54
Journal of Pain and Symptom Management
Vol. 38 No. 1 July 2009
Special Article
Improving Efficiency and Value in Palliative Care with Net Benefit Regression: An Introduction to a Simple Method for CostEffectiveness Analysis with Person-Level Data Jeffrey S. Hoch, PhD Centre for Research on Inner City Health, The Keenan Research Centre, Li Ka Shing Knowledge Institute, St. Michaels Hospital; Pharmacoeconomics Research Unit, Cancer Care Ontario; and Department of Health Policy, Management and Evaluation, University of Toronto, Toronto, Ontario, Canada
Abstract The objective of this article is to illustrate how to do cost-effectiveness analysis (CEA) using net-benefit regression and to explain how this method provides all of the benefits CEA can provide for improving efficiency and value in palliative care. We use a hypothetical data set with person-level data to demonstrate the net-benefit regression framework. Cost and effect data are combined with assumptions about willingness to pay to produce a net-benefit variable for each study participant. This net-benefit variable is the dependent variable in a net-benefit regression. In the simplest formulation, the regression coefficient on the treatment indicator variable estimates the difference in value between extra benefits and extra costs. The estimate and its confidence interval provide policy-relevant information. Net-benefit regression can be used with data from clinical trials or from administrative data sets. The results can be used to help develop policy, with an aim toward improving efficiency and value in health care. J Pain Symptom Manage 2009;38:54e61. Ó 2009 U.S. Cancer Pain Relief Committee. Published by Elsevier Inc. All rights reserved. Key Words Cost-effectiveness analysis, net-benefit regression, incremental net benefit, incremental cost-effectiveness ratio
Introduction State administrations and legislatures are seeking economies now and cannot be satisfied with promises of a better system five years hence. If health professionals cannot or will not contribute their expertise to this issue, others will have to make the Address correspondence to: Jeffrey S. Hoch, PhD, St. Michael’s Hospital, 30 Bond Street, Toronto, Ontario M5B 1W8, Canada. E-mail:
[email protected] Accepted for publication: April 23, 2009. Ó 2009 U.S. Cancer Pain Relief Committee Published by Elsevier Inc. All rights reserved.
hard and painful decisions themselves. Without question, economies will be imposed, one way or another. Comments in 1972 by Earl W. Brian, MD, Director, California Department of Health Care Services.1
More than 35 years later, there are still concerns about budget shortages seemingly predestined by surpluses of expensive new medicines and technologies. Gone are health care’s halcyon days where decision makers could approve only new treatments that were 0885-3924/09/$esee front matter doi:10.1016/j.jpainsymman.2009.04.010
Vol. 38 No. 1 July 2009
Improving Efficiency and Value in Palliative Care
more effective and deny those that were not. Skyrocketing health care costs have made it impossible to provide access to all new treatments that are ‘‘better.’’ Health care budgets are simply not large enough to give everyone the most advanced medical care. The implications of this have been recognized for more than three decades: If the best medical care cannot be given to everyone, then medical care must be rationed in some manner. This is a fact of life. It cannot be altered by wishful thinking or pious statements that the highest quality medical care must be made available to everyone without restriction. We may not be happy with this situation, but it is better to face it directly than to pretend that medical care does not have to be rationed.2 Thus, the challenge is how to say ‘‘no’’ to new things that do not provide good value. Americans are neither fond of rationing nor cost-containment exercises.3e5 However, words like efficiency and value resonate and may be more acceptable in general. Economic ‘‘duality theorems’’ show how cost containment can be reframed as ‘‘efficiency.’’ Some of the earliest mathematical representations of cost-effectiveness analysis (CEA)6,7 portray health care decision makers facing a constrained optimization problem: How to maximize benefit given a fixed budget? At least in theory, CEA has the potential to improve efficiency and value in health care by estimating the extra cost to get one more unit of effect; according to proponents, this ‘‘provides an index by which priorities may be set.’’8 Nevertheless, historically, CEA has not been used frequently in the United States.4 In fact, since its first publication in 1986, only 15 articles in the Journal of Pain and Symptom Management have used the phrase ‘‘cost-effectiveness’’ (results of a PUBMED search on September 5, 2008 of the form (‘‘cost-effectiveness’’) AND (‘‘Journal of Pain and Symptom Management’’[Jour])) out of over 2646 published pieces. However, ‘‘it is likely that cost-effectiveness will receive more attention by policy experts in the near future,’’5 lending credence to the notion that CEA may be transformed from ‘‘an academic curiosity to an essential tool for health care decision making.’’9 By embracing CEA, palliative care researchers could take
55
a leadership role in the production and use of research describing efficiency and value in health care. The objective of this article is to illustrate how to do CEA using net benefit regression,10 a recently developed technique that allows easy calculation of cost-effectiveness estimates and uncertainty measures (e.g., confidence intervals [CIs]). A brief overview of CEA is provided before using a hypothetical data set with individual person-level cost and effect data to demonstrate net benefit regression and comment on how the results can be used to support arguments for palliative care treatments and interventions that improve efficiency or provide good value.
