econometrics: methods & applications

Editor: M. Kenan TERZİOĞLU Co-Editor: Süreyya DAL

ECONOMETRICS: METHODS & APPLICATIONS

ISBN 978-605-344-671-2

9 786053 446712

ECONOMETRICS: METHODS & APPLICATIONS Editor: M. Kenan TERZİOĞLU Co-Editor: Süreyya DAL



© Bu kitabın Türkiye’deki her türlü yayın hakkı Gazi Kitabevi Tic. Ltd. Şti’ne aittir, tüm

hakları saklıdır. Kitabın tamamı veya bir kısmı 5846 sayılı yasanın hükümlerine göre, kitabı yayınlayan firmanın ve yazarlarının önceden izni olmadan elektronik, mekanik, fotokopi ya da herhangi bir kayıt sistemiyle çoğaltılamaz, yayınlanamaz, depolanamaz.

Kapak Tasarımı Nurullah Arda TURAN Düzenleme Özlem Nur KALEÖZÜ Baskı İlksan Matbaası Ltd. Şti. İvedik Org. San. Bölgesi Ağaç İşleri San. Sit. 521. Sok. No: 35 İvedik / ANKARA Tel: (0312) 394 39 55 Baskı Yılı 2018

ISBN 978-605-344-671-2

Gazi Kitabevi Tic. Ltd. Şti. Dögol Caddesi No: 49/B Beşevler, ANKARA Tel: (0 312) 223 77 73 – 213 32 82 • Faks: (0 312) 215 14 50

[email protected] www.gazikitabevi.com.tr

Editor, Dr. M. Kenan Terzioğlu is an assistant professor and co-head of department of the Econometrics Department, Faculty of Economics and Administrative Sciences, at Trakya University, Edirne, Turkey. After graduating from Hacettepe University Statistics Department in 2005, he received his MSc. degree in 2008 from Actuarial Sciences Department, Faculty of Science, Hacettepe University. He studied at Department of Econometrics and Operations, Tilburg University, Netherlands between 2008 and 2009. In 2016, he got his Ph.D. degree in Econometrics from Department of Econometrics at Gazi University, Faculty of Economics and Administrative Sciences. After working in Department of Actuarial Sciences between 2006 and 2008 as a research assistant, he worked as Risk Analyst Assistant (Assistant Specialist) in Market Risk and Asset-Liability and Balance Sheet Risks in Risk Management Department, Ziraat Bank. Between 2009 and 2016, he worked as a Research Assistant at the Department of Econometrics at Gazi University. Since 2016, he has been working as an Assistant Professor in the Department of Econometrics, Trakya University. The fields of his studies are time series analysis, financial-actuarial risk management, Markov Processes and valuation methods.

Co-Editor, Süreyya Dal is a research assistant at the Department of Economics, Faculty of Economics and Administrative Sciences, Trakya University, Edirne, Turkey. She received his Bachelor degree from the Faculty of Economics and Administrative Sciences, Gazi University, Ankara, Turkey, and Master of Research in Econometrics, Gazi University, Ankara, Turkey. Currently, she is a PhD candidate in Econometrics, Uludağ University, Bursa. She was a visiting PhD scholar in Southern Denmark University for 5 months. She has published several articles and conference proceedings. She received many scholarships from international establishments for her studies. Recent work by Dal has looked at issues involving the analysis of categorical data. Previous research has looked at poverty, and methodological issues in the measurement of the poverty.

Contents Preface

vii

Chapter 1

Spatial Heterogeneity in Hedonic Pricing Models: The Housing Market in Istanbul, Turkey Sinem Güler Kangallı Uyar and Nihal Yayla

Chapter 2

The Long-Run Relationship between Hedonic House Prices and Consumer Prices: ARDL Bounds Testing Approach Havvanur Feyza Erdem and Nebiye Yamak

Chapter 3

Introduction to Seasonal Unit Root Processes Mehmet Özmen and Sera Şanlı

Chapter 4

Study on Relative Efficiency with Data Envelopment Analysis: Metropolitan Municipalities in Turkey Mine Aydemir and Nuran Bayram Arlı

Chapter 5

Monitoring of a Production Process with CUSUM and EWMA Quality Control Charts Hakan Eygü

Chapter 6

VAR Model Based Empirical Mode Decomposition: A Frame of Forecasting Performances Emrah Gülay

Chapter 7

Ridge Type Estimation in the Zero-Inflated Negative Binomial Regression Bahadır Yüzbaşı and Yasin Asar

Chapter 8

Cliometric Perspective for Stock Market Reactions to Wars and Political Risks: Evidence from a Falling Empire Avni Önder Hanedar and Elmas Yaldız Hanedar

1-18

19-30

31-46

47-62

63-80

81-92

93-104

105-126

Chapter 9

Determinants of Poverty in Turkey Süreyya DAL

127-140

vii

Preface The importance of experimental economics and econometric methods increase with each passing day as data quality and software performance develops. The transformation towards an information-focused society has begun with the influence of globalization and easy accessibility of technology. Along with this transformation, determining complete and satisfactory statistical results that can be obtained only by implementing appropriate econometric methods is becoming a necessity. Furthermore, new econometric models are developed by diverging from earlier cliché econometric models with the emergence of specialized fields of study. The studies within the scope of this book draw attention to various new developments in econometrics. This reference book is compiled from the articles about research areas of esteemed academician with an emphasis on their fields of interest. This book aims to be an introduction to the new econometric methods as well as a reference for graduate students. The reader is provided with an additional learning outcome by touching upon the discussions on the deficiencies and advantages as well as the basic features of the methods deliberated within the scope of the book. This book, which is expected to be an extensive and useful reference by bringing together some of the developments in the field of econometrics, also contains quantitative example and problem sets. We thank all the authors and reviewers who contributed to this book with their studies that provide extensive and accessible explanations of the existing econometric methods. Publisher and editors are not responsible for the conclusions or ideas contained in the sections of the book. We thank the process manager of Gazi Kitabevi in the successful completion and publishing of this book. Asst. Prof. Dr. Mehmet Kenan TERZİOĞLU Faculty of Economics and Administrative Sciences, Department of Econometrics, Trakya University

Chapter 1

Spatial Heterogeneity in Hedonic Pricing Models: The Housing Market in Istanbul, Turkey Sinem Güler KANGALLI UYAR 1 Nihal YAYLA 2

Abstract. Products in housing markets exhibit heterogeneous properties in that they have different structural, physical and location characteristics. Because of the segmented structure of housing markets, the relationship between home prices and characteristics might vary by location. Therefore, in order to investigate this relationship within the context of the hedonic pricing theory, a geographically weighted regression model was used that allows the relationships to change based on location and allowed for flexibility in determining functional forms. The analysis results indicate that there are nonparametric relationships between home prices and their characteristics. Moreover, these relationships differ for each sub-market. Keywords Hedonic Price Theory, Housing Market, Spatial Heterogeneity, GWR Model

1

Asst. Prof. Dr., Pamukkale University, Faculty of Economics and Administrative Sciences, Department of Econometrics, [email protected], 2 Prof. Dr., Pamukkale University, Faculty of Economics and Administrative Sciences, Department of Economics, [email protected]

2

Spatial Heterogenetiy in Hedonic Pricing Models: The Housing Market in Istanbul, Turkey

I. Introduction Housing markets have a structure that consists of sub-markets as a result of supply and demand segmentation based on home characteristics, and each submarket has a different price structure as a result of unique supply and demand conditions. The segmented structure of housing markets reflects that home prices and their characteristics are closely related, but this situation complicates housing pricing. In order to resolve the complex relationship between the prices and characteristics of heterogeneous goods, such as homes, the hedonic pricing approach was developed. This approach analyses the characteristics of heterogeneous goods so as to determine the effect of each characteristic on price; in other words, it is an approach that enables estimating implicit prices. When the theoretical foundations of the hedonic pricing approach have been established, the number of studies related to this approach for the markets of heterogeneous goods has increased considerably. Particularly, there are many studies on hedonic real estate pricing in the literature. In these studies, it is possible to conclude that housing characteristics are generally classified into three major groups, structural, location, and neighbourhood, although characteristics that influence housing prices vary across countries, territories and even cities (for example, Anglin and Gencay 1996; Pavlov 2000; Farber and Yeates 2006; Alkay 2008; Caglayan and Eban 2009; McMillen and Redfearn 2010; Koramaz and Dokmeci 2012). Some of the studies have found that housing markets are segmented by characteristics (for example, Straszheim 1975; Goodman 1981; Rothenberg et al., 1991; Maclennan and Tu, 1996; Watkins, 2001; Brourassa et al., 2003). The segmented structure of housing markets is important in terms of the econometric method that will be used to investigate the relationship between home prices and their characteristics. This relationship might differ by location because of that segmented market structure. From the econometric point of view, this kind of relationship causes spatial heterogeneity. In this case, instead of using econometric models that assume the same relationship for all locations, econometric models which assume the relationships differ by location might be used. Another important point that requires attention in modelling the relationships between home prices and their characteristics is the spatial dependency that arises from prices as a result of the adjacency effect. This dependency is defined as spatial autocorrelation in econometrics.

Econometrics: Methods & Applications

3

Ordinary least squares (OLS) regression models do not take spatial effects into consideration. In other words, spatial heterogeneity and spatial autocorrelation are not considered during estimation processes. In contrast, spatial lag models, spatial error models and spatial Durbin models, which are generally referred to as parametric spatial regression models, consider spatial effects through a spatial weight matrix. Nevertheless, these models only provide information regarding whether there exists a spatial effect or not, and they do not give any information on how the relationship differs according to location. In other words, these models do not consider the segmented structure of housing markets, and they assume the same relationship coefficient for all locations. One of the nonparametric spatial regression models, the geographically weighted regression (GWR) model, considers the spatial effects defined as spatial heterogeneity and spatial autocorrelation. In the context of housing markets, spatial heterogeneity might arise as the result of the segmented structure of housing markets. On the other hand, the adjacency effect might cause the spatial autocorrelation. The GWR model does not make a priori assumptions about the functional form of the relationships. Rather, that form is determined in a flexible way based on the distribution of the dataset. In this model, as different from parametric spatial models, spatial effects are taken into consideration through the coordinate variables without using a spatial weight matrix. The GWR model is a nonparametric regression model that is specifically constructed to investigate relationships that vary by location. Additionally, it provides the opportunity to interpret the relationships in different segments of the distribution of a dependent variable. The purpose of this study is to acquire more information on the demand side of Istanbul’s housing market. For this aim, the marginal effects of the characteristics of 2838 homes in 39 different counties were estimated for the last quarter of 2013 using the GWR model and the hedonic pricing approach. The GWR model estimation results indicate that the relationships between housing prices and characteristics change by location. In other words, Istanbul’s housing market is segmented by location characteristics. This result reveals that location is one of the most important determinants of home prices. It has been observed that in the most studies related to Istanbul’s housing market use the parametric models which do not take into account spatial effects or, as a consequence, the segmented structure of the market. Unlike the others, this study examines the relationship between home prices and characteristics by taking into account spatial effects. In this case, it can be said that nonparametric

4


regression models that allow the parameters vary by location might be used to estimate the relationship between home prices and characteristics. In this study, the second and third sections explain the hedonic pricing approach and the GWR model, respectively. The fourth section defines the data set and the variables that were used in the analysis, and the fifth section comprises the results obtained through the GWR model within the framework of the hedonic pricing approach. Finally, in the sixth section, along with the analysis results, essential inferences are made about Istanbul’s housing market.

II. Hedonic Pricing Theory Housing is multidimensional because of its characteristics such as immobility, durability, and heterogeneity. Its market structure consists of sub-markets because housing supply and demand are segmented based on these characteristics. This situation indicates that home prices and their characteristics are closely related, and this notably complicates house pricing. In order to resolve this problem, the hedonic housing pricing approach was introduced. Rosen (1974) defines hedonic prices as the implicit prices of individual home characteristics and reports that implicit prices are related to the observed prices of different products and their characteristics. Accordingly, the hedonic pricing approach enables separating out the individual characteristics of goods and determining the impact of each individual characteristic on price. In other words, it is an approach that enables estimating implicit prices. In this approach, instead of estimating goods’ prices as a whole, the prices of their individual characteristics are estimated, and this approach is only applied to heterogeneous goods. The basic form of the standard hedonic pricing model that was developed by Rosen (1974) that is, the functional relationship between the price of the ith heterogeneous good and the good’s characteristics by xi vector can be expressed as: P = Xβ + ε

(1)

If this equation is considered within the scope of housing markets, P refers the home price vector, ε , the independent and normally distributed error terms and , X , the home characteristics vector. The partial derivative of a home’s composite price with respect to each characteristic indicates a marginal contribution of each additional characteristic (Rosen 1974).


5

III.Geographically Weighted Regression Model The ordinary least squares (OLS) model does not take consideration spatial effects. In this model, the variation of parameters by location is not allowed, and the functional forms among variables cannot be determined flexibly. Similarly, the models that consider spatial dependency through spatial weight matrixes and that are generally referred to as parametric spatial models also do not allow for variation in parameters by location. In addition, in these models, functional form among variables is determined based on a priori knowledge of economic theory. These models are also referred to as global models because they do not allow the variation in parameters by location. (Yrigoyen 2007; Charlton and Fotheringham 2009). Thus, the models which consider the spatial effects defined as spatial heterogeneity and spatial autocorrelation are required to use. The GWR model, which was applied by McMillen (1996) and Brunsdon, Fotheringham and Charlton (1996) for the first time, is a recently common tool for modelling spatial heterogeneous processes. The GWR model, which differs from the global models, provides the opportunity to estimate local rather than global parameters because it allows for varying parameters by location. For that purpose, a different set of parameters is estimated for each observation by using the values of the characteristics taken by the neighbouring observations (Ertur and Gallo 2009). Since the GWR model allows the variation of the parameters by location and it determines the functional form based on the distribution of data without making any assumptions, it is classified as a nonparametric spatial model. When the parameters of the global model that are given in equation (1) are revised as a model that varies by location, the GWR model that follows can be obtained: P = Xβ (u ) + ε

(2)

where X represents the vector of home characteristics ,u, the measured coordinate vector in the projected coordinate system and, β (u ) , the coefficients of home characteristics. The expression of parameters based on coordinates indicates that spatial variation is allowed in the values of parameters in the model. The vector of the parameter estimators in the GWR model can be written as:

β (u ) = ( X ' W (u ) X ) −1 X ' W (u ) P

(4)

6


where W (u ) represents a weighting function in which non-diagonal elements are zero and diagonal elements are geographical weights for each observation. The weighting function can be expressed as:

0  w1 (u )  0 w2 (u ) W (u ) =   0 0  0  0

0  0  ... 0   0 wn (u ) 0 0

(5)

The GWR model relies on the approach expressed as the first rule of geography by Tobler (1979): ‘everything is related to everything, but closer things are more related to each other’. Therefore, a weighting matrix is formed in which observations that are closer to the target observation value are assigned more weight. Each observation point is weighted differently according to its current location. Fotheringham et al. (2002) states that, because the data set is weighted differently for each location, the results will be obtained as calculations that are specific to a certain location. Moreover, Brunsdon et al. (1998) state that the GWR model is not influenced by the selected weighting function. The important point here is that it is required to a weighting function that observation values are weighted with decreasing weights as the distance from the target value increases. The function that is generally used in the GWR model is the Gaussian weighting function (Huang and Leung 2002; Yrigoyen et al. 2007; Charlton and Fotheringham 2009). Accordingly, although there are numerous different definitions, the Gaussian-type kernel weighting function can be written as:

wi (u ) = e −0.5( di (u ) / h )

2

(6) when Cartesian coordinates are used as a distance measure, Euclidean distances are used in general (for example, Brundson et al. 1996; Fotheringham et al. 1997; Fotheringham et al. 2001 and Le Sage 2004). On the other hand, when global coordinates are taken into consideration, the great circle distances are used. However, there is no reason to ignore other distance measures (Charlton and Fotheringham 2009). Finally, in order to compare the performances of GWR models, the corrected Akaike Information Criterion (AICc) might be used as a goodness-of-fit measure:


 n + tr ( H )   AICc = 2n log e (σˆ ) + n log e (2π ) + n  n − 2 − tr ( H ) 

7

(7)

where n refers to the number of observations, σˆ , the estimated standard error of the residual value and H, a hat matrix in n x n dimension and is defined as H = X ( X ' X ) −1 X ' . tr(H) is the trace of the matrix H.

IV. Data The data for this study was obtained using a survey and collected during the last quarter of 2013. In this period, 2838 questionnaires were randomly distributed to real estate agents on both the European and Asian sides of Istanbul to identify housing pricing factors. Thus, the home prices are the asking prices. This survey listed the structural, physical, location and payment characteristics of apartment flats in 39 counties in Istanbul. The location characteristics on the list were calculated using a great-circle distance formulation by location details. However, the life quality index (LQI) as a neighbourhood characteristic was obtained from the Istanbul Chamber of Commerce. As a result, the data set contains both primary and secondary data. There are 47 independent variables in the data set. Table 1 shows the sample size of each county, and Figure 1 depicts the distribution of houses on the map. Figure 1. Distribution of Homes in Istanbul

8


Table 1. Sample Sizes by County Adalar 1 Arnavutkoy 3 Atasehir 110 Avcilar 50 Bagcilar 20 Bahcelievler 98 Bakirkoy 25 Basaksehir 113 Bayrampasa 14

Besiktas 24 Eyup Beykoz 7 Fatih Beylikduzu 247 Gaziosmanpasa Beyoglu 7 Gungoren Buyukcekmece 15 Kadikoy Catalca 2 Kagithane Cekmekoy 57 Kartal Esenler 16 Kucukcekmece Esenyurt 525 Maltepe

89 12 34 25 91 27 137 95 170

Pendik Sancaktepe Sarıyer Silivri Sultanbeyli Sultangazi Sile Sisli Tuzla

72 73 9 7 2 54 2 20 10

Umraniye 125 Uskudar 53 Zeytinburnu 22 Total 2463

After determining the sample size for Istanbul’s housing market and submarkets, the relevant literature was examined, especially for Istanbul’s market (for example, Ozus et al. 2007; Keskin 2008; Alkay 2008; Caglayan and Eban 2009; Koramaz and Dokmeci 2012, among others). Then, a group of variables was determined and these are presented as detailed in Table 2. Furthermore, during the model selection, the multicollinearity problem was considered, and the independent variables that could be correlated were examined. However, it was found that independent variables are not highly correlated.


9

Table 2. Variable List Variable Description Dependent Variable: Sales price of a real estate (in TL Currency) Log(Sales Price) Independent Variables Structural Characteristics Size of a real estate m2 Dummy variable (1: if real estate is located at the basement floor; or: 0: if Basement Floor not) Dummy variable (1: if real estate is located on the ground floor; or: 0: if not) Ground Floor Dummy variable (1: if real estate is located on the first floor; or: 0: if not) First Floor Dummy variable (1: if real estate is located on the second floor; or: 0: if not) Second Floor Dummy variable (1: if real estate is located on the third floor; or: 0: if not) Third Floor Dummy variable (1: if the heating system is stove in real estate, or 0: if not) Stove Dummy variable (1: if the real estate is painted, or 0: if not) Painted Dummy variable (1: if there are embedded-in kitchen appliances in real Kitchen estate, or 0: if not) Dummy variable (1: if the real estate is furnished, or 0: if not) Furnished Dummy variable (1: if there is cabin closet in real estate, or 0: if not) Cabin Dummy variable (1: if there is double glazed glass on windows, or 0: if not) Double Glass Dummy variable (1: if there is balcony in real estate, or 0: if not) Balcony Dummy variable (1: if there are electronic home appliances in real estate, or White Appliance 0: if not) Dummy variable (1: if there is air conditioner in real estate, or 0: if not) Air Conditioner Dummy variable (1: if there is laminate flooring in real estate, or 0: if not) Laminate Dummy variable (1: if there is parquet flooring in real estate, or 0: if not) Parquet Dummy variable (1: if there is vinyl flooring in real estate, or 0: if not) Vinyl Dummy variable (1: if there is Jacuzzi in real estate, or 0: if not) Jacuzzi Dummy variable (1: if there is vanity unit in real estate, or 0: if not) Vanity Unit Dummy variable (1: if the real estate has terrace, or 0: if not) Terrace Dummy variable (1: if there is super attendant in the building, or 0: if not) Attendant Dummy variable (1: if there is elevator in the building, or 0: if not) Elevator Age of the real estate Age Number of floor in the building where the real estate is located Number-Floor Dummy variable (1: if there is swimming pool nearby a real estate, or 0: if Swimming Pool not) Dummy variable (1: if there is tennis court nearby a real estate, or 0: if not) Tennis Court Dummy variable (1: if there is private security nearby a real estate, or 0: if Security not) Dummy variable (1: if there is car parking lot nearby a real estate, or 0: if Car Parking not) Payment Characteristics Dummy variable (1: if real estate is good for mortgage, or 0: if not) Mortgage Dummy variable (1: if real estate is sold by owner, or 0: if not) Owner Dummy variable (1: if real estate is sold by a construction company, or 0: if Company not) Locational Characteristics /Distance Variables Latitude of the location of a real estate Latitude Longitude of the location of a real estate Longitude The distance from the closest shopping center (Km) Shopping Center The distance from the closest community health clinic (Km) Health Clinic The distance from the 1st Downtown (1st City Center) Central1 The distance from the closest high school (Km) High School The distance from the closest Airport (Km) Airport

Type Continuous

Continuous Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Discrete Continuous Continuous Discrete Discrete Discrete Discrete

Discrete Discrete Discrete

Continuous Continuous Continuous Continuous Continuous Continuous Continuous

10


Metrobus ISB Bus Terminal Bosphorus Bridge Side Bosphorus View Sea View Nature View

Neighborhood / LQI

The distance from the closest Metrobus Stop (Km) The distance from the closest Istanbul Sea Bus Ferry Terminal (Km) The distance from City Bus Terminal (Km) The distance from the Bosphorus Bridge (Km) Dummy variable (1: if real estate is located in European side, or 0: if not) Dummy variable (1: if real estate has Bosphorus view, or 0: if not) Dummy variable (1: if real estate has sea view, or 0: if not) Dummy variable (1: if real estate has nature view, or 0: if not) Life Quality Indexes of Individual Counties (consisted of sub-scales such as Life Quality Index; Education Index, Health & Life Index, Economic Development Index, Transportation and Accessibility Index, Environmental Status Index, Social Life Index, Demographical Structure Index. Life Quality Index is calculated as average of these 7 sub-scales) Resource: Istanbul Chamber of Commerce, 2011

Continuous Continuous Continuous Continuous Discrete Discrete Discrete Discrete

Continuous

V. Empirical Results The estimation results of GWR and OLS models were given in Table 3. The OLS estimation provides 47 global parameters including intercepts that reflect the effect of each explanatory variable on the dependent variable. The GWR procedure will provide 47 x 2838 parameter estimations, and therefore the summary of estimated regression coefficients is shown in Table 3.

