform creates incentive for public hospitals to maximine revenue in order to be financially self-sufficient. The frontier can be constructed under either constant returns to scale (CRS) or variable returns to scale (VRS) assumptions. Since CRS is only reasonable in industries where production size does not have any effect on productivity, a VRS output oriented model is used here.

4.5.2

Estimation procedure for the meta-frontier and group frontiers

To investigate the impact of reforms on the technical efficiencies of public hospitals, an (overall) meta-frontier is estimated, followed by the constructions of the pre- and post-reform frontiers. The difference between distances from each of these two frontiers to the meta-fronter indicates the degree of progress/regress induced by financial reform. It is also hypothesised that financial reforms do not bring about uniform improvements in efficiency across regions. Measuring such differences is possible with the construction of sub-regional frontiers and the comparison of technical efficiencies under these frontiers with those measured with reference to the overall meta-frontier.

132


4.5.2.1

Estimation of the meta-frontier

The meta frontier is estimated by applying a VRS output-oriented DEA over observations of all 62 hospitals over the six year period. That involves solving 372 individual DEA problems with three outputs and four inputs for individual hospital i (i = 1...62) in period t (t = 1...6). This results in a technical efficiency T EitM for each observation where i and t are hospital and year indexes. These indexes are omitted in the following discussion for notation simplicity. More specifically, it is expressed as: max φit

φit ,λit

s.t. Yλit ≥ φit yit xit ≥ Xλit j0 λit = 1 λit ≥ 1.

(4.12)

where yit and xit are the 3 × 1 vector of output quantities and the 4 × 1 vector of input quantities of hospital ith in the tth period; Y and X are the 3 × 372 matrix of output quantities and the 4 × 372 matrix of input quantities of all 62 hospitals in all 6 years; j0 is an 372 × 1 vector of 1; λit is the 372 × 1 vector of weights14 and φit is the scalar, of which the reciprocal ( φ1it ) is the output-oriented technical efficiency (T EitM ) for hospital i at time t. As φit takes the value no less than 1, the value of technical efficiency T EitM is bounded by zero and one. A value of one indicates that the hospital is fully efficient.

4.5.2.2

Estimation of the two reform frontiers

Pre- and post reform frontiers (three years before reform, 1998-2000 and three years after reform, 2005-2007) are estimated in the same fashion as the meta-frontier. That is, observations of all hospitals during the period 1998-2000 (or 2005-2007) are pooled to estimate the pre-reform (or post-reform) meta-frontier15 . 14

The value of λit provides information on the peers of hospital i in period t. Peers are efficient hospitals that define the surface of the frontier onto which the inputs and outputs of hospital i in period t are projected. 15 One can argue that there may be technological change over a three-year period. Adoption of new medical technologies in a developing country like Vietnam is very gradual due to financial constraints and therefore, the three-year period is quite short for technological progress (if any) to be observed.

133


The only different between these two sets of DEA problem and the meta-frontier DEA problem is the number of observations used. The two matrices Y and X are now 3×186 matrix of output quantities and the 4×186 matrix of input quantities of 62 hospitals in each three-year period (pre- and post-reform); j0 is an 186 × 1 vector of 1; λit is the 186 × 1 vector of weights. Technical efficiencies for individual hospital for each year are defined as φ1it and labelled T E pre and T E post . Technical efficiencies drawn from this exercise are compared to those estimated under the overall meta-frontier T E M in order to measure the effect of reform on technical efficiencies for each hospitals.

4.5.2.3

Estimation of regional (group) frontiers

In order to investigate the hypothesis that financial reforms bring about uniform improvements in efficiency across regions, it is necessary to construct regional (group) frontiers and calculate the technical efficiencies implied by those frontiers. The comparison of those estimates with technical efficiencies calculated under the meta-frontier and two reform frontiers is the indication of the distribution of reform’s gains across region. Vietnam is divided into eight administrative regions of various climate, culture and socio-economic conditions. As health care closely relates to socio-economic situations, a regional variation worth emphasising here is the poverty rate. As shown in Table 4.5, the North has two of the poorest sub-regions in Vietnam, Northeast and Northwest, with close to 30% and 60% of the population living below the poverty line in 2004, respectively. Red River Delta of the North, however, is relatively richer than the rest of the country. As for the Central region, the poverty rates in the Highland and the northern part are relatively higher than that of its southern part, above 30% versus 19% in 2004. The South can also be divided into a relatively richer sub-region - the Southeast - with the poverty rate below 5.5% and the poorer Mekong Delta where close to 20% of the population lives below the poverty line. The diversity in socio-economic and living standards in Vietnam suggests that hospitals operating in different regions are likely to face different environmental constraints, either due to demand patterns, burden of diseases or availability of input factors, especially medical labour. The financial autonomisation reform, therefore, Productivity improvement (or reduction), if any, is likely to be the result of technical efficiency change rather than technological change.

134


Table 4.5: Poverty rate in Vietnam by region

Northeast Northwest

1993

1998

2002

2004

...

...

38.4

29.4

...

...

68.0

58.6

Red river delta

62.7

29.3

22.4

12.1

North central coast

74.5

48.1

43.9

31.9

South central coast

47.2

34.5

25.2

19.0

Central highland

70.0

52.4

51.8

33.1

Southeast

37.0

12.2

10.6

5.4

Source: Vietnam Development Report 2008: Social Protection (WB, 2007)

is expected to be adopted at various degrees in those regions and its potential impact on the productive efficiencies of hospitals are likely to be nonuniform. For each period (pre- and post-reform), four group frontiers are estimated (i.e. in total there are eight group frontiers). These represent sub-regions with different poverty rates: the “Rich” includes Red River Delta and Southeast with 14 hospitals; 12 hospitals in the Mekong Delta and South Central Coast are grouped into the mild-poverty region (the “Middle”); hospitals in the poorest sub-regions are divided into two groups - the “North Poor” (Northeast and Northwest) with 22 hospitals and the “Central Poor” (North Central Coast and Central Highland) with 14 hospitals. The differences in technology gaps between pre- and post-reform regional frontiers with the meta-frontier can answer the question of which regions have experienced efficiency improvement after the reform. The detailed grouping is mapped in Figure 4.8. Technical efficiencies estimated under these group frontiers for pre-reform period pre pre pre are denoted T Erpre , T Em , T Enp and T Ecp for the “Rich”, “Middle”, “North Poor” and “Central Poor” groups, respectively. By symmetry, this applies to the technical efficiencies estimated under these group frontiers for the post-reform period. As the group frontier problem is the sub-set of the meta-frontier problem, the meta-frontier never lies below the group frontiers and the solution of the former cannot exceed that of the latter. That is the value of φit solved under a group frontier problem is no greater than than the value of φit solved under the meta frontier problem. This means that any hospital is no more technically efficient when they are evaluated against the meta-frontier than against the group frontier. Once estimated, technical efficiencies of hospital i at time t with respect to the meta-frontier and with respect to group frontiers can be used to calculate its 135

136

Data: 62 hospitals (all regions) for 3 years after reform (2005-07) Output: TE post = TE measured with respect to the postreform frontier

Post-reform frontier

Data: 62 hospitals (all regions) for 3 years before reform (1998-2000) Output: TE pre = TE measured with respect to the prereform frontier

Pre-reform frontier

Reform frontiers

TE

reform

TE

post

is either TE

pre

or

MTRreform  TE M TE reform with

measured against the meta-frontier

TE M = TE

Output:

Data: 62 hospitals (all regions) for 6 years (19982000 and 2005-07)

Metafrontier

Meta frontier

reform MTR poverty  TE reform TEgroup

Data: 14 hospitals for 3 years (2005-07) - Output: TEcppost =TE w.r.t post-reform frontier for poor Central regions

Post-reform South Poor frontier (North Central and Central Highlands)

post Data: 22 hospitals for 3 years (2005-07) - Output: TEnp = TE w.r.t post-reform frontier for poor North regions

Post-reform North Poor frontier (Northeast and Northwest regions)

Data: 12 hospitals for 3 years (2005-07) - Output: TEmpost =TE w.r.t post-reform frontier for mild poverty regions

Post-reform Middle frontier (Central Coast and Mekong Delta regions)

Data: 14 hospitals for 3 years (2005-07) - Output: TErpost = TE w.r.t the post-reform frontier for rich regions

Post-reform Rich frontier (Red River Delta and Southeast regions)

Data: 14 hospitals for 3 years (1998-2000 - Output: TEcppre =TE w.r.t pre-reform frontier for poor Central regions

Pre-reform South Poor frontier (North Central and Central Highlands)

pre Data: 22 hospitals for 3 years (1998-2000) - Output: TEnp =TE w.r.t pre-reform frontier for poor North regions

Pre-reform North Poor frontier (Northeast and Northwest regions)

Data: 12 hospitals for 3 years (1998-2000) - Output: TEmpre =TE w.r.t pre-reform frontier for mild poverty regions

Pre-reform Middle frontier (Central Coast and Mekong Delta regions)

Data: 14 hospitals for 3 years (1998-2000) - Output: TErpre =TE w.r.t the pre-reform frontier for rich regions

Pre-reform Rich frontier (Red River Delta and Southeast regions)

Group frontiers

reform MTR  TE M TEgroup with “reform” is either pre or post and “group” is one of the four poverty groups

Figure 4.8: Group definition, frontier specification, technical efficiencies and meta-technology ratios


4.6. EMPIRICAL RESULTS AND DISCUSSION

meta-technology ratios as specified in Equation (4.13). It also follows that the meta-technology ratio M T R is the product of the reform meta-technology ratio M T Rref orm and the poverty meta-technology ratio M T Rgroup (see Figure 4.8 for details): MT R =

T EM ref orm T Egroup

=

T EM T E ref orm × T E ref orm T Egroup

= M T Rref orm × M T Rgroup

(4.13)

The decomposition reveals the two sources of changes of the technology gap: (i) the gap between pre- and post-reform frontier and the meta-frontier M T Rref orm and (ii) the gap between group frontiers in each period to the respective period’s (meta) frontier M T Rref orm .

4.6 4.6.1

Empirical results and discussion Meta technology ratios and the effect of reform

The DEA estimates of the meta frontier and group frontiers were computed using the program DEAP 2.1 (Coelli, 1996). Technical efficiencies and meta-technology ratios are estimated for each hospital in each year (for the six years). Tables 4.6 and 4.7 summarise some statistics of the estimates for the four regions under the two periods: before and after reform. The average technical efficiency scores of different regions in Table 4.6 is calculated as the (unweighted) mean of technical efficiencies of individual hospitals in the regions. That is, ¯M = TE

PI

T EM I

i=1

(4.14)

in which I is the number of hospitals in a particular region. The same calculation ref orm ref orm is applied to T¯E and T¯E group . In Table 4.7, the regional meta-technology ratios (M T Rs) are calculated as the ratio of the regional average technical efficiencies M with respect to the meta-frontier (T¯E s) over the regional average technical efficienref orm cies with respect to the regional (group) frontier (T¯E group s). Similarly, M T Rref orm s M ref orm ref orm ref orm are ratios of T¯E s over T¯E , and M T Rgroup s are ratios of T¯E and ref orm T¯E group .

137


As shown in Table 4.6, average technical efficiencies of the two poor regions, North and Central, with respect to the meta-frontier in the post-reform period are lower than those of the pre-reform period (83.54% and 82.81% during the prereform period, and 81.77% and 79.33% during the post-reform period, respectively). However, their average technical efficiencies when against their group frontiers are higher in the post-reform period. This implies lower meta-technology ratios M T R for both groups in the post-reform period, which is shown in Table 4.7. This indicates that after (and during) the financial autonomisation reform, the maximum output that is feasible for the group of poor regions in the North, using their available technology and inputs, is about 90.30% of the output that could have been achieved by the technology available to the public hospital sector in Vietnam. This is 4.34% lower than what was achieved during the pre-reform period. This number is worse for the poor regions in the Central Vietnam, more than 8% lower than before reform16 . Table 4.6: Output oriented technical efficiencies by poverty group Pre-reform

Post-reform

Mean Std. Dev. Mean Std. Dev. M Average technical efficiency with respect to the meta frontier T¯E North Poor (Northwest + Northeast)

0.8354

0.1442

0.8177

0.1732

Central Poor (Central North + Central Highland)

0.8281

0.1652

0.7933

0.1351

Middle (Central Coast + Mekong Delta)

0.8607

0.0890

0.9216

0.0959

Rich (Red River Delta + Southeast)

0.8258

0.1152

North Poor (Northwest + Northeast)

0.8382

0.1450

0.8735 0.1205 ref orm ¯ Average technical efficiency with respect to the reform frontier T E 0.8763

0.1414


0.8287

0.1650

0.8900

0.1127


0.8638

0.0883

0.9296

0.0933


0.8317

0.1171


0.8828

0.1290

0.9132 0.1042 ref orm ¯ Average technical efficiency with respect to the group frontier T E group 0.9056

0.1327


0.9165

0.1419

0.9645

0.0749


0.9546

0.0595

0.9470

0.0904


0.9525

0.0726

0.9828

0.0424

The opposite effect is observed by the mild poverty group: post reform technical efficiencies measured against the group frontiers are lower (95.46% before vs. 94.70% after reform) while those measured with respect to the meta-frontier are higher (86.07% before vs. 92.16% after reform). The wealthiest regions in Vietnam appear to experience more technical efficiency improvement within the group as well as with 16 Mean-comparison tests are performed on the calculated efficiency scores between the two poor regions and the richer regions. They concluded that the average efficiency scores of the poorer and the richer are different

138


respect to the meta-frontier. As a result, both groups experienced positive change in their meta-technology ratio M T R, 7.16% and 2.18% for the mild poverty and rich groups, respectively. The reform meta-technology ratios M T Rref orm are also interesting. As shown in Table 4.7, pre-reform frontier appears to sit higher than the post-reform one, i.e. technology regression is observed during and after the reform. The change is calculated as 5.25%, a fall from 99.62% pre-reform to 94.37% post-reform. Table 4.7: Meta-technology ratios by poverty group Pre-reform

Post-reform

0.9463

0.9030

MTR North Poor (Northwest + Northeast) Central Poor (Central North + Central Highland)

0.9035

0.8225


0.9016

0.9732


0.8670

0.8888


0.9967

0.9331


0.9992

0.8913


0.9964

0.9914


0.9929

0.9565


0.9495

0.9056


0.9042

0.9645

M T Rref orm

ref orm M T Rgroup


0.9049

0.9470


0.8732

0.9828

Recall the analysis in Figure 4.7 (in Section 4.3): reform fails to create a better environment for hospitals in Group 1. This is precisely the situation faced by hospitals in the poor regions. This can be the result of shrinking demand. When hospitals are allowed to charge fee-for-service, they shift their focus to a more expensive style of care. People in these poor regions have very low purchasing power and cannot afford these services, resulting in a delay of care. This is exacerbated by the recent trend of rapid urbanisation and inter-provincial migration for work in Vietnam. Those who can afford to move out of the very poor provinces are usually the near-poor. The urban migration results in a smaller and poorer population in those areas. Another possible explanation is that the reduction in state subsidy may affect some hospitals more than others, especially those that rely heavily on government subsidies for their facility and equipment maintenance and upgrade. Additionally, the incentive scheme that allows hospitals to set salary and bonuses 139


has inevitably made larger facilities, that are usually located in richer regions, more attractive to medical workers. This encourages doctors and nurses to try to work in such facilities, creating imbalances in the system, between poor and rich areas. Notice that for Group 1 (or the poor groups), the distances from pre-reform observations to their respective frontier is further than their post-reform counterparts. This implies that the average TE measured in relation to the pre-reform frontier is lower than that that measured in relation to the post-reform one. This piece of information by itself gives a distorted impression that performance of these hospitals has improved, while in fact, it has not, relative to what is available before and after the reform, i.e. the meta-frontier. Most hospitals in the relatively better-off regions benefit from the reform, similar to those of Group 2 described in Figure 4.7. They move toward the meta-frontier under the expansion of the post-reform technology possibility set. The reform clearly brings about improvement to these groups, probably due to more control over management decisions that were not possible before the reform. They include (but are not limited to) acquiring better medical technologies which used to require years of waiting for approval under state subsidy scheme, opening demand-oriented wards where people can pay out of pocket to receive better treatments, and simply changing the types of services offered. This is reflected by higher meta-technology ratios, precisely what was documented in Table 4.7. In summary, the observed negative change in the meta-technology ratios by some regions between the pre- and post-reform (on average) suggests two things. First, on average, hospital productivity has decreased after the first stage of autonomisation reform was implemented (i.e. Decree 10/2002). The source of productivity deterioration is the contraction of the production possibility set that is reflected by a lower post-reform frontier. This implies the potential for future improvement of technical efficiency is limited unless these constraints are investigated and removed17 . Second, the financial autonomisation scheme appears to benefit some regions more than others. More specifically, the poorest regions in the North and Central of Vietnam have not gained any productivity improvement during the reform period. 17

This is confirmed by the analysis of TFP change which suggests that the Vietnam public hospital industry has experienced productivity deterioration after the financial autonomisation reform was introduced.

140


4.6.2

Sensitivity analysis using the bootstrap method

As noted in the previous section, the main criticism of DEA is its non-parametric non-stochastic feature. Therefore, it is necessary to conduct sensitivity analyses to examine the robustness of the results. This is usually achieved through utilising other estimation methods, alternative model specification, treatment of outliers and bootstrap techniques. This section employs the bootstrap technique for sensitivity analysis. FEAR, an add-on package of the statistics programme R, is used to conduct the bootstrap procedure (see Wilson, 2008). Bootstrapping was introduced first by Efron (1979) and since then has been developed and used extensively in the DEA literature. It is based on the idea of repeatedly simulating the data generation process (DGP) through sampling and applying the original estimator to each simulated sample so that the resulting estimates mimic the sampling distribution of the original estimator (Simar & Wilson, 1998). Simar & Wilson (1998) proposed a four-step procedure to implement the bootstrap in DEA. First, for each hospital observation, compute the estimated efficiency score. Second, use the smooth bootstrap procedure to generate a random sample of efficiency scores. The smooth bootstrap involves the use of a kernel smoothing method to generate a smoothed distribution of estimated efficiencies from the first step. Then, transform the input-output vector (for each observation) using the corresponding re-sampled efficiency scores in the second step. This process creates pseudo data that is then used to estimate efficiency scores, i.e. the bootstrap estimates. It is repeated many times to provide a set of efficiency estimates for individual hospital observation. This allows for the calculation of biased-corrected efficiency scores, its associated standard deviation and confidence intervals. The number of repetitions can vary but Hall (1986) suggests at least 1,000 times to ensure adequate coverage of the confident intervals. The outlined procedure is applied to the estimation of the meta-frontier, two reform frontiers and six regional (group) frontiers. This results into three sets of ref orm corrected technical efficiencies for T E M , T E ref orm and T Egroup , and their associated biases. The bias-corrected technical efficiencies and the biases are calculated

141


by the following formula: Bias = TÊ − T E T˜E = T E − Bias = T E − TÊ + T E = 2 × T E − TÊ " # 12 K X 2 1 Std.Dev. = TÊ k − TÊ K − 1 k=1

(4.15) (4.16) (4.17)

in which TÊ k , TÊ, T˜E and denotes the estimated efficiencies in each pseudo sample, the mean of the bootstrapped technical efficiency and the bias-corrected technical efficiency scores. (for details, see Simar & Wilson, 1998, page 51). K is the number of pseudo samples drawn during the bootstrap process. Some selected bootstrap results are reported in Table 4.818 . They represent hospitals in the four regions of different poverty rates. The original technical efficiencies are those produced by solving the DEA problems in Equation 4.12. The bias-corrected technical efficiencies and the biases are described above. The bootstrap process reveals that some observations are quite sensitive to the re-sampling variation. For instance, hospitals ID-21512, ID-41101, ID-41110 and ID-70117 have biases of around 20%. This is not surprising because many of them are observations that hold up the DEA frontier: they have efficiency scores of unity. The removal of those “influential” observations can alter the shape and position of the frontier, resulting in different efficiency scores. However, the bias degree is rather small for many other influential observations (such as ID-22101, ID-50501 and ID-81701), around 10% or less. The bias-corrected meta-technology ratios (M T R, M T Rref orm and M T Rgroups ) are also computed using the formulae in Equation 4.13. Here, the values of T E M , ref orm T E ref orm and T Egroup are the bias-corrected technical efficiencies estimated under the meta-frontier, the two reform frontiers and the regional group frontiers, respectively. Those bias-corrected meta-technology ratios are then compared to the original values. The calculation reveals that despite the variations of efficiency magnitude, the (average) estimates for each poverty group are quite consistent. The average variations is computed by taking the differences between the regional averages of the original and the bias-corrected meta-technology ratios. The comparison is presented 18

Full results are presented in Appendix C.

142


Table 4.8: Bootstrap results of selected hospitals (1998 and 2007) ±95% CI

Year

ID

Original

Corrected

Biases

Std.Dev

1998

10110

0.700

0.667

0.033

0.014

0.648

0.695

2007

10110

0.669

0.606

0.063

0.041

0.560

0.666

1998

21506

0.707

0.650

0.056

0.030

0.617

0.701

2007

21506

0.635

0.585

0.050

0.025

0.560

0.632

1998

21512

1.000

0.765

0.235

0.227

0.680

0.995

2007

21512

0.667

0.623

0.044

0.023

0.592

0.663

1998

22101

1.000

0.914

0.086

0.039

0.882

0.992

2007

22101

1.000

0.890

0.110

0.074

0.815

0.995

1998

30502

0.939

0.863

0.076

0.050

0.811

0.934

2007

30502

0.912

0.860

0.052

0.027

0.824

0.907

1998

41101

1.000

0.810

0.190

0.141

0.750

0.993

2007

41101

0.832

0.772

0.060

0.041

0.718

0.827

1998

41110

1.000

0.777

0.223

0.192

0.714

0.994

2007

41110

0.693

0.659

0.034

0.016

0.636

0.689

1998

50501

0.868

0.823

0.045

0.023

0.791

0.863

2007

50501

1.000

0.876

0.124

0.069

0.834

0.992

1998

70117

1.000

0.805

0.195

0.139

0.757

0.994

2007

70117

1.000

0.849

0.151

0.112

0.770

0.995

1998

70902

0.919

0.852

0.067

0.032

0.819

0.914

2007

70902

1.000

0.824

0.176

0.128

0.769

0.995

1998

81701

1.000

0.929

0.071

0.031

0.900

0.993

2007

81701

0.722

0.676

0.045

0.023

0.650

0.717

Note: The ID sequences are the hospital identification numbers. The first number in each sequence (from 1 to 8) is the regional code.

in Table 4.9. It can be seen that the differences between pre- and post-reform M T Rs suggest that the two poorest regions have faced more difficulties after the reform was introduced. The Central Coast and Mekong Delta experienced larger positive changes compared to two richest regions of Red River Delta and the Southeast. Table 4.9: Differences between original and corrected pre- and post-reform meta technology ratios Regions

Original

Corrected


-2.5%

-4.8%


-5.3%

-5.8%


8.0%

5.3%


3.9%

1.5%

This calculation suggests that the DEA-based estimates of technical efficiencies 143


and meta-technology ratios are robust and policy implications can be drawn from the estimation.