Economic Evaluation Methods There are many different types of economic evaluations besides CEA, such as cost-benefit analysis, cost-utility analysis, and cost-minimization analysis.11 An easy way to distinguish one from the other is by examining the ‘‘measure’’ of patient outcome. In cost-benefit analyses, there are typically many outcomes and all outcomes are valued in dollars.12 This type of analysis is not very popular in health care. Although cost-benefit analyses assess whether a treatment is worthwhile, CEA determines the relative efficiency of a treatment alternative. In CEA, a single outcome is analyzed (e.g., pain-free days). A second form of CEA is cost-utility analysis, which values outcomes in quality-adjusted life years, equal to the number of life years remaining multiplied by a weighting factor reflecting quality of life.13,14 In cost-minimization analysis, only costs are compared, because patient outcomes are assumed to be identical. Perhaps, the most popular economic evaluation method is CEA. The goal of CEA is to quantify the trade-off between resources used and outcomes gained. The attractiveness of CEA lies in its simplicity; one patient outcome is expressed in its natural units and compared with the ‘‘resources used.’’ As ‘‘resources used’’ are measured in dollars, the summary measure of a CEA involves cost and a patient outcome; typically, the summary measure reported is an incremental cost-effectiveness ratio (ICER). The ICER estimates the ratio of the extra cost to the extra effect (i.e., DC/DE ).
56
Hoch
Estimates of Cost-Effectiveness with PersonLevel Data Considering an economic analysis of ‘‘usual care’’ (t ¼ UC) and ‘‘new treatment’’ (t ¼ TX) using individual person-level data, one can estimate DC and DE using sample means calculated as DC ¼ C TX C UC and DE ¼ E TX E UC , where C t and E t are the sample cost and effect averages for subjects receiving treatment option t. To express the extra cost associated with the extra gain at a per-unit level, analysts b ¼ DC=DE ¼ C TX C UC . estimate the ICER as R E TX E UC
b < WTP, where WTP represents the decision If R maker’s willingness to pay for an additional health outcome or effect, then the new treatment is said to be cost-effective. Although in theory, WTP should be derived from a decision maker’s constrained optimization problem, most would agree that in the real world, WTP is probably affected by political considerations, public attitudes, organizational considerations, and satisfactions of the patients and the health professionals involved in the program.6 A simple, insightful alternative to the ICER is the net benefit approach.15,16 The net benefit calculation determines whether a new treatment meets (or surpasses) the decision maker’s expectations of value for money. These expectations are quantified using WTP for a one-unit gain of outcome. For example, to calculate the average net benefit of new treatment in dollars (i.e., NBTX ), one must subtract the value of the average cost of new treatment (i.e., C TX ) from the monetary value of the average effect (i.e., WTP E TX ). The net benefit for usual care can be made in an analogous fashion as NBUC ¼ WTP E UC C UC . The incremental net benefit (INB) statistic can be estimated as b ¼ NBTX NBUC . Likewise, it is possible to B b ¼ WTP DEDC, where DE and DC write B are the differences in average effect and averb > 0, the extra age cost, respectively. When B outcome produced by new treatment is worth b < 0, the extra the extra cost; conversely, when B benefit is not worth the extra cost. By calcub instead of R b , both estimation and inlating B b is ference are greatly simplified because B b is not) and the CI for B b is easy unbiased ( R b ). to calculate (unlike the one for R Of course, when using the net benefit approach, WTP must be specified; however, this b is also true for any decision based on R
Vol. 38 No. 1 July 2009
because a new treatment is only deemed b < WTP. Thus, one cannot ‘‘cost-effective’’ if R avoid specifying WTP, which plays an explicit role in the INB approach and an implicit role in the ICER approach.