Table 3. Estimation Results Independent Variables

Min.

Intercept

4.718

Airport Kitchen Elevator

-0.062 0.026 -0.002

Shopping Center

-0.02

Balcony

0.01

White Appliance

0.037

Age

-0.004

Basement Floor

-0.278

Bosphorus View

-0.161

Bosphorus Bridge

-0.121

Painted

-0.022

Sea View

-0.019

1.Quartile 4.868*** (180.029) -0.035*** (-45.272) 0.031*** (5.137) 0.002 (0.310) -0.014*** (-6.527) 0.014** (2.189) 0.042*** (4.179) -0.002*** (-4.812) -0.218*** (-7.556) 0.204*** (8.465) -0.024*** (-5.854) -0.019*** (-2.701) -0.012 (-1.271)

Median 4.932*** (182.396) 0.005*** (6.467) 0.049*** (8.119) 0.024*** (3.723) -0.01*** (-4.662) 0.019*** (2.972) 0.048*** (4.776) -0.001** (-2.406) -0.167*** (-5.789) 0.219*** (9.087) -0.023*** (-5.609) -0.018** (-2.559) 0.034*** (3.600)

3.Quartile 5.354*** (198.003) 0.008*** (10.348) 0.057*** (9.444) 0.037*** (5.739) -0.008*** (-3.729) 0.02*** (3.128) 0.053*** (5.274) -0.001** (-2.406) -0.151*** (-5.234) 0.234*** (9.709) -0.019*** (-4.634) -0.017** (-2.416) 0.043*** (4.554)

Max. 6.449 0.013 0.062 0.04 0.001 0.022 0.064 0.000 -0.137 0.309 0.08 -0.013 0.058

Global OLS 4.913*** (181.678) 0.006*** (8.207) 0.054*** (8.908) 0.024*** (3.774) -0.002 (-0.838) 0.017** (2.621) 0.048*** (4.748) -0.001* (-1.892) -0.168*** (-5.809) 0.229*** (9.51) -0.019*** (-4.74) -0.016** (-2.248) 0.04*** (4.258)


11

Table 3. Estimation Results (Continued) Nature View

-0.001

Cabin

0.021

Security

0.032

Vanity Unit

0.011

Central1

-0.009

ISB

-0.017

Double Glass

-0.035

Jacuzzi

-0.019

Attendant

-0.001

Ground Floor

-0.084

First Floor

0.004

Second Floor

0.012

Third Floor

-0.003

Air Conditioner

0.044

Mortgage

0.028

Laminate

-0.031

High School

-0.024

2

m

Vinyl

0.002 -0.089

0.004 (0.678) 0.026*** (3.624) 0.042*** (4.773) 0.014** (2.309) 0.002* (1.892) -0.015*** (-13.477) -0.03*** (-4.555) -0.012 (-1.007) 0.006 (0.926) -0.07*** (-8.801) 0.013 (1.593) 0.024*** (3.053) 0.007 (0.915) 0.049*** (5.831) 0.051*** (5.574) -0.029*** (-4.631) -0.014** (-2.217) 0.002*** (22.344) -0.080*** (-5.913)

0.017*** (2.883) 0.028*** (3.902) 0.062*** (7.046) 0.016*** (2.639) 0.01*** (9.461) -0.014*** (-12.579) -0.014** (-2.126) 0.027** (2.265) 0.01 (1.544) -0.058*** (-7.293) 0.029*** (3.554) 0.036*** (4.579) 0.021*** (2.744) 0.058*** (6.901) 0.068*** (7.433) -0.023*** (-3.673) -0.012 (-1.900) 0.002*** (22.344) -0.031** (-2.291)

0.025*** (4.240) 0.038*** (5.296) 0.08*** (9.092) 0.018*** (2.969) 0.02*** (18.921) 0.007*** (6.289) -0.008 (-1.215) 0.04*** (3.356) 0.029*** (4.478) -0.044*** (-5.533) 0.036*** (4.412) 0.041*** (5.216) 0.025*** (3.267) 0.073*** (8.686) 0.072*** (7.869) 0.004 (0.639) -0.008 (-1.267) 0.003*** (33.516) -0.023* (-1.699)

0.033 0.046 0.09 0.029 0.06 0.011 -0.001 0.045 0.039 -0.032 0.049 0.051 0.033 0.083 0.079 0.008 -0.005 0.003 -0.011

0.015** (2.456) 0.027*** (3.726) 0.046*** (5.177) 0.015** (2.521) -0.001 (-0.885) -0.01*** (-9.3) -0.015** (-2.265) 0.027** (2.283) 0.022*** (3.342) -0.049*** (-6.103) 0.036*** (4.385) 0.040*** (5.03) 0.020*** (2.593) 0.062*** (7.328) 0.074*** (8.106) -0.020*** (-3.116) -0.023*** (-3.689) 0.003*** (29.151) -0.037*** (-2.764)

12


Table 3. Estimation Results (Continued) Independent Variables

Min.

Metrobus

-0.003

CBD

-0.140

Furnished

-0.039

Number-Floor

0.003

Bus Terminal

-0.022

Car Parking

0.021

Parquet

-0.023

Health Clinic

-0.016

Owner

-0.043

Stove

-0.095

Tennis Court

-0.004

Terrace

-0.049

Side

-0.020

LQI

0.112

Swimming Pool

0.053

Observation Number GWR, AICc GWR, SSR GWR, Quasi-Global R2

1.Quartile 0.004*** ((3.749) -0.015*** (-3.667) -0.035*** (-3.302) 0.005 (1.536) -0.017*** (-9.731) 0.027*** (3.872) -0.020*** (-3.849) 0.005 (0.597) -0.027*** (-4.186) -0.067** (-2.258) 0.001 (0.093) -0.044*** (-5.138) -0.013 (-0.464) 0.135*** (13.081) 0.064*** (6.529)

Median 0.010*** (9.372) 0.023*** (5.623) -0.031*** (-2.924) 0.006* (1.843) -0.013*** (-7.441) 0.037*** (5.306) -0.013** (-2.502) 0.014* (1.671) -0.018*** (-2.791) -0.051* (-1.719) 0.013 (1.214) -0.040*** (-4.671) 0.195*** (6.967) 0.179*** (17.345) 0.114*** (11.630)

3.Quartile 0.013*** (12.184) 0.030*** (7.335) -0.014 (-1.321) 0.035*** (10.753) 0.035*** (20.034) 0.040*** (5.736) -0.009* (-1.732) 0.060*** (7.161) -0.013** (-2.015) -0.029 (-0.977) 0.020* (1.867) -0.026*** (-3.036) 0.210*** (7.503) 0.182*** (17.636) 0.119*** (12.140)

Max. 0.015 0.073 -0.002 0.040 0.111 0.045 -0.007 0.076 0.015 -0.022 0.032 -0.016 0.236 0.193 0.133

Global OLS 0.013*** (12.196) 0.017*** (4.142) -0.027** (-2.576) 0.008** (2.498) -0.011*** (-6.182) 0.030*** (4.351) -0.013** (-2.403) 0.005 (0.603) -0.023*** (-3.61) -0.054* (-1.81) 0.017 (1.55) -0.042*** (-4.923) 0.186*** (6.657) 0.194*** (18.747) 0.103*** (10.518)

2838 -3947.966 37.727 0.862

*, **, *** indicate significance at the level 10%, 5% and 1%, respectively. Figures in parenthesis are t-statistics. Dependent variable is logarithmic prices

The coefficients of the dummy variables in the OLS and GWR models, in semilogarithmic form, are calculated and interpreted based on the approach developed by Halvorsen and Palmquist (1980). In Table 4, the 1st quartile shows the lowest 25% of home prices while the 3rd quartile, the highest 25% of home prices. According to these estimation results, in both models, the most significant variables that affect home prices in Istanbul are basement floor, Bosphorus view, LQI and side (Europe or Asia). When the effect of basement floor was interpreted for the 1st quartile, it is observed that the home prices decrease by 19.59%. On the other hand, the


13

negative effect of the basement floor on home prices is 14.01% for the 3rd quartile. The negative effect of basement on home prices is 15.46% in the OLS regression model. Consequently, the basement floor variable negatively affects home prices in both models. The estimation results of the GWR model concerning the Bosphorus view indicate that home prices increase by 22.63% for the 1st quartile. On the other hand, the home prices increase by 26.36% for the 3rd quartile. In the OLS regression model, home prices increase by 25.73% when the home has a Bosphorus view. Accordingly, Bosphorus view has a positive and significant effect on home prices. The estimation results of the GWR model indicate that home prices decrease by 1.29% when the home is on the European side of Istanbul for the 1st quartile. Home prices increase by 23.37% for the 3rd quartile. In the OLS model, home prices increase by 20.44% when they are located on the European side. It is observed that the side variable, which was added to the model in order to investigate the effect of geographical location on home prices, had a significant effect. When life quality increases, it would exhibit a positive effect on home prices. Because enhancing the environmental facilities (education, health, transportation access, etc.) at a location would increase the value of homes at that location, these enhancements would increase demand and, accordingly, home prices. When the other structural and physical characteristics of homes were evaluated in both models, it can be observed that swimming pools, tennis courts, security services, jacuzzis, vanity units and other characteristics positively affected home prices. Another notable point in the estimation results is that the effect of distance variables on home prices. According to these results, there is a nonparametric relationship between distance and home prices. The estimation results of the GWR model indicate that as the distances to the transportation points such as airports, bus terminals, metrobus, ISB, CBD, and central1 increase, home prices decrease in the 1st quartile. However, in the 3rd quartile, as the distance increases, home prices increase also. Based on general results, the GWR model allows us to examine the relationships in the different parts of the distribution of home prices. Moreover, it is possible to observe the relationships varying by location. It can be seen the spatially varying relationship from the map graphics of most significant variables in Figure 2.

14


Figure 2. Map Graphics

Basement Floor

Bosphorus View

Side

LQI

The coefficients of variables vary between the range of -1 and +1 in these graphs. The light colored maps show the negative values while the dark-colored maps have positive values. According to these graphics, the effects of basement floor on home prices is negative in all counties of Istanbul. The effect of Bosphorus view and LQI on home prices are generally positive. However, the effect of side variable on home prices become different. This effect is positive for homes located in all counties in the Anatolian side while the effect is negative for the homes located in counties which is far from the centers of the European side. After estimation of the GWR model, some statistical measures presented in Figure 3. This figure displays the residuals (gwr.e) that were obtained from the GWR model estimates; the predicted housing price values for each model (pred); the standard errors of these values (pred.se); and the local R2 values calculated for each model. In the local R2 value graphic, it can be seen that the explanatory power of the GWR model varied in the range of 0.78 - 0.86 and was concentrated in general on the values of 0.78 and 0.86.


15

Figure 3. Pairwise Graphics of the Diagnostic Measures

Brunsdon, Fotheringham, and Charlton (1999) suggested an analysis of variance (ANOVA) to investigate whether there is a statistically significant difference between the OLS and GWR residuals. The test results in Table 4 show that the F statistic test value (4.534***) was significant at the 1% level. In other words, the GWR model explains the relationships between home prices and their characteristics better than the OLS model. Table 4. ANOVA Results Brunsdon, Fotheringham & Charlton (1999) d.o.f

Sum of squares

47

45.160

GWR Improvement

196.83

11.559

0.059

GWR Residuals

2594.17

33.601

0.013

OLS Residuals

Sum of mean squares

F-value

4.534***

VI. Conclusion In light of the GWR model, it is possible to investigate the nonparametric relationships between home prices and their characteristics and to observe the variation in magnitude and sign of the coefficients by location. In addition, the GWR model enables flexibly determining the functional forms of the relationships. This assumption of the GWR model enables modelling the

16


relationships in conformity with the heterogeneous and segmented structure of the housing market. For this reason, GWR model is used, as different from the other studies related to Istanbul’s housing market. The diagnostic test developed by Brunsdon, Fotheringham, and Charlton (1999) indicates that the GWR model explains the relationship between home prices and their characteristics better than the OLS model. The estimation results for both models show that the most significant variables that affect home prices in Istanbul are basement floor, Bosphorus view, life quality index and side. However, unlike the OLS regression model, the estimation results of the GWR model demonstrate that the relationships between home prices and characteristics change by location.

References Alkay, E. (2008): Housing submarkets in Istanbul. Int. Real Estate Rev. 11(1), 113-127. Anglin, P. M. & Gencay, R. (1996): Semiparametric estimation of a hedonic price function. J. Appl. Econom. 11(6), 633-648. Bourassa, S.C. &Hoesli, M. and Peng, V.S. (2003): Do housing submarkets really matter. J. Hous. Econ. 12, 12-28. Brunsdon, C., Fotheringham, A.S. & Charlton, M.E. (1996): Geographically weighted regression: A method for exploring spatial nonstationary. Geogr. Anal. 28(4), 281-298. Brunsdon, C., Fotheringham, S.& Charlton, M. (1998): Geographically weighted regression-modelling spatial non-stationarity. J. Royal Stat. Soc. 47(3), 431-443. Charlton, M. &Fotheringham, A.S. (2009): Geographically weighted regression. White Paper, National Centre for Geocomputation, National University of Ireland Maynooth. Caglayan, E.& Eban, A. (2009): Determinants of house prices in ıstanbul: A quantile regression approach. Qual.&Quant. 45, 305-317. Ertur, C. & Gallo, J. Le. (2009): Regional Growth and Convergence: Heterogenous reaction versus interaction in spatial econometric approaches. Post-Print. Available at: http://ideas.repec.org/p/hal/journl/hal00463279.html [Accessed February 29, 2016].


17

Farber, S.& Yeates, M. (2006): A comparison of localized regression models in a hedonic house price context. Canadian J. Reg. Sci. XXIX (3), 405-420. Fotheringham A.S., Charlton M. & Brunsdon, C. (2001): Spatial variations in school performance: A local analysis using geographically weighted regression. Geogr. & Environ. Model. 5(1), 43-66. Fotheringham, A. S., Brunsdon, C.& Charlton, M. (2002): Geographically Weighted Regression: The Analysis of Spatially Varying Relationships, John Wiley and Sons Ltd., University of Newcastle, UK. Fotheringham, S. A. (1997): Trends in Quantitative Methods I: Stressing The Local. Prog. Hum. Geogr. 21, 88-96. Goodman, A. C. (1981): Housing submarkets within urban areas: Definitions and evidence. J. Reg. Sci. 21(2),175-185. Huang Y. & Leung Y. (2002): Analysing regional industrialisation in Jiangsu province using geographically weighted regression. J. Geogr. Syst. 4, 233– 249. Keskin, B. (2008): Hedonic analysis of price in the Istanbul housing market. International J. of Strategic Prop. Manag. 12, 125-138. Koramaz, K. T.& Dokmeci, V. (2012): Spatial determinants of housing price values in Istanbul. Eur. Plan. Stud. 20(7). 1222-1237. Maclennan, D. & Tu, Y. (1996): Economic perspectives on the structure of local housing markets. Hous. Stud. 11, 387- 406. McMillen, D. P. & Redfearn, C. L. (2010): Estimation and hypothesis testing for nonparametric hedonic house price functions. J. Reg. Sci. 50(3), s. 712– 733. McMillen, D. P. (2012): Nonparametric Spatial Data Analysis, CRAN. McMillen, D. P. (1996): One hundred fifty years of land values in chicago: A nonparametric approach. J. Urban Econ. 40(0025),100-124. Ozus, E., Dokmeci, V., Kiroglu, G. & Egdemir, G. (2007): Spatial analysis of residential prices in Istanbul. Eur. Plan. Stud., 15(5), 708-721. Pavlov, A. D. (2000): Space-varying regression coefficients: A semi-parametric approach applied to real estate markets. Real Estate Econ. 28(2), 249-263. Rosen, S. (1974): Hedonic prices and implicit markets: Product differentiation in pure competition. J. Polit. Economy. 82(1), 34-55.

18


Rothenberg, J., Galster, G., Butler, R. & Pitkin, J. (1991): The Maze of Urban Housing Markets: Theory, Evidence and Policy. University of Chicago Press, Chicago, IL. Straszheim, M. R. (1975): The role of the housing market. In: Straszheim, M. R. (eds.) An Econometric Analysis of the Urban Housing Market, pp. 1-10. NBER. Tobler, W. (1979): Celluar Geography. In Gale, S., Olsson, G. (eds.) Philosophy in Geography, Dordrecht. Watkins, C.A. (2001): The definition and identification of housing submarkets. Environ. and Plan. A. 33, 2235-2253. Yrigoyen, C.C., Rodriguez, I.G. & Otero, J.V. (2007): Modeling spatial variations in household disposable income with geographically weighted regression. MPRA, 1682, 1-28.

Chapter 2

The Long-Run Relationship between Hedonic House Prices and Consumer Prices: ARDL Bounds Testing Approach Havvanur Feyza ERDEM 1 Nebiye YAMAK 2

Abstract. The aim of this study is to test the long-run relationships between consumer prices index and hedonic house prices index for five regions of Turkey. The data are monthly and cover the period of 2010:01-2017:07. All data come from the Electronic Data Delivery System of the Central Bank of the Republic of Turkey. The data cover five regions of Turkey that include İstanbul; Ankara; İzmir; Samsun, Çorum, Amasya, Tokat, and Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon. In this study, the Autoregressive Distributed Lag (ARDL) bounds testing approach developed by Pesaran and Shin (1999) was used to examine the long-run relationships between consumer prices index and hedonic house prices index. As the result of the study, it was found that there are long-run relationships between consumer and hedonic house prices for five regions of Turkey and also the overall Turkey. Keywords: ARDL; Consumer Price Index; Hedonic Price Index; Turkish Economy.

1

2

Asst. Prof. Dr., Karadeniz Technical University, Faculty of Economics and Administrative Sciences, Department of Econometrics, Turkey, [email protected] Prof. Dr., Karadeniz Technical University, Faculty of Economics and Administrative Sciences, Department of Economics, Karadeniz Technical University, Turkey, [email protected]

20

The Long-Run Relationship between Hedonic House Prices and Consumer Prices: ARDL Bounds Testing Approach

I.

Introduction

Since housing is both consumption good and also store of value especially in developing countries, it can be expected that hedonic housing prices may affect whole consumer prices directly and indirectly. On the other hand, it can be also expected that inflation may influence hedonic housing prices in both channels since it affects the cost of housing and also return of investment. In the applied economics, it is recently argued that there is a strong link between consumer and housing prices. Especially recent global economic crisis in the world have demonstrated that there must be a close relationship between the two variables. In the applied economics, the hedonic measure approach was first introduced and employed to land characteristics by Waugh (1928) while the term “hedonic pricing approach” was first used by Court (1939) in the context of changing price measures for automobiles. Later, this method was augmented by Griliches (1961, 1971) and Rosen (1974) in terms of a quality-adjusted price index that hedonic methods started to receive serious attention (see Schultze and Mackie, 2002 and Triplett, 2004). A hedonic price index is any price index which uses information from hedonic regression on the related commodity or services. The first official hedonic house price index was US Census Bureau’s “One-Family Houses Index” which was first published in 1968 (Triplett, 2004). The aim of this study is to investigate the possible long-run relationships between consumer and hedonic house prices for five regions of Turkey, and also the overall Turkey. The data are monthly and cover the period of 2010:012017:07. In this study, data and methodology, empirical findings, and conclusion are present in Sections 2, 3, and 4, respectively.