4.7

Concluding remarks

This chapter has demonstrated how the meta-frontier framework developed by Battese, et al. (2004) and O’Donnell, et al. (2008) can be used to examine the effect of reform on hospital efficiency. The evaluated reform is the implementation of Decrees 10/2002 (and 43/2006) that gives hospitals more micro-management power over their own revenue, resources and service provision. It is claimed to create a stronger incentive for hospitals to increase productivity and efficiency in providing health care services. The analysis is performed using the non-parametric technique, DEA. The results suggest that the financial autonomisation reform (in the context of existing institutional framework) has not brought about productivity and efficiency improvement. Moreover, its effect across regions is not uniform: the poorest regions are the losers in this reform. This could be the effect of lower demand due to low purchasing power of the poor who now have to face more expensive services while not being fully covered by insurance or exemption schemes. Additionally, this could also be the result of a shrinking medical workforce in the poor regions when the relatively richer can afford to pay more in their facilities and thus are more attractive to medical staff. This finding confirms the argument made by Lieberman & Wagstaff (2009) that in a system where the patient is the main payer like in Vietnam, provider autonomy can spell disaster for the patient and the health system. The ongoing autonomisation of public providers, without removing the out-of-pocket burden, is likely to contribute to Vietnam’s service delivery problems, not help to resolve them. The study has several policy implications. Firstly, market-like reforms in health care, such as this autonomisation reform, do not always lead to productivity and efficiency improvement, while inevitably increasing inequity. Health care services do not share the same properties of the “usual” commodities and its market is usually characterised with market failures, such as information asymmetry. A reform that allows public hospitals to operate in a corporate fashion (i.e. profit-seeking) tends to marginalise the poor and increase the cost of health care. This is because providers can increase volume (supply induced demand), charge on a fee-for-service basis and discriminate against patients who cannot pay (cream-skimming). It can

144


also create difficulties for hospitals in poorer areas with respect to recruiting medical staff and maintaining their demand and thus productivity. This leads to an imbalance of service provisions between regions of different living standards, both in terms of quality and types of available services. These effects are indirect and hard to quantify in advance, they may come from the competition for medical labour resource between regions, which is the consequence of richer hospitals being able to offer better salaries and working conditions. Alternatively, they might come from the interaction between supply (hospitals) and demand (users), where the service prices set by the former is affordable to only a small proportion of the population. Therefore, future rounds of reform should take these equity issues into considerations. Secondly, making public hospitals more accountable for their financial positions is crucial in a developing country with limited government resources. This is precisely the direction of financial autonomisation in Vietnam. However, there should be a monitoring and evaluation system installed together with this reform to not only benchmark hospitals in terms of their financial performance, but also their productive efficiency. The pursuit of profit does not usually coincide with finding the most productive use of resources, especially in the imperfect market of health care. This study demonstrates the possibility of using frontier techniques to evaluate hospital performance for policy making purposes. In fact, regulators around the world have realised the usefulness of frontier techniques and many DEA studies have informed their decision making19 (Bogetoft & Nielsen, 2004). Lastly, it would be fruitful if this exercise could be expanded to cover a larger dataset so that reform effects can be fully studied for hospitals of different governmental levels and classes. Moreover, the effect of the second phase of the autonomisation reform, the implementation of Decree 43/2006, will be better evaluated using the latest data from 2008 and 2009. Like other studies that use the DEA approach, the results are sensitive to outliers, measurement errors and stochastic factors. The bootstrap analysis revealed that some observations that form part of the frontier are more sensitive than others. However, the corrected technical efficiencies, and estimated meta-technology ratios overall lead to the same conclusion that some regions have experienced more diffi19 Recently, DEA has been used directly to define regulatory incentives. In particular, it has been used in incentive regulation of energy utilities. For example, in the regulation of electricity distribution, countries like Norway, Holland, and Finland have introduced DEA based revenue and price cap systems. Furthermore, DEA has been used to determine reasonable cost norms in countries like Australia, England, New Zealand and Sweden. Given this trend, it is natural to discuss the potential use of DEA in the regulation of public hospitals (Bogetoft & Nielsen, 2004).

145


culties than others. The differences between the estimated and corrected values of efficiency scores and meta-technology ratios suggest that caution should be taken for any direct application of efficiency scores. Another drawback of this study lies in the data imperfections, especially with respect to output measures. As described in Section 4.4, output measures are expressed in raw count for eleven categories, of which nine were then aggregated into one variable that captures volume diagnostic tests and treatment procedures. When casemix is not reasonably accounted for, there is a negative bias toward hospitals that treat more seriously ill patients. This issue can only be resolved if better data are available, in which some forms of casemix categorisation is applied. Finally, this analysis cannot address the relationship between quality of care and efficiency due to the deficiency of data on service quality. This is a common problem in the hospital efficiency literature (see discussions in Rosko & Mutter, 2008, for an overview of this issue). Until better data are available (e.g. patient-level data or some general service quality measures), analysis and policy makers still need to handle the efficiency results with care and within the context of individual hospital system.

146

Chapter 5

Conclusions

... (A) large element of the health care sector is provided by nonmarket organisations. Given the complexity and the function undertaken by such institutions, and in the absence of the usual market signals, there is a clear need for instrumentals that offer insights into performance. (Jacobs et al., 2006) This chapter summaries the contributions of the thesis, relevant policy implications and some suggestions on the direction of future research.

5.1

Thesis summary and contributions

The thesis is motivated by the important role that efficiency measurement plays in the improvement of health sector performance. Efficiency is one of the main objectives of the health system, along with equity and quality. Achieving these competing objectives all together is not a trival task in a sector like health care. Good performance cannot therefore be taken for granted (WHO, 2000). Health care is different from other sectors of the economy for two main reasons. First, the existence of uncertainty in the incidence of disease and in the efficacy of treatment leads competitive markets to generate an inefficient allocation of resources (Arrow, 1963). Furthermore, the health care market is characterised by natural imperfections such as externalities, supply-demand dependency, cream skimming, adverse selection, moral hazard and monopoly. Hence, no one should expect freemarket principles to be enough to achieve the objective of efficient allocation of

147

5.1. THESIS SUMMARY AND CONTRIBUTIONS

resources1 . This justifies the need for government intervention to “correct” the potential for distortions, excessive, insufficient or mis-allocated spending. Second, the demand/need for health care services are derived from the desire of good health - a critical component of human well-being. The equity principle, that is equality of access, of medical needs, and of ultimately health outcomes, is therefore a central concern. Pure market based allocation based on willingness to pay for care in the context of income disparity will certainly never guarantee an equitable outcome. Health sector regulations are required to ensure that adequate services are provided in a equal manner. In short, there is a need for an active and complicated role for government (WHO, 2000). Since the health care market cannot be governed by the market mechamism, it should not be evaluated by standard market-based indicators such as profitablity and market share. Performance measurement has emerged from the need for a set of measurable and reliable indicators that take into account the health system’s objectives and priorities. It offers policy-makers an opportunity to secure health system improvement and accountability (Smith et al., 2008). However, the complexity of health care services and the health care market makes the scope of performance measurement enormous. It ranges from examining the state of a country’s health system to assessing activities of health service providers to reflecting the experiences of the individual users. Indeed, governments around the world have tried different health care policies aimed at increasing efficiency. Various schemes and indicators to measure health systems’ performance have been developed and improved over time. Challenges remain still and continue to prevent efficiency measurement from becoming a part of the routine analytical toolbox for decision makers. Amongst these are analytical methodologies, more specifically the reliability of efficiency estimates, and the translation of results into relevant and actionable policy implications. The overall objective of this thesis is to contribute to the existing knowledge of efficiency measurement in the health care sector through the analysis of the reliability of efficiency estimates with respect to various modelling choices. This is achieved through theoretical and empirical applications of different efficiency analysis frameworks in the form of three major studies that have widely different objectives but bear the common theme of measuring hospital efficiency. More specifically, Chapter 2 explores the variation of efficiency estimates under different modelling approaches. This is the first attempt to quantify the effects of 1

As discussed in by Paul Krugman health-care-is-not-a-bowl-of-cherries/.

in

http://krugman.blogs.nytimes.com/2009/06/28/

148


modelling choices on estimated efficiencies in the health care sector using the metaregression technique. It explains the sources of variation in efficiency estimates, and derive statistical comparison between various model choices, including parametric specification, functional form, orientation and efficiency distributions. The estimated coefficients of the meta-regression indicates the degree of bias introduced by omission and aggregation of variables, as well as by using small versus larger samples. It also suggest that efficiency estimates might not differ very much between the choices of parametric versus. non-parametric, or between different efficiency distributions, or even higher functional forms. However, different assumptions on returns to scale and the treatment of panel data are likely to result into various efficiency estimates of the same sample. The results can help analysts in making their modelling decisions, especially with respect to the tradeoff between missing observations and omission (or aggregation) of important variables. The estimated coefficients from the meta-regression can be used to adjust for average estimated efficiencies and to compare efficiency estimates across sample sizes, variable choices and approaches. This ultimately enables comparison and benchmarking across empirical studies of different time frames, sectors, countries, states and regions. Once variations from analytical methodology are reduced, studies can focus on “real life” factors that contribute to the efficiency differences such as funding arrangement, competition regulations, market structure, and ownership, just to name a few. Chapter 3 inquires into the labour efficiency of public hospitals in Queensland. Queensland is chosen for the reason that medical labour shortages have been a critical issue here and an understanding of labour efficiency would be very useful for decision makers. The hospital sector is a labour intensive industry and if productivity improvements are to be made, they are likely to come from the better use of human resource first. The study makes use of a labour requirement function instead of a standard production or cost functions in order to achieve the objective of measuring labour efficiency. The estimation is performed under various model specifications, one of which is the latest generation of frontier model in a panel data framework proposed by Greene (2004, 2005a). It demonstrates that latent heterogeneity should not be ignored when measuring efficiencies, especially when there are reasons to believe that variables controlling for hospital characteristics are not sufficient to account for inefficiency unrelated variations. It also notes that despite the variations of efficiency estimates, different models are quite consistent in determining the relative 149


importance of different outputs and hospital characteristics on labour utilisation. Therefore, these parameters can be used to guide policy designs. This study also highlights (again) the economic cost, due to the lost of scale efficiency, of having a dispersed population. Hospitals in rural and remote areas are usually small and suffer from diseconomies of scale. However, they are crucial to ensure equity of access and health outcome, and therefore, need to carry usual tertiary cares plus added responsibility to provide primary health and aged care services. Opening a new residential town, usually far away from everywhere else means a new(and inefficiently small) hospital or outpatient clinic is needed. Therefore, the social cost of scale inefficiency ought to be one of the arguments against the opening of new residential towns in rural and remote areas The empirical analysis of the impact of financial autonomisation reform on productive efficiency of public hospitals using Vietnam as a case study is presented in Chapter 4. Vietnam’s health sector is in transition and it has experienced various market (and quasi-market) based reforms with the expectation that market-based allocation of health resources would deliver more efficient outcomes. One of the most recent market-based reforms, Decrees 10/2002 and 43/2006, aims at increasing financial autonomy of public providers so that staffing, investment and organisational restructuring decision can be made at the hospital level. This autonomisation reform has been warmly welcomed by both health-sector decision makers, as the pressure of hospital grants/subsidies on the government budget is reduced, and hospital managers, as they enjoy more micro-management flexibility. However, there are concerns about the risk that this flexibility can be exploited in the pursuit of profit under the fee-for-service funding arrangement. This threatens the poor and middle income groups with unaffordable health care. It also presents the possibility of imbalanced service utilisation at different hospital levels, and between rural and urban areas. Furthermore, the scope for providers to induce demand can drives up the cost of health care while resulting in little improvement of health outcomes. The impact evaluation is conducted using the recently developed meta-frontier framework by Battese & Rao (2002) and O’Donnell et al. (2008) on a balanced sample of 62 hospitals over the two periods, before and after reform. The findings carries an important policy message that the reform outcomes have not been shared equally across regions and by different socio-economic groups. Productive efficiency does not necessarily coincide with financial improvement and profitablity. The financial autonomisation that turns public service units into quasi-corporates, in the context of a funding system characterised by out of pocket payment in Vietnam, has 150

5.2. FUTURE RESEARCH

failed to improve either equity or efficiency. This study demonstrates the possibility of using frontier techniques to evaluate hospital performance for policy making in Vietnam, rather than relying only on financial and individual activity indicators. Apart from the policy messages, this study contributes to the literature as a novel approach to study reform’s effects using the meta-frontier framework.

5.2

Future research

There is considerable potential to expand this work in the future. The study in Chapter 2 can be enriched by addressing some other research questions. Once the effect of estimation techniques are isolated, the focus of efficiency comparison can be shifted to operational environment and characteristics of hospitals. For instance, two major questions one can ask are: which financing mechanisms lead to better hospitals performance; and the degree of efficiency variations caused by the use of outputs that are case-mix and quality adjusted, compared to those that are not. The second question is a natural extension because most studies provides a very good description of the variables used, whether outputs are casemix adjusted or not, and if some quality variables are used to control for quality heterogeneity. The first question is more challenging and requires further data collection on funding methods of the study countries during the period of concern. Although quite feasible for OECD countries, this is by no means an easy task for many other countries/regions due to their incomplete records of health system organisation and funding. In addition, there is no doubt that modelling choices would have some effect on the variance of efficiency estimates. This would require more reviews of hospital efficiency studies that report both mean efficiencies and the associated variance. This can improve the confidence placed in the point estimate of efficiency if the study of efficiency variance can result in narrowing the spread of the estimates, The two empirical studies in this thesis (presented in Chapter 3 and 4) do not explicitly address the relationship between quality of care and efficiency due to the deficiency of data on service quality. This is a common problem in the hospital efficiency literature (see discussions in Rosko & Mutter, 2008, for an overview of this issue) and it is not possible to deal with this until better data are available, for instance patient-level data or some general service quality measures.

151


Additionally, improvements can also be made in both studies with better data of output and input measures as well as larger longitudinal data. The labour measures in the Queensland public hospital study can be improved because it lacks detailed separation of staff skill-mix and their respective wages within each labour category. There is usually observed a significant staffing inequality between type of hospitals (large vs. small), and between urban and rural/remote hospitals due to working environment and recruit difficulty in rural and remote regions. The data issue with the Vietnam study lies in its data imperfection with respect to output measures. Output measures are expressed in raw count and when casemix is not reasonably accounted for, there is a negative bias toward hospitals that treat more seriously ill patients. Until a casemix system is implemented and data on hospital activities are reported accordingly, a possible solution is data collection with more detailed break down of services with their assocated prices or variable costs. The Queensland public hospitals are evaluated under the old funding system of historical budget while the new model of casemix funding was introduced in the financial year 2007-08. The new model is believed to positively affect productivity by encouraging providers to favour treatments under lower cost settings while maintaining quality of care (QH, 2008). It would therefore be beneficial to investigate the evolution of efficiency before and after the reform. This requires the collection of (compatible) data for some most recent years (from 2007-08 to 2009-10). Another step further from this study is measuring technical efficiency, instead of labour efficiency only, using input distance function approach. This may be of interest because it can show which inputs are net substitutes or complements, what are their shadow prices and own price elasticities of demand for the inputs. These facts are important to answers questions related to the workforce shortage2 . The study on Vietnam hospitals can also be extended when a larger longitudinal data set becomes available. The last round of financial autonomisation reform started in late 2006 and three years of implementation would be sufficient to give some initial evaluation of the complete effect of both rounds. The inclusion of more cross sections also no doubt improves the reliability of the estimates. With respect to estimation strategy, the meta-frontier can be estimated under the stochastic frontier framework. SFA, in contrast to DEA, is less sensitive to outliers so the comparison of two sets of result (under SFA and DEA) would be useful to shed light on the size of measurement errors and stochastic factors. 2 This direction is suggested by an anonymous reviewer of the Journal of Medical Systems. I am very grateful for the detailed comments and insights by this reviewer.

152


Health economics is the study of how to rationally allocate limited resources to meet seemingly unlimited demand/need of health care services. Absent of other concerns (equity, equality of care etc), rationality means allocating resources in such a way as to achieve highest possible productivity. Thus, efficiency improvement is desired in every health care system and country regardless of income level. Performance measurement offers policy-makers opportunities for monitoring productivity and efficiency of providers, which is fundamental to seek improvement. It provides feedback to clinical facilities and practitioners on their actions and how they compare their peers. Without proper measurement and evaluation, efficiency improvement cannot be systematic and is unlikely to be sustained.

153

Appendix A

List of studies included in the meta-regression analysis

The appendix summarises the 95 studies included in the meta-regression analysis. Those studies are found through an extensive literature search using various databases with key words of “efficiency”, “productivity”, “hospital”, “health center”, “data envelopment analysis”, “stochastic frontier”, “production function” and “cost function”. Amongst more than 220 publications on health care/hospital efficiency, 95 were selected covering the publication period of 1987-2008. Since many studies utilise different methods, and/or use more than one dataset, and/or apply several models to the same dataset, 253 cases were extracted from these studies. The meta-regression analysis is then performed on the meta-data set. See Chapter 2 for the description and results of the meta-regression analysis.

154

155

Bitran & Josep (1987) Grosskopf & Valdmanis (1987) Valdmanis (1992) Grosskopf & Valdmanis (1993)

Lynch & Ozcan (1994) Ozcan & Bannick (1994) Burgess & Wilson (1995)

Burgess & Wilson (1996)

1

5

8

7

6

4

3

2

Author

No

US

US

US

US

US

US

US

US

Country

Nonparametric Input and output distance functions

Nonparametric Input and output distance functions

DEA input oriented DEA output oriented

DEA input oriented DEA input oriented

Estimation method DEA output oriented DEA input oriented

2246

1480

372

1535

49

41

82

Sample size 160

supplies, beds, service mix, provider labour, nursing support labour, other support labour Number of acute care hospital bed weighted by scope of service index, number of long term hospital bed, registered nurses FTE, licensed practical nurse FTE, other clinical labour FTE, nonclinical labour FTE, long term care labour FTE Number of acute care hospital bed weighted by scope of service index, number of long term hospital bed, registered nurses FTE, licensed pratical nurse FTE, other clinical labour FTE, nonclinical labour FTE, long term care labour FTE

capital assets, labour, supplies

acute care, intensive care, surgeries, ambulatory and emergency care physicians, nurses, FTE others, admissions, net plant assets physicians, non physician labours, net plant assets, casemix

FTE, salary dollars, other cost

Input list

acute care inpatient days, case mix weighted acute care inpatient discharges, long term care inpatient days, number of outpatient visits, ambulatory surgical procedures, inpatient surgical procedures acute care inpatient days, case mix weighted acute care inpatient discharges, long term care inpatient days, number of outpatient visits, ambulatory surgical procedures, inpatient surgical procedures

acute care inpatient days, intensive care inpatient days, surgeries, ambulatory plus emergency services adjusted discharges, outpatient visits, training inpatient days, outpatient visits

adult, paediatric, elderly

physicians, FTE non-physician labours, admission, net plant asset

discharges by 15 MDCs

Output list

NA

NA

NA

NA

NA

NA

NA

NA

Control variables

156

Burgess & Wilson (1996)

Magnussen (1996) Morey & Dittman (1996)

9

10

Vitaliano & Toren (1996)

LopezValcarcel & Baber Perez (1996)

White & Ozcan (1996)

12

13

14

11

Author

No

US

Spain

US

US

Norway

US

Country

DEA input oriented

Stochastic cost frontier, CD, exponential DEA input oriented


Estimation method DEA input oriented and cost

170

225

219

105

138

Sample size 360

size, labour, expenses, service complexity

Doctors, other staff, beds

Physicians and nurses, other personnel, beds $ of all nursing services consumed in the year, $ for ancillary services, $ of administrative and general services, number of intensive care beds, number of acute care beds, number of other beds, % of all patient days that are classified as requiring intensive care total cost

number of personnel, number of beds

Input list

patient days, emergency room visits, outpatient clinic visits, case mix index, tech index, occupancy rate medical inpatient days, surgical inpatient days, intensive care inpatient days, obstetric inpatient days, newborn inpatient days, paediatric inpatient days, ambulatory surgical procedures, operations with hospitalisation, upamix, admissions adjusted discharges, outpatient visits

medical and surgical days, long term care days, outpatient visits Number of patient days for patients less than 14, number of patient days for patients from 14-65, number of patient days for patients over 65

Acute days, sub-acute days, intensive days, surgeries performed, discharges, outpatients

Output list

NA

NA

teaching hospitals

NA

Quality, total patient days, occupancy rate, proportion of patients treated as outpatients, intensity of care, public or not, three states (as dummies) NA

Control variables

157

DalmauMatarrodona & Puig-Junoy (1998)

Chirikos (1998a)

17

19

Chang (1998)

16

Chirikos (1998b)

Parkin & Hollingsworth (1997)

15

18

Author

No

Spain

US

US

Taiwan

Scotland

Country

Stochastic cost frontier, TL, CD, half normal, exponential DEA input oriented

Stochastic cost frontier, TL, half normal

DEA input oriented

Estimation method DEA input oriented

94

2232

558

29

Sample size 75

FTE physician (including residents), nurses and equivalents, other non sanitary personnel, inpatient beds

total cost

average number of staffed beds, number of trained, learning and other nurses, number of professional, technical, admin, and clerical staff, junior and senior non nursing medical and dental staff, cost of drug supply, hospital’s capital charge FTE physicians, FTE nurses and medical supporting personnel, FTE general and admin personnel Total cost

Input list

clinic visits (including regular and emergency), weighted patient days, gere, A and I, CHRO post admission patient days for which Medicare is primary payer, post admission patient days for which Medicaid is primary payer, post admission patient days for which payer is either Blue Cross, other private payer or self-pay patient, casemix Case mix adjusted admission, post-admission patient days corresponding to three different payer groups, two outpatient indices Case mix adjusted discharged patients, inpatient days in acute and sub-acute, inpatient days in intensive, inpatient days in long term and other services, surgical interventions, hospital day care services, ambulatory visits, resident physicians

medical acute discharges, surgical acute discharges, accident and emergency attendances, outpatient attendances, obstetrics and gynaecology discharges, other specialty discharges

Output list

NA

NA

NA

NA

NA

Control variables

158

Author

Linna (1998)

Mobley & Magnussen (1998)

O’Neil (1998)

Webster et al. (1998)

Linna & Hakkinen (1998)

No

20

21

22

23

24

Finland

Australia

US

Norway, US

Finland

Country

DEA input oriented and cost, Stochastic production and cost frontiers, CD, translog, half normal, truncated DEA input oriented and cost, Stochastic cost function, CD, half normal, exponential

DEA input oriented

Estimation method DEA input oriented and cost, Stochastic cost function, linear, truncated DEA input oriented

48

301

27

228

Sample size 43

total cost

technological services, beds, FTEs, supply (operational expenses excluding payroll, capital and depreciation) FTE professional medical offers, total contract value of VMO, nurses FTE, other staff FTE, beds, materials (non labour cost)

FTE physicians and residents, FTE other labours, beds

total cost

Input list

Emergency visits, scheduled visits, admissions, bed days, residents, nurse education, student, research

Number of patient days in three age groups, number of outpatient visits, case mix index for patient 65+ Adjusted inpatient medical, adjusted inpatient surgical, adjusted outpatient, residents trained Acute care inpatient days, surgery inpatient days, non inpatient occasion of services, nursing home type inpatient days, accident/emergency

Emergency visits, outpatient visits, DRG inpatients, bed days (applied for inpatient episodes exceeding a certain cut off point), residents trained, on the job training nurses, research

Output list

NA

NA

NA

Hospital types

NA

Control variables

159

Athanassopoulos Greece et al. (1999)

Fuiji & Ohta (1999)

Kerr et al. (1999)


Maniadakis et al. (1999)

26

27

28

29

30

Scotland

Finland

Northern Ireland

Japan

Jordan

Al-Shammari (1999)