Net Benefit Regression Recent methodological work has developed a regression framework for the net benefit approach.10 Under this framework, each subject’s net benefit (NBi) is computed from the observed data as WTP effecti$ costi where effecti and costi are the data for the ith person’s effect and cost, respectively, and WTP is a willingness-to-pay number that must be specified. In its simplest form, net benefit regression involves fitting the following simple linear regression model: NBi ¼ b0 þ bTX TXi þ ei
ð1Þ
th
where TXi is the i person’s treatment indicator (TXi ¼ 1 for new treatment and 0 for usual care) and ei is a stochastic error term. Equation (1) is typically fit several times, each time with a different value of WTP (in theory, decision makers know their WTP and consider the results with this value of WTP). To explore if cost-effectiveness varies by patient subgroup, Equation (1) can be enhanced with a vector of subject characteristics (Xi) and interaction terms between subject characteristics and the treatment indicator (Xi$TXi). Alternatively, stratified regressions can be run (e.g., one for people with cancer-related fatigue and one for people without). The ordinary least squares (OLS) estimate of bTx in Equation (1) is simply b b TX ¼ NBTX NBUC ¼ ðWTP E TX C TX Þ ðWTP E UC C UC Þ b: ¼ WTP DEDC ¼ B Thus b b TX , the estimate of bTX in Equation (1), estimates the INB. Figure 1 provides intuition behind why the intercept term b b 0 ¼ NBUC and why the slope term b b TX ¼ NBTX NBUC . From the algebra, it is clear that b b TX ¼ NBTX . When the assumptions for b0þb OLS hold, one can use the probability that b b TX > 0 for the y -axis and WTP for the x-axis to generate a cost-effectiveness acceptability
Vol. 38 No. 1 July 2009
Improving Efficiency and Value in Palliative Care
57
Net benefits are measured on a scale from small (on the left) to large (on the right).
a
Net benefits are reverse scaled from large (on the left) to small (on the right).
b
Average net benefit for TX is levitated…
the graph rotates…
and regression will fit a line with a slope equal to the difference in the average net benefits for the UC and TX groups. The y-intercept term equals the average net benefit for UC, the slope equals the incremental net benefit and the sum of the intercept and the slope equals the average net benefit for TX.
Fig. 1. a) Net benefit for the new treatment (TX) and usual care (UC) groups, first presented with usual scaling followed by reverse scaling. b) Net benefit for the TX and UC groups, rotated to allow a regression to fit straight line from the average net benefit for UC to the average net benefit for TX.
curve (CEAC).17e19 The CEAC, highlighting the relationship between the probability of cost-effectiveness and the unknown WTP, has been debated as a way of summarizing the results of a CEA.20e25
Hypothetical Example Data We created a hypothetical data set of 18 people (nine receiving new palliative care treatment and nine receiving ‘‘usual care’’) to demonstrate net benefit regression. Cost and effect data were combined with assumptions about WTP to produce a net benefit variable for each study participant. For example, when WTP ¼ $50,000, the net benefit for Person 1 who received usual
care (see the table below Fig. 2) is equal to $50,000 0.15 $13,000 ¼ $7500 $13,000 ¼ $5500. Hence, for Person 1, the extra costs outweigh the extra benefits by $5500. For net benefit regression, we created each subject’s net benefit variable to use as the dependent variable.
Results The sample cost and effect data are reported and illustrated in Table 1 and Fig. 2. The new treatment is estimated to cost $5000 more and yield 0.10 more units of effect. If we assume that the patient outcome chosen for ‘‘effect’’ was life years, then the ICER estimate is b ¼ DC/ $50,000 per additional year of life ( R DE ¼ $5000/0.10 ¼ $50,000). With this ICER estimate, there are unanswered questions about its statistical uncertainty and whether
58
Hoch
Vol. 38 No. 1 July 2009
Table 1 Sample Statistics Based on Hypothetical Data
20000
_ CTX
Treatment Arm
Mean
SD
SE
Comparison (n ¼ 9) Cost Effect
$13,000 0.17
1224.75 0.012
408.25 0.004
New treatment (n ¼ 9) Cost Effect
$18,000 0.27
1224.75 0.012
408.25 0.004
$5000 0.10
d d
577.35 0.01
15000
cost
_ CUC
10000
5000
0 .15
_ EUC
.2
.25
_ ETX
.3
Incrementsa Cost difference (DC ) Effect difference (DE )
effect o = Usual Care
= New Treatment
Data stratified by treatment allocation (TX = 0 is usual care; TX = 1 is new treatment)
Incremental net benefits ($)
Net monetary benefit in $ (SE)
Willingness to pay 0 50,000 100,000
5,000 (577.35) 0 (645.50) 5000 (816.50)
SD ¼ standard deviation; SE ¼ standard error. a Differences are calculated as new treatment values minus comparison values.