I.

Data and Methodology

In this study, the Autoregressive Distributed Lag (ARDL) bounds testing approach developed by Pesaran and Shin (1999) was used to investigate the long-run relationships between consumer (CPI) and hedonic house prices indices (HHP). The ARDL approach does not require prior knowledge on the order of integration of the variables. It can be easily used for the variables with different orders of integration. At this point, it should be noted that all variables must be I(0) or I(1), but not higher than I(1). The ARDL approach has some certain advantages in comparison with other conventional co-integration


21

methods such as Engle-Granger and Johansen-Juselius methods. Among others, the most important advantage of this technique is that it gives the possibility of short and long run parameters of the model simultaneously by using the unrestricted ARDL error correction model. The ARDL bounds testing methodology to co-integration involves estimating the following regression. ∆HHPt =α0 + ∑ki=1 βi ∆HHPt-i + ∑ki=1 γi ∆CPIt-i +δ1 HHPt-1 + δ2 CPIt-1 +ε1t

(1)

where the coefficients 𝛽𝑖 and γi represent the short-run dynamics of the variables and the coefficients 𝛿1 and 𝛿2 represent the long-run relationship between consumer and hedonic house prices. After estimation of the above regression, the following null hypothesis of no co-integration is tested against the alternative hypothesis of the presence of co-integration by using F-statistics. H0 : 𝛿1 = 𝛿2 = 0 H1 : δ1 ≠0, 𝛿2 ≠ 0 The data are monthly and cover the period of 2010:01-2017:07. All data come from the Electronic Data Delivery System of the Central Bank of the Republic of Turkey. The data cover five regions of Turkey that include İstanbul, Ankara; İzmir; Samsun, Çorum, Amasya, Tokat; Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon, and also the overall Turkey. All variables were measured in real terms and seasonally adjusted using Census X-12 process. After seasonal adjustment, a logarithmic transformation was done on the data. The details and symbols of all variables are given in Table 1. Table 1. Symbols Used for Variables HHP

Hedonic House Price Index –Turkey

HHP1

Hedonic House Price Index –İstanbul

HHP2

Hedonic House Price Index –İzmir

HHP3

Hedonic House Price Index- Ankara

HHP4

Hedonic House Price Index –Samsun, Çorum, Amasya, Tokat

HHP5

Hedonic House Price Index-Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon.

CPI

Consumer Price Index –Turkey

CPI1

Consumer Price Index –İstanbul

CPI2

Consumer Price Index –İzmir

CPI3

Consumer Price Index -Ankara

CPI4

Consumer Price Index –Samsun, Çorum, Amasya, Tokat

CPI5

Consumer Price Index -Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon.

22

II.


Empirical Findings

Even though the ARDL approach does not require prior knowledge on the order of integration of the variables, the order of integration must be determined for each variable in order to decide whether the use of the ARDL is appropriate. For this purpose, the Augmented Dickey-Fuller (ADF) 3 and Phillips-Perron (PP) 4 unit root tests were first performed for the level and first difference of each variable. Tables 2 and 3 present the results of the ADF and PP test statistics with and without the inclusion of a trend detecting a unit root in the levels and first differences of the variables. As seen from the Tables 2 and 3, the ADF and PP test statistics calculated for the levels of the variables indicate that the non-stationary of the levels of the variables cannot be rejected at any significant level. However, the first difference of each variable appears to be stationary according to the ADF and PP test statistics. In Tables 2 and 3, ***, ** and * denote significance level of 1%, 5% and 10%, respectively. Table 2: Unit Root Test Results for Hedonic House Price Index ADF Intercept

Trend-Intercept

Intercept

Trend-Intercept

HHP

1.447572

-2.864035

2.496855

-3.467194**

∆HHP

-4.796123***

-5.088865***

-4.680872***

-4.949719***

HHP1

-1.110914

-1.445077

0.779910

-2.246075

∆HHP1

-4.937405***

-1.527426

-4.735916***

-4.920640***

HHP2

4.970686

-2.062113

5.185807

-2.098343

∆HHP2

-7.418844***

-9.001449***

-7.418844***

-9.278932***

HHP3

0.408443

-2.364065

0.940099

-2.491087

∆HHP3

-7.280363***

-7.243252***

-7.068004***

-6.992159***

HHP4

3.540013

-0.872212

5.558100

-0.677929

∆HHP4

-7.416851***

-8.120419***

-7.331337***

-8.728030***

-1.094820

3.714282

-1.065506

-8.531898***

-7.569181***

-8.497445***

HHP5 ∆HHP5 3

3.522760 -7.569181***

Dickey and Fuller (1979). Phillips and Perron (1988).

4

PP


23

Table 3: Unit Root Test Results for Consumer Price Index ADF Intercept

PP Trend-Intercept

Intercept

Trend-Intercept

CPI

0.997180

-2.691953

1.098685

-2.857726

∆CPI

-8.374049***

-8.496090***

-8.357432***

-8.466947***

CPI1

1.294837

-2.836334

1.245177

-2.923390

∆CPI1

-8.300355***

-8.484550***

-8.300355***

-8.463866***

CPI2

1.032226

-2.554301

1.589928

-2.554301

∆CPI2

-10.58976***

-10.78925***

-10.65732***

-11.40983***

CPI3

0.818074

-3.590705**

∆CPI3

-8.012503***

-8.067617***

-7.909683***

-7.976799***

CPI4

0.912551

-4.147705***

0.955895

-2.945694

∆CPI4

-7.547772***

-7.589309***

-7.396829***

-7.552772***

CPI5

0.632809

-3.166368*

0.670800

-2.525738

∆CPI5

-7.485631***

-7.515583***

-7.306886***

-7.344421***

0.881846

-2.954576

Tables 4, 5, 6, 7, 8, and 9 show the bound test statistics. As seen as from all tables, the value of the F-statistics is greater than the upper critical value bounds for all regions. Therefore, there is a long-run relationship between hedonic house and consumer prices for all regions and also the overall Turkey. According to the results of tables, in a common long-run equilibrium, hedonic house and consumer prices are co-integrated. As seen as from • Table 4, the estimated long-run coefficient of CPI1 is 1.9645 and the coefficient is statistically significant at 1% level. The coefficient implies that hedonic house price increases (decreases) by 19.645 percent if consumer price increases (decreases) by 10 percent for İstanbul. • Table 5, in the long-run hedonic house price index equation for İzmir, the estimated long-run elasticity coefficient is 2.0549 and the coefficient is statistically significant at 1% level. The elasticity coefficient shows that

24


house price increases (decreases) by 41.08 percent if consumer price increases (decreases) by 20 percent. • Table 6, the estimated long-run elasticity coefficient of CPI3 is 1.2278 and the coefficient is statistically significant at 1% level. Hedonic house price increases (decreases) by 6.135 percent if consumer price increases (decreases) by 5 percent for Ankara. • Table 7, it can be argued that hedonic house price increases (decreases) by 13.24 percent if consumer price increases (decreases) by 10 percent for Samsun, Çorum, Amasya, Tokat. The estimated long-run coefficient of CPI4 is 1.3248 and the coefficient is statistically significant at 1% level. • Table 8, the estimated long-run coefficients of CPI4 is found to be 1.572 for Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon and the coefficient is statistically significant at 1% level. In this region, hedonic house price increases (decreases) by 15.72 percent if consumer price increases (decreases) by 10 percent. • Table 9, in the long-run hedonic house price index equation for the overall Turkey, the estimated long-run elasticity coefficient is 1.4982. The elasticity coefficient implies that hedonic house price increases (decreases) by 14.98 percent if consumer increases (decreases) by 10 percent. In addition, the diagnostic test results of all models such as autocorrelation and heteroscedasticity are given in the following tables. As seen as from the tables, there are no autocorrelation and heteroscedasticity problems in any model. The figures in the tables also present CUSUM and CUSUM of SQUARES of the estimated ARDL models. It can be seen in the figures, all estimated coefficients in the ARDL models are stable. There is no structural break in the models in terms of coefficients.


Table 4. ARDL Test Results for Istanbul

Table 5. ARDL Test Results for Izmir

25

26


Table 6. ARDL Test Results for Ankara

Table 7. ARDL Test Results for Samsun, Çorum, Amasya, Tokat


Table 8. ARDL Test Results for Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon

Table 9. ARDL Test Results for Turkey

27

28


III. Conclusion The purpose of this study is to examine the possible long-run relationships between consumer and hedonic house prices indices for five regions of Turkey and also the overall Turkey. The data are monthly and cover the period of 2010:01-2017:07. The data cover five regions of Turkey that include İstanbul; Ankara; İzmir; Samsun, Çorum, Amasya, Tokat and Artvin, Giresun, Gümüşhane, Ordu, Rize, Trabzon. In this study, the long-run relationship between two variables was investigated by implementing ARDL bounds test. According to the ARDL bound test results, there is a strong long-run relationship between house prices and consumer prices for all regions an also the overall Turkey. The long run relationship is found to be statistically significant at least at 1% level. The results imply that house prices and consumer prices are linked to each other in the long-run in Turkey. In terms of long-run elasticity, the biggest coefficient was estimated for region of İzmir while the smallest one is for the region of Ankara. So, in İzmir, hedonic price increases (decreases) by 20.54 percent if consumer price increases (decreases) by 10 percent. In Ankara, hedonic price increases (decreases) by 12.27 percent if consumer price increases (decreases) by 10 percent.

References Court, A. T. (1939). Hedonic Price Indexes with Automotive Examples, in The Dynamics of Automobile Demand, The General Motors Corporation, New York, 99-117. Dickey, D. A. and W. A. Fuller (1979). Distribution of the Estimators for Autoregressive Time Series with a Unit Root, Journal of the American Statistical Association, 74(336a), 427–431. Griliches Z. (1961). Hedonic Price Indexes for Automobiles: An Econometric Analysis of Quality Change, in The Price Statistics of the Federal Government, G. Stigler (chairman). Washington D.C.: Government Printing Office. Griliches Z. (1971). Introduction: Hedonic Price Indexes Revisited, in Z. Griliches (ed), Price Indexes and Quality Change, 3-15. Cambridge MA: Harvard University Press. Pesaran, M. Hashem and Y. Shin (1999). Autoregressive Distributed Lag Modelling Approach to Cointegration Analysis, In: Storm S, editor.


29

Econometrics and Economic Theory in the 20th Century: the Ragnar Frisch Centennial Symposium. Cambridge University Press; [chapter 1]. Phillips, P. C. and P. Perron (1988). Testing for A Unit Root in Time Series Regression, Biometrica, 75(2), 335-346. Rosen S. (1974). Hedonic Prices and Implicit Markets: Product Differentiation in Pure Competition, Journal of Political Economy, 82(1), 34-55. Schultze C. S., and C. Mackie (2002). At What Price? Conceptualizing and Measuring Cost-of-Living and Price Indexes, in Panel on Conceptual, Measurement, and Other Statistical Issues in Developing Cost-of- Living Indexes, Committee on National Statistics, Division of Behavioral and Social Sciences and Education. Washington, DC, National Academy Press. Triplett, J. E. (2004). Handbook on Hedonic Indexes and Quality Adjustments in Price Indexes: Special Application to Information Technology Products, STI Working Paper, 9. Waugh F. V. (1928). Quality Factors Influencing Vegetable Prices, Journal of Farm Economics, 10.

30


Chapter 3

Introduction to Seasonal Unit Root Processes

1, 2

Mehmet ÖZMEN 3 Sera ŞANLI 4

Abstract. Most macroeconomic time series are measured at seasonal frequencies and in the case of including seasonal unit roots, since applying the transaction of differencing to these type of series as the number of repeating unit roots will give rise to a nonstationary structure in series; it is of high importance to have a knowledge about whether the unit roots in series are seasonal or not. In this chapter, apart from the seasonal integration testing sequence proposed by Ilmakunnas (1990) who takes Dickey and Pantula (1987) approach as basis, HEGY seasonal unit root approach in which auxiliary regression models including combinations of various deterministic components like trend, constant term and seasonal dummies have been employed and that is proposed by Hylleberg, Engle, Granger and Yoo fundamentally for quarterly data as the most popular approach has been handled. Besides HEGY procedure; DHF, OCSB and Kunst seasonal integration tests have been covered. Apart from presenting the knowledge about at which seasonal frequencies series contain seasonal unit roots or not, seasonal unit root analyses which are carried out by taking HEGY testing equations for quarterly and monthly data as basis have a significant importance with respect to revealing that it is required to apply filters corresponding to the frequencies in which the series in question has unit roots in order to make that series stationary.

Keywords: DHF, HEGY, Seasonal Unit Roots, Seasonal Integration.

1

This study has been derived from the Master Thesis that has been prepared in consultancy of Assoc. Prof. Mehmet Ozmen called “The Econometric Analysis of Seasonal Time Series: Applications on Some Macroeconomic Variables (Sanli, 2015)”. 2 This study has been supported by TUBITAK (The Scientific and Technological Research Council of Turkey) – BIDEB (Scientist Support Department) within the scope of 2211-E Direct National Scholarship Programme for PhD Students. 3 Prof. Dr., Cukurova University, Faculty of Economics and Administrative Sciences Department of Econometrics, Balcali, Adana ,[email protected] 4 Res. Ass., Cukurova University, Faculty of Economics and Administrative Sciences Department of Econometrics, Balcali, Adana, [email protected]

32


I. Introduction This chapter introduces a number of testing procedures in order to test the null hypothesis of seasonal integration by giving its definition and thus handles the implications of seasonal unit root processes. Economic time series are generally recorded at some fixed intervals. When we are dealing with macroeconomic time series, seasonal models are mostly available at monthly or quarterly frequencies and time series plotting of these frequencies as a separate curve is of crucial importance with respect to giving a useful insight about describing the seasonal behaviour of a series (Ghysels & Osborn, 2001, p.3). This seasonal behaviour could be viewed as deterministic or stochastic. For deterministic and stochastic seasonality in details, see Ghysels and Osborn (2001). A seasonal series is a series with a spectrum having distinct peaks at the 2πj seasonal frequencies ws = , j = 1,........, s / 2 where s is the number of s time periods in a year supposing that s is an even number. For example, if we are dealing with quarterly data, s is equal to 4 and for monthly data s is 12. For the classes of time series models prevalently used in order to model seasonality, see Hylleberg, Engle, Granger and Yoo (1990) (Hylleberg et al., 1990, pp. 215220).

Testing Seasonal Integration As well known, a series generated by a unit root process can wander widely over time not having any inclination to return to its underlying mean value and thus not having any tendency to return to a deterministic pattern. In that case, with the values wandering to a great extent for the seasons, any basic relationships between the expected values for the different seasons remain beside the point in practice. A seasonal unit root process requires seasonal differencing. If seasonal differencing is not applied to the series having stochastic seasonality, the series continues to be nonstationary. Therefore, it is of great importance to determine which type of seasonality the series in question displays because nonstationarity and non-invertibility situations create difficulties in parameter estimation and forecasting (Tam & Reinsel, 1997, p.725). Sometimes, to discriminate between deterministic seasonality and a seasonal unit root process may be hard. Since


33

according to Bell (1987), the two competing processes which are the simple deterministic

seasonality

model

y sτ = γ s + ε sτ ,

and

the

process

of

∆ S y sτ = (1 − θ S LS )ε sτ , for s = 1,........., S , τ = 1,........., Tτ that states a seasonal

unit root process are equivalent in the special case of θ S = 1 . There is a prevalent view about the cancellation of the seasonal differencing operator ∆ S and the noninvertible MA operator 1 − LS with θ S = 1 . The extension of this logic also occurs when there is “near cancellation” situation with θ S close to but less than unity. As a result, applying seasonal differencing to a deterministic seasonal process prompts the existence of seasonal differencing operator in the MA operator and this will lead to non-invertibility of MA operator (Ghysels & Osborn, 2001, pp. 28-29; 42-43). Engle, Granger and Hallman (1989) make the definition of a seasonally integrated series as follows: Definition: A nonstationary series is said to be seasonally integrated of order

(d , D) , denoted SI s (d , D) , If it can be transformed to a stationary series by applying s-differences D times and then differencing the resulting series d times using first differences. A seasonal difference is the difference between an observation and its value for the corresponding season one year before. If the series is measured s times per annum and it displays a seasonal pattern, then the differencing to remove seasonality should be s rather than one. So, the type of operator to be applied here is xt − xt − s instead of xt − xt −1 and the transaction to get these variables is called seasonal differencing or s-differencing. Taking s-differences transforms a linear trend with an additive seasonal effect to a constant and if this transaction is applied to a quadratic trend (where the trend is nonlinear) with additive seasonality, it brings about a series still including a trend component but with no seasonal pattern. So, in order to make such a series is stationary, first differencing of the s-differences may be required (Charemza & Deadman, 1992, pp. 53, 129-130). Ilmakunnas (1990) has tried to illustrate a testing sequence in order to test the appropriate order of differencing in quarterly data. Introducing this testing

34


sequence requires two alternative definitions of seasonal integration. According to the first definition which is the one defined by Osborn, Chui, Smith & Birchenhall (OCSB) (1988), a time series is said to be integrated of order (d , D) , denoted I (d , D) if the series becomes stationary subsequent to firstdifferencing d times and seasonally differencing D times. In other saying; if

(1 − L) d (1 − LS ) D xt = ∆d ∆DS xt becomes stationary, xt is said to be I (d , D) . Since Ilmakunnas (1990) focuses on the quarterly time series ( s = 4) , it is concerned with the case where I (1,1) is the maximum order of integration. The second alternative definition for seasonal integration comes from Engle et al. (1989) that has already been mentioned above. To this definition; if

(1 − L) d S ( L) D xt = ∆d S ( L) D xt is stationary, xt is said to be seasonally integrated of orders d and D expressed as SI (d , D) where S (L) is a seasonal filter used in transforming the variables to moving sums. In the case of quarterly data, seasonal filter is stated as S ( L) = 1 + L + L2 + L3 and it takes place in the decomposition of ∆ 4 = (1 − L) S ( L) = (1 − L)(1 + L)(1 + iL)(1 − iL). Since ∆∆ 4 is decomposed as (1 − L) 2 S ( L) or (1 − L)[(1 − L) S ( L)] = (1 − L)(1 − L4 ) ,

SI (2,1) and I (1,1) are the same. In the same manner, SI (1,0) is the same as I (1,0) and also SI (1,1) and I (0,1) are the same. To illustrate the testing sequence for quarterly data, starting point is taken as the maximum order of seasonal integration, i.e. the case SI (2,1) . The testing sequence of Ilmakunnas pursues the view proposed by Dickey and Pantula (1987). According to their view, determining the suitable integration order is based on the starting point of the testing sequence (Ilmakunnas, 1990, pp. 7981). Ilmakunnas (1990) mentions about how to handle unit root testing in a seasonal context considering the initial test of the SI ( 2,1) null hypothesis. In the study, it is expressed that SI (2,1) is tested against SI (2,0) , SI (1,1) and SI (1,0) alternatives using the HEGY test regression. In case we reject the null hypothesis in favour of either SI (1,1) or SI (1,0) alternatives, we have to check the presence of zero frequency unit root against SI (0,1) or SI (0,0) processes, respectively continuing for testing against lower orders of integration (Ghysels & Osborn, 2001, p. 76). Table 1 has been showing which hypotheses can be tested with each given test in the testing sequence:


35

Table 1. Some Seasonal Integration Tests for Different Hypotheses Description of the tests ADF for

∆ 4 series: t-statistics of β in p

∆∆ 4 X t = β∆ 4 X t −1 + ∑ α j ∆∆ 4 X t − j + u t

Null Hypothesis

Alternative Hypothesis

SI(2,1)

SI(1,1)

SI(1,1)

SI(0,1)

SI(1,1)

SI(0,0)

SI(2,1)

SI(1,0)

j =1

ADF for

S (L)

β

series: t-statistics of

in

p

∆ 4 X t = βS ( L) X t −1 + ∑ α j ∆ 4 X t − j + u t j =1

DHF: t-statistic for

β

in p

∆ 4 X t = βZ t −4 + ∑ α j ∆ 4 X t − j + u t j =1

p

where

Z t = X t − ∑θ j X t − j

θj

and

is the coefficient of

j =1

∆ 4 X t − j from a regression of ∆ 4 X t DHF for

∆

series: t-statistic for

β

on its p lagged values. in

p

∆∆ 4 X t = βZ t*− 4 + ∑ α j ∆∆ 4 X t − j + u t j =1

p

where

Z t = ∆X t − ∑ θ j ∆X t − j

θj

is the coefficient

∆∆ 4 X t

on its p lagged

and

j =1

of

∆∆ 4 X t − j

values.

from a regression of

Remarks

36


HEGY: t-statistics for

π3 = π4 = 0 and

π1

and

π2

π 2 ,π 3 ,π 4

and F-statistic for testing

(or t-statistics sequentially for

π4

(two-sided test) SI(1,1)

tested;

SI(1,0)

π1 = 0

π 3 ) in

∆ 4 X t = π 1 Z 1,t −1 + π 2 Z 2,t −1 + π 3 Z 3,t − 2 + π 4 Z 3,t −1

SI(1,1)

π 2 ,π 3 ,π 4

SI(0,0)

tested;

p

+ ∑ α j ∆ 4 X t − j + ut j =1

SI(1,1)

SI(0,1)

SI(0,1)

SI(0,0)

π 1 tested;

π2 = π3 = π4 = 0

p

Z 1t = S ( L)( X t − ∑ θ j X t − j ) ,

where

j =1

p

Z 2t = −(1 − L + L − L )( X t − ∑ θ j X t − j ) 2

π 2 ,π 3 ,π 4

3

tested;

π1 ≠ 0

j =1

p

Z 3t = −(1 − L2 )( X t − ∑ θ j X t − j )

and

θj

are obtained

j =1

as in DHF. OCSB: t-statistics for

β1 and β 2

in p

∆∆ 4 X t = β1 Z 4,t −1 + β 2 Z 5,t − 4 + ∑ α j ∆∆ 4 X t − j + u t j =1

SI(1,0)

β2

SI(2,1)

SI(1,1)

β1 tested; β 2 = 0

SI(1,1)

SI(0,0)

β2

p

where

Z 4t = ∆ 4 X t − ∑ θ j ∆ 4 X t − j

,

tested; β 1

SI(2,1)

=0

j =1

p

Z 5t = ∆X t − ∑ θ j ∆X t − j j =1

and

θj

are obtained as in DHF for

∆

tested;

β1 ≠ 0

series.