25

Country

Author

No

Nonparametric Input distance function, Malmquist

DEA input oriented and cost, Stochastic cost frontier, linear, half normal

Stochastic cost frontier, TL, truncated DEA output oriented

DEA input oriented

Estimation method DEA output oriented

Average working hours of doctors, other employees, total cost of materials, equipment and other costs

Doctor, nurse, other personnel, bed, cubic meter, admission for stroke, fractured neck of femur, myocardial infraction

75

Nurses, consultants, administration, ancillary, beds

Doctors in general medicine, doctors in surgical, doctors in labs, management and nursing staff, hospital beds total cost

bed days, physicians FT, health personnel)

Input list

95

33

2781

98

Sample size 15

total number of inpatients and outpatient per day, Ratio of inpatient/outpatient, number of examination/100 patients Surgical, medical, obstetrics and gynaecology, accident and emergency Number of emergency visits, scheduled and follow up visits, DRG weighted number of total admission, number of bed days exceeding a cut off point, number of residents receiving 1 year training, total number of on the job training weeks of nurses, number of A&E attendances, adjusted inpatients, adjusted day cases, adjusted outpatients, standardises survivals after admission for stroke, fractured neck of femur, and myocardial infraction

Patient days, minor surgical operations, major surgical operations Patients general medicine, patients surgical, lab tests, clinical examinations

Output list

NA

NA

dummies for emergency hospital and general type of hospital, nursing standard There are control variables but not specified

NA

NA

Control variables

160

Chirikos & Sear (2000)

Yong & Harris (1999)

33

35

Rosko & Chilingerian (1999)

32

Chern & Wan (2000)

Rosko (1999)

31

34

Author

No

US

US

Australia

US

US

Country

DEA output oriented, Stochastic cost function, CD, TL, half normal

Stochastic cost frontier, TL, half normal Stochastic cost frontier, CD, half normal, exponential DEA input oriented

Estimation method Stochastic cost frontier, TL, truncated, half normal, exponential

186

80

35

195

Sample size 3262

Beds, service complexity, non-physicians FTE, operating expenses (not including payroll, capital or depreciation) wage and salary for personnel engaged in inpatient care activities, wage and salary for personnel assigned to non patient care, other expenses, capital costs, adjusted depreciation charges for fixed and movable equipment, other non patient cost

total cost

total cost

total cot

Input list

case mix weighted admission, three post-admission patient days variables, test and procedures, level of activity in ambulatory centre

weighted inliers equivalent separation (case mix adjusted), on campus medical clinical occasion of services, emergency/casualty occasion of services Case mix adjusted discharges, outpatient visits

inpatient discharge, outpatient visit

outpatient visits, inpatient discharges, post-admission days

Output list

NA

NA

teaching, A1 hospital

case mix index, ER visit/total outpatient, dummy for hospitals are member of teaching hospitals, dummy for teaching hospital that are noth COTH member NA

Control variables

161

Frech & Mobley (2000)

37

40

39

Harris et al. (2000) Maniadakis & Thanassoulis (2000) Prior & Sol (2000)

Deily et al. (2000)

36

38

Author

No

Spain

UK

US

US

US

Country

DEA input oriented DEA input oriented and cost, Malmquist DEA input oriented

Stochastic cost frontier, CD, half normal

Estimation method Stochastic cost frontier, TL, half normal

132

75

20

378

Sample size 4739

health staff, other staff, bed, purchase of materials

service mix, size, employees, operational expenses Doctors, nurses, other personnel, beds cubic metres per 100

total cost, net plant property and equipment at beginning of period (measured by depreciation and amortisation), number of licensed physicians with admitted privileges

total cost

Input list

medicine inpatient days, surgery inpatient days, obstetrics and gynaecology inpatient days, paediatric inpatient days, psychiatric inpatient days, long stay inpatients, intensive care inpatients, external visits

adjusted discharges, outpatient visits A&E attendances, adjusted inpatient, adjusted day stays, adjusted outpatients

Total inpatient discharges in each of 6 payoff categories, number of outpatient visits, number of FTE interns and residents per staff bed (teaching output)

admission, inpatient days, outpatient visits

Output list

NA

NA

hospital accreditation, number of FTE residents per bed, % intensive bed care, number of inpatient surgical operations per admission, % outpatient visits that are surgical, % outpatient visits that are emergency, index of high technology 5 case mix indices, proportions of outpatient visits that are non-surgical, sub-acute, newborns, medical surgical acute care, intensive care, expenditure on charity care and donation, scope of service index, worker age index, income per capita in the hospital NA

Control variables

162

Sommersguter- Austria Reichmann (2000) Athanassopoulos Greece & Gounaris (2001) Folland & US Hofler (2001)

Fuiji (2001)

43

46

45

44

Sahin & Ozcan (2000)

42

Japan

Turkey

Spain

Puig-Junoy (2000)

41

Country

Author

No

Stochastic cost frontier, cubic, truncated, half normal

DEA input oriented, Malmquist DEA input oriented and cost Stochastic cost frontier, CD, TL, truncated

DEA input oriented


Total cost

1661

955

beds, specialists, GP, nurses, other allied professionals, revolving funds expenditure FTE labour, beds, expenses for external medical services

FTE physicians, FTE nurses and equivalents, FTE other non-salary personnel, inpatient beds

Input list

total cost (sum of one cost component with known prices and one with unknown prices) Cost

98

22

80

Sample size 94

general medical surgical, paediatrics, obstetrics/gynaecology, all other inpatient, (all measured by annual inpatient days), and outpatient visits inpatients/day, outpatients/day, number of clinical examinations/100 patients

medical patients, surgical patients, medical examinations, lab tests

Outpatient, number of credit points times a steering factor

case-mix adjusted discharged patients, inpatient days in acute and sub-acute services, inpatient days in intensive care, inpatient days in long term care and other services, surgical interventions, ambulatory visits, resident physicians outpatient visits, discharged patients, hospital mortality rate

Output list

dummy for teaching, general, meeting some standard of nursing, standard of meal, standard of bed, inverse of bed occupancy rate, dummy for being subsidised, and urban

% board certified, reservation quality

NA

NA

NA

NA

Control variables

163

Grosskopf et al. (2001)

Jacobs (2001)

48

49

Rosko (2001a)

Giokas (2001)

47

50

Author

No

US

UK

US

Greece

Country

Stochastic cost frontier, TL, truncated

Nonparametric Input distance function DEA input oriented and cost, Stochastic cost function, linear, half normal


1498

232

792

Sample size 91

Total cost

Cost index

beds, med staff, med residents and interns, registered nurses, licensed practical nurses, FTE other labours

Total cost

Input list

Episodes per spell, transfer per spell, transfer out per spell, emergency, finished consultant episodes, non primary outpatient attendance, emergency index, proportion under 15, proportion 60+, proportion of female, students, research, market force factor DRG weighted inpatient discharges, outpatient visits

inpatient days medical, inpatient days surgical, outpatient visits, ancillary services (including anaesthesiology, lab, x-ray) patients, inpatient surgical, outpatient surgical, ER visits, outpatient visit

Output list

dummy for being member of COTH, dummy for teaching hospitals not being a member of COTH, emergency/OPV, Outpatient surgeries/OPV, HMO enrolment/pop, Medicare discharges/total discharges, Medicaid discharges/total discharges, dummy for investor owned

NA

NA

NA

Control variables

164

Rosko (2001b)

Zere et al. (2001)

Hofmarcher et al. (2002) McKay et al. (2002)

51

52

53

54

Author

No

US

Austria

South Africa

US

Country

DEA output oriented, Malmquist DEA input oriented Stochastic cost frontier, TL, half normal

Estimation method Stochastic cost frontier, TL, half normal

4075

93

86

Sample size 1966

medical staff, para-medical staff, admin, beds Total cost

beds, recurrent expenditure

Total cost

Input list

admission, inpatient days, outpatient visits

patient days, discharges

outpatient visits, inpatient days

inpatient discharges, outpatient visits, days in long term units

Output list

dummy for accredited hospitals, number of FTE residents per bed, % intensive care beds, number of inpatient surgical operations per admission, % outpatient visits that are surgical, % outpatient visits that are emergency, high tech index

NA

dummy for being member of COTH, dummy for teaching hospitals not being a member of COTH, medicare patient casemix index, emergency/OPV, Outpatient surgeries/OPV, HMO enrolment/pop, Medicare discharges/total discharges, Medicaid discharges/total discharges NA

Control variables

165

Biorn et al. (2003) Carey (2003)

56

58

Street (2003)

Street & Jacobs (2002)

55

57

Author

No

UK

US

Norway

UK

Country

DEA input oriented Stochastic cost frontier, hybrid TL, half normal Stochastic cost frontier, linear, half normal, exponential

Estimation method Stochastic cost frontier, linear, half normal, truncated, exponential

226

1209

432

Sample size 217

Total cost

FTE physicians, FTE other labours, medical expenses Total cost, beds

Total cost

Input list

case-mix adjusted inpatients, first outpatient attendances weighted by specialty, first addicent and emergency attendances, transfer into hospital per spell, transfer out of hospital per spell, emergency admission per spell, finished consultant episode int

inpatient services, outpatient services adjusted admission, adjusted patient days

transfer to hospital per spell, transfer out of hospital per spell, emergency admission per spell, finished consultation episode inter-specialty transfer per spell, episode per spell

Output list

non-primary outpatient attendances per inpatient spell, standardised index of unexpected emergency admission/total emergency admissions, proportion of patient less than 15, more than 60, female, student whole time teaching equivalent per spell % of total revenue

case mix, HMO, HHI, for profit, teaching, system

non-primary outpatient attendances per inpatient spell, standardised index of unexpected emergency admission/total emergency admissions, HRG weight (casemix index), proportion of patient less than 15, more than 60, female, student whole time teaching equivalent per spell NA

Control variables

166

Bilodeau et al. (2004)

Chang et al. (2004)

Grosskopf et al. (2004)

Kirigia et al. (2004)

Martinussen & Midttun (2004) Dervaux et al. (2004)

Valdmanis et al. (2004)

Harrison et al. (2004)

59

60

61

62

63

65

66

64

Author

No

US

Thailand

US and France

Norway

Keynia

US

Taiwan

Canada

Country

DEA input oriented

Nonparametric Input distance function DEA output oriented

DEA input oriented

DEA input oriented

DEA input oriented

DEA output oriented


525

68

1080

153

32

252

483

Sample size 1359

beds, doctors, nurses, other staff, allowance, drug expenses, other operating expenses operating expenses, FTE, services, beds

beds, physicians, nurses, other labours

hours and expenses on labour, expenditure on supplies, foods and meals prepared for patients, total expenditure of drugs, engery, and others categories; equipment, building and physicians patient beds, physicians, nurses, supporting medical personnel (including ancillary service personnel) fully licensed and staffed beds, FTE physicians, FTE registered nurses, FTE licensed practice nurses, FTE medical residents, FTE other personnel clinical officers and nurses, physiotherapist and the like, lab technicians, admin staff, non-wage expenditure, beds FTE physicians, other labours, beds

Input list

NA

NA

OP visits for non-poor, poor, IP weight for non-poor, poor admissions, outpatient visits

NA

NA

NA

NA

NA

NA

Control variables

admissions, births, inpatient surgeries, outpatient surgeries, emergency visits, outpatient visits, medical interns

inpatient care, outpatient care

three groups of diseases treated and general outpatient visits

inpatients, inpatient and outpatient surgeries, outpatient visits

patient days, clinic or outpatient visits, surgical patients

inpatient days, outpatient visits, lab exams performed for pay, laundry and cafeteria services, and teaching

Output list

167

Renner et al. (2005)

Chen et al. (2005)

Ramanathan (2005)

Rosko & Proenca (2005)

Harrison & Ogniewski (2005) Liu & Mills (2005) Gannon (2005) Osei et al. (2005)

Bates et al. (2006)

67

68

69

70

71

75

74

73

72

Author

No

US

Ghana

Ireland

China

US

US

Oman

US

Sierra Leone

Country

DEA input oriented

DEA output oriented DEA input oriented DEA input oriented

Stochastic cost frontier, TL, truncated DEA input oriented

DEA output oriented

DEA input oriented

Estimation method DEA output oriented

306

17

44

120

252

1368

20

89

Sample size 37

FTE registered nurses, FTE licensed practical nurses, FTE other salaried personnel, beds, expenditures on materials, supplies, active physicians

Doctors/dentists, technical staff, subordinate staff, bed

doctors, nurses, fixed assess value, hospital beds, supplies value staff, beds, non-medical staff

Operating expenses, FTEs, beds

Total cost

general service cost, routine and special case cost, cumulative capital investment, ancillary service cost beds, doctors, others

Technical staff, sub-ordinate technical staff, materials and supplies, capital inputs

Input list

admissions, outpatient visits, surgical operations DRG adjusted inpatients, outpatients, day cases Maternal and child health care visits, deliveries, inpatient discharges inpatient days, emergency room outpatient visits, nonemergency room outpatient visits, surgeries, births

outpatient visits, inpatients, major surgical procedures, minor surgical procedures adjusted inpatient discharges, outpatient visits, days in long-term visits, % emergency, % outpatient surgery inpatient days, surgical procedures, outpatient visits

Antenatal and post natal care, babies delivered, nutrition/growth monitoring visits, family planning visits, under 5 immunised and pregnant women immunised, health education routine care bed days, special care bed days

Output list

NA

NA

NA

NA

NA

COTH member, teaching hospital

NA

NA

NA

Control variables

168


Prior (2006)

79

80

Staat (2006)

Yaisawarng & Burgress Jr (2006) Zere et al. (2006)

82

83

84

Rebba & Rizzi (2006)

81

78

KontodimopoulosGreece et al. (2006) Linna et al. Norway, (2006) Finland

77

Namibia

US

Germany

Italia

Spain

Finnish

Scotland

Ferrari (2006b)

76

Country

Author

No

Stochastic cost frontier, log linear, half normal DEA output oriented

DEA input oriented

DEA input oriented DEA input oriented and cost Stochastic cost frontier, CD, half normal Nonparametric output distance function DEA input oriented

Estimation method DEA output oriented, Malmquist

recurrent expenditure, beds, nursing staff

131

30

total cost, avop bed, icu score

160

physicians, other staff, beds, materials

Total cost, Variable cost, capital cost

Cost

doctors, nurses, beds

total capital charges, medical staff FTE, nurses FTE, other staff FTE, beds

Input list

physicians, nurses, other employees, hospital beds, acute care admissions (proxy for hospital demands) per diem, beds

85

29

48

98

17

Sample size 53

outpatient visits, inpatient days

cases, reciprocal length of stay, casemix for medicine, surgery and fields of specialisations basic1, complex, nonvest1

DRG weighted inpatient cases, treatment days, emergency service cases

acute inpatient days, long stay care, intensive, outpatient visits

inpatient surgery, inpatient medical, inpatient other, outpatient day cases and day patients patient admissions, outpatients, preventive medicine services DRG weighted discharges, weighted day cares, outlier days, weighted outpatient visits DRG weighted discharges, outpatient visits, bed-days, teaching, research

Output list

NA

3 access variables, urban, teaching, mental

NA

NA

NA

emerg, dead, priceind, home, operative, non-operative

NA

NA

NA

Control variables

169

Aletras et al. (2007)

Arocena & Garcia-Prado (2007)

Goncalves et al. (2007)

Goncalves et al. (2007)

Hajialiafzali et al. (2007)

Kibambe & Kocht (2007) Masiye (2007)

Smet (2007)

85

86

87

88

89

90

92

91

Author

No

Belgium

South Africa Zambia

Iran

Brazil

Italia

Costa Rica

Country

Stochastic cost frontier, TL, truncated


DEA input oriented

Nonparametric Input distance function, Malmquist Nonparametric output distance function DEA output oriented


184

30

39

53

27

3186

113

Sample size 51

active beds, medical doctors and specialists, nurses nonlabour cost, medical doctors, nurses and the like, admin and other staff Total cost

FTE medical doctors, FTE nurses, FTE other personnel, staffed beds

mortality rate, mean length of stay in hospital

Beds, beds for day hospital, physicians, nurses, teaching staff, other personnel

physicians, nurses, beds, expenditure in goods and services

physicians, other staff, beds

Input list

ambulatory care, inpatient MCH (no. of delivery), lab tests, xray and theatre operations admissions, patient days for 7 categories

% of admission relating to three chapter of IDG with the greatest mortality rate, mean value paid through the hospital admission authority outpatient visits, emergency visits, medical interventions, ratio of major surgeries to total surgeries (for complexity) total admissions

inpatient days, discharged patients, day hospital treatment, emergency room treatment

patient days, inpatient cases, surgeries, outpatient visits, average length of stay, occupancy rate, roemer index discharges, outpatient services (all adjusted for case-mix) (as good outputs)

Output list

university dummy, queuing indicator, occupancy rate, 2 dummies for regions

NA

NA

NA

NA

NA

NA

NA

Control variables

170

Friesner et al. (2008)

Kirigia et al. (2008) Lee et al. (2008)

93

94

95

Author

No

US

Angola

US

Country

DEA output oriented DEA input oriented

Estimation method DEA input oriented and cost

435

28

Sample size 1076

service complexity, hospital size (beds), amount of labour used, medical supply expenses

Doctor nurses, Drugs other, beds

beds, area of hospital, number of paid hours per hospital

Input list outpatient visits, Medicare inpatient days, Medicaid inpatient days, non-Medicare non-Medicaid inpatient days, Medicare casemix index, Medicaid casemix index, non-Medicare, Medicaid casemix index OPDANC visits, Patient admission Casemix adjusted number of discharged, number of outpatient visits, number of FTE trainees

Output list

NA

NA

NA

Control variables

Appendix B

The construction of cost weights for output aggregation for Vietnamese public hospitals

In the Vietnam Hospital Inventory dataset, apart from inpatient (IP) and outpatient (OP), there are nine output categories: surgeries, deliveries, three diagnostic tests (blood, bio-chemical and micro-organisms) and four diagnostic procedures (xray, ultra sound, endoscopy and CT MRI). All data on output categories have one common problem: they are all raw counts of cases/ tests/ procedures. The use of the unweighted sum of outputs will create bias in efficiency estimates. For instance, hospitals that treat more complicated casemix will be identified as less inefficient compared to those that treat simple cases (especially when the sample contains large general hospitals and small hospitals). Due to a large number of outputs, aggregation of some outputs is required to facilitate the analysis. Unweighted aggregation is not desirable because it would certainly introduces bias to estimated results.

B.1

Aggregation issues

Several weight-construction options are available. One can (i) use output (market) prices as weights, (ii) construct weight from outputs and their respective revenue shares, or (iii) calculate the unit cost for each output category and use those as weights. The first two options are not feasible due to data limitation. Neither revenues by output group nor actual output prices are observed. It is noted that the scheduled fees for different output categories are available but the price setting mechanism is problematic. First, the fees do not include labour costs. Second, 171

B.1. AGGREGATION ISSUES

it has not been updated regularly to keep up with either inflation or actual price changes of material costs and technology depreciation. The last option, i.e. using the unit cost as weight, is feasible because some hospital cost data are available for the construction of unit cost. The data, however, is less than ideal for the following reasons: (i) Output for each group is a raw count of cases/tests/procedures performed. A specialised blood test, such as the arterial blood gas, requires different medical skills and procedure, and thus is quite different than a regular glucose test, but they are grouped in the same category “blood test”. (ii) Not all hospitals in the final dataset have cost data. Nonetheless, all regions have several representative hospitals. This might help even out regional differences on cost. (iii) For some hospitals, some output categories are not produced in designated departments (especially for diagnostic tests and procedures). For instance, some hospitals have separate administrative units to provide blood, bio-chemical and micro-organism tests while some others have a single (combined) diagnostic test department that perform all those tests. Hence, separation of labour and capital costs devoted to each category is not straightforward. (iv) The measure of capital is “depreciation value” which is less than ideal. In principal, the capital used should be measured as the capital consumed in the current period as input to the production process. Hence, the accounting measure of depreciation of physical stock usually offers little meaningful indication of capital consumed. However, attributing capital use to any particular period is challenging even when some measures of capital stock are available. (v) Medical consumption (such as drugs, chemicals and other medical goods) are either missing or recorded under the General and Administration Department. Therefore, it is impossible to separate medical consumption by different specialised departments. (vi) Output of the “outpatient department” (Khoa Kham Benh) is consultation and scheduled outpatient visits, which are slightly different. The hospital inventory data contains separate variables for consultations and outpatient visits while the hospital costing data has only one category for outpatient department. Thus, separation of labour and capital costs for consultation and outpatient visit is problematic. 172

B.2. CALCULATION OF UNIT COST PER SERVICE GROUP

(vii) Medical departments (i.e. not including general and administration, surgeries, deliveries, diagnostic test and procedure departments) produce mostly inpatient care although some might offer outpatient services. However, these numbers are not reported separately in the hospital costing data and grouped under outpatient visits (of the “outpatient department”) in the hospital inventory data.

B.2

Calculation of unit cost per service group

As mentione before, nine outputs of procedures and tests will be aggregated into one. Therefore, the problems pointed out in points (v) and (vi) above will not affect the calculation although it should be noted that problem (i) remains a major source of bias. With respect to problem (ii), the average costs of available hospitals will be used as the indicator for the relative level of resource consumption by each output group. Hospitals in the costing sample are located in different regions of Vietnam so the average will be less affected by the extreme value of one region. However, the use of national-average unit-cost will disadvantage hospitals that produce more capital-intensive outputs (such as expensive tests and diagnostic procedures) if those hospitals face a relatively lower price of labour. Approaches to solve problem (iii) differ from labour to capital. It can be assume that labour is proportionate to output volume; that is the amount of work required to produce a blood test is the same as that for a CT scanning case. People working in those departments are generally technicians with similar skill level. Furthermore, in modern hospitals, more complicated diagnostic procedures are usually assisted by sophisticated technologies/equipments; hence the variation of labour amount spent in each procedure can be small. This assumption will under-estimate output that uses more labour and thus, hospitals producing more labour-intensive outputs. For capital, it is possible to separate some equipment used for different tests and procedures thanks to the relatively complete list of equipments per department. For instance, a hospital may have one image-based diagnostic department in charge of providing Xray, CT scan, ultra-sound and so on. Specialised equipment/machines used for those procedures are different and it is not hard to assign them into inputfor-xray or input-for-CT groups. General facilities like beds, tables, cupboards and so on are divided proportionately to the volume of respective outputs. 173


However, the difficulty is measuring attributing capital (flow) from the stock of available capital (problem (iv) above). Therefore, the concept of user cost (described in OECD, 2009) is utilised here in order to estimate the capital contribution. This concept appears to be relevant to this exercise as it captures the opportunity cost of capital. It correspond to the marginal returns generated by the asset during one period of production, i.e. equal to the value as the rental that the capital ower could achieve if he rented out the asset during one period for use in production. The user cost per unit of asset that is t years old for period of interest is given as:

f t = Pt × r + Pt

i 1+ 2

δt − Pt × i 1 + 2

(B.1)

in which ft denotes user cost; Pt is the asset price at the beginning of period; r and i are the interest rate and rate of price change of the asset, respectively, between the beginning and the end of the period; the rate of depreciation of the asset is expressed by δt . Note that the rate of price change i and the depreciation rate δt are both divided by 2 for the reason that the estimation is corrected for half year value since the asset price used is obtained at the beginning of the period. It is clear that Equation B.1 has three components: a return on capital (Pt × r), a depreciation charge (Pt 1 + 2i ) and a revaluation of the asset, i.e. holding gains or losses (Pt × i 1 + δ2t ). All together, these components capture the opportunity cost of the asset, and thus its shadow price. This formula relies on some explicit assumptions. Firstly, for every asset, there is only one set of depreciation rates, which is calculated by a straight line depreciation method using service life information. Secondly, the rate of price change coincides with the general inflation rate during the period of calculation (between 1998-2007), which was 3.5%. The average unit-cost for output group ith can be calculated by either N 1 X wn,i + fn,i wi = N n=1 Yn,i

or

(B.2)

PN wi =

wn,i + fn,i PN n=1 Yn,i

n=1

(B.3)

where wn,i , fn,i , Yn,i are the total labour wage, the sum of capital costs of all relevant assets (as calculated above) and the volume of the output ith in hospital n, 174


respectively. The former calculates the average from the cost weights of individual hospitals while the latter aggregates the costs attributed to output i across all hospitals and divides it by the sum of volumes of output i across all hospitals. The implicit assumption of Equation B.2 is that it assigns equal weights for all hospitals in the sample. In practice, hospitals produce very different amounts of each output category. If a hospital produces a large amount of output i (i.e. accounting for a large share of the total output i of all hospitals) then the average unit cost should be closer to the unit cost of that hospital. Equal weight will pool the average unit cost toward those of hospitals that produce a smaller quantity of outputs than it should be. The second formula, specified in Equation B.3, on the other hand, does not suffer from this problem. For this reason, it is used to estimate the average cost weight for output category. The estimated cost weights and standardised weights (using outpatient as the base of 1) for nine output categories are produced in Table B.11 . Table B.1: Cost weights for hospital outputs

Surgical operations

Average cost

Weight

262,450.663

102.939

Deliveries

28,902.678

11.336

Biochemical tests

2,618.585

1.027

Blood tests

1,734.021

0.680

Micro-organism tests

5,168.475

2.027

X-ray

11,549.273

4.530

CT-scan

94,555.675

37.087

Ultra sound

15,576.117

6.109

Endoscopy

62,683.160

24.586

Inpatient days

30,189.585

11.841

2,549.569

1.000

Outpatient visits

The weight estimates are meaningful. First, surgical operations are the most costly procedures, which is reasonable because they requires not only highly skilled labours but also special equipment and mecidine. Second, an inpatient day costs more or less the same as a delivery case. It is noted that inpatient days capture only the hotelling expenditure and some basic nursing services. They exclude procedures and tests required for inpatient treatment. This is the same for delivery cases that require mostly accomodation and medical services. Some deliveries are simple and inexpensive, but births by caesarian usually consume more medical resources which 1

Note that the unit cost calculated here does not include medical consumption because it is impossible to separate consumptions by different departments (refer to problem (v) above).