Fig. 2. Cost and effect data by treatment-allocation status. o ¼ usual care and x ¼ new treatment. The table below the figure gives the data stratified by treatment-allocation status (tx ¼ 0 is usual care and tx ¼ 1 is new treatment).
the new treatment is cost-effective. Facilitating the answers to these questions are the results from net benefit regression. Table 2 presents regression results for the simple linear regression specified in Equation (1) as NB ¼ b0 þ bTX TX þ e where TX ¼ 1 if the person received new treatment and 0 if usual care, and e is the error term. To illustrate the sensitivity of our estimate of bTx with respect to our choice of WTP, we graphed b b TX (the INB estimate) and its confidence interval against WTP in Fig. 3. Both in Table 2 and in Fig. 3, we show WTP values that produce confidence intervals that just include zero.
Discussion of the Example The regression results in Table 2 show WTP ranges in which the economic conclusions are sensitive (and WTP ranges in which the conclusions are not). For example, for WTP < $37,000, the confidence interval for b b TX (the INB
estimate) only contains negative values. This means that new treatment is not cost-effective if decision makers value an extra unit of patient outcome less than $37,000. For WTP values > $65,000, the confidence interval for bTX (the INB estimate) only contains positive values. This means that new treatment is cost-effective if decision makers value an extra unit of patient outcome more than $65,000. Thus, new treatment is cost-effective if decision makers value an extra year of life more than $65,000, but not cost-effective if decision makers value an extra year of life less than $37,000. For WTP values between $37,000 and $65,000, the 95% confidence interval includes zero. This means the data do not support a definitive conclusion about whether the new treatment is cost-effective. Figure 3 illustrates the estimate of the INB with the upper and lower 95% confidence limits included. The value of b b TX increases with WTP, consistent with DE > 0. The y-intercept is negative, consistent with DC > 0. Thus, new treatment is both more costly and more effective. The estimate b ¼ DC=DE) is marked by of the ICER (i.e., R the x-intercept at $50,000; this point occurs b . The WTP values for the upwhere WTP ¼ R per and lower 95% confidence limits that just include $0 (which are $37,000 and $65,000, respectively in this case) indicate a parametric b (equivalent to using confidence interval for R Fieller’s theorem).
59
0.1705 4.49 0.0500
The data best support the hypothesis that the ICER is $50,000 for one more unit of patient outcome. The 95% confidence interval for that estimate is $37,000e$65,000. The conclusions from our results are not sensitive to WTP values less than $37,000 (do not use new treatment) or WTP values more than $65,000 (do use new treatment).
Figures rounded to the nearest whole number. a Statistically significant at the 5% level.