(Source: Ilmakunnas, 1990, pp. 82-83).

II.

Dickey-Hasza-Fuller (DHF) Test

One of the simplest testing procedures for seasonal integration possibly belongs to the one proposed by Dickey, Hasza and Fuller (1984) and modified by Osborn et al. (1988), denoted DHF. It can be regarded as a generalization of the Augmented Dickey Fuller test (ADF) and it is the first test of the null hypothesis y t ~ SI (1) .Using DHF test for seasonal integration is identical to


37

testing for stochastic seasonality. If we assume a [ y t = φ s y t − s + ε t ] process, DHF test can be parameterized as ∆ s y t = α s y t − s + ε t where α s = −(1 − φ s ) . Here the null hypothesis of seasonal integration (presence of seasonally integrated process) is α s = 0 and the alternative of a stationary stochastic seasonal process implies α s < 0 (Baltagi, 2001, p. 661). For a detailed DHF test steps starting form higher orders of seasonal differencing, see Charemza and Deadman (1992, pp. 136-140).

III. HEGY Test The analysis of seasonal unit roots is fundamentally conducted with the most popular approach developed by Hylleberg et al. (1990) called HEGY by working with different models that include trends, constants and seasonal dummies in order to determine the type of seasonality and one apparent advantage of HEGY procedure over DHF is that it enables to test for unit roots at each frequency separately without maintaining that there are unit roots at some or all other frequencies (Ghysels, Lee & Noh, 1994, p. 416). Hylleberg et al. (1990) have introduced a factorization of the seasonal differencing polynomial ∆ 4 = (1 − L) for quarterly data using lag operator L and developed 4

a testing procedure in the most general form for seasonal unit roots that could be estimated by OLS in the following way: 3

k

i =1

i =1

∆ 4 y t = α + βt + ∑ α i Di ,t + π 1Y1,t −1 + π 2Y2,t −1 + π 3Y3,t −2 + π 4Y3,t −1 + ∑ ci ∆ 4 y t −i + ε t (1) where k is the number of lagged terms included to ensure that residuals are white noise, the Di ,t are seasonal dummy variables and the Yi ,t variables are constructed from the series on y t as:

Y1,t = (1 + L)(1 + L2 ). y t = y t + y t −1 + y t − 2 + y t −3

(2)

Y2,t = −(1 − L)(1 + L2 ). y t = − y t + y t −1 − y t − 2 + y t −3

(3)

Y3,t = −(1 − L)(1 + L). y t = − y t + y t − 2

(4)

Y4,t = −( L)(1 − L)(1 + L). y t = Y3,t −1 = − y t −1 + y t −3

(5)

(Charemza & Deadman, 1992, p. 141).

38


The null hypothesis of the HEGY test is that the variable in question is seasonally integrated and the factorization of the expression ∆ 4 = (1 − L)

4

could say something relating to roots:

(1 − L) 4 = (1 − L)(1 + L)(1 + L2 ) = (1 − L)(1 + L)(1 − i ⋅ L)(1 + i ⋅ L) where i is an imaginary part of a complex number such that i 2 = −1 . When looked at this factorization, it is seen that a quarterly stochastic seasonal unit root process has four roots of modulus one. One root (1 − L) described as being at ‘zero frequency’ (in the case of π 1 = 0 ) removes the trend. The other three roots which remove the seasonal structure imply stochastic cycles of biannual and annual periodicity (Charemza & Deadman, 1992, pp. 141-142). For HEGY procedure, we can test the hypotheses which are H 0 : π 1 = 0 against H 1 : π 1 < 0 (t statistic); H 0 : π 2 = 0 against H 1 : π 2 < 0 (t statistic) and H 0 : π 3 = π 4 = 0 against H 1 : π 3 ≠ π 4 ≠ 0 (F statistic). If we do not reject the null hypothesis of

π 1 = 0 , π 2 = 0 and π 3 = π 4 = 0 ; it is concluded about the existence of nonseasonal, biannual and annual unit roots respectively. For a series to include no seasonal unit roots, both π 2 = 0 and the joint F test which is π 3 = π 4 = 0 should be rejected. On the other hand, in conclusion to find out that a series is stationary and thus includes no unit roots at all (including at zero frequency), we must establish that each of the π ’s is different from zero (Hylleberg et al., 1990, pp. 221-223). In Table 2, a summary of long-run and seasonal frequencies has been presented for quarterly data: Table 2 Long Run and Seasonal Frequencies for Seasonal Unit Root Tests in Quarterly Data Tested hypothesis Frequency

Period

Cycles/year

Root

Filter

H 0 : Unit Root

0 Long run

∞

0

1

(1 − L)

π1 = 0

1; 3

±i

(1 + L2 )

π3 ∩π4 = 0

2

-1

(1 + L)

π2 = 0

π 3π 2

,

2

Annual

π

Semiannual

4;

4 3

2

Note. The information on first five columns have been obtained from Diaz-Emparanza & López-de-Lacalle (2006, p.7).


39

For the Monte Carlo values for the one sided t tests on π 1 , π 2 and the joint F test on

π 3 ∩ π 4 = 0 , see Hylleberg et al. (1990) (pp. 226-227).

Beaulieu and Miron (1992) have examined the HEGY testing procedure for monthly data. In that case, the differencing operator ∆ 12 will have 12 roots lying on the unit circle ( 1 − L12 = 0 ). For monthly data; the seasonal unit roots, their cycles and corresponding frequencies have been presented in table 3. While Beaulieu and Miron (1992) present a HEGY test equation for monthly case as in follows 12

ϕ ( L) * y13t = ∑ π k y k ,t −1 + ε t

(6)

k =1

where

ϕ * ( L)

is a polynomial associated with roots that are outside the unit

circle (see Beaulieu & Miron (1992, pp. 2-4)), The test equation for the presence of seasonal unit roots given in (6) takes a somewhat different form in Franses (1991) as follows

ϕ * ( L )y8 ,t = π 1 y1,t −1 + π 2 y2 ,t −1 + π 3 y3 ,t −1 + π 4 y3 ,t − 2 + π 5 y4 ,t −1 + π 6 y4 ,t − 2 +π 7 y5 ,t −1 + π 8 y5 ,t − 2 + π 9 y6 ,t −1 + π 10 y6 ,t − 2 + π 11 y7 ,t −1 + π 12 y7 ,t − 2 + µt + ε t where

ϕ * ( L)

is some polynomial function of L,

(7)

µ t represents the

deterministic component which may include a constant, seasonal dummies or a trend, and

y1,t = (1 + L)(1 + L2 )(1 + L4 + L8 ) y t = (1 + L + L2 + ...... + L11 ) y t y 2,t = −(1 − L)(1 + L2 )(1 + L4 + L8 ) y t y 3,t = −(1 − L2 )(1 + L4 + L8 ) y t y 4,t = −(1 − L4 )(1 − 3L + L2 )(1 + L2 + L4 ) y t y 5,t = −(1 − L4 )(1 + 3L + L2 )(1 + L2 + L4 ) y t

y 6,t = −(1 − L4 )(1 − L2 + L4 )(1 − L + L2 ) y t

y 7 ,t = −(1 − L4 )(1 − L2 + L4 )(1 + L + L2 ) y t

(8)

40


y8,t = (1 − L12 ) y t (Franses, 1991, p. 100; Maddala & Kim, 1998, p. 368). It is remarkable to say that in order to make the residuals white noise, augmented lagged values of y8,t should be used in (7). As in quarterly case, with these transformations of y t in (8) the seasonal unit roots are excluded at given frequencies while they are preserved at remaining frequencies. In Table 3, the outline of long run and seasonal frequencies has been presented. Table 3. Long Run and Seasonal Frequencies for Seasonal Unit Root Tests in Monthly Data Frequency

0 Long run

π 11π ,

6

6

Period

Cycles/year

Root

Filter

Tested hypothesis

∞

0

1

(1 − L)

π1 = 0

12; 1.09

1; 11

1 ( 3 ± i) 2

(1 − 3L + L2 )

π 11 ∩ π 12 = 0

6; 1.2

2; 10

1 (1 ± 3i ) 2

(1 − L + L2 )

π7 ∩π8 = 0

±i

(1 + L2 )

π3 ∩π4 = 0

(1 + L + L2 )

π5 ∩π6 = 0

(1 + 3L + L2 )

π 9 ∩ π 10 = 0

(1 + L)

π2 = 0

H 0 : Unit Root

Annual

π 5π 3

,

3

Semiannual

π 3π ,

2 2 2π 4π , 3 3

4;

4 3

3; 9

3; 1.5

4; 8

1 − (1 ± 3i ) 2

2.4; 1.7

5; 7

−

2

6

Quarterly

5π 7π , 6 6

π

Bimonthly

1 ( 3 ± i) 2 -1

Note. The information on first five columns have been obtained from Diaz-Emparanza & López-de-Lacalle (2006, p.7).

Applying OLS procedure to (7) gives estimates of the

π i . By the same logic in

quarterly case; if π 2 through π 12 are significantly different from zero (the case in which the null hypothesis of stochastic seasonality is not true), then there will be no seasonal unit roots and the pattern that the data display becomes


41

deterministic or constant seasonal. In this situation the dummy variable representation can be applied for modelling this pattern. Thus, in case there are seasonal unit roots, the corresponding

π i will be zero. Due to the fact that pairs

of complex unit root are conjugates, these roots will exist only in case pairs of π ' s are jointly equal to zero. If π 1 through π 12 are all unequal to zero, we experience a stationary seasonal pattern and seasonal dummy variables can be used to model such a pattern. At the same time, when the coefficient for a given π is statistically not different from zero, then it can be said that data have a varying seasonal pattern. If π 1 = 0 , we cannot reject the presence of root 1 with long-run frequency and if all

π i are equal to zero, it becomes suitable to

apply the (1 − L12 ) filter. If only some pairs of π ' s are zero, the relevant operators can be used. Since, all critical values of the test have a non-standard distribution; critical values are generated by Monte Carlo simulations (Franses, 1991, p. 101; Maddala & Kim, 1998, p. 370; Sørensen, 2001, p. 77 ). See also Feltham and Giles (1999) for semi-annual data and Rubia (2001) for weekly data in order to test seasonal unit roots with HEGY.

IV. Kunst Test Kunst (1997) suggests a general AR (s ) model in order to test (1 − L ) and it is s

described as

y t − y t − s = θ1 y t −1 + θ 2 y t − 2 + ...... + θ s y t − s + ε t

(9)

Kunst’s test bears resemblance to the DHF test in that it only detects the presence of all seasonal unit roots. The augmented Kunst model can be expressed as

∆ s yt = θ1 y t −1 + θ 2 y t −2 + ...... + θ s y t − s + κ 1 ∆ s y t −1 + ...... + κ p ∆ s y t − p + ε t

(10)

When s = 4 (that is, for quarterly time series), the Kunst test regression is given as

∆ 4 y t = α 1 y t −1 + ...... + α 3 y t −3 + δy t − 4 + ε t , ( t = 1,......, T ) which is an F-type test given as

(11)

42


Fαˆ* ,....,αˆ 1

3 ,δ

ˆ

= (T − 4)(εˆ0′ εˆ0 − εˆ ′εˆ ) /(εˆ ′εˆ )

(12)

where εˆ0 and εˆ1 vectors represent the estimated residuals under the null

H 0 : α 1 = ... = α 3 = δ = 0

and

alternative

hypotheses

respectively

(El

Montasser, 2011, p. 27; Zhang, 2008, pp. 9-10).

V.

OCSB Test

Osborn et al. (1988) have tried to investigate whether (1 − L) or (1 − L ) s

operators or both of them or none of them should be applied to data. The OCSB regression model in the original form is expressed as

∆ 1 ∆ s y t = β1 ∆ s y t −1 + β 2 ∆ 1 y t − s + ε t

(13)

This model can be generalized with deterministic components as follows:

η ( L)∆ 1 ∆ s y t = µ t + γ 1 ∆ s y t −1 + γ 2 ∆1 y t − s + ε t

(14)

where η (L) is an AR polynomial (lag polynomial with roots outside the unit circle), ∆ s = (1 − L ), ∆ 1 = (1 − L) and s

S −1

S −1

s =1

s =1

µ t = α 0 + ∑ α s D s ,t + β 0 t + ∑ β s D s ,t t

(15)

Here, t is a deterministic trend. In the original study, the seasonal trend is not given place in µ t i.e. β s = 0 for ∀ s. In order to find out which filter is suitable for y t , the significances of γ 1 and γ 2 are tested. When both γ 1 and

γ 2 are equal to zero ( γ 1 = γ 2 = 0 ), using ∆ 1 ∆ s filter is suitable. When

γ 1 = 0 and γ 2 ≠ 0 , ∆ 1 filter should be selected; when γ 1 ≠ 0 and γ 2 = 0 , ∆ s filter is suitable. If both γ 1 and γ 2 are unequal to zero ( γ 1 ≠ γ 2 ≠ 0 ), in that case no differencing filter is required. By the same logic given above, in the case of quarterly data OCSB testing regression is given as k

∆1∆4 yt = α 0 + α1 D1,t + α 2 D2 ,t + α 3 D3 ,t + γ 1∆4 yt −1 + γ 2 ∆1 yt −4 + ∑ φi ∆1∆4 yt −i +ε t (16) i =1


43

The necessary joint hypothesis about the usefulness of the ∆ 1 ∆ 4 operator is

α 1 = α 2 = α 3 = γ 1 = γ 2 = 0 . If γ 2 = 0 with γ 1 < 0 , the ∆ 4 filter is needed and if γ 1 = 0 and γ 2 < 0 , the ∆ 1 filter should be applied to data. For these three hypotheses, critical values are available only for a sample size of 136 in Osborn (1990) (Franses, 1998, p. 563; Maddala & Kim, 1998, p. 366; Zhang, 2008, p. 11; Platon, 2010, pp. 2-3). See Franses and Hobjin (1997) for critical values considering the extensions of HEGY and OCSB procedures.

VI. Conclusion This chapter has introduced a testing sequence for seasonal integration proposed by Ilmakunnas (1990) and seasonal unit root testing procedures: DHF, HEGY, Kunst and OCSB. In addition, it is of high importance to express that since most economic and financial time series data display seasonal variations at different frequencies; if a series has unit roots at which frequencies, differencing filters corresponding to those frequencies should be applied to the series in interest in order to make it stationary and these filters have been given place in Table 2 and Table 3 for quarterly and monthly data.

References Baltagi, B. (Ed.). (2001). A companion to theoretical econometrics. Oxford: Blackwell Publishers. Beaulieu, J. J., & Miron, J. A. (1992). Seasonal unit roots in aggregate U.S. data (NBER Technical Paper No. 126). Cambridge: National Bureau of Economic Research. Bell, W. R. (1987). A note on overdifferencing and the equivalance of seasonal time series models with monthly means and models with (0,1,1)12 seasonal parts when Θ = 1 . Journal of Business and Economic Statistics, 5, 383-387. Charemza, W. W., & Deadman, D. F. (1992). New directions in econometric practice: General to specific modelling, cointegration and vector autoregression (1st ed.). Aldershot, UK: Edward Elgar Publishing Limited.

44


Diaz-Emparanza, I., & López-de-Lacalle, J. (2006). Testing for unit roots in seasonal time series with R: The uroot package. Retrieved May 10, 2015, from http://www.jalobe.com:8080/doc/uroot.pdf. Dickey, D., Hasza, D., & Fuller, W. (1984). Testing for unit roots in seasonal time series. Journal of the American Statistical Association, 79, 355367. Dickey, D. A., & Pantula, S. G. (1987). Determining the order of differencing in autoregressive processes. Journal of Business and Economic Statistics, 5(4), 455-461. El Montasser, G. (2011). The overall seasonal integration tests under nonstationary alternatives. Journal of Economics and Econometrics, 54, 2439. Engle, R. F., Granger, C. W. J., & Hallman, J. J. (1989). Merging short and long run forecasts: An application of seasonal cointegration to monthly electricity sales forecasting. Journal of Econometrics, 40, 45-62. Feltham, S. G., & Giles, D. E. A. (1999). Testing for unit roots in semi-annual data (Econometrics Working Paper EWP No. 9912). Department of Economics, University of Victoria. Franses, P. H. (1991). Model Selection and Seasonality in Time Series. Doctoral dissertation, Erasmus University Rotterdam, Netherlands. Retrieved from http://hdl.handle.net/1765/2047. Franses, P. H. (1998). Modeling seasonality in economic time series. In A. Ullah & D.E.A. Giles (Eds.), Handbook of Applied Economic Statistics (pp. 553-577). New York: Marcel Dekker. Franses, P. H., & Hobijn, B. (1997). Critical values for unit root tests in seasonal time series. Journal of Applied Statistics, 24(1), 25-48. Ghysels, E., Lee, H. S., & Noh, J. (1994). Testing for unit roots in seasonal time series: Some theoretical extensions and a Monte Carlo investigation. Journal of Econometrics, 62, 415- 442. Ghysels, E., & Osborn, D. R. (2001). The econometric analysis of seasonal time series. Cambridge: Cambridge University Press. Hylleberg, S., Engle, R., Granger, C., & Yoo, S. (1990). Seasonal integration and cointegration. Journal of Econometrics, 44, 215-238.


45

Ilmakunnas, P. (1990). Testing the order of differencing in quarterly data: An illustration of the testing sequence. Oxford Bulletin of Economics and Statistics, 52, 79-88. Kunst, R. M. (1997). Testing for cyclical non-stationarity in autoregressive processes. Journal of Time Series Analysis, 18, 123-135. Maddala, G. S., & Kim, I. M. (1998). Unit Roots, cointegration and structural change. Cambridge: Cambridge University Press. Osborn, D. R. (1990). A survey of seasonality in UK macroeconomic variables. International Journal of Forecasting, 6, 327-336. Osborn, D.R., Chui, A. P. L., Smith, J. P., & Birchenhall, C. R. (1988). Seasonality and the order of integration for consumption. Oxford Bulletin of Economics and Statistics, 50, 361-377. Platon, V. (2010). Application of seasonal unit roots tests and regime switching models to the prices of agricultural products in Moscow 1884-1913. Retrieved, January 4, 2015, from http://www.hse.ru/data/2010/10/22/1222675037/Seasonal%20unit%20r oots%20and%20regime%20switch.pdf. Rubia, A. (2001). Testing for weekly seasonal unit roots in daily electricity demand: Evidence from deregulated markets (IVIE Working Paper No. 2001-21). Instituto Valenciano de Investigaciones Económicas, S.A. Sanli, S. (2015), The Econometric Analysis of Seasonal Time Series: Applications on Some Macroeconomic Variables, Master’s Thesis, Cukurova University, Adana. Sørensen, N. K. (2001). Modelling the seasonality of hotel nights in Denmark by county and nationality. In T. Baum & S. Lundtrop (Eds.), Seasonality in tourism (pp. 75-88). Oxford: Elsevier. Tam, W. K., & Reinsel, G. C. (1997). Tests for seasonal moving average unit root in ARIMA models. Journal of the American Statistical Association, 92(438), 725-738. Zhang, Q. (2008). Seasonal unit root tests: A comparison. Doctoral Dissertation. North Carolina State University, Raleigh.