175


make the overall cost weight for a delivery roughly the same as an inpatient day. Third, CT MRI scans (and endoscopy procedures) are shown as very expensive diagnostic procedures compared to other tests and procedures. The reason is that the equipment used in these procedures are expensive while the medical labour cost in Vietnam (measured by salary) is much cheaper. The weight reflects the capital cost involved. This can make labour intensive output like birth deliveries look cheaper. Similarly, other tests (blood, biochemical and micro-organism) and procedures (x-ray and ultra sound) are relatively cheapter but still have weights that are more or less correlated with the cost of equipment and facility. Last, oupatient visits consist of short and simple consultations and follow-up visits which are performed mostly by GPs. Since the labour cost is relatively low in Vietnam, it is not surprising that outpatient visit costs less than many other tests and procedures.

176

Appendix C

The bootstrap results for estimated technical efficiency of Vietnamese public hospitals

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

10110

1998

0.700

0.667

0.033

0.014

0.648

0.695

10110

1999

0.696

0.663

0.033

0.016

0.641

0.691

10110

2000

0.689

0.647

0.041

0.019

0.624

0.684

10110

2005

0.654

0.597

0.056

0.033

0.561

0.650

10110

2006

0.656

0.601

0.056

0.033

0.561

0.652

10110

2007

0.669

0.606

0.063

0.041

0.560

0.666

10113

1998

0.853

0.780

0.072

0.038

0.738

0.847

10113

1999

0.715

0.669

0.046

0.021

0.646

0.711

10113

2000

0.724

0.674

0.050

0.023

0.651

0.720

10113

2005

0.755

0.687

0.067

0.037

0.653

0.750

10113

2006

0.984

0.898

0.085

0.045

0.853

0.978

10113

2007

1.000

0.903

0.097

0.052

0.856

0.993

10701

1998

0.991

0.920

0.071

0.038

0.876

0.986

10701

1999

0.781

0.723

0.059

0.035

0.683

0.778

10701

2000

0.773

0.735

0.038

0.019

0.708

0.769

10701

2005

0.856

0.806

0.050

0.023

0.779

0.850

10701

2006

0.787

0.750

0.037

0.017

0.726

0.785

10701

2007

0.781

0.744

0.037

0.017

0.718

0.779

11301

1998

0.978

0.912

0.066

0.037

0.874

0.974

11301

1999

0.817

0.765

0.052

0.029

0.725

0.812

11301

2000

0.761

0.715

0.046

0.028

0.677

0.757

11301

2005

0.786

0.746

0.039

0.017

0.723

0.780

11301

2006

0.831

0.795

0.036

0.017

0.772

0.826

177

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

11301

2007

0.987

0.931

0.056

0.025

0.901

0.983

11501

1998

0.866

0.807

0.058

0.034

0.763

0.861

11501

1999

0.836

0.782

0.055

0.031

0.741

0.832

11501

2000

0.784

0.730

0.054

0.029

0.693

0.779

11501

2005

1.000

0.902

0.098

0.058

0.847

0.995

11501

2006

1.000

0.880

0.120

0.067

0.839

0.996

11501

2007

1.000

0.797

0.203

0.154

0.751

0.994

20101

1998

0.683

0.634

0.048

0.033

0.588

0.679

20101

1999

0.609

0.560

0.049

0.034

0.515

0.604

20101

2000

0.676

0.630

0.045

0.028

0.589

0.671

20101

2005

0.584

0.543

0.041

0.024

0.514

0.581

20101

2006

0.552

0.525

0.026

0.011

0.510

0.548

20101

2007

0.512

0.480

0.032

0.018

0.455

0.508

20901

1998

0.766

0.728

0.038

0.017

0.705

0.760

20901

1999

0.676

0.639

0.037

0.015

0.620

0.670

20901

2000

0.764

0.742

0.022

0.010

0.726

0.760

20901

2005

0.669

0.625

0.044

0.021

0.603

0.664

20901

2006

0.656

0.625

0.031

0.013

0.609

0.652

20901

2007

0.611

0.572

0.039

0.019

0.550

0.608

21101

1998

0.881

0.819

0.062

0.037

0.776

0.876

21101

1999

0.741

0.705

0.035

0.014

0.687

0.736

21101

2000

0.770

0.743

0.028

0.011

0.727

0.766

21101

2005

0.834

0.785

0.049

0.020

0.764

0.829

21101

2006

0.880

0.837

0.043

0.018

0.814

0.875

21101

2007

0.818

0.772

0.046

0.020

0.747

0.814

21501

1998

1.000

0.796

0.204

0.166

0.735

0.995

21501

1999

0.938

0.876

0.062

0.035

0.837

0.933

21501

2000

0.936

0.890

0.046

0.021

0.863

0.931

21501

2005

1.000

0.864

0.136

0.081

0.817

0.993

21501

2006

1.000

0.840

0.160

0.103

0.800

0.993

21501

2007

1.000

0.901

0.099

0.050

0.858

0.992

21502

1998

1.000

0.850

0.150

0.095

0.802

0.996

21502

1999

0.962

0.907

0.055

0.031

0.864

0.954

21502

2000

0.887

0.839

0.048

0.025

0.805

0.882

21502

2005

0.999

0.918

0.082

0.040

0.880

0.992

21502

2006

0.988

0.925

0.063

0.033

0.884

0.983

178

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

21502

2007

0.962

0.907

0.054

0.030

0.861

0.954

21504

1998

0.798

0.737

0.061

0.040

0.686

0.794

21504

1999

0.805

0.759

0.045

0.029

0.717

0.800

21504

2000

0.857

0.808

0.050

0.022

0.784

0.853

21504

2005

1.000

0.832

0.168

0.103

0.801

0.992

21504

2006

0.955

0.875

0.080

0.043

0.832

0.951

21504

2007

0.748

0.689

0.059

0.030

0.660

0.744

21505

1998

1.000

0.762

0.238

0.231

0.674

0.994

21505

1999

1.000

0.873

0.127

0.120

0.762

0.994

21505

2000

0.711

0.653

0.059

0.035

0.611

0.708

21505

2005

0.881

0.848

0.033

0.014

0.828

0.877

21505

2006

0.806

0.760

0.046

0.020

0.737

0.803

21505

2007

0.766

0.728

0.038

0.017

0.707

0.762

21506

1998

0.707

0.650

0.056

0.030

0.617

0.701

21506

1999

0.660

0.620

0.040

0.020

0.593

0.655

21506

2000

0.496

0.469

0.028

0.014

0.450

0.493

21506

2005

0.658

0.628

0.030

0.012

0.612

0.653

21506

2006

0.551

0.512

0.040

0.024

0.481

0.549

21506

2007

0.635

0.585

0.050

0.025

0.560

0.632

21508

1998

0.908

0.831

0.077

0.041

0.788

0.901

21508

1999

0.910

0.828

0.082

0.045

0.784

0.905

21508

2000

0.941

0.870

0.070

0.041

0.817

0.935

21508

2005

1.000

0.787

0.213

0.170

0.728

0.992

21508

2006

1.000

0.852

0.148

0.102

0.787

0.995

21508

2007

0.909

0.843

0.066

0.046

0.780

0.905

21509

1998

0.867

0.766

0.100

0.073

0.691

0.862

21509

1999

1.000

0.836

0.164

0.114

0.777

0.995

21509

2000

1.000

0.780

0.220

0.200

0.711

0.995

21509

2005

0.758

0.716

0.042

0.018

0.698

0.754

21509

2006

0.901

0.843

0.057

0.025

0.817

0.897

21509

2007

0.599

0.562

0.037

0.017

0.542

0.595

21510

1998

1.000

0.865

0.135

0.079

0.824

0.995

21510

1999

0.734

0.663

0.071

0.042

0.627

0.730

21510

2000

0.779

0.715

0.064

0.038

0.675

0.774

21510

2005

1.000

0.925

0.075

0.031

0.902

0.995

21510

2006

0.996

0.933

0.062

0.033

0.889

0.990

179

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

21510

2007

0.905

0.860

0.046

0.023

0.826

0.901

21511

1998

0.684

0.624

0.060

0.033

0.588

0.680

21511

1999

0.654

0.609

0.045

0.025

0.574

0.650

21511

2000

0.521

0.478

0.043

0.024

0.451

0.517

21511

2005

0.707

0.655

0.052

0.022

0.636

0.702

21511

2006

0.734

0.698

0.036

0.014

0.681

0.730

21511

2007

0.625

0.593

0.032

0.012

0.578

0.620

21512

1998

1.000

0.765

0.235

0.227

0.680

0.995

21512

1999

0.853

0.747

0.106

0.091

0.665

0.849

21512

2000

1.000

0.755

0.245

0.241

0.673

0.994

21512

2005

1.000

0.771

0.229

0.223

0.679

0.995

21512

2006

1.000

0.775

0.225

0.220

0.679

0.994

21512

2007

0.667

0.623

0.044

0.023

0.592

0.663

21513

1998

1.000

0.832

0.168

0.113

0.776

0.994

21513

1999

0.997

0.891

0.106

0.077

0.809

0.991

21513

2000

1.000

0.758

0.242

0.249

0.668

0.995

21513

2005

0.482

0.455

0.027

0.016

0.433

0.479

21513

2006

0.465

0.428

0.037

0.021

0.404

0.463

21513

2007

0.457

0.429

0.027

0.015

0.409

0.453

21514

1998

1.000

0.771

0.229

0.215

0.689

0.995

21514

1999

0.952

0.847

0.105

0.089

0.754

0.947

21514

2000

1.000

0.768

0.232

0.217

0.688

0.995

21514

2005

0.596

0.531

0.065

0.051

0.482

0.592

21514

2006

0.546

0.516

0.030

0.014

0.498

0.543

21514

2007

0.618

0.571

0.047

0.023

0.549

0.616

21701

1998

0.679

0.649

0.030

0.014

0.629

0.677

21701

1999

0.663

0.633

0.030

0.013

0.615

0.659

21701

2000

0.630

0.608

0.022

0.009

0.595

0.627

21701

2005

0.931

0.850

0.080

0.045

0.804

0.924

21701

2006

1.000

0.827

0.173

0.112

0.789

0.993

21701

2007

1.000

0.865

0.135

0.074

0.823

0.995

21707

1998

0.848

0.754

0.094

0.068

0.690

0.844

21707

1999

0.904

0.814

0.090

0.053

0.768

0.900

21707

2000

0.516

0.466

0.050

0.029

0.440

0.512

21707

2005

0.659

0.613

0.046

0.022

0.590

0.655

21707

2006

0.723

0.685

0.038

0.019

0.660

0.718

180

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

21707

2007

0.774

0.728

0.045

0.025

0.697

0.769

22101

1998

1.000

0.914

0.086

0.039

0.882

0.992

22101

1999

0.967

0.910

0.057

0.029

0.873

0.962

22101

2000

0.984

0.943

0.040

0.016

0.920

0.977

22101

2005

0.986

0.915

0.071

0.042

0.866

0.981

22101

2006

1.000

0.838

0.162

0.109

0.790

0.995

22101

2007

1.000

0.890

0.110

0.074

0.815

0.995

22501

1998

0.962

0.878

0.084

0.052

0.820

0.953

22501

1999

0.739

0.686

0.053

0.035

0.641

0.735

22501

2000

0.715

0.670

0.045

0.027

0.631

0.711

22501

2005

0.991

0.929

0.062

0.035

0.885

0.986

22501

2006

0.832

0.774

0.058

0.033

0.733

0.828

22501

2007

1.000

0.884

0.116

0.069

0.830

0.993

22502

1998

1.000

0.941

0.059

0.024

0.914

0.994

22502

1999

0.804

0.775

0.029

0.014

0.754

0.799

22502

2000

0.818

0.790

0.028

0.013

0.770

0.814

22502

2005

1.000

0.871

0.129

0.073

0.829

0.995

22502

2006

0.983

0.907

0.076

0.042

0.862

0.978

22502

2007

0.842

0.787

0.055

0.030

0.743

0.836

22505

1998

0.811

0.748

0.063

0.032

0.715

0.807

22505

1999

0.648

0.608

0.040

0.020

0.581

0.644

22505

2000

0.726

0.673

0.053

0.029

0.636

0.722

22505

2005

0.822

0.766

0.056

0.025

0.738

0.818

22505

2006

0.835

0.796

0.039

0.015

0.779

0.829

22505

2007

0.762

0.726

0.036

0.016

0.706

0.757

30502

1998

0.939

0.863

0.076

0.050

0.811

0.934

30502

1999

0.858

0.803

0.055

0.031

0.764

0.853

30502

2000

0.805

0.771

0.034

0.015

0.750

0.802

30502

2005

0.936

0.868

0.068

0.036

0.826

0.931

30502

2006

0.953

0.884

0.068

0.034

0.848

0.947

30502

2007

0.912

0.860

0.052

0.027

0.824

0.907

40101

1998

1.000

0.853

0.147

0.093

0.807

0.996

40101

1999

0.828

0.747

0.081

0.058

0.694

0.824

40101

2000

0.735

0.686

0.049

0.029

0.650

0.731

40101

2005

0.883

0.831

0.051

0.023

0.806

0.877

40101

2006

0.893

0.852

0.042

0.019

0.826

0.888

181

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

40101

2007

0.999

0.942

0.057

0.027

0.905

0.993

40305

1998

0.924

0.867

0.057

0.026

0.836

0.920

40305

1999

0.815

0.777

0.038

0.017

0.753

0.810

40305

2000

0.910

0.877

0.033

0.016

0.853

0.904

40305

2005

0.849

0.779

0.070

0.042

0.734

0.842

40305

2006

0.823

0.770

0.052

0.025

0.740

0.817

40305

2007

0.759

0.712

0.047

0.024

0.677

0.754

40502

1998

0.965

0.920

0.045

0.020

0.895

0.959

40502

1999

0.872

0.827

0.045

0.019

0.803

0.868

40502

2000

0.833

0.794

0.038

0.016

0.775

0.829

40502

2005

0.683

0.640

0.043

0.023

0.611

0.678

40502

2006

0.690

0.640

0.049

0.024

0.610

0.685

40502

2007

0.705

0.655

0.050

0.025

0.626

0.701

41101

1998

1.000

0.810

0.190

0.141

0.750

0.993

41101

1999

0.873

0.779

0.094

0.073

0.709

0.867

41101

2000

1.000

0.756

0.244

0.229

0.684

0.994

41101

2005

0.864

0.805

0.058

0.035

0.758

0.860

41101

2006

0.872

0.805

0.067

0.042

0.753

0.868

41101

2007

0.832

0.772

0.060

0.041

0.718

0.827

41103

1998

0.910

0.809

0.101

0.074

0.735

0.905

41103

1999

0.673

0.614

0.059

0.035

0.577

0.669

41103

2000

0.603

0.554

0.050

0.025

0.530

0.600

41103

2005

0.896

0.850

0.046

0.021

0.820

0.890

41103

2006

0.855

0.802

0.054

0.026

0.770

0.852

41103

2007

0.705

0.654

0.051

0.023

0.632

0.701

41104

1998

1.000

0.866

0.134

0.079

0.822

0.994

41104

1999

0.893

0.803

0.090

0.052

0.756

0.888

41104

2000

1.000

0.833

0.167

0.109

0.785

0.992

41104

2005

0.960

0.867

0.093

0.060

0.805

0.956

41104

2006

0.881

0.812

0.069

0.034

0.775

0.876

41104

2007

0.966

0.883

0.083

0.044

0.838

0.961

41105

1998

1.000

0.811

0.189

0.141

0.760

0.994

41105

1999

0.839

0.752

0.087

0.062

0.693

0.834

41105

2000

0.677

0.622

0.055

0.037

0.581

0.673

41105

2005

0.849

0.808

0.041

0.020

0.779

0.845

41105

2006

0.590

0.562

0.029

0.013

0.546

0.588

182

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

41105

2007

0.648

0.618

0.030

0.014

0.599

0.645

41106

1998

1.000

0.761

0.239

0.233

0.677

0.995

41106

1999

0.698

0.616

0.083

0.059

0.560

0.693

41106

2000

0.685

0.628

0.058

0.028

0.601

0.681

41106

2005

0.959

0.877

0.082

0.043

0.835

0.955

41106

2006

1.000

0.844

0.156

0.111

0.775

0.994

41106

2007

1.000

0.859

0.141

0.092

0.802

0.993

41107

1998

1.000

0.758

0.242

0.228

0.683

0.993

41107

1999

1.000

0.759

0.241

0.236

0.673

0.994

41107

2000

0.856

0.785

0.072

0.053

0.715

0.850

41107

2005

0.702

0.639

0.062

0.039

0.597

0.698

41107

2006

1.000

0.872

0.128

0.077

0.816

0.993

41107

2007

0.606

0.562

0.044

0.023

0.537

0.603

41108

1998

1.000

0.854

0.146

0.096

0.798

0.994

41108

1999

0.852

0.769

0.083

0.059

0.709

0.847

41108

2000

1.000

0.861

0.139

0.089

0.799

0.994

41108

2005

0.706

0.658

0.047

0.026

0.625

0.702

41108

2006

0.754

0.694

0.060

0.032

0.660

0.749

41108

2007

0.769

0.722

0.047

0.024

0.692

0.765

41109

1998

0.718

0.642

0.076

0.060

0.578

0.714

41109

1999

0.541

0.496

0.045

0.031

0.457

0.538

41109

2000

0.937

0.859

0.078

0.048

0.804

0.932

41109

2005

0.514

0.490

0.024

0.010

0.476

0.511

41109

2006

0.509

0.483

0.026

0.014

0.462

0.506

41109

2007

0.716

0.671

0.045

0.020

0.649

0.711

41110

1998

1.000

0.777

0.223

0.192

0.714

0.994

41110

1999

0.694

0.626

0.068

0.039

0.589

0.690

41110

2000

0.504

0.456

0.048

0.028

0.432

0.502

41110

2005

0.739

0.677

0.062

0.035

0.637

0.735

41110

2006

0.722

0.690

0.032

0.015

0.668

0.718

41110

2007

0.693

0.659

0.034

0.016

0.636

0.689

50103

1998

0.822

0.764

0.058

0.036

0.713

0.818

50103

1999

0.879

0.826

0.053

0.026

0.792

0.874

50103

2000

0.911

0.854

0.057

0.027

0.822

0.904

50103

2005

1.000

0.871

0.129

0.073

0.830

0.996

50103

2006

1.000

0.870

0.130

0.078

0.809

0.995

183

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

50103

2007

1.000

0.842

0.158

0.108

0.784

0.993

50301

1998

0.817

0.773

0.043

0.025

0.736

0.813

50301

1999

0.708

0.676

0.032

0.018

0.648

0.706

50301

2000

0.849

0.823

0.026

0.014

0.798

0.845

50301

2005

1.000

0.898

0.102

0.054

0.856

0.994

50301

2006

1.000

0.847

0.153

0.103

0.787

0.993

50301

2007

1.000

0.842

0.158

0.114

0.771

0.994

50303

1998

0.846

0.791

0.055

0.028

0.760

0.842

50303

1999

0.873

0.821

0.053

0.024

0.794

0.867

50303

2000

0.767

0.728

0.039

0.017

0.706

0.762

50303

2005

0.886

0.821

0.065

0.040

0.778

0.882

50303

2006

0.983

0.916

0.067

0.037

0.875

0.979

50303

2007

1.000

0.905

0.095

0.048

0.867

0.993

50501

1998

0.868

0.823

0.045

0.023

0.791

0.863

50501

1999

0.760

0.718

0.042

0.020

0.690

0.755

50501

2000

0.807

0.782

0.025

0.010

0.766

0.801

50501

2005

1.000

0.838

0.162

0.097

0.803

0.992

50501

2006

1.000

0.867

0.133

0.080

0.819

0.994

50501

2007

1.000

0.876

0.124

0.069

0.834

0.992

50704

1998

0.899

0.830

0.069

0.036

0.790

0.893

50704

1999

0.832

0.786

0.046

0.024

0.754

0.828

50704

2000

0.922

0.886

0.036

0.014

0.865

0.916

50704

2005

0.945

0.884

0.061

0.027

0.853

0.939

50704

2006

0.944

0.893

0.051

0.022

0.861

0.938

50704

2007

0.927

0.853

0.074

0.042

0.807

0.922

50902

1998

0.830

0.793

0.037

0.016

0.773

0.825

50902

1999

0.828

0.790

0.037

0.018

0.764

0.822

50902

2000

0.766

0.745

0.021

0.008

0.729

0.760

50902

2005

0.903

0.817

0.085

0.053

0.766

0.898

50902

2006

0.835

0.780

0.054

0.026

0.749

0.830

50902

2007

0.751

0.716

0.035

0.014

0.696

0.747

51101

1998

0.824

0.772

0.051

0.025

0.735

0.818

51101

1999

0.750

0.713

0.037

0.016

0.689

0.746

51101

2000

0.723

0.682

0.042

0.019

0.658

0.719

51101

2005

1.000

0.870

0.130

0.073

0.828

0.992

51101

2006

0.980

0.909

0.071

0.046

0.854

0.975

184

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

51101

2007

0.994

0.942

0.052

0.026

0.902

0.987

60201

1998

0.647

0.591

0.056

0.037

0.548

0.644

60201

1999

0.450

0.414

0.036

0.025

0.384

0.447

60201

2000

0.503

0.470

0.033

0.020

0.443

0.500

60201

2005

0.839

0.787

0.052

0.028

0.755

0.835

60201

2006

0.765

0.709

0.057

0.026

0.683

0.760

60201

2007

0.532

0.503

0.030

0.015

0.484

0.530

60305

1998

0.965

0.905

0.060

0.028

0.870

0.958

60305

1999

0.724

0.679

0.045

0.020

0.656

0.719

60305

2000

0.656

0.610

0.046

0.025

0.581

0.652

60305

2005

0.938

0.871

0.066

0.038

0.823

0.934

60305

2006

0.806

0.739

0.067

0.038

0.698

0.802

60305

2007

0.846

0.791

0.055

0.031

0.746

0.842

70114

1998

0.866

0.803

0.063

0.038

0.752

0.861

70114

1999

0.877

0.821

0.056

0.037

0.769

0.872

70114

2000

0.907

0.854

0.053

0.034

0.800

0.903

70114

2005

0.762

0.718

0.044

0.024

0.685

0.757

70114

2006

0.725

0.675

0.049

0.026

0.644

0.722

70114

2007

0.731

0.676

0.055

0.034

0.632

0.726

70115

1998

0.746

0.699

0.046

0.024

0.668

0.742

70115

1999

0.685

0.642

0.043

0.020

0.621

0.680

70115

2000

0.768

0.724

0.044

0.019

0.700

0.762

70115

2005

0.654

0.611

0.042

0.023

0.579

0.650

70115

2006

0.712

0.656

0.056

0.034

0.613

0.708

70115

2007

0.658

0.609

0.050

0.026

0.578

0.655

70116

1998

0.851

0.812

0.039

0.016

0.792

0.846

70116

1999

0.812

0.756

0.056

0.025

0.730

0.806

70116

2000

0.919

0.864

0.056

0.028

0.831

0.913

70116

2005

0.920

0.855

0.065

0.035

0.814

0.915

70116

2006

0.987

0.925

0.062

0.039

0.876

0.981

70116

2007

0.993

0.907

0.086

0.062

0.836

0.986

70117

1998

1.000

0.805

0.195

0.139

0.757

0.994

70117

1999

1.000

0.887

0.113

0.085

0.799

0.993

70117

2000

1.000

0.830

0.170

0.115

0.776

0.994

70117

2005

1.000

0.915

0.085

0.064

0.825

0.994

70117

2006

1.000

0.875

0.125

0.092

0.791

0.993

185

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

70117

2007

1.000

0.849

0.151

0.112

0.770

0.995

70301

1998

0.784

0.748

0.035

0.015

0.728

0.780

70301

1999

0.870

0.837

0.033

0.016

0.814

0.866

70301

2000

0.845

0.816

0.030

0.013

0.797

0.841

70301

2005

0.872

0.807

0.065

0.041

0.756

0.867

70301

2006

0.942

0.876

0.066

0.041

0.821

0.937

70301

2007

0.843

0.789

0.054

0.027

0.755

0.838

70501

1998

0.945

0.858

0.087

0.056

0.792

0.940

70501

1999

0.622

0.549

0.072

0.046

0.510

0.619

70501

2000

0.535

0.479

0.055

0.029

0.454

0.530

70501

2005

0.948

0.879

0.069

0.035

0.839

0.943

70501

2006

0.886

0.823

0.064

0.035

0.782

0.880

70501

2007

0.842

0.799

0.043

0.018

0.778

0.837

70502

1998

0.823

0.764

0.059

0.028

0.732

0.819

70502

1999

0.703

0.655

0.048

0.027

0.621

0.699

70502

2000

0.642

0.609

0.033

0.016

0.588

0.637

70502

2005

0.964

0.900

0.064

0.040

0.851

0.960

70502

2006

0.895

0.846

0.049

0.019

0.824

0.889

70502

2007

0.858

0.819

0.039

0.015

0.799

0.854

70902

1998

0.919

0.852

0.067

0.032

0.819

0.914

70902

1999

0.841

0.797

0.044

0.021

0.768

0.836

70902

2000

1.000

0.895

0.105

0.052

0.864

0.994

70902

2005

0.998

0.938

0.060

0.027

0.909

0.993

70902

2006

0.905

0.846

0.060

0.030

0.812

0.902

70902

2007

1.000

0.824

0.176

0.128

0.769

0.995

71102

1998

0.949

0.874

0.075

0.047

0.822

0.944

71102

1999

0.954

0.882

0.072

0.052

0.816

0.950

71102

2000

0.856

0.786

0.070

0.046

0.733

0.851

71102

2005

0.960

0.894

0.065

0.035

0.851

0.953

71102

2006

0.890

0.834

0.056

0.027

0.801

0.884

71102

2007

0.995

0.939

0.057

0.030

0.898

0.991

80102

1998

0.901

0.821

0.080

0.057

0.754

0.896

80102

1999

1.000

0.748

0.252

0.231

0.684

0.993

80102

2000

0.867

0.802

0.065

0.045

0.749

0.862

80102

2005

0.871

0.792

0.079

0.044

0.751

0.867

80102

2006

0.778

0.742

0.037

0.017

0.717

0.774

186

Hosp ID

Year

Original

Bias-corrected

Bias

Std.dev

L-CI95%

U-CI95%

80102

2007

0.732

0.688

0.044

0.023

0.655

0.728

80303

1998

0.933

0.895

0.038

0.017

0.872

0.927

80303

1999

0.790

0.753

0.037

0.017

0.732

0.784

80303

2000

0.776

0.745

0.031

0.014

0.726

0.772

80303

2005

0.974

0.915

0.059

0.026

0.884

0.968

80303

2006

0.886

0.845

0.040

0.016

0.826

0.881

80303

2007

0.780

0.749

0.031

0.013

0.733

0.775

80902

1998

1.000

0.820

0.180

0.134

0.756

0.993

80902

1999

1.000

0.863

0.137

0.112

0.773

0.993

80902

2000

1.000

0.830

0.170

0.121

0.771

0.995

80902

2005

1.000

0.815

0.185

0.144

0.742

0.995

80902

2006

1.000

0.814

0.186

0.139

0.753

0.995

80902

2007

1.000

0.833

0.167

0.129

0.757

0.995

81701

1998

1.000

0.929

0.071

0.031

0.900

0.993

81701

1999

0.768

0.732

0.035

0.018

0.706

0.765

81701

2000

0.750

0.726

0.024

0.010

0.708

0.746

81701

2005

0.771

0.726

0.045

0.022

0.701

0.767

81701

2006

0.726

0.688

0.038

0.014

0.670

0.718

81701

2007

0.722

0.676

0.045

0.023

0.650

0.717

82302

1998

1.000

0.923

0.077

0.036

0.882

0.994

82302

1999

0.967

0.902

0.065

0.033

0.864

0.961

82302

2000

0.951

0.889

0.061

0.029

0.855

0.946

82302

2005

0.928

0.866

0.062

0.030

0.832

0.922

82302

2006

0.928

0.868

0.060

0.026

0.839

0.919

82302

2007

0.934

0.875

0.059

0.028

0.841

0.928

187

Appendix D

Review of provider payment methods (PPMs)

D.1

Classification of provider payment methods

Provider payment method (PPM) is one of two major elements of health care financing, together with fund pooling mechanism. A payment method usually involves the combination of two parameters, price and quantity, which shape the incentives to providers - the supply side of health care services. Funding pooling mechanisms, on the other hand, influence consumers’ behaviour - the demand side. There are various ways to classify PPMs. For instance, input-based payment methods include line-item budget, fee-for-service and historical budget while capitation, case-based and per-diem are representatives of the output-based payment approach. Case-based reimbursement, per-diem and fee-for-service often let money follow patients in their choice of providers while global budget solutions (such as historical budget and capitation) tend to make patients follow the money. Providers will bear all the risk of marginal cost over-run in the case of “hard” line-item and global budgets, but bear no risk at all under soft budget or when payments are made as fee-for-service. The risk is shared between payers and providers when the per-diem method is employed. In the case of prospective payment approach, payments are made or committed to providers before services are rendered. Payments made after the provision of services are called retrospective. The latter can involve pre-determined service prices/rates (e.g. per-diem) or actual cost (e.g. fee-for-service) while for the former, the prices/rates of services are set prospectively by default. According to PWC (2008), while payment systems have historically been retrospective, they are becoming increasingly prospective so as to allow providers to plan their services and 188