0.8132 75.00 F
2023a (3053, 993) 1457a (0, 2914) 6718a (7641, 5796) 1305a (2610, 0) 4000a (2776, 5224) 5000a (3269, 6731) Constant term New treatment indicator (TX ¼ 1 0 Yes; TX ¼ 0 0 No)
13,000a (13,865, 12,135) 5000a (6224, 3776)
4500a (5468, 3532) 0 (1368, 1368)
WTP ¼ $64,569 WTP ¼ $100,000
WTP ¼ $50,000
WTP ¼ $36,952
Improving Efficiency and Value in Palliative Care
WTP ¼ $0 Variables
‘‘Confidence Intervals ¼ $0’’ WTPdINB Estimate (95% CI) ‘‘Extreme’’ WTP ValuesdINB Estimate (95% CI)
‘‘ICER Estimate’’ b ¼ WTPdINB Estimate R (95% CI)
Net Benefit Regression Results by WTP
Table 2 Simple Linear Regression Results with Hypothetical Data (n ¼ 18)
Vol. 38 No. 1 July 2009
How would one know the value of a new treatment or intervention without knowing the extra cost in relation to the extra benefit? CEA answers this question based on the simultaneous analysis of both cost and effect data. The contention that palliative care treatments generate precious benefits that are difficult to measure condemns the efforts to use economic evidence to support the use of these novel interventions. The problem is that if one cannot show evidence of benefit (because the data do not exist), why would anyone want to pay extra for it? On the other hand, if one can save money giving patients less health care in a cheaper setting, who cares?dunless the patient is better off. Health maintenance organizations developed a nasty reputation for cost cutting. It may not be in the best interest of palliative care advocates to appear to be emulating this practice. Clearly, data from randomized controlled trials fit neatly into the net benefit regression framework, but so do the data from large administrative data sets. Regression techniques to analyze observational data can be used directly in net-benefit regression. Of course, more complex analytical strategies are available for the analysis of cost-effectiveness data, and there is growing interest in improving the quality of economic evaluations done with individual person-level data.26(Chap.8) Already, there has been substantial methodological work to improve the quality of CEAs done with person-level data. Researchers have studied how best to handle missing data,27 skewed cost data,28,29 between-center differences,30 multilevel models,31 and seemingly unrelated regression models.32 An important advantage of net benefit regression is that it allows all of the methods that have been developed for regression analysis (e.g., model fit
Incremental net benefit (INB)
60
Hoch
10000
INB estimate 95% confidence intervals 5000
5000
4000 3000 2914 2000 1000 0 -1000
0 -2000
0 0
-3000 -2610
-4000
-5000
Vol. 38 No. 1 July 2009
researchers have the opportunity to lead the way in exploring efficiency and value in health care, informing policy debates about how scarce health care resources can best be spent. Efficiency is important and net benefit regression is a simple way to create cost-effectiveness results that can be used to inform policy.
-5000
0
20000
40000
60000
80000
100000
Willingness to pay (WTP)
Fig. 3. Estimates of INB and its 95% confidence interval as a function of different willingness-to-pay values.
diagnostics, advanced estimation and inference techniques, and others) to be directly applied to economic evaluation. Another advantage of net benefit regression in its simplest form is that most researchers are familiar with OLS regression. If key methods can be explained in a familiar regression framework, it seems likely that their widespread adoption might be facilitated. Efforts to explain how to do CEA with simple linear regression using OLS represent an attempt at more universal knowledge transfer.
Conclusion Scarcity in health care forces choice about what resources will be used to treat whom. By examining both the extra costs and extra effects of a novel treatment, economic evaluations estimate the ‘‘value for money.’’ It may seem harsh to deny treatments that are more effective on the basis of the results of a CEA. Is it worse, however, than wasting precious health care resources on inferior investments in health care? The costs to consider are not only the money that is spent, but also the potential patient outcomes forgone. Commentators have opined that ‘‘limits on health-care resources mandate that resourceallocation decisions be guided by considerations of cost in relation to expected benefits.’’8 Moreover, Garber9 suggests the erosion of commercial health insurance, and the growing burden of public health insurance programs may encourage greater use of CEA to complement evidence-based criteria. Palliative care
References 1. Brian EW. Further thoughts on patient incentives. Health Serv Res 1972;7(2):150e153. 2. Kaplan RS, Lave LB. Further thoughts on patient incentives. Health Serv Res 1972;7(2): 148e150. 3. Teixeira R. Happy with health care? Int J Health Serv 2000;30(3):581e584. 4. Neumann PJ. Using cost-effectiveness analysis to improve health care: opportunities and barriers. New York: Oxford University Press, 2005. 5. Ginsburg ME. Cost-effectiveness: will the public buy it or balk? Health Aff (Millwood) 2004; Suppl Web Exclusives; W4-297eW4-299. 6. Torrance GW, Thomas WH, Sackett DL. A utility maximization model for evaluation of health care programs. Health Serv Res 1972;7(2):118e133. 7. Weinstein M, Zeckhauser R. Critical ratios and efficient allocation. J Public Econ 1973;2(2): 147e157. 8. Weinstein MC, Stason WB. Foundations of costeffectiveness analysis for health and medical practices. N Engl J Med 1977;296(13):716e721. 9. Garber AM. Cost-effectiveness and evidence evaluation as criteria for coverage policy. Health Aff (Millwood) 2004; Suppl Web Exclusives; W4-284eW4-296. 10. Hoch JS, Briggs AH, Willan AR. Something old, something new, something borrowed, something blue: a framework for the marriage of health econometrics and cost-effectiveness analysis. Health Econ 2002;11(5):415e430. 11. Hoch JS, Dewa CS. An introduction to economic evaluation: what’s in a name? Can J Psychiatry 2005;50(3):159e166. 12. Weisbrod BA. Benefit-cost analysis of a controlled experiment: treating the mentally ill. J Hum Resour 1981;16(4):523e548. 13. Gold MR, Stevenson D, Fryback DG. HALYS and QALYS and DALYS, oh my: similarities and differences in summary measures of population health. Annu Rev Public Health 2002;23:115e134. 14. McGregor M, Caro JJ. QALYs: are they helpful to decision makers? Pharmacoeconomics 2006; 24(10):947e952.