46


Chapter 4

Study on Relative Efficiency with Data Envelopment Analysis: Metropolitan Municipalities in Turkey* Mine AYDEMİR 1 Nuran BAYRAM ARLI 2

Abstract. Resource use and allocation is an important issue to focus on. In order to achieve the most efficient use of resources, these efficiency levels can be reviewed on a per-unit basis, suggesting improvements. At the same time, these units can be compared to other reference units. There are many analysis that can be used for efficiency measurements. One of these is data envelopment analysis, which is used to measure relative efficiency, which can work with multiple inputs and multiple outputs. Calculating the financial terms of the metropolitan municipalities in Turkey and efficiency level relative comparison is the objective of this study. 11 of the 16 metropolitan municipalities compared with the data envelopment analysis were found to be efficient while 5 were not efficient. It has been determined that inefficient municipalities may become more efficient by interfering with some inputs of the input-oriented model. Reference sets have been established for those inefficient in 16 municipalities, which are relatively comparable.

Keywords: Efficiency, Financial Efficiency, Data Envelopment Analysis, Turkish Metropolitan Municipalities

Res.Ass. Uludağ University, Faculty of Economics and Administrative Sciences, Department of Econometrics, Görükle Campus, 16059 Nilüfer/Bursa, [email protected] 2 Prof., Dr., Uludağ University, Faculty of Economics and Administrative Sciences, Department of Econometrics, Görükle Campus, 16059 Nilüfer/Bursa, [email protected] * This article has been written using a part of the master's thesis. 1

48

Study on Relative Efficiency with Data Envelopment Analysis: Metropolitian Municipalities in Turkey

I.

Introduction

In today's conditions, it has become a necessity to prevent wastage and idle usage and to carry out performance measurements to achieve certain goals such as ensuring social welfare and development, the proper growth of the economy and making necessary investments. To ignore them will lead to the collapse of the economies, the emergence of crises, and the failure of institutions and also to the withdrawal from social welfare as a result of the incorrect or careless usage of sufficient resources. Changes are experienced in economy, politics, management and technology in all fields of life. A different understanding, perspective or a different technique emerges regardless of the public or private sector, and the operations that are not the solution to current situations are desired to be changed. The rearrangements in the public sector and the beginning of the development of different aspects coincide with the 1980s. These years are the beginning of a progress from conventional method to more contemporary methods. The most important concept at this point is performance (Akçakaya, 2012: 172). The determination of the point where an institution is located and being able to take it to the best point that it can reach are associated with the determination of its performance. To take right decisions depends on carrying out performance measurements with the correct data. Performance measurements, the first applications of which were observed in the private sector, have also begun to be used by the public institutions over time and even have become a necessity. They should use their resources in the most efficient way even if they do not seek profit. It will also be a good practice to combine performance measurements with the budget within this goal (Kaygısız, 2011: 1). Performance measurement is required for the production of better quality goods and services and is a very important part of management (Yenice, 2002: 58-59). Resources are aimed to be used within the private or public economy to meet social requirements. The best usage of resources primarily requires the examination of the present conditions and then the determination of methods to make them the most efficient. Accordingly, with respect to the determination of efficiency, various analyses are used regarding to what extent and to which part the improvement should be made. These analyses are divided into ratio


49

analyses, parametric analyses and non-parametric analyses within themselves. The technical efficiency and the efficiency measurement of decision-making units were mentioned in the article written by Farrell in 1957, and then the foundations of the Data Envelopment Analysis (DEA), which performs the efficiency measurement, were laid by using multiple inputs and multiple outputs in another article written by Charnes, Cooper and Rhodes in 1978. Thus, the DEA which is a linear programming-based non-parametric method using multiple input and output variables at the same time started to be used in the efficiency measurement. In many countries of the world, state institutions and organisations constitute a significant part of total economic activities (Sharman, 1981: 1.5). There is a similar situation in Turkey. Due to their important position within the economy, it is necessary that the resources of the public sector are used properly, the values are kept under control, and the outputs are the best of all. The most important one among the local administrations in Turkey is municipalities. The purpose of this study is to measure the financial efficiency of the Metropolitan Municipalities, which are discussed among municipalities, with the DEA, and to obtain results on how each ineffective municipality can become efficient by determining efficient municipalities and making them a reference set for the others.

II.

Method

Decision-Making Unit (DMU) is defined as an institution, organisation, business, etc. from which information is desired to be obtained about its performance. DMUs are the units that are examined in the studies carried out and the efficiency of which is desired to be measured. Universities, courts, municipalities and branches can be shown as an example for DMU. The DMU term was first used in the CCR model, which was created by Charnes, Cooper and Rhodes and is shown with their initials. DMU converts inputs into outputs, and the most critical point in its evaluation is the determination of inputs and outputs. The outputs used in this evaluation should reflect all useful results. The inputs should include all resources that can have an effect on outputs. Besides, there may be environmental factors that may affect the conversion (Thanassoulis, 2001: 22). The inputs and outputs used for the realisation of processes are different in most cases. Their types, quantities, units and many

50


other features are different. This situation may cause troubles in performing the efficiency and productivity analyses and interpreting the results. The Data Envelopment Analysis, which is included in nonparametric analysis methods, was used in this study. Non-parametric methods do not have a functional form. These methods are based on optimisation under certain constraints. Their assumptions are more flexible compared to parametric methods, and they can perform measurements using multiple inputs and outputs at the same time (Kecek, 2010: 53). In non-parametric methods, all deviations from the efficient frontier are considered to be inefficiency (Tarım, 2001: 46). It is not possible to include all inputs and outputs that are involved in the process in the analysis while the efficiency of the organisation is investigated. The use of multiple inputs and outputs that can explain the efficiency in the best way and are selected within the scope of certain rules and performing evaluations in relation to the organisation increase the possibility of finding the points that should be addressed (Charnes et al., 1981: 669). The DEA makes comparisons based on the best rather than average and determines whether DMUs are efficient or not. The DEA creates an efficiency score for each DMU. This efficiency score is found by the ratio of the weighted outputs to the weighted inputs and is between zero and one. DMU, the efficiency score of which equals to one, works fully efficiently. Weighting can be realised as the optimisation of the profit maximisation in profit-oriented organisations, but other methods are needed for weighting in non-profit organisations (public administrations) and the DEA comes into play at this point (Tarım, 2001: 49). Virtual weights, the value of which is not yet known, are assigned to inputs and outputs, and an attempt to make the outputs to inputs ratio maximum is made using linear programming. Weights vary from one DMU to another, and the DEA uses them by identifying the best ones (Cooper et al., 2006: 21). The point that should be taken into account is that the calculated efficiency scores are only for the DMUs addressed. This means that the values obtained are not absolute but relative efficiency values. The DMU that makes the highest production until a better DMU is added to the set will be accepted as technical-efficient (Tarım, 2001: 203). The model in which the DEA was first used is the CCR model which was presented by Charnes, Cooper and Rhodes and works under the assumption of Constant Returns to Scale (CRS). In later times, Banker et al. (1984) proposed a model that works under the assumption of a Variable Returns to Scale (VRS),


51

and this model took the initials of Banker, Charnes and Cooper and began to be used as the BCC model (Güneş, 2006: 3). The DEA works under the assumptions of constant returns to scale (CCR, CRS) or a variable return to scale (BCC, VRS). CRS should be used if all DMUs are working at the optimum scale, but VRS should be used if there is a reverse situation. Measurement is also performed under both assumptions to determine it. If the VRS values are found to be greater than the CRS values, this means that DMUs do not work under the assumption of CRS. Furthermore, Scale Efficiency is measured in the VRS model. The scale efficiency shows whether the DMU is working at the optimum scale but does not give its cause. Scale inefficiency can be caused by increasing returns to scale or decreasing returns to scale (Düzgün, 2011: 101). The output-oriented model creates the input combination that can acquire the maximum output. An attempt to acquire output using minimum input is made in the input-oriented model. The output-oriented CCR model investigates how much the output combination needs to be increased while the input level remains the same. The ratio of weighted input to weighted output is minimised. In the CCR input-oriented model, the output variables are kept constant and how much input components should be decreased is examined to achieve the same output level (Metthews&Ismail, 2006: 7). If the purpose in a CCR model is to reach the desired outputs by minimising the inputs, this is an input-oriented model. Here, an attempt to maximise the output is made with the available inputs without using more input. The DEA has a very flexible methodology. It includes many models that can be used in the process of converting inputs into outputs in an organisation. The researcher should spend enough time and make enough effort to understand the production process in any application. The most appropriate model can be chosen only in this way (Banker, 1992: 344). The first point in the selection of DMUs is homogeneity. Each DMU should be similar to another and should act with similar purposes. Another point should be the DMUs that produce similar outputs using similar inputs. Sizes and units can be different, but each DMU should make a production with similar inputs and outputs to be able to perform the measurement. When these are provided, the efficiency measurements can be performed for each DMU, and ultimately, the comparison of inefficient DMUs can be made with the reference sets to be

52


created. It will not be healthy to make this comparison in a case when there is not homogeneity (Ramanathan, 2003: 173). There are no fixed rules for determining inputs and outputs. Good information about the function and structure of the DMU, the efficiency of which is desired to be measured, is required during this determination. To know the structure and functioning will help to determine the effects and to identify the input-output relationships (Banker, 1992: 344). The increase in the number of inputs and outputs weakens the discrimination power of the DEA. DMUs that are actually inefficient may appear efficient due to insignificant inputs or outputs (Yıldırım, 2010: 144).

III.

Municipalities

How to make production and distribution in society, which methods will be used, by whom the decisions will be taken are important questions. People in society have common needs and there is a need for central and other administrations to meet these needs and ensure welfare. Administrative organisation is classified as central administration and local administration. The structures, types, numbers and authorities of local administrations are different in each country. This is due to the social and legal structures of these countries (Nadaroğlu, 2001: 3-4). The services provided by local administrations in Turkey can be simply grouped under three headings. The first one includes the goods and services offered in full public nature (bridge construction, city library, public security services, street lighting, marriage services, etc.). The production of these goods and services is made by tax revenues. The second one includes the goods and services offered in semi-public nature (garbage collection, paid parking, water disposal, etc.). In these services, the beneficiaries are involved in service charges. The third one includes the goods and services offered in the form of private goods (veterinary services, transport services, bread and bakery products, water, electricity, etc.). Such goods are intended for the market and can be priced and marketed. The services offered by the local administrations with the central administration, in other words, the services possible to be offered by both sides have been determined as the health services, education and cultural services, environmental services, public works and zoning services, and social services (Çetinkaya, 2012: 50-51).


53

The most important of the local administrations in Turkey is municipalities. The number of the Metropolitan Municipalities is 16 by the year 2013. Municipalities have been established to meet the needs of the inhabitants. Important sources of revenue have been allocated. The budget revenues and expenditures of 16 Metropolitan Municipalities realised in 2013 are presented in Appendix Table 1. Metropolitan Municipalities, the revenues of which remained below their expenditures, are İstanbul, Kocaeli and Sakarya Metropolitan Municipalities. The expenditures of the other 13 Metropolitan Municipalities did not exceed their revenues. While the municipal budgets are prepared, some of the services required to be realised due to resource constraints are preferred, the others may be delayed due to constraints. The goal should be to produce the maximum output with the best use of resources. Municipalities are one of the first authorities to apply for the solution to a problem that may occur since they are one of the closest administrative units to the public. In particular, the fact that people take an active role in determining the needs and services will also provide comfort for the solution of problems. When Appendix Table 2 was examined, as a result of the expenses of metropolitan municipalities according to the economic classification and the realization of the revenues they obtained in 2013, capital expenditures (42.8%) had the greatest share among expenditures, and the interests, shares and fines (75.9%), also referred to as other revenues, had the greatest share among revenues. It is observed that the budget balance was negative.

IV.

Findings

DMUs are the units to be investigated in terms of their efficiency. In the study, Metropolitan Municipalities in Turkey are the decision-making units, the relative efficiency of which is desired to be investigated. The primary assumption in the selection of DMUs is that they are homogeneous. When the structure of the Metropolitan Municipalities in Turkey was taken into account, and as it was aimed to reveal the financial efficiency in the analysis, 3 input (population, current expenditures (personnel expenditures+social security expenditures+purhcases of Goods and Services), investment expenditures (Interest Expenditures+Capital Payments+Capital Transfers)) and 3 output (tax revenues, enterprise and ownership revenues, other revenues) variables were

54


determined. The municipalities in Turkey are weak in terms of income. The large part of their revenues consists of shares received from the central administration. Municipalities will have limited intervention in revenue due to this negative aspect in revenues. This situation is observed when Appendix Table 3 is examined. In 2012 and 2013, the items of Interests, Shares and Fines of the municipalities constituted an average of 50% of the total revenues. While revenues are determined by the central administration, services are provided by the local administrations. The shares transferred from the centre are more compared to own revenue (Tekeli, 2002: 79). By taking into consideration them, population, current expenditures and investment expenditures were selected as the input variables, and tax revenues, other revenues and enterpriseownership revenues were selected as the output variables. When Appendix Table 2 is examined, it is observed that the highest two percentages in the revenues section are Interests, Shares and Fines (75.9%), that were included in the study with the name of Other Revenues, and Enterprise and Ownership Revenues (14%). Since this study focused on the financial structure, these two variables were determined as the output variables. Besides, it is observed that the Tax Revenues variable which is frequently used as an output variable in the literature has a small value of 1.6%. This variable is expected to have a small role on efficiency by the structure in Turkey. DEA models can be created as input or output-oriented in reaching the efficient frontier. It can be used by reducing the amounts of inputs without changing the outputs or by increasing the outputs without changing the inputs (Li&Liu, 2011: 51). With the variables used, Metropolitan Municipalities are expected to increase their efficiency by decreasing the number of inputs without changing their outputs. Municipalities in Turkey cannot generate sufficient own revenues. As a result of this, there is a dependency on the centre. Starting from this point of view, it can be said that the autonomy of the municipalities is limited. The ratio of transfers in total revenues indicates the dependence on the centre. For all municipalities, this ratio was calculated to be 52.5%, 52.3% and 52.7% in 2011, in 2012 and in 2013, respectively. This ratio reaches approximately 60% for Metropolitan Municipalities (Local Administrations Activity Report, Year 2013). The metropolitan municipalities in Turkey, which are dependent on the central unit in terms of revenue, will not be able to intervene at this point to differentiate efficiency or get the results they want. The input-oriented model


55

should be preferred by taking into account these circumstances. The models to be created under the assumptions of CRS and VRS will be examined separately and one of them will be preferred. As a result of the analysis performed with three input and three output variables specified, it was determined that the tax revenue variable did not differentiate efficiency in the models, it was also observed that the efficiency scores were the same when the tax revenue variable was removed, therefore, it was excluded from the model. The analysis was maintained with 3 inputs and 2 outputs. In the model used as input-oriented, population, current expenditures and investment expenditures were taken as the input variables, and other revenues and enterprise and ownership revenues were taken as the output variables. The relative efficiency scores obtained when the DEA was performed are presented in Appendix Table 3. When efficiency scores are examined, it is observed that 8 Metropolitan Municipalities (Adana, Antalya, Diyarbakır, Istanbul, Izmir, Kayseri, Kocaeli, Sakarya) were efficient and 8 Metropolitan Municipalities (Ankara, Bursa, Erzurum, Eskişehir, Gaziantep, Konya, Mersin, Samsun) were inefficient in the efficiency scores calculated under the assumption of CRS, and that 11 Metropolitan Municipalities (Adana, Ankara, Antalya, Diyarbakır, Erzurum, Eskişehir, Istanbul, Izmir, Kayseri, Kocaeli, Sakarya) were efficient and 5 Metropolitan Municipalities (Bursa, Gaziantep, Konya, Mersin, Samsun) were inefficient in the efficiency scores calculated under the assumption of VRS. As a result of the efficiency measurement, it was observed that some Metropolitan Municipalities were operating fully efficiently while some of them were not operating efficiently. For Metropolitan Municipalities, that are not functioning efficiently, to become efficient, Metropolitan Municipalities that are most suitable for their own structures were designated as a reference. Reference sets are presented in Appendix Table 4. The municipalities in the reference set are examples for the inefficient municipality to become efficient. For example, when the DEA is examined, Bursa Metropolitan Municipality can take Antalya, Izmir, Kayseri, Kocaeli and Sakarya Metropolitan Municipalities as a reference to become efficient. The Metropolitan Municipality that should primarily take reference is Antalya Metropolitan Municipality. It can also turn to the other Metropolitan Municipalities in the reference set if its values that should be differentiated to become more efficient are incompatible with this municipality

56


and if it wants to prefer a different decrease or increase path to become efficient more precisely. The DEA makes a comparison in terms of efficiency and also identifies targets, and it says that the efficiency score can be improved by achieving these goals (Talluri, 2000: 9). Based on this, the potential improvement ratios can be calculated. The inefficient DMU should decrease the variable value in the ratio which is found if the calculated improvement ratio is negative, should increase in the ratio which is found if it is positive and should not perform any operation if it is zero (Bakırcı&Babacan, 2010: 227). Potential improvement ratios for inefficient DMUs are presented in Appendix Table 5. All of the values in the input variables are marked as negative. This means that it is possible to increase efficiency by making a reduction in the variables in various percentages. The efficiency values can be differentiated by making changes in the input variables since the input-oriented model is used. Therefore, the ratios of change in the input variables should be examined, and it should also be investigated which variable should be addressed.

V.

Conclusion

The concept of globalisation manifests itself in all fields of life. This concept, which affects almost every situation, has also led to the necessity of keeping pace and being adapted. Countries, institutions and individuals should take this into account while taking decisions and be able to determine their requirements correctly. Certain rules and regulations within the system have expired with globalisation, and new formations have been needed for the flow to continue. Along with the change in the understandings of management and functioning, performance has come to the forefront to reveal them, and the need for the tools such as efficiency, productivity and quality has further increased. Efficiency, from among these concepts each of which represents a different dimension of performance, refers to reaching the best outputs with the available inputs. In this study, 16 metropolitan municipalities in Turkey were designated as DMUs and the data in the year 2013 were used. Following the literature review, the input and output variables used in the measurement of financial efficiency were determined. For the financial efficiency measurement, 3 input variables were determined to be population, current expenditures and investment expenditures while 3 output variables were determined to be tax revenues, enterprise and ownership revenues and other revenues. As a result of the


57

analysis performed as input-oriented, it was found out that the tax revenues variable did not contribute to efficiency. This result is consistent with the expectation. Tax revenues should be increased at very high ratios for inefficient units to become efficient. For this reason, the tax revenues variable was excluded from the study. This result is an expected result when the revenueexpenditure structures of the municipalities in Turkey are considered. The metropolitan municipalities, which are financially dependent on the centre, have troubles in creating their own revenues. In other words, they show a weak structure in terms of revenues. In the study carried out by Kaygısız (2011), it was determined that Adana, Antalya, Gaziantep, Kayseri, Kocaeli and Samsun Metropolitan Municipalities did not work efficiently and that the other Metropolitan Municipalities were operating fully efficiently, in the financial DEA model established with the data of 2014. In the study carried out by Kutlar et al. (2012) with 25 municipalities consisting of both metropolitan and provincial municipalities using the data of 2008, it was concluded that Samsun, Izmir, Erzurum, Ankara and Diyarbakır metropolitan municipalities did not work efficiently. Çağlar (2003) examined 32 provincial and Metropolitan Municipalities using the data of 2001 and concluded that Adana, Diyarbakır, Istanbul and Izmir Metropolitan Municipalities were operating fully efficiently while the other municipalities did not work fully efficiently. Ertuğ&Girginer (2015) were found inefficient municipalities (Bursa, Konya, Mersin) same as our study. In this study in which the efficiency of the Metropolitan Municipalities was investigated, it was observed that Bursa, Gaziantep, Konya, Mersin and Samsun Metropolitan Municipalities did not work fully efficiently in the financial DEA model established with the data of 2013. These results are consistent with the results of other studies. To be able to obtain the best outputs with the resources in hand in the public sector, as well as in other sectors, will make that institution successful and will move the community it serves forward. It will not be enough to ensure efficiency only by the fact that the figures are in the place where they should be in a mechanical manner. Municipalities should determine the services to be provided by taking into account the wishes and needs of people. Based on these determinations, they should distribute their resources and should choose the ways when the efficient result can be achieved while using the resources they have. Only in this case, they will have ensured that people to whom they offer

58


service reach welfare and their satisfaction increases, as well as providing financial efficiency.