D.1. CLASSIFICATION OF PROVIDER PAYMENT METHODS

processes around the payment they expect to receive and for payers to better control their budgets and outlays. Langenbrunner and Liu’s classification through the resource allocation and purchasing arrangement (RAP) perspective looks first at whether a PPM involves a “middle man”, and if so how the financial risk is shared among them. For instance, direct payment from patients to providers without reimbursement implies retrospective methods such as per-diem and fee-for-service and so risk is shared between patients (also buyers) and providers. The direct payment from patients to providers with full or partial reimbursement of expenses incurred through the RAP mechanism involves three players but is still retrospective. The only difference is that some financial risk is now shifted to payers (the “middle man”). However, if the payment is made to providers through the RAP mechanism, with the patient bearing only a limited co-payment or informal charge, the PPM may be either retrospective or prospective (Langenbrunner & Liu, 2004). The classification approach taken in this paper will be based on price and quantity decisions: whether the rate and/or volume of services are set prospectively or retrospectively, which in turn determines if the total budget is fixed or variable. This allows us to investigate the source of productivity and efficiency gain/reduction (if any). Under fixed budget method, providers receive a “lump sum” and thus, producing extra units of outputs will not lead to increased revenue. That is, the marginal benefit of providing more services is zero. As a result, providers have no other choice but to reduce marginal costs. Therefore, the source of efficiency improvement (if any) comes from input conservation. In contrast, providers under the variable budget model receive incremental income for production of extra outputs. This is a powerful motivation for providers to increase production until marginal revenue equals marginal cost. Here, marginal revenue is determined by the “price” of services paid to providers. Generous service prices will induce them to increase service volumes. Low fee schedules tend to force providers to conserve inputs. Productivity and efficiency improvement therefore can be the results of either input conservation or output expansion. This approach also allows discussion on the compatibility of different methods in mixed modalities under different funding sources in order to achieve the multiple (and often conflicting) objectives of efficiency, equity and quality of care. Figure D.1 illustrates the classification of different PPMs.

189


Figure D.1: Classification of provider payment methods

FLOATING BUDGET

FIXED BUDGET

Price

Pre-determined

(rate, fee, unit cost)

Quantity (service units

Actual

Pre-determined

Actual

Actual

Line-item

Case-based

Fee for service

Historical cost

Per-diem

Provider Payment Methods

Scheduled fee

Capitation

Pay for performance

D.1.1

Fixed budget

This group of modalities requires the provider budget to be prospectively fixed, determined by pre-set quantities and prices (or service rates). The ability to amend the budget is an exception rather than a rule. The pre-set prices can be “per patient” fees, or schedule fees for procedures and services provided. Quantity can be understood as the expected volume of services, estimated by the population in a particular geographical area, or number of participants in an insurance scheme. Quantity can also be understood as the predicted amount of expenditure to provide a targeted volume of service. This total cost can be broken down into sub-categories such as labour, equipment and facilities, drugs and medical goods, and administrative expenditure. Two major fixed budget modalities reported in the literature are line-item and global budget. They share a major common goal of cost containment. Administration costs involved in managing these modalities are usually lower than other methods because individual case reporting to the payer is not required. Efficiency, 190


either allocative or technical, might be achieved, depending on the modality’s design although it is argued that this objective is unlikely to be realised. In practice, cost cutting usually comes in the form of under-service, i.e. decreased volume of services, or under-treatment, i.e. quality deterioration, rather than productivity and efficiency enhancement.

D.1.1.1

Line item budget

The line-item budget method has been historically used in centralised health systems where governments control all aspects of financing and provision of services, and is still in use in many developing countries. This method intends to control resource allocation and spending. It is considered an effective method to limit the consequences of weak local management (Barnum et al., 1995), a characteristic of developing countries. It was common in the Eastern bloc, former Soviet republics, and some African countries. It is found today in government-run systems in all regions of the world (such as Bahrain, Bangladesh, Mozambique, and Saudi Arabia) (Langenbrunner et al., 2009). Line-item budgeting is an input-based prospective payment method. The budget is based on norms or historical patterns (Barnum et al., 1995). Basic norms are major health service inputs, such as unit-cost or per-hospital-bed or operating expenses of the previous year’s budget. Number of beds in turn define the number and type of staff required, drug and medical supplies, equipment, furniture and maintenance. The budget is set prospectively and switching funds across line items is usually prohibited, unless approval is received from the central authorities (Langenbrunner et al., 2009). This method is usually criticised for disincentivising productivity and efficient use of resources. Firstly, payment is not linked to output and results and therefore the providers have limited accountability for performance. The line item budget is usually accompanied by medical staff compensated on a salary basis. The relatively weak tie between individual performance and compensation leads to low productivity (Barnum et al., 1995). Furthermore, patients under this method “follow the money” and thus have limited choice as to quality and health service mix. The pressure to satisfy consumers’ demand tends to be insufficient to induce service improvement. Secondly, there is a weak incentive to produce services at minimum cost. The existence of unspent funds at the end of the budget period is often interpreted as 191


an indicator of excessive allocation, leading to a reduction in the following period. Providers thus have incentive to spend their funds rapidly without regard for efficiency in order to ensure next year’s allocation (Barnum et al., 1995; Langenbrunner et al., 2009). Lastly, the line budget is rigid in its use of resources and discourages the use of the least costly combination of inputs for providers to produce services (Barnum et al., 1995). It is argued that under this method, technical and allocative efficiency of health services can be theoretically optimised by manipulating the government budget lines over time to increase the delivery of more cost effective health interventions and decrease the delivery of less cost effective health interventions. However, this assumes governments understand the long-term dynamics of disease patterns and choose the right combination to achieve these outputs (Langenbrunner & Liu, 2004). In practice, line-item budget tends to lock inputs to an inefficient input mix (Barnum et al., 1995), which can only be adjusted periodically with significant delay. On balance, the productivity and efficiency loss under this method tend to outweigh the benefit brought by effective control of resource allocation, spending, and limiting the consequences of weak local management. Therefore, this method has been gradually abandoned in more developed countries while retained in developing countries where budget control is the first priority. Figure D.2: Efficiency incentives by line-item budget

Increased quality Total cost containment Longer length of stay Complex casemix Increased volume Service intensity per case Efficient input mix Unit cost containment

Note: Bars to the left: tend to reduce; bars to the right: tend to increase.

192


D.1.1.2

Global budget

A global budget is an overall spending limit or target that constrains the price and quantity of the service provided (Dredge, 2004). The budget, usually negotiated and agreed by the funding agency for a defined period of time (such as a fiscal year), acts as an overall spending target or limit that constrains the price and sets the quality of the services to be provided. A formal contract can be used to set out detailed expectations of provider performance (including quality standard, efficiency and health status indicators). The global budget is effectively a one-line item budget for providers but allows more discretion. Similar to line-item budget, the main purpose of a global budget scheme is to control the aggregate expenditure on a particular health care program, service or institution. It is in contrast to “open-ended” reimbursement systems such as those that utilise cost reimbursement (Feldman & Lobo, 1997) because the budget is determined ex ante through pre-set price (or fee) and a defined volume of services (either by service units or number of patients) (refer to Figure D.1). Most global budget schemes share two features. The first is a fund holder1 that determines the budget for each hospital during a stated period of time. Because it determines the budget for every provider in the system, the fund holder controls the total amount spent on hospital services as well as the allocation of funds across hospitals and regions of the country. The second feature is that the hospital has considerable flexibility as to how its budget is spent within each period. Because hospitals exercise this control, the central authority can absolve itself from the responsibility of “micro-managing” the production of health care services. The job is turned over to professionals who are better qualified to determine the optimal quality and quantity of care to patients (Feldman & Lobo, 1997). It is argued that global budgets can encourage the development of changes to service delivery patterns, and the inclusion of appropriate incentives can improve efficiency, good clinical practice, and quality. Similar to line-item budgeting, because payment is both set and made prospectively, the potential gains of this method are cost containment, funding certainty, cheaper administration, and improved coordination in service planning (Dredge, 2004). It provides a stronger incentive for efficiency through offering flexibility to managers in allocating funds across expenditure categories. Barnum et al. (1995) argued that theoretically a global budget 1

The fund holder can be a central authority/organisation or a private entity. There can be more than one fund holder in the market.

193


scheme can be designed to achieve both short and long run efficiencies. The long run effect (over many budget periods) can be achieved if budgets for each period are adjusted to reflect the service load (based explicitly or implicitly on units of services), and number of cases or capitation (population served). The short-run effect can be realised when managerial flexibility allows economic responsibility to be decentralised to the level where providers and patients meet. The fixed budget coupled with flexibility in input choice provides a mechanism to improve efficiency. In practice, both short and long-term efficiency effects can be compromised due to deterioration of quality of care and excessive budget flexibility. If quality standard, and requirements for regular collection and assessment of information on quality, is not explicitly built in the payment scheme, this method will provide little incentive for quality. Hospitals might pursue profitable behaviours such as under-treatment of admitted patients, reduction in the number of admissions or they may admit the cheapest case-mix possible. The degree of cap flexibility (i.e. the ability to amend the budget over the set period) also adversely affects the goal of efficiency and cost containment. Under “hard” budgets, it is difficult (sometimes impossible) to revise the budget, thus the financial risk is transferred to the provider which acts as a powerful efficiency incentive. However, it can affect quality of services negatively if the budget allocated is inadequate. “Soft” budgets, on the other hand, impel purchasers to assume the overruns. This partly undermines the purpose of global budgets: enforcing provider’s accountability and cost control effort. The types of global budget vary significantly with the basis upon which the budget is made, the budget flexibility, and types and number of providers (Langenbrunner et al., 2009). It can come close to being a line-item budget if the budget is based on spending categories, and can go as far as retrospective payment when “soft” budgeting is allowed. In practice, the budget is usually based on (i) historical activities with demand growth adjustment, (ii) inputs, such as number of actual beds (bed norm) or number of staff, (iii) a target population, i.e. capitation, or (iv) volume of services and type of cases. Each has various efficiency implications.

Historical cost

Historical cost is the global budget modality set on the basis of the previous period’s budget, adjusted for demonstrated and predicted changes in demand of health services. Unwanted changes in quantity of production can increase or decrease the budget. The effects of inflation and possible new investment in equipment and 194


personnel is usually added. Strictly speaking, historical cost does not fit in either of the categories “predetermined” or “actual” quality and price (as described in Figure D.1) because the budget is based on the previous period. However, some determinants of the budget relates to demand and historical unit-costs. As such, it still fits as a prospective payment modality. The global budget based on historical cost has the virtue of simplicity and stability over an extended period of time while retaining the main features of a global budget modality: efficiency can be gained as providers have considerable discretion over the use of the funds in the budget allocated to them. However, it has been argued by (Barnum et al., 1995; Aas, 1995) that this modality does not always lead to cost containment and long-term efficiency. Firstly, allocation budgets on the basis of historical practice creates an institutional inertia that tends to lock-in existing patterns of resource use. Secondly, there is a risk that becoming more cost efficient (i.e. saving resources) will lead to a lower budget for the coming years, thus discouraging providers from striving for more efficient ways to deliver services. Furthermore, it does not explicitly take into account equity of service provision. Main centres of population may have a disproportionate centralisation of services, thus receiving a large amount of financial resources while rural and remote areas do not have sufficient budget to maintain health care services (Dredge, 2004). The historical cost method retains a common and disadvantageous feature of a global budget: the risk of under-serving and under-treatment to keep within budget. The most profitable behaviour includes either reduction in the number of admissions, admission of the cheapest case-mix possible or downgrading service quality (Aas, 1995).

Capitation

Under the capitation formula, the providers receive a pre-set “periodic fee” for each enrollee to cover a pre-determine package of health care services. It can be applied to a specific geographic area (called geographic capitation) in which money is allocated to a group of providers to take care of health services for the population in the area covered. Alternatively, the patient list capitation method applies to a list of enrollees and money is allocated to a provider to take care of all individuals enrolled on the list (Aas, 1995). The fee may be flat for each provider or risk-

195


Figure D.3: Efficiency incentives by historical cost payment



adjusted, based on the relative risk of the registered population (Langenbrunner & Liu, 2004). An example of patient-list capitation is the Dutch system where sickness fund patients enrol on the list of a provider of their choice. This provider receives a captiations payment for each patient on its list (Jegers et al., 2002). Several variables have been considered in connection with the capitation formula. They include scale and scope of services (e.g. number of individuals, degree of urbanisation and distances to hospital), and risk factors (e.g. age, sex, family structure, socio-economic conditions, number of individuals disabled and under rehabilitation, mortality and morbidity). The number and type of providers, provider’s choice and the coverage of service packages involved in the capitation scheme have various effects on efficiency gain. Some beneficial features of capitation discussed by Barnum et al. (1995) include decentralisation, increased competition and cost efficiency, predictable budgets and thus total cost of health care services. Theoretically, since providers are responsible for delivering the contracted package of services for the fixed payment, they are motivated to innovate in cost reducing technology, use lower cost alternative treatment settings, and provide cost-effective care. In practice, like other capped budgets, there exists the risk of premature discharge, lower quality of care and cost shifting to reduce per patient expenses. This effect can be lessened if the capitation 196


formula can be designed to introduce a substantial amount of competition among providers, which can lead to efficiency gain while maintaining quality. Geographical capitation has another advantage over the historical budget model. It tends to lead to a more equitable distribution of resources because population size, rather than historical demand and utilisation of services, enter the formula. A region with high purchasing power tends to receive higher budgets under historical formula than a poor region because its willingness-to-pay is historically higher. Willingness to pay fails to reflect need for health care services, which is better proxied by population size and other demographic factors. The success of capitation modality depends on factors such as provider choice, referral constraint, degree of market competition and risk selection practices. There has been a growing recognition that free provider choice is critical in capitation formula design. Restricted provider choice forces the patients to “follow the money”, hence supply becomes less responsive to demand. By contrast, flexibility in provider choice allows the “money to follow the patients”. This improves the market’s contestability, which results in increasing competition among providers. This is a powerful force for efficiency and quality gain. Geographic capitation limits provider choice because the providers get paid for providing services to people living a defined geographic area. This can create a barrier for the treatment of patients from other regions (Aas, 1995), which substantially reduces the competitive pressure on providers within a region. It should be noted that hospitals have natural monopoly power within a geographic location due to economies of scale and scope. Therefore, if patients cannot move freely between regions for health services, providers within the defined location will enjoy a significant monopoly power. The patient-list capitation encounters the same problem if the health packages do not allow enrollees to shop among available providers. Referal constraint can generate strong incentives affecting the pattern of resource use, and thus overall efficiency. For instance, if the payment is to primary care providers and covers only the services provided at that level, the providers have incentive to minimise costs by referring patients to more specialised providers even though the referral is not necessary (Langenbrunner & Liu, 2004). This results in allocative inefficiency of outputs produced and higher health care cost. If the defined services covered by the capitation formula create no incentive to shift cost, efficiency gains can be achieved. Regulations and practices with regard to preferred risk selection can have an

197


effect on both efficiency and equity. This is more relevant for patient-list capitation than geographic capitation because cream-skimming is only possible if fund holders are allowed to select enrollees. In this case, if capitation payment is not adjusted for individual risks, there is an incentive to avoid high risk individuals and only include the most profitable segments of the population (Aas, 1995). This can result in deleterious consequences for both efficiency and equity. With capitation, there is inherently an incentive to make care efficient for the whole episode. But it also means the lack of incentive for activities. Hospitals are motivated to provide insufficient or reduced quality service in order to minimise costs (both treatment and preventive work) to be more economically profitable (Barnum et al., 1995; Langenbrunner & Liu, 2004). For instance, the global budget system in France was found to lead to slower growth in overall hospital expenditures, but this was the result of a lower volume of services rather than a reduction in the cost per service (Redmon and Yakoboski 1995) in (Langenbrunner et al., 2009). Barnum et al. (1995) suggested that reimbursement by capitation can be followed up by other measures of regulations in order to reduce the adverse effect on reduction of quality and volume of services. Figure D.4: Efficiency incentives by capitation method



198


D.1.2

Floating budget

This group of modalities involves a payment formula that has least one floating variable, either quantity or price. They evolve from pre-fixed budget methods, aiming at creating incentives for efficient behaviours while still maintaining a certain aspect of cost containment, either through unit cost or volume restrictions. Some, if well designed, can entail quality enhancement incentives. Compared to fixed budget modalities, these methods usually require more administrative resources due to reporting and monitoring requirements. They include the “per diem” payment method where price is set per bed-day, “per service” where payments are calculated on a set of standard charges for specific services, the “case-based” payment method where rates are set for individual groups of cases that consume a similar amount of resources, and the “fee-for-service” where the fee is set retrospectively. Except for the last method (F4S), the explicit incentive for other modalities is reducing unit cost while maximising the volume of health services provided.