Vol. 38 No. 1 July 2009
Improving Efficiency and Value in Palliative Care
15. Stinnett AA, Mullahy J. Net health benefits: a new framework for the analysis of uncertainty in cost-effectiveness analysis. Med Decis Making 1998; 18(2 Suppl):S68e80. 16. Tambour M, Zethraeus N, Johannesson M. A note on confidence intervals in cost-effectiveness analysis. Int J Technol Assess Health Care 1998; 14(3):467e471. 17. Hoch JS, Rockx MA, Krahn A. Using the net benefit regression framework to construct costeffectiveness acceptability curves: an example using data from a trial of external loop recorders versus Holter monitoring for ambulatory monitoring of ‘‘community acquired’’ syncope. BMC Health Serv Res 2006;6(1):68. 18. Fenwick E, Marshall DA, Levy AR, Nichol G. Using and interpreting cost-effectiveness acceptability curves: an example using data from a trial of management strategies for atrial fibrillation. BMC Health Serv Res 2006;6(1):52.
61
24. Fenwick E, Briggs A. Cost-effectiveness acceptability curves in the dock: case not proven? Med Decis Making 2007;27(2):93e95. 25. Groot Koerkamp B, Hunink MG, Stijnen T, et al. Limitations of acceptability curves for presenting uncertainty in cost-effectiveness analysis. Med Decis Making 2007;27(2):101e111. 26. Drummond MF, Sculpher MJ, Torrance GW, O’Brien BJ, Stoddart GL. Methods for the economic evaluation of health care programmes, 3rd ed. Oxford: Oxford University Press, 2005. 27. Manca A, Palmer S. Handling missing data in patient-level cost-effectiveness analysis alongside randomised clinical trials. Appl Health Econ Health Policy 2005;4(2):65e75. 28. Basu A, Arondekar B, Rathouz P. Scale of interest versus scale of estimation: comparing alternative estimators for the incremental costs of a comorbidity. Health Econ 2006;15(10):1091e1107.
19. Lothgren M, Zethraeus N. Definition, interpretation and calculation of cost-effectiveness. Health Econ 2000;9(7):623e630.
29. Briggs A, Nixon R, Dixon S, Thompson S. Parametric modelling of cost data: some simulation evidence. Health Econ 2005;14(4):421e428.
20. Briggs A, Fenn P. Confidence intervals or surfaces? Uncertainty on the cost-effectiveness plane. Health Econ 1998;7(8):723e740.
30. Nixon R, Thompson S. Methods for incorporating covariate adjustment, subgroup analysis and between-centre differences into cost-effectiveness evaluations. Health Econ 2005;14(12):1217e1229.
21. Fenwick E, Claxton K, Sculpher M. Representing uncertainty: the role of cost-effectiveness acceptability. Health Econ 2001;10(8):779e787. 22. Fenwick E, O’Brien B, Briggs A. Cost-effectiveness acceptability curvesdfacts, fallacies and frequently. Health Econ 2004;13(5):405e415. 23. van Hout B, Al M, Gordon G, Rutten F. Costs, effects and C/E-ratios alongside a clinical trial. Health Econ 1994;3(5):309e319.
31. Manca A, Rice N, Sculpher M, Briggs A. Assessing generalisability by location in trial-based costeffectiveness analysis: the use of multilevel models. Health Econ 2005;14(5):471e485. 32. Willan A, Briggs A, Hoch J. Regression methods for covariate adjustment and subgroup analysis for non-censored cost-effectiveness data. Health Econ 2004;13(5):461e475.