References Akçakaya, M. (2012). Kamu Sektöründe Performans Yönetimi ve Uygulamada Karşılaşılan Sorunlar. Journal of Black Sea Studies, 32, 171-202. Bakırcı, F. & Babacan, A. (2010). İktisadi ve İdari Bilimler Fakültelerinde Ekonomik Etkinlik. Atatürk Üniversitesi İktisadi ve İdari Bilimler Dergisi, 24(2), 215-234. Banker, R. D. (1992). Selection of Efficiency Evaluation Models. Contemporary Accounting Research. 9(1), 343-355. Charnes, A., Cooper, W.W. & Rhodes, E. (1981). Evaluating Program and Managerial Efficiency: An Application of Data Envelopment Analysis to Program Follow Through. Management Science, 27, USA. Cooper, W.W., Seinford, L. M. & Tone, K. (2006). Intoduction to Data Envelopment Analysis and Its Uses. Springer, USA. Çağlar, A. (2003). Veri zarflama analizi ile belediyelerin etkinlik ölçümü. Hacettepe Üniversitesi Fen Bilimleri Enstitüsü, Yayımlanmamış Doktora Tezi, Ankara. Çetinkaya, Ö. (2012). Mahalli İdareler Maliyesi, 3. b., Ekin Basın Yayın Dağıtım, Bursa. Deniz, N. (2009). Türkiye’deki İllerin Kaynak Kullanımına Göre Göreli Etkinliklerinin Klasik ve Bulanık Veri Zarflama Analizi Yöntemleriyle Belirlenmesi. Anadolu Üniversitesi Fen Bilimleri Enstitüsü, Yayımlanmamış Yüksek Lisans Tezi, Eskişehir. Düzgün, M. (2011). Veri Zarflama Analiziyle Elektrik Dağıtım Şirketlerinin Etkinlik ve Verimlilik Analizi, Ankara Üniversitesi Sosyal Bilimler Enstitüsü, Yayımlanmamış Yüksek Lisans Tezi, Ankara. Ertuğ, Z. K. & Girginer, N. (2015). Financial Effıciency Analysis Of Metropolitan Municıpalities with Integrated DEA And GRA: The Case of Turkey, Uluslararası İktisadi ve İdari İncelemeler Dergisi, 15: 411-428. Güneş, T. (2006). Bulanık Veri Zarflama Analizi, Ankara Üniversitesi Fen Bilimleri Enstitüsü, YayımlanmamışYüksek Lisans Tezi, Ankara.


http://www.migm.gov.tr/Dokumanlar/2008_faaliyet_raporu[1].pdf tarihi: 17/05/2015)

59

(Erişim

http://www.mta.gov.tr/mevzuat/duyurular/duyanalitik-butce-3.pdf (Erişim tarihi 19/06/2015) https://portal.muhasebat.gov.tr (Erişim tarihi: 18/06/2015) Kaygisiz, Z. (2011). Belediyelerin Performanslarının Maliyet Analizi Yaklaşımlarıyla Değerlendirilmesi. Eskişehir Osmangazi Üniversitesi Fen Bilimleri Enstitüsü, Yayımlanmamış Doktora Tezi, Eskişehir. Kecek, G. (2010). Veri Zarflama Analizi Teori ve Uygulama Örneği, Siyasal Kitabevi, Ankara. Kutlar, A., Bakırcı, F. & YükseL, F. (2012). An Analysis on the Economic Effectiveness of Municipalities in Turkey. African Journal of Marketing Management, 4(3), 80-98. Li, Y. & Liu, C. (2011). Construction Capital Productivity Measurement Using a Data Envelopment Anaysis. International Journal of Construction Management, 11(1), 49-61. Matthews, K. & Ismail, M. (2006). Efficiency and Productivity Growth of Domestic and Foreign Commercial Banks in Malaysia, Cardiff Economics Working Paper, No. E2006/2. Nadaroğlu, H. (2001). Mahalli İdareler, 7.b., Beta Basım, İstanbul. Ramanathan, R. (2003). An Introduction to Data Envelopment Analysis A Tool for Performance Measurement, Sage Publications, New Delhi. Sharman, H. D. (1981). Measurement of Hospital Technical Efficiency: A Comparative Evaluation of Data Envelopment Analysis and Other Efficiency Measurement Techniques for Measuring and Location Inefficency in Health Care Organizations, Harvard University, Phd Thesis. Talluri, S. (2000).Data Envelopment Analysis: Models and Extensions, Production. Operations and Management, 31(3), 8-10. Tarım, A. (2001). Veri Zarflama Analizi, Matematiksel Programlama Tabanlı Göreli Etkinlik Ölçüm Yaklaşımı, Sayıştay Yayınları, Ankara. Tekeli, R. (2002). The Design and Effects of Intergovenmental Transfers: The Case of Turkish Municipalities, University of Leicester, Phd Thesis.

60


Thanassoulis, E. (2001). Introduction to the Theory and Application of Data Envelopment Analysis, A Foundation Text with Integrated Software, Springer Science, New York. Yenice, E. (2002). Kamu Kesiminde Performans Ölçümü ve Bütçe İlişkisi. Sayıştay Dergisi, 61. Yıldırım, İ. E. (2010). Veri Zarflama Analizinde Girdi ve Çıktıların Belirlenmesindeki Kararsızlık Problemi İçin Temel Bileşenler Analizine Dayalı Bir Çözüm Önerisi. İstanbul Üniversitesi İşletme Fakültesi Dergisi, 39(1), 141-153. Appendix Table 1. Metropolitan Municipality Population, Budget Revenue and Budget Expenditures Metropolitian Municipality

Population

Total Budget Revenues Adana Metropolitan Municipality 2.149.260 616.763.152 Ankara Metropolitan Municipality 5.045.083 5.181.213.416 Antalya Metropolitan Municipality 2.158.265 423.776.646 Bursa Metropolitan Municipality 2.740.970 883.330.544 Diyarbakır Metropolitan Municipality 1.607.437 226.989.082 Erzurum Metropolitan Municipality 766.729 145.761.051 Eskişehir Metropolitan Municipality 799.724 374.774.582 Gaziantep Metropolitan Municipality 1.844.438 649.524.436 İstanbul Metropolitan Municipality 14.160.467 8.596.356.876 İzmir Metropolitan Municipality 4.061.074 2.096.697.731 Kayseri Metropolitan Municipality 1.295.355 495.640.698 Kocaeli Metropolitan Municipality 1.676.202 1.268.401.851 Konya Metropolitan Municipality 2.079.225 890.807.510 Mersin Metropolitan Municipality 1.705.774 414.607.706 Sakarya Metropolitan Municipality 917.373 183.320.920 Samsun Metropolitan Municipality 1.261.810 238.090.942 Source:http://www.tuik.gov.tr, Metropolitan municipality annual reports (2013)

Total Budget Expenditures 523.436.166 3.714.345.961 390.593.659 802.673.540 206.863.269 130.906.935 218.419.306 525.997.949 8.939.033.998 2.086.778.138 476.019.910 1.392.944.334 514.227.030 412.348.384 205.006.663 209.151.144


61

Table 2. Economic Classification (2013) Metropolitan Municipalities (Million TL) Expenditures 22.682 Personnel Expenditures 1.971 Social Security Expenditures 308 Purhcases of Goods and Services 6.097 Interest Expenditures 594 Current Transfers 1.270 Capital Payments 9.716 Capital Transfers 722 Loans 2.004 Revenues 20.720 Tax Revenues 326 Enterprise and Ownership Revenues 2.907 Grands and Aids and Other Special Revenues 138 Interests, Shares and Fines 15.730 Capital Revenues 1.535 Recovery of Debts 85 Budget Balance -1.962 Source: http://www.muhasebat.gov.tr

Percent 8.7 1.4 26.9 2.6 5.6 42.8 3.2 8.8 1.6 14.0 0.7 75.9 7.4 0.4 100

Table 3. Efficiency Scores (2013) Decision Making Units Adana Metropolitan Municipality Ankara Metropolitan Municipality Antalya Metropolitan Municipality Bursa Metropolitan Municipality Diyarbakır Metropolitan Municipality Erzurum Metropolitan Municipality Eskişehir Metropolitan Municipality Gaziantep Metropolitan Municipality İstanbul Metropolitan Municipality İzmir Metropolitan Municipality Kayseri Metropolitan Municipality Kocaeli Metropolitan Municipality Konya Metropolitan Municipality Mersin Metropolitan Municipality Sakarya Metropolitan Municipality Samsun Metropolitan Municipality

CRS Efficiency Scores 1.0000 0.8111 1.0000 0.7772 1.0000 0.5807 0.5152 0.5257 1.0000 1.0000 1.0000 1.0000 0.5663 0.8570 1.0000 0.6619

VRS Efficiency Scores 1.0000 1.0000 1.0000 0.7777 1.0000 1.0000 1.0000 0.6482 1.0000 1.0000 1.0000 1.0000 0.5817 0.8625 1.0000 0.8326

Table 4. DEA Reference Sets (2013) Decision Making Units

Reference Sets

Bursa Metropolitan Municipality Gaziantep Metropolitan Municipality Konya Metropolitan Municipality Mersin Metropolitan Municipality Samsun Metropolitan Municipality

Antalya, İzmir, Kayseri ,Kocaeli, Sakarya Erzurum, İzmir, Kayseri, Kocaeli Adana, Kayseri, Kocaeli, Sakarya Antalya, İzmir, Kocaeli, Sakarya Erzurum, İzmir, Kayseri, Sakarya

62


Table 5. Potential Improvement Ratios for Inefficient DMUs Inefficient Decision Making Units

Population (%)

Current Expenditures (%)

Investment expenditures (%)

Bursa Metropolitan Municipality Gaziantep Metropolitan Municipality Konya Metropolitan Municipality Mersin Metropolitan Municipality Samsun Metropolitan Municipality

-22.23 -35.17 -41.83 -13.75 -29.00

-22.23 -35.17 -61.76 -15.94 -16.74

-22.24 -43.28 -41.82 -13.75 -16.74

Chapter 5

Monitoring of a Production Process with CUSUM and EWMA Quality Control Charts Hakan EYGÜ 1

Abstract. Control charts are statistical methods used with aim of seeing output data dealing with production and to detect the change occurring in the process during production. In the case of deviation of the standards predetermined of the products manufactured in production process, to prevent of the possible losses which may exist during intervention are of great importance as regards time and cost. So, statistical process control used commonly in foreign literature in today. The aim of this study is to show the use of control charts which are the quality tools for increasing the levels of products or services. In the conducted study, in order to determine the level of aimed quality level, cumulative sum quality control charts (CUSUM) and exponentially weighted moving average control charts (EWMA) are used. In the application, data set including varieties affecting sugar quality produced in that firm were used. Also, which of the charts in what situations are used has been mentioned. The control chart for the sugar industry will be helpful for the enterprise applications which will be held and theoretical studies.

Keywords: Quality, Statistical Quality Control, CUSUM and EWMA Control Charts

1

Asst. Prof. Dr., Ataturk University, Faculty of Economics and Administrative Sciences Department of Econometrics, [email protected] , https://orcid.org/0000-0002-41042368

64 Monitoring of A Production Process with Cusum and Ewma Quality Control Charts

I.

Introduction

It can be said that the concept of quality is the most important characteristic of any product. Due to this feature, the development in the field of quality control continues in the first quarter of the twenty-first century. Because consumers' awareness is increasing, it shows that consumer demands are inevitable. This situation has increased competition among firms. Manufacturers have begun to produce high quality products at minimum cost. It is inevitable to measure the quality of products in many industries. For this reason, researches have been carried out on the development of methods related to this subject. Statistical process control methods allow a product to be produced to meet the most economic and needs. For this purpose the collected data are used at all stages of production using statistical techniques. In this phase, control methods help Statistical Process Control (SPC) practitioners to identify the time of a change after a control chart generates a signal. Using change point estimation with the monitoring tool surely improves the special cause detection ability of the monitoring system. We have assumed the researcher is knowledgeable about univariate statistical estimation and control procedures (such as Shewhart charts). Control charts which were first developed by Shewhart is widely used in order to detect the causes of variability. This control graphics are not sensitive to small shifts in the process. Alternatively, cumulative sum quality control charts (CUSUM) and exponentially weighted moving average control charts (EWMA) were developed. This charts are often used in detecting unusual changes in variables that are independent and thus not influenced by the behavior of other variables. These changes occur often in industrial settings. These may range from research to maintaining the quality of production. Thus, the practitioner is not well prepared to face the problems encountered when applying a control chart procedure to a real process situation. These problems are further compounded by the lack of adequate computer software to do the required complex computations.

II.

Literature review

There are many studies on control charts in the literature. Shewhart, in a bell telephone laboratories memorandom dated 1924, presented the first sketch of a


65

univariate control charts (Duncan, 1986). Although his initial chart was for monitoring the percentage defective in a production process, he later extended his idea to control charts for the average and standart deviation of a process. Shewhart chart designed the mean, of a group of perocess observations taken on a process variable at the same time point. Drawn on the chart are the upper control limit (UCL) and the lower control limit (LCL). Shewhart charts are often used in detecting unusual cahanges in variables (Mason and Young, 2002). Mendel (1969) showed that the combination of control graphs and regression analysis is effective in controlling the change around the mean. Oktay (1994) indicated that the varieties of Shewhart, Cusum and Ewma control charts procedure have been compared. Two very effective alternatives to the Shewhart control chart may be used when small process shifts are of interest: the cumulative sum (cusum) control chart, and the exponentially weighted moving average (ewma) control chart. Cusum and EWMA control charts are excellent alternatives to the Shewhart control chart for phase process monitoring situations (Montgomery, 2009). Nicolay et al. (2012) suggested that statistical quality control (SQC), a process is used as a tool to continuously monitor and identify potential problems early and to assess the impact of quality improvement interventions. Patel and Divecha (2011) proposed a modified Ewma control chart using the Markov Chain approximation under the assumption of normal distribution in order to more accurately determine the variations and errors in the process. Yang et al. (2012) compared the Shewart and Cusum control chart types for selecting the appropriate control chart to change the mean and variance of a variable. Abbasi et al. (2012) indicated that Shewart, Ewma with Cusum control charts a model that increases the sensitivity of the control chart. For subgroups based data, several authors worked on the performance of the Shewhart, CUSUM and EWMA charts (Mukherjee and Sen, 2015). Several authors have proposed CUSUM type non-parametric charts, which showed more sensitivity to the shifts in process location or dispersion (Graham et al., 2014; Mukherjee and Marozzi, 2017; Mukherjee et al., 2013). Ugaz et al. (2017) showed that new adaptive EWMA control charts based on the assessment of a potential misadjustment, which is translated into a timevarying smoothing parameter. The resulting control charts can be seen as a smooth combination between Shewhart and EWMA control charts, which could be efficient for a wide range of shifts. Qui (2017) indicated that statistical process control (SPC) charts could be a


useful tool, although conventional SPC charts need to be modified properly in some cases. In the study, introduced some basic SPC charts and some of their modifications, and describe how these charts can be used for monitoring different types of processes.

III.

Methodology

Statistical Quality Control Charts Statistical process control (SPC) is a powerful collection of problem-solving tools useful in achieving process stability and improving capability through the reduction of variability. SPC is one of the greatest technological developments of the twentieth century because it is based on sound underlying principles, is easy to use, has significant impact, and can be applied to any process (Montgomery, 2009:180). SPC techniques industry to develop and monitor processes is used Various control charts have been developed to monitor the variables in the process and to detect uncontrolled conditions that degrade the quality of the products (Noorossana and Vaghefi, 2006:191). Commonly used control charts Shewart control chart, Moving Sum, Cumulative Sum Quality Control Charts (CUSUM), Exponentially Weighted Moving Average (EWMA) (Russo et., 2012:36). The most commonly used application in recent years Shewart, Cusum and Ewma control charts were taken. These control graphics are used in different fields in the literature with different methods. Woodall (1985) revealed that the economic design of control charts to minimise the expected total cost. It is important, however, to consider the statistical propeties of economically designed control charts. Noorossana et al. (2009) provided a maximum likelihood estimator in order to identify the time of a step change in high-yield processes. They studied the change point estimation for a geometric process as the number of items until the occurrence of the first non-conforming item can be modeled by a geometric distribution. The add-on procedure was used with the geometric chart and provided accurate and precise estimations for different magnitudes of shifts in the process non-conformity proportion.

Cumulative Sum Quality Control Charts (CUSUM) Application of the cumulative sum quality-control method (hereafter called “cusum”) has been limited in clinical chemistry, even though the method appears to have advantages in detecting systematic changes in the analytical


67

process. This lack of acceptance is partly due to the additional effort required to calculate and maintain the cusum control chart, but also due to the qualitative manner in which the cusum chart is generally interpreted (Westgard et., 1977:1881). A cumulative statistic is generally more efficient than singlesample data points in control chart applications. This is generally true for cusum control charts for a variety of statistics, including individual measurements, sample means, sample variances, ranges and statistics using attribute data (DeVor et., 1992:385). Cumulative sum control chart Shewart control charts are one of four graphics developed as an alternative to these graphics because of their small but continuous insensitivity to the sample in the averages of the samples. Cusum control chart showed that cumulative sum of standard deviations from target value of sample values, it takes into account all the information in the sample data by marking it on the graphic. This superiority is particularly significant for the n = 1 sample subgroups more effective. It should also be remembered that Shewart control charts will be more effective if the process average is a larger slip (1.5σ or 2σ) than a small shift of 0.5σ (Işığıçok, 2012:289). The cumulative sum (or cusum) control chart is a good alternative when small shifts are important. The cusum chart directly incorporates all the information in the sequence of sample values by plotting the cumulative sums of the deviations of the sample values from a target value. For example, suppose that samples of size n ≥ 1 are collected, and 𝑥̅𝐽 is the average of the jth sample. Then if µ 0 is the target for the process mean, the cumulative sum control chart is formed by plotting the quantity 𝐶𝑖 = ∑𝑖𝑗=1(𝑋�𝑗 − 𝜇0 )

(1)

against the sample number i. Ci is called the cumulative sum up to and including the ith sample. Because they combine information from several samples, cumulative sum charts are more effective than Shewhart charts for detecting small process shifts. Furthermore, they are particularly effective with samples of size n = 1. This makes the cumulative sum control chart a good candidate for use in the chemical and process industries where rational subgroups are frequently of size 1, and in discrete parts manufacturing with automatic measurement of each part and on-line process monitoring directly at the work center (Montgomery, 2009:402).


An alternative procedure to the use of a tabular cusum is the V-mask control scheme proposed by Barnard (1959). The V-mask is applied to successive values of the cusum statistic,

𝐶𝑖 = ∑𝑖𝑗=1 𝑦𝑖 + 𝐶𝑖−1

(2)

where yi is the standardized observation 𝑦𝑖 = (𝑥𝑖 − 𝜇0 )/𝜎. A typical V-mask is shown in Figure 1.

The decision procedure consists of placing the V-mask on the cumulative sum control chart with the point O on the last value of Ci and the line OP parallel to the horizontal axis. If all the previous cumulative sums, C1, C2, . . . , Ci lie within the two arms of the V-mask, the process is in control. However, if any of the cumulative sums lie outside the arms of the mask, the process is considered to be out of control. In actual use, the V-mask would be applied to each new point on the cusum chart as soon as it was plotted, and the arms are assumed to extend backward to the origin. The performance of the V-mask is determined by the lead distance d and the angle θ shown in Fig. 1. The tabular cusum and the V-mask scheme are equivalent if

𝑘 = 𝐴𝑡𝑎𝑛𝜃

(3)

and

ℎ = 𝐴 𝑑 𝑡𝑎𝑛(𝜃))𝑑𝑘

(4)

Figure 1. A typical V-mask

In these two equations, A is the horizontal distance on the V-mask plot between successive points in terms of unit distance on the vertical scale. Refer to Fig. 1. For example, to construct a V-mask equivalent to the tabular cusum scheme used in equations 1, where and h = 5, we would select A = 1 (say), and then equations (3) and (4) would be solved as follows. Johnson (1961) has suggested


69

a method for designing the V-mask; that is, selecting d and q. He recommends the V-mask parameters ∆

𝜃 = 𝑡𝑎𝑛−1 �2𝐴� and 2 1−𝛽 𝑑 = �𝛿2 � 𝑙𝑛 � 𝛼 �

(5) (6)

where 2a is the greatest allowable probability of a signal when the process mean is on target (a false alarm) and b is the probability of not detecting a shift of size d. If b is small, which is usually the case, then

𝑑 = −2

ln (𝛼) 𝛿

(7)

We strongly advise against using the V-mask procedure. Some of the disadvantages and problems associated with this scheme are as follows: 1. The headstart feature, which is very useful in practice, cannot be implemented with the V-mask. 2. It is sometimes difficult to determine how far backward the arms of the Vmask should extend, thereby making interpretation difficult for the practitioner. 3. Perhaps the biggest problem with the V-mask is the ambiguity associated with α and β in the Johnson design procedure (Montgomery, 2009:416). Exponentially Weighted Moving Average (EWMA) The exponentially weighted moving average (EWMA) control chart is also a good alternative to the Shewhart control chart when we are interested in detecting small shifts. Pham (2006) stated that EWMA is a statistic for monitoring the process that averages the data in away that gives less and less weight to data as they are further removed in time. Fort he EWMA control technique, the decision depends on the EWMA statistics, which is an exponentially weighed average of all prior data, inclufdin the most recent measurement. By the choice of weighting factor, λ, the EWMA control procedure can be made sensitive to a small or gradual drift in the process, whereas the Shewhart control procedure can only react when the last data point is outside a control limit.


The statistic that is calculated is, 𝑧𝑖 = λ𝑥𝑖 + (1 − λ)𝑧𝑖−1

(8)

where 0 < l ≤ 1 is a constant and the starting value (required with the first sample at i = 1) is the process target, so that z 0 =µ 0. Sometimes the average of preliminary data is used as the starting value of the EWMA, so that 𝑧0 = 𝑥̅ (Montgomery, 2009:419).

IV.