D.1.2.1

Per diem

The per diem (or per bed day) approach calculates the budget by a standard daily charge multiplied by the number of actual bed days produced during a period of time. The per-diem is to cover all expenses related to hotelling services (labour, medical goods, and depreciation of facility and equipment) and treatment cost (medical labour and drugs). Here, the dominant incentive is to raise the number of hospital days and increase bed occupancy which is usually achieved through boosting hospital admissions and average length of stay. More hospital admission is possible through increasing bed capacity and shifting outpatient and community-based rehabilitation services to the hospital setting. However, the incentive to lengthen stay is likely to be stronger than the incentive to raise admissions. Hospital days tend to be more expensive early in a stay than later (Aas, 1995) as the marginal cost of an extra bed day diminishes quickly after the first few days. Therefore, more admissions, on average, leads to higher cost than low admission with longer stay. In either case, inefficiency remains the major problem: allocative inefficiency in the former (boosting hospital admissions) and technical inefficiency the latter (increasing average length of stay).

199


Apart from the incentive to “upgrade” less acute cases (such as outpatients and people undergoing rehabilitation) to acute treatment (inpatient), the per diem method also provides an incentive to cream skim because it does not distinguish between expensive and inexpensive cases. Simple case-mixes will be preferred over complicated and more expensive ones to reduce unit cost. Complicated cases, once admitted, might face under-treatment because there is an incentive to reduce the intensity of services for each bed-day, which directly affects quality of care. However, quality might not be a serious problem if length of stay for those cases increase, i.e. complicated cases still receive the same amount of treatment overall. The problem is now technical inefficiency. In developed health systems, the per diem payment is usually a complementary payment scheme rather than the major hospital payment methods. It is used for non-acute patients, patients in nursing homes or acute patients with outlier lengths of stay. Many developing countries still use the per diem method in paying hospitals and tertiary care. Figure D.5: Efficiency incentives by per-diem reimbursement



200


D.1.2.2

Per service (or scheduled fee)

The per service payment method, also called the negotiated fee schedule, calculate payments on a set of standard charges for specific services. The standard charges are usually set or negotiated between providers and payers. A regulated fee schedule is a set of standard fees managed by the government. This practice exists in some countries with centrally planned health care systems (such as China and Vietnam) and applies for some service types in developed countries. The Belgian nomenclature-system regulates fees for an extensive list of medical, diagnostic and therapeutic interventions. This system is also in use for the financing of physiotherapists, speech-trainers, dentists and nursing home care (Jegers et al., 2002). It is argued that this method does not promote efficient medical practices and fails to contain cost because because providers can unbundle or modify their services in the face of the standard rates and requires more frequent services to maintain their income (Barnum et al., 1995). This incentive is similar to that under the per diem payment as both tend to increase intermediate outputs per case. Increased intensity per case cannot always be interpreted as rising quality of care because many services are unnecessary and potentially harmful to the patients. It also has a harmful allocative effect of switching to more profitable, rather than clinically necessary, services. The consequence is rising health care costs, independent of number of cases treated and quality of care. These effects on cost and quality might be lessened if the payment method is accompanied by some cost-sharing schemes whereby patients need to pay out of pocket some medical costs, and a more competitive provider market.

D.1.2.3

Episode of care (case-based payment)

The case-based payment is a modality in which providers are paid a pre-determined amount covering all services per case or episode of illness. The basic method of case-based reimbursement is to bundle services into distinct case categories that are reasonably homogeneous with respect to resource use and reimburse a fixed amount per category (Barnum et al., 1995). Rate setting requires the use of case-mix groups established across patients and hospitals. The fundamental assumption behind case-mix grouping is that the resource required for a case is determined by the demographic, diagnostic, and treatment profile. The rates reflect historical costs of both individual hospitals and the 201


Figure D.6: Efficiency incentives by scheduled fee payment



entire network of hospitals. The main objectives of case-based payment is creating incentives for providers to increase efficiency, control unit cost and reduce the inpatient length of stay. It can be used to change service patterns, and encourage providers to favour treatments under lower cost settings while maintaining quality of care. The most popular, and probably the most ubiquitous international case mix grouping system, is the diagnosis-related groups (DRGs). This system was initially designed in the United States for the Medicare and Medicaid health insurance programs in the late 1960s. Subsequently, it has been imported into many other nations. Since the mid-1990s, the DRG-based payment system for hospitals has been tried in countries like Australia and Sweden. Other countries employ the system to allocate funds from central budgets to local health purchasers (e.g. Norway) or develop hospital budgets (e.g. New Zealand). More recently, a number of high and middle income countries (such as Korea, Taiwan, and Hungary) started experimenting with case-based hospital payment as a major part of their health system reform programs. Diagnosis-related groups classify each case according to the diagnosis and other characteristics of the case. The payment rate for each DRG then uses a resourcebased formula that reflect varying degrees of complexity and refinements to account for differences in the nature of the treated cases and the resources required to di202


agnose and treat them. This design allows for the following attributes (i) grouping is medically meaningful (i.e. patients from homogeneous diagnostic categories are grouped under the same DRG); (ii) variance between cases in the same group is kept at minimum; (iii) the number of classes is manageable; and (iv) each DRG has an associated a “price” (or weight). There is an inevitable trade-off between within-DRG homogeneity and the number of DRGs. A small number of DRGs requires aggregation of sub-groups and thus, introduces more heterogeneity (with respect to resource consumption) into each group. Thus, legitimate differences in costs between DRGs are not reflected. There is an incentive to cream-skim lower resource cases within each DRG to reduce unit cost. On the other hand, a large number of DRGs allow for fairly homogeneous cases within each DRG. The cost variation across cases then is small, hence less cream skimming incentive. However, the cost estimates from a small number of cases per DRG may not be statistically stable, and the system can be administratively burdensome. In addition, the greater the number of groups, the closer the payment system comes to fee-for-service, and the efficiency incentives may decrease (Langenbrunner et al., 2009). The transition to a cost-based reimbursement method represents the main thrust of hospital financing reforms over the past 20 years (Telyukov, 2001). This payment mechanism attracts the attention of reformers who want their national hospital sectors to be more productive, responsive to consumer expectations, and adaptable to changes in the demand for services. However, evidence to support this is somewhat controversial. PWC (2008) argues that the move to case rates has improved efficiencies. Introduction of case-based payments appears to reduce inpatient utilisation. Average length of stay and number of hospital beds per capita have dropped in much of the OECD as the number of countries implementing DRG-like payments has increased. Growth rate of total cost was found to cease. However, there is no evidence of significant impact on quality and outcome while other unintended effects include “up-coding” and “DRG creeping” practices (Pauly, 2001). Coulam and Gaumer (1992) found that errors in coding were not random and systematically favoured high weight DRG. It is suggested that providers have an incentive to diagnose patients into highly paid case categories and code medical records in such a way as to increase payment. Administrative costs under case-based funding are usually high due to the complex payment formula. Investment in infrastructure, record keeping and monitoring is needed. Extensive management of information on patient protocols and their 203


associated costs is required to establish the case categories and appropriate reimbursement rates, especially as new technologies and drugs become available. On the other hand, this case-based information system, once in place, can generate benefits beyond simple reimbursement. The collected information can be used to make evaluations on cost effectiveness, performance, and facilitate comparisons of provider quality and efficiency (Barnum et al., 1995). Figure D.7: Efficiency incentives by case-based reimbursement



D.1.2.4

Fee for service

Fee for service (F4S) is essentially a form of retrospective method in which payment is calculated after the provision of services and usually equals the actual cost incurred (Barnum et al., 1995). It usually applies to the delivery of specific items such as doctor consultations, specific x-ray tests, specific surgical operations, and other services. F4S can also include itemised charges for medical products and goods, because medical labour services are often provided with material products (Langenbrunner & Liu, 2004). F4S has one obvious advantage: reflecting the actual work done and efforts actually made. This encourages providers to deliver services with quality and stay more in touch with demand. This might also encourage providers to be more internally 204


efficient, hence increasing productive efficiency (Liu & Mills, 2007), although evidence for this effect is mixed. Another positive element of F4S is that it is relatively easier for hospital management to track resource flow, activities, and input utilisation. The billing system is also easy to run under this scheme, which is suitable for many low-income countries with weak information and management systems. Hospitals can estimate service costs and compare them with revenues, seeking to raise productivity(de Roodenbeke, 2004). Price inflation usually follows. Although creating a highly responsive supply side and being simple to implement, this approach is not necessary cheaper than other PPMs and is certainly not an appropriate tool for promoting either social efficiency or cost containment. All costs will be passed down to payers, either to the patients if they pay out-of-pocket or to the third party payers, such as insurance companies. As the revenue comes from both volume and price of services retrospectively, supply induced demand is an unavoidable phenomenon. There is a strong incentive for providers to deliver unnecessary services, particularly when the workload is low and treatment effectiveness is ambiguous (Liu & Mills, 2007), which is bad for social efficiency. F4S simply results in an explosion of medical cost due to volume and intensity of service per case, especially toward expensive, and not necessarily more effective interventions. According to Langenbrunner & Liu (2004), many countries have experienced a correlation between F4S implementation and pronounced increases in health expenditure. Another problem that has been identified is that providers may try to increase the quantity of services by reducing the length of time spent with each patient or delegating work to less-qualified medical staff, particularly if the workload is heavy Liu & Mills (2007). Quality deterioration in this case is usually unavoidable. Lastly, F4S can require high administration costs due to extra resources devoted to reporting and monitoring individual cases. This is another disadvantage that makes F4S unlikely to be the option for purchasing general primary and tertiary care. On the other hand, it can be an acceptable payment modality for some selective, non-critical care that complements necessary services, and sometimes hospital ambulatory activities and outpatient services.

D.1.2.5

Performance based payment

Under this modality (also called “value-based purchasing” or “pay for performance” - P4P), providers are rewarded for meeting pre-set targets for service delivery, which usually involves achieving targeted quality of care. This payment scheme 205


Figure D.8: Efficiency incentives by fee-for-service payment method



appears necessary because existing payment mechanisms fail to reward providers for high quality. When providers are paid for each service performed, it gives them a strong financial incentive to perform as many services as possible regardless of quality. When a fixed budget is applied, it gives providers the incentive to cut down both quantity and quality of care. It is expected P4P will change providers’ behavior to improve quality, patient safety and to reduce costs through efficiency improvement. A limited number of states/countries (mostly in the US and the UK) have applyed this method. This is because most current health systems lack basic requirements to undertake P4P. The first difficulty involves the ability to measure performance. Basing P4P on just a few indicators such as admissions or length of stay may compromise other objectives such as improved quality of care. As more objectives and indicators are addressed (such as risk-adjusted mortality, post-operative complications, and readmission), of which some are even conflicting, indicators multiply, adding administrative complexity and discouraging transparency Langenbrunner & Liu (2004). When P4P is applied to individual physicians, the problem expands to measure individual attribute to overall performance. As Langenbrunner & Liu (2004) pointed out, in health care, medical personnel have to work together to improve quality, and better performance is usually the outcome of joint efforts. Since individual performance is hard to separate from group performance, it is difficult to

206


measure. Another difficulty relates to administration and data gathering. As the number of objectives and indicators increase, data collection, monitoring and evaluation becomes time consuming and labour intensive, and thus very costly. Equity is also a potential problem. P4P can create a two-tier system where people able to pay more will have access to better services. Despite intense interest in and optimism about P4P in health care among both policy makers and payers, the number of studies looking at its impact on performance indicators such as quality of care, hospital safety, reduction of adverse outcomes, and service utilisation are still quite modest. As Rosenthal et al. (2005) pointed out, most of these programs are in the early stages of trial and evaluation and as a result, there is little published research and only a few studies demonstrating that P4P leads to improved quality of care. Some other reviews, such as Borenstein et al. (2004), found that quality improvement due to P4P was small and inconsistent and thus, no definite conclusion could be drawn. Figure D.9: Efficiency incentives by pay-for-performance method



D.1.3

The mix approach

Overall, floating budget methods clearly offer incentives to conserve unit cost and increase service volumes, and are thus likely to result in over-utilisation of services. 207


There is scope for productivity and efficiency improvement if cost conservation effort is achieved through better ways of delivering services, rather than cutting quality of care through pre-mature discharges and under-treatment. However, this group of PPMs usually fails to contain health care expenditures. Some create adverse incentives of cream-skimming and up-coding when providers try to maximise revenues. In contrast, fixed budget methods are effective in control health expenditure through containment of both unit-cost and service volume. Compared to floating budget methods, they offer weaker incentives for productivity, efficiency and quality improvement. Moreover, they can endanger access to care if they create incentives to exclude certain patient groups, especially the most costly ones (higher risk patients) from the providers perspective (Jegers et al., 2002). However, fixed payment methods are relatively cheaper and requires less sophisticated infrastructure and management capacity. It is thus clear that all methods generate both adverse and beneficial incentives. A mix of different provider payment methods has been suggested as a venue to offset the disadvantages while retaining the desired characteristics of individual modes. Even so, no system is completely free from unwanted side effects although gains can be achieved depending on the mixing design. Some mixes can retain desirable features of individual PPMs while others proves to exacerbate adverse incentives. The payment system can be mixed over several dimensions. First, different types of payment methods are used for different types of providers, i.e. general practitioners vs. specialists or primary care centres vs. hospitals or private vs. public providers. In France, for instance, public hospitals are funded under a global budget while private facilities are reimbursed per diem prices. Second, different cost categories can be linked to different payment methods. For instance, investments in space and equipment (fixed cost) can be subsidised (i.e. retrospectively, up to a ceiling) while the provider receives a global budget to cover its annual operating costs (e.g. nurses, materials, replacement of equipment). This mix is seen in Belgium. Third, different types of services provided by an individual provider can be funded by separate methods. Inpatient services can be funded under a global budget while outpatient and emergency visits receive fee-for-service. The focus here is on PPMs for provider organisations (such as hospital activities) rather than reimbursement for individual doctors working in the hospital setting2 . 2

It is noted that PPMs for provider organisations and doctors can be constructed to counteract or reinforce each other. If an overall objective is to control total costs, payment methods for the organisation and the doctors should be consistent in order to achieve the goal. For instance, capitation at the hospital level might go with salaried doctors rather than fee-for-service doctors (Aas, 1995).

208

D.2. FUNDING FOOLING MECHANISMS AND PROVIDER PAYMENT METHODS

One possible payment arrangement for provider organisations is to construct the cap payment to cover all but some types of care (such as outpatient and emergency services). The major goal is control over total costs (global budget) and to promote productivity (case-based or fee-for-service). A hospital with responsibility for a given population receiving capitation payment for inpatient care has an incentive to reduce hospital production costs. Alternatively, the payment method for inpatient care can be designed as an integrated method between a case-based payment with a global budget. At the end of each year, the global budget is compared with what has been produced. If production has been low, payment for the new year is reduced by a corresponding amount. Different hospital categories may be reimbursed with different rates per case which should be sufficient to cover marginal costs. This setting gives incentive to contain unit cost per case and volume of service as well as length of stay. Meanwhile, outpatient and emergency services and inpatient outliers can be reimbursed under fee for services (either with scheduled fee or actual cost). Fee for service is a good purchasing instrument in this case because paying the actual cost incurred can help maintain quality standards. Especially for expensive inpatient cases, it might mitigate the incentive for quick discharge or up-coding. It also gives hospital incentives to reduce unnecessary admissions if the medical conditions can be treated in the outpatient settings (because the inpatient budget is capped while the outpatient budget is floating). Another possible mix is between per-diem for hospital stays (inpatients) and feefor-service or scheduled-fee for specific services. This combination, commonly seen in developing countries, proves to be administratively easy, yet offers no incentives for either productivity and efficiency improvement or cost containment. Per-diem payments applied to hospital stays encourage hospitals to extend length of stay while fee for service for individual items (such as X-rays and lab tests) induces excessive uses of billed services, especially the profitable ones. The result is cost explosion and over-utilisation of services, which is productively inefficient and sometimes clinically harmful.

D.2

Funding fooling mechanisms and provider payment methods

Provider payment systems, in association with fund pooling mechanisms, makes up the two major functions in health care financing. They determine the volume, pattern and distribution of health care consumption, which in turn define how well 209


a health system achieves its objectives of equity, efficiency and quality of care. In health care, evidence suggests that PPS - the supply side - is likely to have more significant impact on service than fund pooling mechanism - the demand side. The reason is that the demand side (patients/consumers) needs to rely on the supply side (providers) to make consumption decisions. A fund pooling mechanism consists of two health care financing functions: revenue collection and pooling resources. Revenue collection concerns raising money from households, businesses and external sources. Pooling deals with the accumulation and management of revenues so that members of the pool share collective health risk, thereby protecting individuals from large and unpredicted health expenditures. Together with purchasing function (i.e. paying providers), these two functions translate into three major objectives (i) raising sufficient and sustainable revenues in an efficient and equitable manner, (ii) managing these revenues to equitably and efficiently pool health risks and (iii) ensuring the purchase of health services in an allocative and technically efficient manner. Major funding mechanisms include taxation, social health insurance, private health insurance and out of pocket payment. These mechanisms vary in the degree which they share and pool risk pooling and sharing, from taxation and social health insurance (very high) to out of pocket payment (none at all)3 . While funds mainly derive from individuals and businesses, the collection and management process and its interaction with provider payment systems is where health systems’ objectives of efficiency, equality and health outcomes can be realised. Figure D.10 describes the matrix of possible outcomes under different combinations of provider payment methods and funding mechanisms. PPMs, the horizontal direction, affects provider behaviours - the supply side while the vertical direction presents funding mechanisms that mostly affect user behaviour - the demand side. Each cell in this matrix represents a health financing method, such as capitation funded by taxation or F4S scheme covered by out of pocket payment. For each cell, the likelihood of achieving the objectives of equity, efficiency, cost containment, quality of care and administrative simplicity is expressed in four levels: very likely (top, yellow), likely (upper middle, green), unlikely (lower middle, orange) and very unlikely (bottom, red). It is noted that not all PPMs and funding mechanisms are compatible. For instance, fixed budget (including line item 3

Community based health insurance is another funding mechanism. However, it is considered a transitional rather than a long-term mechanism for health care funding. It has emerged due to high out of pocket spending and large irregulated private health insurance sector in developing countries.

210


and global budgeting) cannot be financed by out of pocket payment by individuals. Those cells are blacked out. In some systems, supply and demand side incentives can cancel each other out. For instance, out-of-pocket payment restricts utilisation and therefore can reduce the supply-induced-demand effect under free-for-service or per-service payments. In this case, supply’s incentive are partly offset by demand’s constraints. Likewise, demand’s incentive for over-utilisation is limited if a provider faces a budget cap. Some combinations of PPMs and funding mechanisms, on the other hand, reenforce each other’s incentives. For instance, when providers receive fee for service funded by taxation, they have incentive to induce patients who faced zero effective price to order extra services. The influence of supplier-induced demand and moral hazard results in over-utilisation and fuel health expenditure. The matrix in Figure D.10 addresses five major objectives of a health system: efficiency, quality of care, equity, expenditure control and administration feasibility. The incentive and likelihood of achieving those are rated on the scale of high, medium high, medium low, and low. It can be seen that PPMs funded by taxation or social health insurance are likely to achieve equity of access while achievements on efficiency, quality and expenditure containment vary. Although demand’s moral hazard exists under taxation and social health insurance systems, its effects on service volume are likely to be offset under fixed budget payments. Floating budgets, on the other hand, are likely to re-enforce this demand’s incentive through supplier induced demand, especially if fee-for-service is the option. It is reasonable to assume that some payment methods are not implemented under a private health insurance system. Insurance organisations, as payers, are profit maximizers. They want to maximise revenue through attracting more enrollees by offering competitive health packages while minimising cost by contracting with cost-efficient providers. Line-item budgets, historical payment and geographical capitation, which tend to offer low efficiency and quality of care incentives, are the least likely to help them achieve these two objectives simultaneously. Patient list capitation, however, is a common practice. For instance, the first cell represent the line-item budget method under tax funding. This combination offers fairly low incentive for efficiency and quality of care. Providers are concerned with maintaining their budgets and thus are likely to cut

211

212

Effecting users behaviour

Out-of-pocket payment

Private (voluntary) health insurance Efficiency Quality of care Equity

Cost containment Admin simplicity


Efficiency Quality of care Equity

Efficiency Quality of care


Equity of care



Cost containment Admin simplicity Equity of care




Equity Cost containment Admin simplicity

Patient list capitation

Equity of care





Social Health Equity of care Insurance Efficiency Quality of care

Taxation

Historical payment

Line item budget

Geographic capitation

FIXED BUDGET

Cost containment

Efficiency Quality of care Equity Cost containment

Admin simplicity Cost containment Efficiency Quality of care Equity

Efficiency Quality of care Equity Cost containment

Admin simplicity Cost containment Efficiency Quality of care Equity

Admin simplicity

Cost containment


Admin simplicity

Equity Admin simplicity Efficiency Quality of care

Equity of care

Admin simplicity

Cost containment

Cost containment

Equity


Cost containment

Admin simplicity

Quality of care Equity Cost containment Admin simplicity

Efficiency

Cost containment

Equity Efficiency Quality of care Admin simplicity

Cost containment

Quality of care Admin simplicity

Efficiency

Administration Efficiency Quality of care

Equity

Case-based

Equity

Per service

FLOATING BUDGET


Equity Admin simplicity

Per diem

Effecting providers behaviour

Figure D.10: Matrix of provider payment methods and fund pooling mechanisms

Equity (lowest)

Admin simplicity Quality of care Cost containment Efficiency

Quality of care Equity Cost containment Admin simplicity

Quality of care

Cost containment

Efficiency Admin simplicity

Equity

Quality of care

Cost containment

Admin simplicity

Efficiency

Equity Quality of care

Fee for service



down service intensity and length of stay per case to keep unit cost low (rather than being more innovative in efficiency improvement). However, this is effective in controlling expenditure. This system is relatively cheap and requires simple infrastructure and management skill. It is financed by taxation and thus is likely to result in equity of access to citizens. The combination of out-of-pocket funding and a fee-for-service method provides a somewhat opposite picture: high incentive for efficiency and quality of care while inequity of access is the major problem. Expenditure in this case is largely contained by the patients’ ability to pay but high expenditure growth might still be realised. Administration can be easy and relatively cheap when there is little requirement for reporting and monitoring individual cases.

213

Bibliography

I. M. Aas (1995). ‘Incentives and financing methods’. Health Policy 34:205–220. M. Abreu, et al. (2005). ‘A Meta-Analysis of Beta-Convergence: The Legendary Two-Percent’. Discussion Papers 05-001/3, Tinbergen Institute. ABS (2003). ‘Australian Hospital Statistics’. Tech. rep., Australian Bureau of Statistics. ABS (2007). ‘Australian Hospital Statistics’. Tech. rep., Australian Bureau of Statistics. S. J. Adams (2005). ‘Vietnams Health Care System: A Macroeconomic Perspective’. Tech. rep., International Monetary Fund Hanoi, Vietnam. AIHW (2006). ‘Medical labour force 2004’. Tech. rep., Australian Institute of Health and Welfare. AIHW (2008). ‘Medical labour force 2005’. Tech. rep., Australian Institution of Health and Welfare. M. Al-Shammari (1999). ‘A multi-criteria data envelopment analysis model for measuring the productive efficiency of hospitals’. International Journal of Operations and Production Management 19:9:879–890. V. Aletras, et al. (2007). ‘The short-term effect on technical and scale efficiency of establishing regional health systems and general management in Greek NHS hospitals’. Health Policy 83(2-3):236–245. Aletras, Vassilios Kontodimopoulos, Nick Zagouldoudis, Athanasios Niakas, Dimitris. Allen (2006). ‘Health Economist Review: Final report’. Tech. rep., The Allen Consulting Group Pty Ltd.

214

BIBLIOGRAPHY

P. Arocena & A. Garcia-Prado (2007). ‘Accounting for quality in the measurement of hospital performance: evidence from Costa Rica’. Health Economics 16(7):667– 685. K. Arrow (1963). ‘Uncertainty and the welfare economics of medical care’. The American Economic Review 53:941–973. A. Athanassopoulos, et al. (1999). ‘A descriptive assessment of the production and cost efficiency of general hospital in Greece’. Health Care Management Science 2(2). A. Athanassopoulos & C. Gounaris (2001). ‘Assessing the technical and allocative efficiency of hospital operations in Greece and its resource allocation implications’. European Journal of Operational Research 133(2):416–431. V. Bankauskaite & G. Dargent (2007). ‘Health systems performance indicators: methodological issues’. Presupuesto y Gasto Pblico 49:125–137. R. Banker, et al. (1984). ‘Some models for estimating technical and scale inefficiencies in Data Envelopment Analysis’. Management Science 30:1078–1092. G. P. Barbetta, et al. (2007). ‘Behavioral differences between public and private not-for-profit hospitals in the Italian National Health Service’. Health Economics 16(1):75–96. Barbetta, Gian Paolo Turati, Gilberto Zago, Angelo M. D. Barnum & J. Gleason (2005). ‘Technical efficiency bias caused by intra-input aggregation in data envelopment analysis’. Applied Economics Letters 12(13):785– 788. H. Barnum, et al. (1995). ‘Incentives and provider payment methods’. International Journal of Health Planning and Management 10:23–45. L. J. Bates, et al. (2006). ‘Market structure and technical efficiency in the hospital services industry: A DEA approach’. Medical Care Research and Review 63(4):499–524. Bates, Laurie J. College, Bryant Mukherjee, Kankana Santerre, Rexford E. G. Battese & T. Coelli (1992). ‘Frontier Production Functions, Technical Efficiency and Panel Data: With Application to Paddy Farmers in India’. Journal of Productivity Analysis 3:1/2:153–169. G. Battese & D. Rao (2002). ‘Technology gap, Efficiency, and a stochastic metafrontier function’. International Journal of Business and Economics 1(2):87–93. 215

BIBLIOGRAPHY

G. E. Battese, et al. (2004). ‘A Metafrontier Production Function for Estimation of Technical Efficiencies and Technology Gaps for Firms Operating Under Different Technologies’. Journal of Productivity Analysis 21:91–103. D. Bilodeau, et al. (2004). ‘Measuring hospital performance in the presence of quasifixed inputs: An analysis of Quebec Hospitals’. Journal of Productivity Analysis 21:183–199. E. Biorn, et al. (2003). ‘The effects of activity-based finacning on hospital efficiency: a panel data analysis of DEA efficiency scores 1992-2000’. Health Care Management Science 6:271–283. G. R. Bitran & V.-S. Josep (1987). ‘Some mathematical programming based measures of efficiency in health care institutions’. Advances in mathematical programming and financial planning 1:61–84. P. Bogetoft & K. Nielsen (2004). ‘DEA Based Yardstick Competition in Natural Resource Management’. J. Borenstein, et al. (2004). ‘The association between quality improvement activities performed by managed care organizations and quality of care’. The American Journal of Medicine 117(5):294–304. Z. Boutsolie (2010). ‘Demand variability, demand uncertainty and hospital costs: a selective survey of the empirical literature’. Glogal Journal of Health Science 2 (1). K. Boyle, et al. (1994). ‘What Do We Know About Groundwater Values? Preliminary Implications from a Meta Analysis of Contingent-Valuation Studies’. American Journal of Agricultural Economics 76(5):1055–1061. B. Bravo-Ureta, et al. (2007). ‘Technical efficiency in farming: a meta-regression analysis’. Journal of Productivity Analysis 27(1):57–72. P. Brokett, et al. (1997). ‘Data transformation in DEA cone ratio envelopment approach for monitoring bank performance’. European Journal of Operational Research 98:250–265. M. Brons, et al. (2005). ‘Efficiency of urban public transit: a meta analysis’. Transportation 32 (1):121. R. Brouwer, et al. (1999). ‘A meta-analysis of wetland contingent valuation studies’. Regional Environmental Change 1:47–57. 216

BIBLIOGRAPHY

H. S. Brown (2003). ‘Managed care and technical efficiency’. Health Economics 12(2):149–158. J. Burgess & P. Wilson (1995). ‘Decomposing hospital productivity changes, 19851988: A nonparametric malmquist approach’. Journal of Productivity Analysis 6(4):343–363. 105th Annual Meeting of the American-Economic-Association, ANAHEIM, CA, JAN 05-07, 1993. J. F. Burgess & P. W. Wilson (1996). ‘Hospital ownership and technical inefficiency’. Management science 42:1:110–123. J. Butler (1995). Hospital cost analysis. Boston: Kluwer Academic Publishers. K. Button & P. Rietveld (2000). ‘A meta-analysis of the impact of infrastructure policy on regional development’. In H. Kohno, P. Nijkamp, & J. Poot (eds.), Regional cohesion and competition in the age of globalization. Cheltenham: Edward Elgar. J. Cameron (1985). ‘The indirect costs of graduate medical education.’. The New England Journal of Medicine 312:1233–1238. D. Card & A. Krueger (1995). ‘Time-series minimum-wage studies: A metaanalysis’. American Economic Review 85:238–243. K. Carey (2003). ‘Hospital cost efficiency and system membership’. Inquiry-the Journal of Health Care Organization Provision and Financing 40(1):25–38. K. Cavlovic, et al. (2000). ‘A meta-analysis of environmental Kuznets studies’. Agriculture and Resource Economics Review 29(1):3242. C. Cazals, et al. (2002). ‘Non-parametric frontier estimation: a robust approach’. Journal of Econometrics 106(1):1–25. H. Chang, et al. (2004). ‘Hospital ownership and operating efficiency: evidence from Taiwan’. European Journal of Operational Research 159:513–527. H. H. Chang (1998). ‘Determinants of hospital efficiency: the case of central government-owned hospitals in Taiwan’. Omega-International Journal of Management Science 26(2):307–317. A. Charnes, et al. (1978). ‘Measuring the efficiency of decision making units’. European Journal of Operational Research 2:429–444. A. Charnes, et al. (1990). ‘Polyhectal Cone-ratio models with an illustrative application to large commercial banks’. Journal of Econometrics 46:73–91. 217

BIBLIOGRAPHY

A. Charnes, et al. (1989). ‘Cone ratio data envelopment analysis’. International Journal of Systems Science 20(7):1099–1118. A. Chaudhuri & K. Roy (2008). ‘Changes in out-of-pocket payments for healthcare in Vietnam and its impact on equity in payments, 19922002’. Health Policy 88 (1):38–48. A. Chen, et al. (2005). ‘Measurement and sources of overall and input inefficiencies: Evidences and implications in hospital services’. European Journal of Operational Research 161(2):447–468. J.-Y. Chern & T. T. Wan (2000). ‘The impact of the prospective payment system on the technical efficiency of hospitals’. Journal of Medical System 24:3:159–172. T. N. Chirikos (1998a). ‘Further evidence that hospital production is inefficient’. Inquiry-the Journal of Health Care Organization Provision and Financing 35(4):408–416. T. N. Chirikos (1998b). ‘Identifying efficiently and economically operated hospitals: The prospects and pitfalls of applying frontier regression techniques’. Journal of Health Politics Policy and Law 23(6):879–904. T. N. Chirikos & A. M. Sear (2000). ‘Measuring hospital efficiency: A comparison of two approaches’. Health Services Research 34(6):1389–1408. L. Christensen & W. Greene (1976). ‘Economies of scale in United States electric power generation’. Journal of political economy 84(4):655–676. T. Coelli (1996). ‘A guide to DEAP Version 2.1: Data Envelopment Analysis (computer) program.’. Working paper 96/08, Center for Productivity and Efficiency Analysis: University of New England, Australia. T. J. Coelli, et al. (2005). An introduction to efficiency and productivity analysis. Springer. N. H. P. Committee (2004). ‘National Report on Health sector performance indicator 2003’. Tech. rep., Australian Institute of Health and Welfare. C. Cornwell, et al. (1990). ‘Production frontiers with cross sectional and time series variation in efficiency levels’. Journal of Econometrics 46(1-2):185–200. P.-Y. Cremieux & P. Ouellette (2001). ‘Omitted variable bias and hospital costs’. Journal of Health Economics 20(2):271–282.

218

BIBLIOGRAPHY

R. Croson & M. Marks (2000). ‘Step Returns in Threshold Public Goods: A Metaand Experimental Analysis’. Experimental Economics 2(3):239–259. J. Cunningham, et al. (2006). ‘Bed occupancy, turnover interval and MRSA rates in Northern Ireland’. British Journal of Nursing 15:6:324–328. J. Dalhuisen, et al. (2001). ‘Price and income elasticities of residential water demand: Why empirical estimates differ,’. Discussion paper TI 2001-057/3, Amsterdam: Tinbergen Institute. E. Dalmau-Matarrodona & J. Puig-Junoy (1998). ‘Market structure and hospital efficiency: evaluating potential effects of deregulation in a national health service’. Review of Industrial Organisation 13:447–466. H. Dao, et al. (2008). ‘User fees and health service utilisation in Vietnam: how to protect the poor?’. Public Health 122:1068–1078. G. Davies (2005). ‘Queensland Public Hospitals Commission of Inquiry Report’. Tech. rep., Queensland Government. R. de Mooij & S. Ederveen (2001). ‘Taxation and Foreign Direct Investment: A Synthesis of Empirical Research’. Cesifo working paper series, CESifo Group Munich. E. de Roodenbeke (2004). ‘Purchasing Inpatient and Outpatient Care Through Hospitals’. Tech. rep., World Bank. J. DeCoster (2004). ‘Meta-analysis.’. In K. Kempf-Leonard (ed.), The Encyclopedia of Social Measurement. San Diego, CA: Academic Press. M. A. Deily, et al. (2000). ‘Exit and inefficiency: the effects of ownership type’. The Journal of human resources 35:4:734–747. B. Dervaux, et al. (2004). ‘Comparing French and US hospital technologies: a directional input distance function approach’. Applied Economics 36(10):1065– 1081. DHA (2007a). ‘Health and Aging Factbook’. Tech. rep., Australian Government Department of Health and Aging. DHA (2007b). ‘State of our public hospitals’. Tech. rep., Australian Government: Department of Health and Aging.

219

BIBLIOGRAPHY

DHA (2008). ‘Report on the Audit of Health Workforce in Rural and Regional Australia, April 2008’. Tech. rep., Australian Government Department of Health and Ageing, Commonwealth of Australia, Canberra. C. Doucouliagos (1995). ‘Worker participation and productivity in labor-managed and participatory capitalist firms: A meta-analysis’. Industrial and Labor Relations Review 4:58–77. C. Doucouliagos (1997). ‘The aggregate demand for labour in Australia: A metaanalysis’. Australian Economic Papers 36:224–242. H. Doucouliagos & M. Paldam (2005). ‘Aid Effectiveness on Accumulation. A Meta Study’. Econonomic working paper 2005-12, School of Economics and Management, University of Aarhus. H. Doucouliagos & M. Paldam (2006). ‘Aid effectiveness on accumulation: A meta study’. Kyklos 59(2):227–254. R. Dredge (2004). ‘Hospital Global Budgeting’. Tech. rep., World Bank. M. Drummond, et al. (1997). Methods for the Economic Evaluation of Health Care Programmes. Oxford: Oxford University Press. S. J. Duckett (1998). ‘Casemix funding for acute hospital inpatient services in Australia’. MJA 196:S17–21. S. J. Duckett (2004). The Australian Health Care system. Oxford University Press: South Melbourne. B. Efron (1979). ‘Bootstrap methods: Another look at the Jackknife’. Annals of Statistics 7:1–26. E. . Erlandsen (2008). ‘Improving the Efficiency of Health-Care Spending: What Can be Learnt from Partial and Selected Analyses of Hospital Performance’. Economic Studies No. 44, OECD: Paris. M. Espey (1998). ‘Gasoline demand revisited: An international meta-analysis of elasticities’. Energy Economics 20:273–295. M. Espey & H. Kaufman (2000). ‘The impact of airport noise and proximity on residential property values’. Growth and Change 31:341–352. M. Espey & D. Thilmany (2000). ‘Farm labor demand: A meta-regression analysis of wage elasticities’. Journal of Agricultural and Resource Economics 25:252–266. 220

BIBLIOGRAPHY

R. Fare, et al. (1983). ‘The relative efficiency of Illinois electric utilities’. Resources and Energy 5:349–367. R. Fare, et al. (1985). The measurement of efficiency of production. Boston: Kluwer Academic Publisher. R. Fare, et al. (1997). ‘The effect of risk based capital requirements on profit efficiency in Banking’. Discussion paper series No 97-12, Department of Economics, Southern Illinois University at Carbondale. R. Fare, et al. (2004). ‘Aggregation bias and its bounds in measuring technical efficiency’. Applied Economics Letters 11(10):657–660. R. Fare & C. Lovell (1978). ‘Measuring the technical efficiency of poduction’. Journal of Economic Theory 19/1:150:162. R. Fare & D. Primont (1995). Multi-output production and duality: theory and applications. Kluwer Academic Publishers. M. J. Farrell (1957). ‘The measurement of productive efficiency’. Journal of the Royal Statistical Society 120:253–281. M. Farsi, et al. (2005). ‘Efficiency measurement in network industries: Application to the swiss railway companies’. Journal of regulatory economics 28(1):69–90. M. Farsi, et al. (2003). ‘Unobserved heterogeneity in stochastic cost fronter models: a comparative analysis’. Department of Economics, University of Lugano and Swiss Federal Institute of Technology, Swizerland. R. Feldman & F. Lobo (1997). ‘Global budgets and excess demand for hospital care’. Health Economics 6:187–196. A. Ferrari (2006a). ‘The internal market and hospital efficiency: a stochastic distance function approach’. Applied Economics 38:2121–2130. A. Ferrari (2006b). ‘Market oriented reforms of health services: a non-parametric analysis’. Service Industries Journal 26(1):1–13. G. Ferrier & J. Hirschberg (1997). ‘Bootstrapping confidence intervals for linear programming efficiency scores: with an illustration using Italian bank data’. Journal of Productivity Analysis 8:19–33. G. Ferrier & C. Lovell (1990). ‘Measuring cost efficiency in banking: econometric and linear programming evidence’. Journal of Econometrics 46:229–245. 221

BIBLIOGRAPHY

G. D. Ferrier & V. Valdmanis (1996). ‘Rural hospital performance and its correlates’. Journal of Productivity Analysis 7:63–80. R. Florax, et al. (2002). ‘Meta-analysis: A tool for upgrading inputs of macroeconomic policy models’. Working paper TI 2002-041/3., Amsterdam: Tinbergen Institute Discussion Paper,. S. Folland & R. Hofler (2001). ‘How reliable are hospital efficiency estimates? Exploiting the dual to homothetic production’. Health Economics 10(8):683–698. H. Frech & L. Mobley (2000). ‘Efficiency, growth, and concentration: An empirical analysis of hospital markets’. Economic inquiry 38(3):369–384. Annual Meeting of the Allied-Social-Science-Association, SAN FRANCISCO, CALIFORNIA, JAN 05-07, 1996. A. Frick, et al. (1985). ‘Case-mix and cost differences between teaching and nonteaching hospitals’. Medical Care 23:4:283–295. H. O. Fried, et al. (2008). ‘Efficiency and Productivity’. In The measurement of productive efficiency and productivity growth. Oxford University Press: New York. D. Friesner, et al. (2008). ‘Are hospitals seasonally inefficient? Evidence from Washington State’. Applied Economics 40(6):699–723. Friesner, Dan Roseman, Robert McPherson, Matthew Q. A. Fuiji (2001). ‘Determinants and probability distribution of inefficiency in the stochastic cost frontier in Japanese hospitals’. Applied Economics Letters 8:807– 8012. A. Fuiji & M. Ohta (1999). ‘Stochastic cost frontier and cost inefficiency of Japanese hospitals: a panel data analysis’. Applied Economic Letters 6:527–532. J. Fuller & K. Hester (1998). ‘The effect of labor relations climate on the union participation process,’. Journal of Labor Research 19:173–187. B. Gannon (2005). ‘Testing for variation in technical efficiency of hospitals in Ireland’. The Economics and Social review 36:3:273–294. U.-G. Gerdtham, et al. (1999a). ‘Internal markets and health care efficiency: a multiple-output stochastic frontier analysis’. Health Economics 8(2):151–164. U.-G. Gerdtham, et al. (1999b). ‘The Impact of Internal Markets on Health Care Efficiency: Evidence from Health Care Reforms in Sweden’. Applied Economics 31:935–945. 222

BIBLIOGRAPHY

D. I. Giokas (2001). ‘Greek hospitals: how well their resources are used’. OmegaInternational Journal of Management Science 29:73–83. M. Goddard, et al. (1998). ‘The NHS Performance Framework: Taking Account of Economic Behaviour’. University of York: Center for Health Economics 158. A. C. Goncalves, et al. (2007). ‘Data envelopment analysis for evaluating public hospitals in Brazilian state capitals’. Revista De Saude Publica 41(3):427–435. W. Greene (2003a). ‘Simulated likelihood estimation of the normal-gamma stochastic frontier function’. Journal of productivity analysis 19:179–190. W. Greene (2003b). ‘Simulated Likelihood Estimation of the Normal-Gamma Stochastic Frontier Function’. Journal of Productivity Analysis 19:179190. W. Greene (2004). ‘Distinguishing between heterogeneity and inefficiency: stochastic frontier analysis of the World Health Organization’s panel data on national health care systems’. Health Economics 13:959–980. W. Greene (2005a). ‘Fixed and random effects in Stochastic frontier models’. Journal of productivity analysis 23:7–32. W. Greene (2005b). ‘Reconsidering heterogeneity in panel data estimators of the stochastic frontier model’. Journal of Econometrics 126:269–303. W. Greene (2008). ‘The econometric approach to efficiency analysis’. In The measurement of productive efficiency and productivity growth. Oxford University Press: New York. O. Groene, et al. (2008). ‘The World Health Organization Performance Assessment Tool for Quality Improvement in Hospitals (PATH): An Analysis of the Pilot Implementation in 37 Hospitals’. International Journal for Quality in Health Care 20(3):155–161. W. Groot & H. M. van den Brink (2000). ‘Overeducation in the labor market: A meta-analysis’. Economics of Education Review 19:149–158. S. Grosskopf, et al. (2001). ‘Comparing teaching and non-teaching hospitals: a frontier approach (teaching vs. non-teaching hospitals)’. Health Care Management Science 4:83–90. S. Grosskopf, et al. (2004). ‘Competitive effects on teaching hospitals’. European Journal of Operational Research 154(2):515–525.

223

BIBLIOGRAPHY

S. Grosskopf & V. Valdmanis (1987). ‘Measuring hospital performance - a non parametric approach’. Journal of Health Economics 6(2):89–107. S. Grosskopf & V. Valdmanis (1993). ‘Evaluating hospital performance with casemix adjusted outputs’. Medical Care 31(6):525–532. H. Hajialiafzali, et al. (2007). ‘Efficiency measurement for hospitals owned by the Iranian Social Security organisation’. Journal of Medical Systems 31(3):166–172. Hajialiafzali, Hossein Moss, J. R. Mahmood, M. A. P. Hall (1986). ‘On the number of bootstrap simulation required to construc a confidence interval’. Annals of Statistics 14:1453–1462. S. Hao & C. Pegels (1994). ‘Evaluating relative efficiencies of veterans affairs medical centers using data envelopment, ratio, and multiple regression analysis’. Journal of Medical Systems 18:5567. J. Harris (1977). ‘The Internal Organization of Hospitals: Some Economic Implications’. The Bell Journal of Economics 8(2):467–482. J. Harris, et al. (2000). ‘Do Mergers Enhance the Performance of Hospital Efficiency?’. The Journal of the Operational Research Society] 51(7):801–811. J. P. Harrison, et al. (2004). ‘Efficiency of Federal Hospitals in the United States’. Journal of Medical System 28:5:411–422. J. P. Harrison & R. J. Ogniewski (2005). ‘An efficiency analysis of Veterans Health Administration hospitals’. Military Medicine 170(7):607–611. M. M. Hofmarcher, et al. (2002). ‘Measuring hospital efficiency in Austria: A DEA approach’. Health Care Management Science 5:7–14. B. Hollingsworth (2003). ‘Non-parametric and parametric applications measuring efficiency in health care’. Health Care Management Science 6:203–218. B. Hollingsworth (2008a). ‘The measurement of efficiency and productivity of health care delivery’. Health Economics 17(10). B. Hollingsworth (2008b). ‘The measurement of efficiency and productivity of health care production’. Health economics 17:1107–1128. B. Hollingsworth & S. J. Peacock (2008). Efficiency measurement in Health and Health care. Routledge: London, New York.

224

BIBLIOGRAPHY

B. Hollingsworth & A. Street (2006). ‘The market for efficiency analysis of health care organisations’. Health Economics 15(10):1055–1059. B. Hollingsworth & J. Wildman (2003). ‘The efficiency of health production: reestimating the WHO panel data using parametric and non-parametric approaches to provide additional information’. Tech. rep., Health Economics. C. Hollingsworth & D. Parkin (1995). ‘The efficiency of Scottish acute hospitals: An application of data envelopment analysis’. Mathematical Medicine and Biology 12(3-4):161–173. Z. Huang & S. Li (2001). ‘Stochastic DEA models with different types of input output disturbances’. Journal of Productivity Analysis 15(2):95–113. M. D. Intriligator (1978). Econometric Models,Techniques and Application. Prentice: New York. R. Jacobs (2001). ‘Alternative methods to examine hospital efficiency: Data envelopment analysis and stochastic frontier analysis’. Health Care Management Science 4:103–115. R. Jacobs, et al. (2006). Measuring efficiency in Health care: analytic techniques and health policy. Cambridge University Press. M. Jegers, et al. (2002). ‘A typology for provider payment systems in health care’. Health Policy 60:255273. R. Kane, et al. (2005). ‘Estimating the patient care costs of teaching in a teaching hospital’. Amedican Journal of Medicine 118:7:767–772. A. Keegan (2008). ‘Hospital Bed Occupancy’. C. A. Kerr, et al. (1999). ‘Best-practice measures of resource utilisation for hospitals: a useful complement in performance assessment’. Public Administration 77:3:639– 650. J. N. Kibambe & S. F. Kocht (2007). ‘DEA applied to a Gauteng sample of public hospitals’. South African Journal of Economics 75(2):351–368. Kibambe, Jacques Ngoie Kocht, Steven F. J. M. Kirigia, et al. (2008). ‘A performance assessment method for hospitas: the case of municipal hospitals in Angola’. Journal of Medical System 32:509–519.