Materials and Methods

Purpose of the Application The study utilized sugar production data for 2016-2017 in the Erzurum Sugar Factory provinces in Turkey. The corresponding data were obtained from factory under the responsibility of the sugar production data. It is aimed to use the statistical quality control tools included in the process to reduce the variability of the production process and to propose the control chart selection that will provide the maximum benefit. Method of the Application During application of the study, in a firm manufacturing sugar, cusum and ewma contorl charts were used and application of SQC was realized. In the application, data set including 60 unit belonging to variant affecting sugar quality produced in that firm used. With this data different control chart methods are compared and suggestions were made to increase the quality level in terms of production. It is used to draw Minitab 15.0 package program calculations and check charts in the application section. Computation for Cusum Cusum control chart parametres; 𝑥̅𝐽 is the average of the jth sample if the sample is taken as the average 𝑘 = 𝑥̿ . Ci is called thecumulative sum up to and including the ith sample. The average values of the calculated samples for cusum control chart, the deviations from the sample mean and the cumulative values of the deviations from the sample mean are shown in Table 4.1.


71

Table 4.1. Values Calculated for Cusum Control Chart Sample no

1

2

3

4

5

6

7

8

9

10

xi

15.8

16.2

16.4

17.8

17.0

17.7

� 𝒙𝒊 − 𝒙

-1.39

-0.99

-0.79

0.61

-0.19

0.51

17.8 0.61

17.7 0.51

19.0 1.81

18.0 0.81

-1.39

-2.38

-3.17

-2.56

-2.75

-2.24

-1.61

-1.12

0.69

1.50

Ci

Sample no

11

12

13

14

15

16

17

18

19

20

xi

23.0 5,81

18.0 0,81

19.0 1.81

18.0 0.81

16.0 -1.19

15.8 -1.39

15.14 -2.05

17.09 -0.10

17.0 -0.19

17.66 0.47

7.31

8.12

9.93

10.74

9.55

8.16

6.11

6.01

5.82

5.35

� 𝒙𝒊 − 𝒙 Ci

Sample no

21

22

23

24

25

26

27

28

29

30

xi

16.0 -1.19

18.0 0.81

17.0 -0.19

18.0 0.81

17.4 0.21

14.42 -2.77

21.01 3.82

16.02 -1.17

18.64 1.45

16.0 -1.19


4.14

4.97

4.78

5.59

5.80

3.03

6.85

5.68

7.13

5.94

Sample no

31

32

33

34

35

36

37

38

39

40

xi

18.0 0.81

11.8 -5.39

16.6 -0.59

16.8 -0.39

17.04 -0.51

16.46 -0.73

19.09 1.90

17.21 0.02

15.0 -2.19

18.6 1.41


6.75

1.36

0.77

0.38

-0.13

-0.86

1.04

1.06

-1.13

0.28

Sample no

41

42

43

44

45

46

47

48

49

50

xi

17.43 0.24

17.0 -0.19

16.86 -0.33

16.34 -0.85

15.42 -1.77

16.40 -0.79

16.85 -0.34

17.2 0.01

16.0 -1.19

17.7 0.51


0.52

0.33

0

-0.85

-2.62

-3.41

-3.75

-3.74

-5.13

-4.62

Sample no

51

52

53

54

55

56

57

58

59

60

xi

16.0 -1.19

17.2 0.01

18.26 1.07

17.48 0.29

16.4 -0.79

17.0 -0.19

16.0 -1.19

17.8 0.61

17.3 0.11

17.56 0.37

-5.81

-5.50

-4.43

4.14

-4.93

-5.12

-6.31

-5.70

-5.59

-5.22


If the mean 𝑥̅ is represented by the standart deviation 𝜎𝑥̅ , it is assumed that α = 0,01 and β = 0, as in similar studies in the literature. In this case, with the help of the factor values used for the control variables d 2 =2.534 and V-mask equivalent to the tabular cusum scheme used in equations (1), where k=1, and then equations (3) and (4) would be solved as follows; 𝑅� 1.6718 𝜎= = = 0.659 𝑑2 2.534 ∆= 𝑘𝜎 = 1(0.659) = 0.659 𝜎𝑥̅ =

𝜎

√𝑛 − 1

=

0.659

√60 − 1

= 0.0858


𝛿2 = �

∆ 2 0.659 2 � =� � = 59 𝜎𝑥̅ 0.0858

𝑑 = −2

𝑙𝑛𝛼 𝑙𝑛0.01 = −2 = 0.156 2 𝛿 59

it is calculated as. The angle θ must be present for drawing the separated large a V mask θ angle with the horizontal axis arm. In order to find this angle, the value A must be found in addition to the values already found. 𝐴 = 2𝜎𝑥̅ = 2(0.0858) = 0,1716

found as and angle of θ, 𝜃 = 𝑡𝑎𝑛−1 �

0.659 ∆ � = 𝑡𝑎𝑛−1 � � = 620 2(0.1716) 2𝐴

Figure 2. Cusum Control Chart Cusum control chart of the observations on the variables are presented in Figure 2. In order to more reliably determine whether the process is under control V mask is used. The Cusum control chart of 60 samples evaluated within the scope of application is calculated by the common usage method in the literature. According to the result of calculation and graphic chart Cusum control it was


73

observed under control of the production process. It has been observed that the values of the V mask resultant applied to the void control chart are not outside the upper and lower control limits. Computation for Ewma The volume of the sample subgroups n, and 𝑋�𝑗 (𝑗 = 1,2, … , 𝑡, … , 𝑘) the averages

of exponential weighted moving average values of period t, λ=0.5 are taken as the weighting coefficients of the samples for the same or greater detection of large and small slips in the process. Where λ is a correction coefficient with 00} �−𝑙𝑛(𝑌𝑖 !) − (𝑌𝑖 + 𝛼 −1 )𝑙𝑛�(1 + 𝛼𝜇𝑖 )� + 𝑌𝑖 𝑙𝑛(𝛼) + 𝑌𝑖 𝑙𝑛(𝜇𝑖 )� ,

where 𝜆𝑖 = 𝑒𝑥𝑝(𝒒′𝑖 𝜽) such that 𝒒′𝑖 is the ith observation vector in matrix 𝑸. It is known that the asymptotic distribution of the maximum likelihood estimator � = �𝜷 �, 𝜽 � � is multivariate normal distribution which is given by (MLE) 𝝑

� − 𝝑)~𝑁 �𝟎′𝑝 +𝑝 , � (𝝑 1 2

𝜕2 𝐿

− 𝜕𝜷𝜕𝜷′ 𝜕2 𝐿

− 𝜕𝜷𝜕𝜽′

𝜕2 𝐿

− 𝜕𝜷𝜕𝜽′ 𝜕2 𝐿

− 𝜕𝜽𝜕𝜽′

−1

� �,

where 𝟎′𝑝1 +𝑝2 the vector of zeros with dimension 𝑝1 + 𝑝2 . Also, see Cameron 𝜕2 𝐿

𝜕2 𝐿

𝜕2 𝐿

and Trivedi (2013) for descriptions of 𝜕𝜷𝜕𝜷′, 𝜕𝜷𝜕𝜽′ and 𝜕𝜽𝜕𝜽′.

As a modification to the MLE, we propose to use Ridge regression method and � RR , 𝜽 � RR = �𝜷 � RR � as follows obtain 𝝑 � RR = (𝑿′𝑾 � �𝑿 + 𝑘𝑐 𝑰)−1 �𝑿′𝑾 �𝑿�𝜷 𝜷 and

� RR = �𝑿′𝑾 �, �𝑿 + 𝑘𝑧 𝑰�−1 �𝑿′𝑾 �𝑿�𝜽 𝜽

� is a diagonal matrix whose ith where 𝑘𝑧 and 𝑘𝑐 are positive constants and 𝑾 diagonal element is 𝜇̂ 𝑖 .

Estimating the ridge parameters

The estimation of Ridge parameters has been a prominent issue and there is various amount of estimation methods of Ridge parameter in the literature. In this study, we use five of the early proposed techniques in ZINB ridge estimator. The first estimator was proposed by Schaefer et al. (1984) for the ridge logistic estimator as

Econometrics: Methods & Applications 1

,

�2 𝜎

.

𝑘𝑐1 = 𝜏�2

𝑚𝑎𝑥

97

� such that 𝚮 is the matrix whose columns are the eigenvectors of where 𝝉� = 𝚮𝜷 �𝑿. Another estimator which was firstly defined by Hoerl and Kennard 𝑿′𝑾 (1970) in the linear model, is defined as follows 𝑘𝑐2 = 𝜏�2

𝑚𝑎𝑥

Now, following Hoerl and Kennard (1970), we consider 𝑘𝑐𝑗 =

�2 𝜎 𝜏�𝑗2

and propose

to use the arithmetic mean, geometric mean and harmonic mean of 𝑘𝑐𝑗 to propose the following estimators respectively, 1

𝑝

𝑝

𝑝

�2 𝜎

𝑘𝑐3 = 𝑝 ∑𝑗=1 𝜏�2 , 𝑗

𝑘𝑐4 = �� 𝑘𝑐5 =

𝑗=1

𝑝

𝜎� 2 , 𝜏̂𝑗2

2. 𝑝 𝜏̂𝑗 ∑𝑗=1 2 𝜎�

� RR . In an analogous manner, we may One of these estimators may be used in 𝜷 � RR which we skip the propose to use logistic analogues of these estimators in 𝜽 exact definitions.

In the following section, we perform a Monte Carlo simulation study to compare the performances of MLE and RR.

III.

A Monte Carlo Simulation Experiment

In this section, we present the details of the Monte Carlo experiment. The crucial factors of this experiment are the degree of correlation 𝜙 2 among the count-component variables, the degree of correlation 𝜌2 among the zerocomponent variables and the sample size 𝑛. Following Kibria et al. (2013), we generate the data matrices 𝑿 and 𝑸 by using the equations 𝑥𝑖𝑗 = (1 − 𝜌2 )1/2 𝑧𝑖𝑗 + 𝜌𝑧𝑖𝑝 and

98

Ridge Type Esitimation in the Zero-Inflated Negative Binomial Regression

𝑞𝑖𝑗 = (1 − 𝜙 2 )1/2 𝑤𝑖𝑗 + 𝜙𝑤𝑖𝑝 ,

where 𝑧𝑖𝑗 and 𝑤𝑖𝑗 are pseudo-random numbers following the standard normal distribution such that 𝑖 = 1,2, … , 𝑛 and 𝑗 = 1,2, … , 𝑝 . We consider that 𝜌 = 0.7, 0.9, 0.99 and 𝜙 = 0.5, 0.9. The sample size changes as 𝑛 = 100, 200 and 500. Using the equations (1) and (2) we generate the dependent variable which is distributed as zero-inflated negative binomial. At first, we generate a binary variable from binomial distribution using 𝒑 = 𝑒𝑥𝑝(𝑸𝜽)/(1 + 𝑒𝑥𝑝(𝑸𝜽)), and then, we generate a random variable form negative binomial distribution and update the vector 𝒑 by changing values coming from the negative binomial 𝑝 distribution. Moreover, the slope parameters are taken as ∑𝑗=1 𝛽𝑗2 = 1 and ∑𝑝𝑗=1 𝜃𝑗2 = 1 . We calculate the simulated MSE values using the following equation

1000

1 � − 𝝎)′(𝝎 � − 𝝎), �) = � (𝝎 𝑀𝑆𝐸(𝝎 1000 𝑟=1

� represents each estimator considered in the experiment for both zerowhere 𝝎 component and count component and 𝝎 is the real value of the parameter vectors, therefore we compute the simulated MSEs for zero and count components separately. Table 1. The simulated MSE values when ∅ = 𝟎. 𝟓 𝝆 n

100

200

500

0.7

MLE RR(k1) RR(k2) RR(k3) RR(k4) RR(k5) MLE RR(k1) RR(k2) RR(k3) RR(k4) RR(k5) MLE RR(k1) RR(k2) RR(k3) RR(k4) RR(k5)

Count 0.82087 0.72267 0.67042 0.89019 0.80674 0.71401 0.55802 0.52912 0.61629 0.94382 0.87576 0.75506 0.18111 0.24854 0.48288 0.98970 0.95659 0.75823

Zero 97.33731 4.34962 4.43661 1.24119 1.58435 2.76234 4.69231 1.89953 2.36950 1.11495 1.24074 1.77319 3.83275 1.52628 1.44139 1.00230 1.01043 1.12878

Count 2.75961 1.52876 1.72105 1.02632 1.10589 1.35605 1.32622 0.93396 0.74358 0.97479 0.92358 0.76597 0.49701 0.50161 0.66618 0.99302 0.97017 0.84452

0.9 Zero 6.80732 2.77144 3.97794 1.28078 1.77011 2.89835 4.08965 1.53582 1.93694 1.05463 1.11691 1.40940 3.61701 1.75126 1.60497 1.00382 1.01950 1.21203

0.99 Count 27.40508 11.45590 13.61700 1.11549 4.35665 9.04677 16.49285 7.59407 5.84119 1.00498 1.74008 3.19792 6.70923 2.78758 1.27119 1.00004 0.99652 1.00075

Zero 6.79280 2.46553 3.54018 1.11576 1.37802 2.31363 4.90791 1.74743 2.65675 1.15597 1.27943 1.91059 3.78866 1.92438 1.57899 1.00444 1.01715 1.17840


99

Table 2. The simulated MSE values when ∅ = 𝟎. 𝟗 𝝆 n

100

200

500

0.7

MLE RR(k1) RR(k2) RR(k3) RR(k4) RR(k5) MLE RR(k1) RR(k2) RR(k3) RR(k4) RR(k5) MLE RR(k1) RR(k2) RR(k3) RR(k4) RR(k5)

Count 0.82948 0.65095 0.68765 0.89752 0.83484 0.73455 0.52703 0.52416 0.59668 0.93704 0.86622 0.72842 0.17201 0.24081 0.48390 0.99109 0.96270 0.76751

0.9

Zero 11.76612 2.90407 6.12218 1.25711 2.00650 4.19695 6.07982 2.08081 2.62470 1.02873 1.13471 1.86275 4.46795 1.50414 1.19639 1.00072 1.00439 1.04990

Count 3.35617 1.77467 2.09739 1.02635 1.20263 1.63299 1.28192 0.95134 0.82768 0.94764 0.89796 0.84201 0.51727 0.47891 0.65880 0.99307 0.96493 0.83253

Zero 10.30652 4.68183 4.95923 1.22157 1.76397 3.44899 6.67635 1.80962 2.73169 1.01305 1.11907 1.86743 4.11103 1.49533 1.38334 1.00552 1.03220 1.13628

0.99 Count 27.32481 8.36064 13.23081 1.83878 4.89687 8.88424 15.25964 6.55909 4.14492 1.00387 1.21522 2.19617 7.13242 3.63720 1.51394 0.99823 0.98991 1.06924

Zero 13.56217 2.74190 6.11302 2.09604 2.62683 4.10021 6.91082 2.35694 2.82006 1.00965 1.10017 1.81516 4.11211 1.87247 1.29296 1.00228 1.01205 1.08404

We present the simulated MSE values of the estimators in Tables 1-2. According to the tables, we observe that the increase in the sample size makes a positive effect in the MSE values generally. It is seen that the MSE of count variables are increased when 𝜌 increases, however, the MSE of zero variables have deficiencies in the same situation. Moreover, if we increase the degree of correlation among the zero-variables ∅ , we see that there is a tendency of increase in the MSE values, especially for MLE.

According to the results, MLE has the worst performance in most of the cases in the simulation. Although RR(k3) has the least MSE value in most of the situations, there is no best estimator for each case. However, the listed ridge estimators have better performances than MLE in general. Also, we can say that when the sample size is high (𝑛 = 500) and the correlation among the variables is low (𝜌 = 0.7), MLE has the best performance.

IV.

A Real Data Application

In this section, we consider the Expenditure and Default data sets which is analyzed by Greene (2003) and which is freely available in the package of AER in R software. In this data, the dependent variable of interest is the number of

100


major derogatory reports recorded in the credit history of a sample of applicants for a type of credit card.

Figure 1. Histogram of the response variable

As it is indicated in Greene’s study, this dependent variable will be equal to zero or will take positive values when the costumers have missed several revolving credit payments at any time. Furthermore, we notice that the histogram plot of the response variable which is given in Figure 1 shows that the number of zeros is a lot more than counts. Hence, it is considered that the response variable is inflated by zeros. Table 3. Descriptions of variables for the Expenditure and Default Data set Variables Response Variable Reports Predictors Card Age Income Share expenditure owner factor selfemp factor dependents Months majorcards Active

Descriptions Number of major derogatory reports. Factor. Was the application for a credit card accepted? Age in years plus twelfths of a year. Yearly income (in USD 10,000) Ratio of monthly credit card expenditure to yearly income Average monthly credit card expenditure Does the individual own their home? Is the individual self-employed? Number of dependents Months living at current address Number of major credit cards held Number of active credit accounts


101

In Table 3, we also describe the variables of the Expenditure and Default data. In Table 4, we show the formula of the fitted model.

Table 4. The formula of count and zero components reports ~ 𝛽𝑐0 + 𝛽𝑐1 income+𝛽𝑐2 majorcards+𝛽𝑐3 age+𝛽𝑐4 share+𝛽𝑐5 expenditure | 𝛽𝑧0 + 𝛽𝑧1 age+𝛽𝑧2 owner+𝛽𝑧3 selfemp+𝛽𝑧4 dependents

In the following formula,

reports ~ 𝛽𝑐0 + 𝛽𝑐1 income + 𝛽𝑐2 majorcards + 𝛽𝑐3 age + 𝛽𝑐4 share + 𝛽𝑐5 expenditure

the count model that is a negative binomial regression is given, and the logit model is given as reports ~ 𝛽𝑧0 + 𝛽𝑧1 age + 𝛽𝑧2 owner + 𝛽𝑧3 selfemp + 𝛽𝑧4 dependents

In Table 5, we report the coefficient estimates of MLE and Ridge via five different ridge parameters. The standard errors in parentheses are based on bootstrap samples re-sampled 1000 times. One might see that the suggested method has smaller standard errors. We also report the prediction error (PE) values to ease the comparison. Pes can be computed as

� − 𝒀�′�𝒀 � − 𝒀�, PE = ∑𝑛𝑖=1�𝒀 where

�

� � �1 − 𝑒𝑥𝑝�𝑸𝜽� �. � = 𝑒𝑥𝑝�𝑿𝜷 𝒀 �� 1+𝑒𝑥𝑝�𝑸𝜽

Overall, the suggested estimations have better prediction accuracy than MLE.

102


Table 5. The coefficients and standard errors of each method in application MLE

RR(k1)

RR(k2)

RR(k3)

RR(k4)

RR(k5)

0.42327

0.39140

0.39580

-0.00007

0.00312

0.30162

(0.00968)

(0.00930)

(0.00950)

(0.00001)

(0.00221)

(0.00895)

-0.01700

-0.01041

-0.01132

0.00056

0.03568

0.00815

(0.00084)

(0.00083)

(0.00083)

(0.00002)

(0.00081)

(0.00081)

0.05107

0.04995

0.05011

-0.00001

0.00499

0.04682

(0.00269)

(0.00268)

(0.00268)

(0.00001)

(0.00101)

(0.00267)

-0.01514

-0.01499

-0.01501

-0.00532

-0.00939

-0.01455

(0.00026)

(0.00026)

(0.00026)

(0.00008)

(0.00015)

(0.00026)

-8.85722

-8.12689

-8.22784

-0.00004

-0.00305

-6.0708

(0.12184)

(0.1148)

(0.11572)

(0.00000)

(0.00056)

(0.09519)

0.00052

0.00032

0.00035

-0.00183

-0.00187

-0.00026

(0.00002)

(0.00002)

(0.00002)

(0.00002)

(0.00002)

(0.00002)

2.87562

2.58602

2.63359

0.00220

0.07150

1.96779

(0.26558)

(0.07198)

(0.07095)

(0.00005)

(0.00493)

(0.05215)

-0.12839

-0.11744

-0.11924

-0.0129

-0.01779

-0.09398

(0.01366)

(0.00317)

(0.00271)

(0.00014)

(0.00033)

(0.00206)

0.82843

0.79096

0.79724

0.00355

0.08632

0.70310

(0.24545)

(0.05363)

(0.03673)

(0.00009)

(0.00373)

(0.01561)

0.10007

0.08604

0.08832

-0.00006

-0.00053

0.05751

(0.07505)

(0.01409)

(0.01729)

(0.00003)

(0.00296)

(0.01385)

-0.03205

-0.03181

-0.03187

0.00094

0.01367

-0.02997

(0.17175)

(0.03246)

(0.01950)

(0.00012)

(0.00206)

(0.00552)

1.87091

1.83310

1.83473

1.86010

Count (Intercept)

income

majorcards

age

share

expenditure

Zero (Intercept)

age

owneryes

selfempyes

dependents

PE 1.87589 1.86685 The values in parenthesis are the standard errors.

In Figure 2, we plot the densities of both methods. We just consider the Ridge estimates with 𝑘𝑐4 parameter since the rest gives similar pattern. According to Figure 2, both methods gives a satisfactory performance in the sense of prediction response.