225

BIBLIOGRAPHY

J. M. Kirigia, et al. (2004). ‘Using data envelopment analysis to measure the technical efficiency of public health centers in Kenya’. Journal of Medical Systems 28(2):155–166. A. Kneip, et al. (1998). ‘A note on the convergence of nonparametric DEA estimators for production efficiency scores’. Econometric Theory 14:783–793. N. Kontodimopoulos, et al. (2006). ‘Balancing efficiency of health services and equity of access in remote areas in Greece’. Health Policy 76(1):49–57. S. Kumbhakar (1990). ‘Production frontiers, panel data and time varying technical inefficiency’. Journal of Econometrics 46(1-2):201–211. S. Kumbhakar & L. Hjalmarsson (1995). ‘Decomposing technical change with panel data - an application to the public sector’. Scandinavian Journal of Economics 97(2):309–323. Econometric-Society European Meeting, UPPSALA, SWEDEN, AUG, 1993. S. Kumbhakar & C. A. K. Lovell (2000). Stochastic frontier analysis. Cambridge University Press (New York). S. C. Kumbhakar & E. G. Tsionas (2006). ‘Estimation of stochastic frontier production functions with input oriented technical efficiency’. Journal of Econometrics 133:71–96. J. Ladinsky, et al. (2000). ‘Changes in the Health Care System of Vietnam in Response to the Emerging Market Economy’. Journal of Public Health Policy 21(1):82–98. J. C. Langenbrunner, et al. (2009). Designing and implemeting Health care provider payment system: How-to manuals. World Bank, Washington, D.C. J. C. Langenbrunner & X. Liu (2004). ‘How to pay? Understanding and using incentives’. Tech. rep., World Bank: Resource Allocation and Purchasing Project. J. Lauer, et al. (2004). ‘World health system performance revisited: the impact of varying the relative importance of health system goals’. BMC Health Services Research 4(19). K.-H. Lee, et al. (2008). ‘The association between hospital ownership and technical efficieny in managed care environment’. Journal of Medical system 33:307–315. Y. Lee & P. Schmidt (1993). ‘A Production Frontier Model with Flexible Temporal Variation in Technical Inefficiency’. In C. L. H.O. Fried & e. S.S. Schmidt (eds.), 226

BIBLIOGRAPHY

The Measurement of Productive Efficiency: Techniques and Applications. New York: Oxford University Press. T. Li & R. Rosenman (2001). ‘Cost efficiency in Washington hospitals: A stochastic frontier approach using panel data’. Health Care Management Science 4:73–81. S. Lieberman & A. Wagstaff (2009). ‘Health financing and delivery in Vietnam: Looking forward’. Tech. rep., The World Bank. M. Linna (1998). ‘Measuring hospital cost efficiency with panel data models’. Health Economics 7:415–427. M. Linna & U. Hakkinen (1998). ‘A comparative application of econometric frontier and DEA method for assessing cost efficiency of Finnish hospitals’. Health, the medical profession and regulation pp. 169–187. M. Linna & U. Hakkinen (1999). ‘Determinants of cost efficiency of Finnish hospitals: a comparison of DEA and SFA’. Working paper, National research and development center for welfare and health. M. Linna & U. Hakkinen (2006). ‘Reimbursing for the costs of teaching and research in Finnish hospitals: a stochastic frontier analysis’. International Journal of Health Care Finance Economics 6:83–97. M. Linna, et al. (2006). ‘Comparing hospital cost efficiency between Norway and Finland’. Health Policy 77(3):268–278. Linna, Miika Hakkinen, Unto Magnussen, Jon. X. Liu & A. Mills (2005). ‘The effect of performance related pay of hospital doctors on hospital behaviour: a case study from Shandong, China’. Human resources for health 3:11. X. Liu & A. Mills (2007). ‘Payment mechanisms and provider behaviour’. In A. S. Preker (ed.), Public ends, private means: strategic purchasing of health services. World Bank. J. Loomis & D. White (1996). ‘Economic benefits of rare and endangered species: Summary and meta analysis’. American journal of agricultural economics 78(5):1407. G. Lopez-Casanovas & M. Saez (1999). ‘The impact of teaching status on average costs in Spanish hospitals’. Health Economics 8:7:614–651.

227

BIBLIOGRAPHY

G. Lopez-Valcarcel & P. Baber Perez (1996). ‘Changes in the efficiency of Spanish public hospitals after the introduction of program-contracts’. Investigaciones Economicas 20:3:337–402. G. Lordan (2009). ‘Considering endogeneity and heterogeneity-a hierarchical random parameters approach to measuring efficiency’. Applied Economics 41(26):3411– 3423. C. K. Lovell (1996). ‘Applying efficiency measurement techniques to the measurement of productivity change’. Journal of Productivity Analysis 7:329–340. C. K. Lovell (2006). ‘Frontier Analysis in healthcare’. International Journal of Healthcare technology and management 7:1/2:5–14. D. Luong, et al. (2007). ‘Vietnam during economic transition: a tracer study of health service access and affordability.’. International Journal of Health Services 37(3):573–588. J. R. Lynch & Y. A. Ozcan (1994). ‘Hospital closure - an efficiency analysis’. Hospital & Health Services Administration 39(2):205–220. J. MacKinnon, et al. (1983). ‘Tests for model specification in the presence of alternative hypotheses: Some further results’. Journal of Econometrics 21:5370. J. Magnussen (1996). ‘Efficiency measurement and the operationalization of hospital production’. Health Services Research 31(1):21–37. N. Maniadakis, et al. (1999). ‘The impact of the internal market on hospital efficiency, productivity and service quality’. Health Care Management Science 2:75– 85. N. Maniadakis & E. Thanassoulis (2000). ‘Assessing productivity changes in UK hospitals refecting technology and input prices’. Applied Economics 32:1575–1589. P. E. Martinussen & L. Midttun (2004). ‘Day surgery and hospital efficiency: empirical analysis of Norwegian hospitals, 1999-2001’. Health Policy 68:183–196. F. Masiye (2007). ‘Investigating health system performance: an application of data envelopment analysis to Zambian hospitals’. BMC Health Services Research 7(58):(25 April 2007). N. McKay, et al. (2002). ‘Ownership and changes in hospital inefficiency, 19861991’. Inquiry - The journal of health care organisation provision and financing

228

BIBLIOGRAPHY

39(4):388–399. Annual Meeting of the Southern-Economic-Association, BALTIMORE, MARYLAND, NOV 08-10, 1998. L. R. Mobley & J. Magnussen (1998). ‘An international comparison of hospital efficiency: does institutional environment matter?’. Applied Economics 30(8):1089– 1100. MOH (2002). ‘National Health Survey’. Tech. rep., Ministry of Health, Vietnam. R. Morey, et al. (1995). ‘Estimating the hospitalwide cost differentials warranted for teaching hospitals: an alternative to regression approaches’. Medical Care 33(5):531552. R. C. Morey & D. A. Dittman (1996). ‘Cost pass-through reimbursement to hospitals and their impacts on operating efficiencies’. Annals of Operations Research 67:117–139. L. R. Murillo-Zamorano (2004). ‘Economic efficiency and fronter techniques’. Journal of Economics surveys 18:1:33–77. J. Newhouse (1994). ‘Frontier estimation: how useful a tool for health economics?’. Journal of health economics 13:317–322. K. Nguyen & T. Giang (2007). ‘Non-parametric Analysis of Efficiency Performance for Hospitals and Medical Centers in Vietnam’. In K. Nguyen & T. Giang (eds.), Technical Efficiency anf Productivity Growth in Vietnam: Parametric and Nonparametric Analyses. Publishing House of Social Labour. Q. Nguyen, et al. (2008). ‘Joint Annual Health Review: Health financing in Vietnam’. Tech. rep., Ministry of Health and Health Partnership Group. P. Nijkamp & J. Poot (2003). ‘Meta-Analysis of the Impact of Fiscal Policies on Long-Run Growth’. Discussion Papers 02-028/3, Tinbergen Institute. T. Nunamaker (1985). ‘Using data envelopment analysis to measure the efficiency of non-profit organizations: a critical evaluation’. Managerial and Decision Economics 6:50–58. C. O’Donnell, et al. (2009). ‘Efficiency analysis in the presence of uncertainty’. School of Economics: University of Queensland . C. O’Donnell, et al. (2008). ‘Metafrontier frameworks for the study of firm-level efficiencies and technology ratios’. Empirical economics 34:231–255.

229

BIBLIOGRAPHY

OECD (2002). ‘Measuring up: improving the performance of health systems in OECD countries.’. Tech. rep., The Organisation for Economic Co-operation and Development. OECD (2004). ‘Towards high performing health systems’. Tech. rep., Organisation for economic cooperation and development. OECD (2009). Measuring Capital: OECD manual 2009. Organisation of Economic Cooperation and Development. L. O’Neil (1998). ‘Multifactor efficiency in Data Envelopment Analysis with an application to urban hospitals’. Health Care Management Science 1:19–27. D. Osei, et al. (2005). ‘Technical efficiency of public distric hospitals and health centers in Ghana: a pilot study’. Cost Effectiveness and Resource Allocation 3:9. Y. A. Ozcan & R. R. Bannick (1994). ‘Trends in Department of Defence hospital efficiency’. Journal of Medical Systems 18(2):69–83. M. Pang, F. Mike Drummond & F. Song (1999). ‘The use of meta-analysis in economic evaluation’. Working Papers 173chedp, Centre for Health Economics, University of York. H. Pardes (1997). ‘The future of medical schools and teaching hospitals in the era of managed care’. Academic Medicine 72:2:97–102. D. Parkin & B. Hollingsworth (1997). ‘Measuring production efficiency of acute hospital in Scotland, 1991-94: validity issue in data envelopment analysis’. Applied Economics 29:1425–1433. PC (2009). ‘Public and Private Hospitals: Research Report’. Tech. rep., Productivity Commission. F. Pedraja-Chaparro, et al. (1997). ‘On the role of wight restrictions in data envelopment analysis’. Journal of Productivity Analysis 8:215–230. M. Pitt & L. Lee (1981). ‘The measurement and sources of technical technical inefficiency in the indonesian weaving industry’. Journal of development economics 9(1):43–64. D. Prior (1996). ‘Technical Efficiency and Scope Economies in Hospitals’. Applied Economics 28(10):1295–1301.

230

BIBLIOGRAPHY

D. Prior (2006). ‘Efficiency and total quality management in health care organizations: A dynamic frontier approach’. Annals of Operations Research 145:281–299. Prior, Diego. D. Prior & M. Sol (2000). ‘Technical efficiency and economies of diversification in health care’. Health Care Management Science. 3(4):299–307. J. Puig-Junoy (2000). ‘Partitioning input cost efficiency into its allocative and technical components: an empirical DEA application to hospitals’. Socio-Economic Planning Sciences 34:199–218. J. Puig-Junoy & V. Ortun (2004). ‘Cost efficiency in primary care contracting: a stochastic frontier cost function approach’. Health Economics 13(12):1149–1165. PWC (2008). ‘You get what you pay for.’. PricewaterhouseCoopers’ Health Research Institutute . QH (1998). ‘Hospital funding model for Queensland public hospitals’. Tech. rep., Queensland Health. QH (1999). ‘Hospital funding model for Queensland public hospitals’. Tech. rep., Queensland Health. QH (2000). ‘Hospital funding model for Queensland public hospitals’. Tech. rep., Queensland Health. QH (2005). ‘Queensland Health Systems Review (Final review)’. Tech. rep., Queensland Health. QH (2006). ‘Monthly activity collection and Queensland hospital admitted patient data collection’. Tech. rep., Queensland Health. QH (2007). ‘Health action: delivery reform’. Tech. rep., Queensland Health. QH (2008). ‘Casemix funding in Queensland’. Tech. rep., Queensland Health: Budget Service Unit. R. Ramanathan (2005). ‘Operations assessment of hospitals in the Sultanate of Oman’. International Journal of Operations & Production Management 25(1):39– 54. V. Rebba & D. Rizzi (2006). ‘Measuring hospital efficiency through Data Envelopment Analysis when Policy makers’ references matter’. Tech. Rep. No. 13/WP/2006, Ca’ Foscari University of Venice: Department of Economics. 231

BIBLIOGRAPHY

P. Reid, et al. (2005). Building a better delivery system: A new engineering/health care partnership. The National Academy Press. A. Renner, et al. (2005). ‘Technical efficiency of peripheral health units in Pujehun district of Sierra Leone: a DEA application’. BMC health service research 5:77. J. Richardson, et al. (2003). ‘A critique of theWorldHealthOrganisations evaluation of health systemperformance’. Health Economics 12:355366. Y. Roll, et al. (1991). ‘Controlling factor weights in data envelopment analysis’. IIE transactions 23:2–9. R. Rosenman & D. Friesner (2004). ‘Scope and scale inefficiencies in physician practices’. Health Economics 13:11:1091–1116. M. B. Rosenthal, et al. (2005). ‘Early Experience With Pay-for-Performance: From Concept to Practice’. Journal of The American Medical Association 294:1788– 1793. M. Rosko & J. Chilingerian (1999). ‘Estimating Hospital Inefficiency: Does Case Mix Matter?’. Journal of Medical Systems 23(1):57–71. M. D. Rosko (1999). ‘Impact of internal and external environmental pressure on hospital inefficiency’. Health Care Management Science 2:63–74. M. D. Rosko (2001a). ‘Cost efficiency of US hospitals: A stochastic frontier approach’. Health Economics 10:539–551. M. D. Rosko (2001b). ‘Impact of HMO penetration and other environmental factors on hospital X-efficiency’. Medical Care Research and Review 58:430–454. M. D. Rosko & J. A. Chillingerian (1999). ‘Estimating hospital inefficiency: does case-mix matter?’. Journal of Medical Systems 23:1:57–71. M. D. Rosko & R. L. Mutter (2008). ‘Stochastic frontier analysis and hospital inefficiency: A review of empirical issues and an assessment of robustness’. Medical Care Research and Review 65:131–166. M. D. Rosko & J. Proenca (2005). ‘Impact of network and system use on hospital X-inefficiency’. Health Care Management Review 30(1):69–79. P. Rouse & R. Swales (2006). ‘Pricing public health care services using DEA: Methodology versus politics’. Annals of Operations Research 145:265–280. Rouse, Paul Swales, Robert. 232

BIBLIOGRAPHY

I. Sahin & Y. A. Ozcan (2000). ‘Public sector hospital efficiency for provincial markets in Turkey’. Journal of Medical Systems 24(6):307–320. N. Sari (2003). ‘Efficiency outcomes of market concentration and managed care’. International Journal of Industrial Organization 21(10):1571–1589. P. Schmidt & R. Sickles (1984). ‘Production frontiers and panel data’. Journal of business and economic statistics 2:99–138. M. Segall, et al. (2002). ‘Economic transition should come with a health warning: the case of Vietnam’. Journal of Epidemiology and Community Health 56(7):497505. J. Sengupta (1990). ‘Transformation of stochastic DEA models’. Journal of Econometrics 46(1-2):7–38. J. Sengupta (2000). ‘Efficiency analysis by stochastic data envelopment analysis’. Applied Economic Letters 7(6):379–383. A. Sepehri, et al. (2003). ‘If they get sick, they are in trouble: health care restructuring, user charges and equity in Vietnam’. International Journal of Health Services 33(1):137 – 161. A. Sepehri, et al. (2006). ‘Does non-profit health insurance reduce financial burden? Evidence from the Vietnam living standard survey panel’. Health Economics 15:603–316. L. Simar & P. Wilson (1998). ‘Sensitivity analysis of efficiency scores: how to bootstrap in non-parametric frontier models’. Management Science 44(11):49– 61. L. Simar & P. Wilson (1999a). ‘Estimating and bootstrapping Malmquist indeces’. European Journal of Operations Research 115:459–471. L. Simar & P. Wilson (1999b). ‘Of course we can bootstrap DEA scores! But does it mean anything?’. Journal of Productivity Analysis 11:67–80. L. Simar & P. Wilson (2008). Statistical inference in non-parametric frontier models: recent developments and perspectives, chap. Chapter 4, pp. 421–521. Oxford University Press: New York. F. A. Sloan, et al. (1983). ‘Effects of teaching on hospital costs’. Journal of Health Economics 2:1–28.

233

BIBLIOGRAPHY

M. Smet (2007). ‘Measuring performance in the presence of stochastic demand for hospital services: an analysis of Belgian general care hospitals’. Journal of Productivity Analysis 27(1):13–29. Smet, Mike. P. Smith (1997). ‘Model misspecification in Data Envelopment Analysis’. Annals of Operations Research 73:233–252. P. Smith, et al. (2008). ‘Performance measurement for health system improvement: experiences, challenges and prospects’. Tech. rep., World Health Organization and European Observatory on Health Systems and Policies. M. Sommersguter-Reichmann (2000). ‘The impact of the Austrian hospital financing reform on hospital productivity: empirical evidence on efficiency and technology changes using a non-parametric input-based Malmquist approach’. Health Care Management Science 3:309–321. P. C. Sprivulis, et al. (2006). ‘The Association Between Hospital Overcrowding and Mortality Among Patients Admitted via Western Australian Emergency Departments’. Medical Journal of Australia 184:5:208–212. M. Staat (2006). ‘Efficiency of hospitals in Germany: a DEA-bootstrap approach’. Applied Economics 38(19):2255–2263. Staat, Matthias. L. Steinmann, et al. (2004). ‘Measuring and comparing the (in)efficiency of German and Swiss hospitals’. The European Journal of Health Economics 5(3):216–226. A. Street (2003). ‘How much confidence should we place in efficiency estimates?’. Health Economics 12(11):895–907. A. Street & R. Jacobs (2002). ‘Relative performance evaluation of the English acute hospital sector’. Applied Economics 34(9):1109–1119. S. Surrao, et al. (2002). ‘Hospital funding and services in Queensland’. Annual health review 24:1:99–120. L. Tauer (2001). ‘Input Aggregation and Computed Technical Efficiency’. Applied Economics Letters 8(5):295–297. A. Telyukov (2001). ‘Prospective Case-Based Payment for Hospitals: A Guide with Illustrations from Latin America’. Tech. rep., Health sector reform Initiative: LACHRS. A. Thiam, et al. (2001). ‘Technical efficiency in developing country agriculture: a meta-analysis’. Agricultural Economics 25:235243. 234

BIBLIOGRAPHY

R. Thompson, et al. (1990). ‘The role of multiplier bounds in efficiency analysis with application to Kansas Farming’. Journal of Econometrics 46(1/2):93–108. R. Thompson, et al. (1986). ‘Comparative site evaluations for locating a high energy physics labs in Texas’. Interfaces 16(6):35–49. A. Thorson & E. Johansson (2004). ‘Equality or equity in health care access: a qualitative study of doctors explanations to a longer doctors delay among female TB patients in Vietnam’. Health Policy 68(1):37–46. UN (2006). ‘World Population Prospects: The 2006 Revision’. Tech. rep., Department of Economic and Social Affairs: United Nation. UNDP (2003). ‘Health Care Financing for Vietnam’. Working paper 2, United Nations Vietnam: Hanoi. P. G. Uslu & T. Pham (2008). ‘Effects of changes in public policy on efficiency and productivity of general hospitals in Vietnam’. Tech. rep., University of East Anglia. V. Valdmanis (1992). ‘Sensitivity analysis for DEA models: an empirical example using public vs. NFP hospitals’. Journal of Public Economics 48:185–205. V. Valdmanis, et al. (2004). ‘Capacity in Thai public hospitals and the production of care for poor and nonpoor patients’. Health Services Research 39(6):2117–2134. Part 2. D. Vitaliano & M. Toren (1996). ‘Hospital Cost and Efficiency in a Regime of Stringent Regulation’. Eastern Economic Journal 22(2):161–175. A. Wagstaff & E. van Doorslaer (2003). ‘Catastrophe and impoverishment in paying for health care: with applications toVietnam1993–1998’. Health economics 12:921934. WB (2007). ‘Vietnam Development Report 2008: Social protection’. Tech. rep., Joint Donor Report to the Vietnam Consultative Group Meeting by the World Bank, the Asian Development Bank, the Department for International Development, the European Commission and German Development Cooperation. WB (2008a). ‘Health financing and delivery in Vietnam: the short and medium term policy agenda’. Tech. rep., The World Bank. WB (2008b). ‘Taking stock: An Update on Vietnam’s Recent Economic Developments’. Tech. rep., The World Bank: Vietnam. 235

BIBLIOGRAPHY

R. Webster, et al. (1998). ‘Comparing techniques for measuring the efficiency and productivity of Australian Private hospitals’. WORKING PAPER 98/3, Australian Bureau of Statistics. K. R. White & Y. A. Ozcan (1996). ‘Church ownership and hospital efficiency’. Hospital & Health Services Administration 41(3):297–310. A. Whitehead (2002). Meta-analysis of controlled clinical trials. John Wiley and Sons, Ltd. WHO (2000). ‘Health systems: improving performance’. Tech. rep., World Health Organisation. WHO (2006). ‘The World Health Report: Working together for health’. Tech. rep., The World Health Organisation. WHO (2008). ‘World Health Statistics’. Tech. rep., The World Health Organisation. WHO (2009). ‘World Health Statistics’. Tech. rep., The World Health Organisation. P. Wilson (2008). ‘FEAR: A software package for frontier efficiency analysis with R’. Socio-Economic Planning Sciences 42:247–254. A. Worthington (2000). ‘An empirical survey of frontier efficiency measurement techniques in health care services’. Working paper, School of Economics and Finance; Queensland University of Technology. A. C. Worthington (2004). ‘Frontier efficiency measurement in health care: A review of empirical techniques and selected applications’. Medical Care Research and Review 61(2):135–170. S. Yaisawarng & J. F. Burgress Jr (2006). ‘Performance based budgeting in public sector: an illustration from the VA health care system’. Health Economics 15:295– 310. K. Yong & A. Harris (1999). ‘Efficiency of Hospitals in Victoria under casemix funding: a stochastic frontier approach’. Working paper 92, Center for Health Program Evaluation. E. Zere, et al. (2006). ‘Technical efficiency of district hospitals: Evidence from Namibia using Data Envelopment Analysis’. Cost Effectiveness and Resource Allocation 4:5.

236

BIBLIOGRAPHY

E. Zere, et al. (2001). ‘Technical efficiency and productivity of public sector hospitals in three south African provinces’. South African Journal of Economics 69:2:336– 358. Y. Zhang & R. Bartels (1998). ‘The effect of sample size on the mean efficiency in DEA with an application to electricity distribution in Australia, Sweden and New Zealand’. Journal of Productivity Analysis 9:187–204. S. Zuckerman, et al. (1994). ‘Measuring hospital efficiency with frontier cost functions’. Journal of health economics 13(3):255–280.

237