103

Figure 2. The densities of the methods

V.

Conclusion

This paper is concern with the Ridge-type estimation in ZINB regression models. In this purpose, we suggest the Ridge regression estimation for both count and zero parts, simultaneously. We also demonstrated a Monte Carlo simulation study to investigate the performance of suggested method in the presence of multicollinearity. Furthermore, we analyzed a real-life application. In summary, the proposed method has better performance than MLE as expected.

References Asar, Y., Ahmed, S.E. & Yüzbaşı, B. (To appear in 2018). Efficient and Improved Estimation Strategy in Zero- Inflated Poisson Regression Models. Proceedings of the Twelfth International Conference on Management Science and Engineering Management. Lecture Notes on Multidisciplinary Industrial Engineering. Springer Cameron, A. C., & Trivedi, P. K. (2013). Regression analysis of count data (Vol. 53). Cambridge university press. Hoerl, A. E., Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55-67.

104


Kibria, B. G., Månsson, K., & Shukur, G. (2013). Some ridge regression estimators for the zero-inflated Poisson model. Journal of Applied Statistics, 40(4), 721-735. Greene, W.H. (1994). Accounting for Excess of Zeros and Sample Selection in Poisson and Negative Binomial Regression Models, New York University. Greene, W.H. (2003). Econometric Analysis, 5th edition. Upper Saddle River, NJ: Prentice Hall. Lambert, D. (1992). Zero-inflated Poisson regression, with an application to defects in manufacturing. Technometrics, 34(1), 1-14. Schaefer, R. L., Roi, L. D., & Wolfe, R. A. (1984). A ridge logistic estimator. Communications in Statistics-Theory and Methods, 13(1), 99-113. Sheu, M. L., Hu, T. W., Keeler, T. E., Ong, M., & Sung, H. Y. (2004). The effect of a major cigarette price change on smoking behavior in California: a zero‐inflated negative binomial model. Health Economics, 13(8), 781-791. So, S., Lee, D. H., & Jung, B. C. (2011). An alternative bivariate zero-inflated negative binomial regression model using a copula. Economics Letters, 113(2), 183-185. Staub, K. E., & Winkelmann, R. (2013). Consistent Estimation of Zero‐Inflated Count Models. Health Economics, 22(6), 673-686. Wang, P. (2003). A bivariate zero-inflated negative binomial regression model for count data with excess zeros. Economics Letters, 78(3), 373-378. Wang, P., & Alba, J. D. (2006). A zero-inflated negative binomial regression model with hidden Markov chain. Economics Letters, 92(2), 209-213. Zhu, F. (2012). Zero-inflated Poisson and negative binomial integer-valued GARCH models. Journal of Statistical Planning and Inference, 142(4), 826839.

Chapter 8

Cliometric Perspective for Stock Market Reactions to Wars and Political Risks: Evidence from a Falling Empire Avni Önder HANEDAR 1 Elmas Yaldız HANEDAR 2

Abstract. In this paper, based on cliometric methodology we use new historical data on the most popular stocks traded at the İstanbul bourse between 1910 and 1914, to examine the effect of wars on stock market prices. During this period, the Ottoman Empire was involved in the Turco-Italian and the Balkan wars, leading to massive land losses and risks for the companies before the First World War. The data are manually collected from the available volumes of a daily Ottoman newspaper, Tanin. Our findings are surprising, as we observe only a temporary and small drop of prices, indicating little perceived risk by stock investors of the İstanbul bourse. Keywords: Cliometrics; The İstanbul Stock Exchange; Stocks; The Turco-Italian War; The Balkan Wars; Structural Breaks

1

Asst. Prof. Dr., Sakarya University, Faculty of Political Sciences, Department of Econometrics, Esentepe Kampüsü 54187 Sakarya–Turkey , [email protected] 2 Asst. Prof. Dr., Sakarya University, Faculty of Business, Department of Business Sakarya– Turkey, [email protected]

106

Cliometric Perspective for Stock Market Reactions to Wars and Political Risks: Evidence from a Falling Empire

I.

Introduction

History provides qualified and detailed information on economic, social, political lives of the past. In recent years, the history has been widely interested in the tools of economics in order to understand today’s realities. For instance, studying the comparative history of developed and underdeveloped economies can provide policy tools to poor countries in order to catch up with the developed ones. Moreover, to identify a causal effect, the history guides economists by suggesting exogenous variables. Another methodological development in economics emerged about the use of mathematics and statistics in economic modelling which allows reproducible findings and practical solutions for policy implications. Similarly, economic historians adopted such methods especially by the 1970s, to kill the myths of traditional historians (Abramitzky, 2015). The path breaking research that combines history and econometrics is Fogel (1964)’s study on the impacts of railroads on economic growth in the USA, providing quantitative findings based on a counterfactual method making possible to identify an event’s impact in case of its absence. This kind of applications is being increasingly popular and not previously possible in traditional historical methodology which follows narrative of the unique historical events. The transformation of economic history from a narrative world to quantitative format is known as Cliometrics. The Cliometrics differs from the traditional practise of history by bringing economic theory into historical analysis. On the other hand, the application of the formal models to the historical data was heavily criticised due to the absence of research on underlying structures (Haupert, 2006, pp. 3–33). One of the most popular topics in cliometric research is about the impact of political events on financial outcomes. Ferguson (2006) and Mauro et al. (2006) explain how the political crises leading to the First World War were reflected at the London Stock Exchange from a quantitative perspective. For the effects of political events on financial markets in the Ottoman Empire, there were many myths and unsettled discussions on which historians debated. The historians rarely collect data on economic outcomes of Ottoman Empire and therefore do not use econometrics to provide answers to historical discussions with some exceptions (See Özcumur & Pamuk, 2002). On the other hand, by combining economic theory and quantitative methods, historical debates can be solved and


107

moreover ignored topics can be examined. For instance, Hanedar et al. (2015) verify that the private investors of the Ottoman bonds were more capable to anticipate efficiently gathered information on the threats of wars against Italy and the Balkan countries by 1914, as compared to the Ottoman statesmen. On the other hand, Hanedar et al. (2016) finds that some events between 1918 and 1925 were not perceived as important by the Ottoman bond investors in contrast to the historians’ arguments. The aim of this chapter is to understand the stock market reactions to military conflicts. This topic is especially important nowadays as Middle Eastern economies have been witnessing wars since the last decade. History provides good insights at this point for today’s investors by creating a large literature. However the literature on Middle Eastern—and even for the transition and emerging—economies is limited (See Önder & Şimga-Muğan (2006) for a discussion). Our chapter is the first study to provide a historical narrative to explain the changes of Ottoman stock returns due to the wars on the eve of the First World War (WWI). A historical examination of the İstanbul bourse during the Turco-Italian and Balkan wars is important to shed light on the different effects of conflicts. We use unique data on stock prices of 9 popular domestic joint-stock companies traded at İstanbul Bourse from 1910 to 1914. This period provides an interesting case to study the risk perceptions of investors, as the conflicts caught the Ottoman Empire unprepared (Hall, 2000, p. 14; Erickson, 2001, p. 3; Childs, 2008, p. 72; Giolitti, 2012, p. 59)—which could have led to higher uncertainty on the stock exchange market—. The remainder of the chapter is organized as follows; Section 2 discusses the related literature. Section 3 provides information on the Turco-Italian and the Balkan wars. Section 4 explains our data, while Section 5 covers the methodology applied. Section 6 presents the empirical results. Section 7 discusses these results, and finally section 8 concludes.

II.

Literature Review

Many financial studies address the negative reactions to political risks on different stock markets. Zussman et al. (2008) focus on asset prices during Israeli-Palestinian conflicts since the late 1980s and find that asset prices in both Israeli and Palestinian markets increase when a peace initiative takes place and decrease in case of conflicts. Similarly, Franck & Krausz (2009) show that

108


the conflicts between 1945 and 1960 were strongly reflected on the Israeli Stock Exchange, as the end of conflicts was approaching. Choudhry (2010), Charles and Darné (2014), Hudson & Urquhart (2015), Mathy (2016), and Urquhart & Hudson (2016) point out lower prices of US and British Stocks due to news during the Second World War. Rigobon & Sack (2005), Schneider & Troeger (2006), Kollias et al. (2010), and Dimic et al. (2015) indicate the presence of higher risk for stock exchanges in various countries during many conflicts after 1990. Several papers show that the impact of political risks on stock markets is related to various factors. Recently, Hobbs et al. (2016) suggest that the negative impacts of wars vary by industry of the stocks traded on the US Stock Exchange from 1963 to 2012. Amihud & Wohl (2004) indicate that the negative relationship between the Iraq War and stock prices on the US Stock Exchange was related to the lengths of these conflicts. Le Bris (2012) implies that the correlation between wars from 1870 to 1945 and the volatility in the stock prices at the French Stock Exchange depends on the financing methods of wars, which could lead to higher public expenditures, inflation, and disruptions of the public services. There are only a couple of recent studies about the impacts of political risks on the İstanbul Stock Exchange. Önder & Şimga-Mugan (2006) consider the effects of political and economic news on the volatility of stock returns and the trading volume in Turkey and Argentina from 1995 to 1997, and address the presence of a statistically significant relationship between stock market indicators and political risks. Based on data from Turkey between 1995 and 2008, İkizlerli & Ülkü (2012) point out that the negative effect of political risks on stock market outcomes depends on industry and the origin of investors. Another strand of the literature—concentrated on US Stock Exchanges— indicate positive or insignificant impacts of wars on stock markets (Cutler et al. (1989), Guidolin & La Ferrara (2010), Kollias et al. (2013), Brune et al. (2014), Charles & Darne (2014)).

III.

Historical background for the Turco-Italian and the Balkan wars

During the period from 1910 to 1914, the Turco-Italian and the Balkan wars, preceding WWI and its catastrophic results hit the Ottoman economy. The


109

Turco-Italian war began on 29 September 1911, as the First and Second Balkan wars broke out on 17 October 1912 and 29 June 1913, respectively. Using data for Ottoman trade with several of its trading partners between 1830 and 1913, Hanedar (2016) shows that the Balkan wars had negative impacts on Ottoman exports, whereas the Turco-Italian war had no such disruptions on Ottoman foreign trade. As shown in Figure 1, the Ottoman foreign trade experienced a temporary reduction between 1911 and 1913. Al & Akar (2014, pp. 296–297) argue positive reactions at the İstanbul Bourse while the TurcoItalian and Balkan wars were reaching to an end. Before the outbreak of the Turco-Italian war, a commentary in İkdam, a widely read Ottoman newspaper, reported Italy’s desire to occupy Libya, which resulted in a negative response on the İstanbul bourse, while several companies announced bankruptcy (İkdam, 25 September 1911, p. 1). Another commentary of İkdam pointed out price fluctuations at the İstanbul bourse because of the news about the Italian occupation in Libya (İkdam, 29 September 1911, p. 1). During the First Balkan war, Tercüman-ı Hakikat reported price increases in the İstanbul Bourse as discussions on the end of conflicts with Balkan states (Tercüman-ı Hakikat, 1 February 1913, p. 1). Hanedar et al. (2015) evince that the outbreak of the Turco-Italian and Balkan wars were correlated with a lower likelihood of Ottoman debt repayments.

Figure 1. The Ottoman Foreign Trade, 1830–1913 (British Pounds) Note: Total Ottoman foreign trade in real terms (base year 1913) expressed in thousand British pounds. Source: Hanedar (2016).

Geyikdağı (2011, p. 54) and Geyikdagi & Geyikdagi (2011, pp. 388–389) state that stores and commercial ships were damaged, as business activities, railway traffic and operations were disrupted during the hostilities of 1911–13. For instance, the Anatolian Railway, İstanbul Ferry, and the Dersaadet Tramway

110


companies’ activities were interrupted by the Turco-Italian and Balkan wars, since the ships were bombed by enemies and the Ottoman state bought the horses of trams. Moreover the skilled workers from enemy countries were fired due to political hostilities (Tutel, 1997, pp. 161–162; Kayserilioğlu, 2003, p. 197; McMeekin, 2010, pp. 239–240). As shown in Figure 2, the Imperial Ottoman Bank’s profits and foreign bank entry were decreasing during the hostilities. The Ottoman state commandeered carriages for commercial transportation, as lighthouses did not work, while many trade routes and ports were damaged (Quataert, 1996, pp. 767–768; 2005, p. 126). Beehler (1913, pp. 69, 72–73) provides information on complaints from foreign traders because of restrictions on shipping imposed by the Ottoman state and its adversaries, which could have led to lower trade for the Ottoman Empire. On the other hand, Geyikdagi & Geyikdagi (2011) argue that the effects of wars on political risks for firms were negligible due to political and economic support of their home countries and concessions provided by the Ottoman state.

Figure 2. The Imperial Ottoman Bank’s Profits and Number of Foreign Bank Branches, 1895–1914 Notes: The average net profits are expressed in gold Liras (the graph on the left). The foreign bank branches are the average number of foreign bank branches in the Ottoman Empire (the graph on the right). Sources: Eldem (1999, pp. 510–512) and Hanedar (2015).


IV.

111

Data

To examine the impact of the wars on the eve of WWI, we use the closing prices of 10 stocks traded on the İstanbul bourse from 1910 to 1914. As the prices might not be stationary, we estimate the stock returns the following way: R t =ln(P t /P t-1 )

(1)

where P t is daily price of each stock at time t. The stocks in sample were issued by the Ottoman General Insurance company (Osmanlı Sigorta Şirket-i Umûmiyesi) (OGI), the Regie (Tobacco) company (Tütün Rejisi) (R), the Imperial Ottoman Bank (Bank-ı Osmanî-i Şâhâne) (IOB), the Balya-Karaaydın mining company (Balya-Karaaydın Maden Şirketi) (BK), the Kessendre mining company (Kessendre-Kassandra Madenleri Osmanlı Anonim Şirketi) (KM) 3 , the Ereğli mining company (Ereğli Şirketi) (EM), the Anatolian Railway company (Anadolu Demiryolu Şirketi) (AR), the İstanbul Ferry company (Şirket-i Hayriyye) (IF), and the Dersaadet Tramway company (Dersaadet Tramway Şirketi) (DT). All these companies played crucial roles for the Ottoman economy and operated in the most attractive sectors, i.e., banking, mining, agriculture, and transportation. IF and DT were the oldest joint-stock companies in the Ottoman Empire, established in 1856 and 1869 respectively. IF and DT were jointly owned by local investors and the Ottoman state, while the other companies were founded by foreigners and non-Muslim citizens of the Ottoman Empire (Kazgan et al., 1999, pp. 340–342; Akyıldız, 2011; Yılmaz, 2012). In 1899, the value of stocks issued by the IOB was the largest, i.e., 4,490,000 Liras, as compared to other companies. R issued 2,464,000 Liras of stocks. Other firms such as OGI and DT issued relatively lower amounts, i.e., 164,000 and 200,000 Liras, respectively (Fertekligil, 2000, pp. 44–45; Al & Akar, 2014, pp. 120–127). Our data source is Tanin, which is the only newspaper that provides information on stock prices, as shown in Figure 3. As a pro-government newspaper, Tanin was a widely circulated daily Ottoman newspaper in İstanbul. The National Library of Turkey and the Beyazıt State Library have digital copies of these newspapers. The prices of 6 stocks are denominated in the Turkish Liras, while

3

KM had two different stocks traded at the İstanbul bourse, KM1 and KM2.

112


those of KM, EM, and AR are reported in French Francs. 4 Although there are other stocks traded on the İstanbul bourse, such as the Ottoman Navigation company (Osmanlı İttihad Şirketi), the ten stocks that we use are the most important stocks in terms of book value. Moreover, data for other stocks were not continuously available to allow econometric analysis. 5

Figure 3. Data on Stock Prices of Several Companies in the Ottoman Empire, 1911 Note: Figure 3 shows the name of companies and stock prices in Tanin. Source: Tanin (1911). 14 November, p. 2.

Table 1 shows the descriptive statistics for stock returns from 1910 to 1914. Jarque-Bera and kurtosis statistics indicate that all series have asymmetric leptokurtic distributions. AR, IF, EM, DT, and R’s stock returns have right tailed distribution, as those of the other companies have left tailed distributions.

4

5

The Ottoman Empire and France adopted gold standard by the outbreak of WWI, leading to stability in the foreign exchange rates (Pamuk, 2000, pp. 216–221; Tuncer & Pamuk, 2014, pp. 182–183). There can be several reasons for these non-available observations. There could be no transaction for the stock whose price is not reported, or the newspaper could not have enough space to report the price of a negligible stock.


113

Table 1. Descriptive Statistics for Stock Returns, 1910–14

OGI

Mean

Maximum

Minimum

Std. Dev.

Skewness

Kurtosis

0.00066

0.52

-0.50

0.05

0.00

52.97

0.00013

R IOB BK KM1 KM2 EM AR

-0.00050 0.00002 -0.00100 -0.00090 -0.00053 -0.00025

0.23 0.08 0.31 0.23 0.21 0.46 0.33

-0.27 -0.15 -0.46 -0.22 -0.22 -0.46 -0.32

0.03 0.02 0.06 0.03 0.04 0.05 0.05

0.07 -1.71 -1.44 -0.09 -0.41 0.51 0.04

32.89 21.60 30.74 15.70 14.42 55.33 34.30

IF

-0.00022

0.42

-0.40

0.04

0.44

96.68

DT

0.00187

0.53

-0.53

0.06

0.06

42.50

JB 49002.13 17459.83 6986.76 13967.98 3093.82 2510.95 52859.13 17755.72 107885.50 22555.32

Obs. 471 469 469 431 460 460 463 435 295 347

Note: All Jarque-Bera (JB) statistics are statistically different from zero at %1 percent.

V.

Methodology

To examine the reactions of stock investors at the İstanbul bourse to wars, we use an Iterative Sums of Squares (ICSS) approach proposed by Inclan and Tiao (1994). In contrast to alternative methods developed by Bai and Perron (1998, 2003), ICSS is a method to identify sudden breaks in the volatility of financial assets’ outcomes and does not impose an assumption on the exogenous selection of the maximum number of break points. To estimate the number of breaks and the changes in volatility or variance, ICSS estimates the cumulative sum of squares as follows: (Ck / Ct ) − k / T , Dk =

1,..., T with D0 = 0 k= Dk =

(2)

where Ck = ∑ t =1 ε t2 is the cumulative sum of squares, starting from the k

beginning of the series to the k th point in time. As a sudden change exists in variance, the plot of Dk locates out of specified boundaries. If the absolute value of Dk is greater than the critical value, the null hypothesis for absence of the sudden change in variance would be rejected.

114


VI.

Results

Table 2 indicates break times in volatility of the stock returns and corresponding changes in the stock prices. Almost none of the break points are related to the dates of the wars on the days before WWI. We observe several break points in R and AR, which could be related to events about the outbreak of the Turco-Italian war. On 14 September 1911, there is a break in the volatility of AR’s stock return two weeks prior to the outbreak of the Turco-Italian war, resulting in lower price in the long-run. The price fall is the highest one, as compared to those of the other break points. We observe two break points in R on 26 and 30 September 1911 before the outbreak of the Turco-Italian war, leading to a higher price in the long-run. During September 1911, the tension between Italy and the Ottoman Empire increased due to the campaigns of the Italian government, press, and diplomats, as they were preparing to invade Libya (Tercüman-ı Hakikat, 12 September 1911, p. 1; 18 September 1911, p. 1; Tanin, 12 September 1911, p. 1; 19 September 1911, p. 1; Hüseyin Cahid, 22 September 1911, p. 1). This implies that investors expected higher risk for AR’s stocks and lower risk for those of R’s stocks, as the outbreak of the Turco-Italian war was approaching. We identify two break points for KM1 on 6 and 12 October 1911. These dates bring lower prices and correspond to two months after the outbreak of the Turco-Italian war. On 6 and 13 October 1911, different commentaries in Tanin reported great disappointment and reactions of the Ottomans against Italy because of the Ottoman defeat (Hüseyin Cahid, 6 October 1911, p. 1; Tanin, 13 October 1911, p. 1), while Italy occupied Tripoli few days after the outbreak of the war (Beehler, 1913, p. 20). On 12 October 1911, Tanin reported propositions for sanctions against Italians living in the Ottoman Empire (Hüseyin Cahid, 12 October 1911, p. 1). Based on these results shows we can argue that investors have expected higher risks for KM due to the Turco-Italian war. On the other hand, the risk decreases while the conflicts could be shortlived. Other breaks in KM1, AR, and DT are observed before the end of the TurcoItalian war. The breaks in KM1 are identified on 25 March, 21 June, and 23 August 1912 and brought lower prices in the long-run. On 25 March 1912, Tercüman-ı Hakikat reported the Italian occupation in several islands


115

(Tercüman-ı Hakikat, 25 Mart 1912, p. 1). On 21 June 1912, Tanin disseminated news about the new taxes that the Ottoman government imposed to finance the increasing war expenditures (Tanin 21 June 1912, p. 1). On 24 August 1912, Tanin discussed that the end of the conflicts was approaching (Tanin, 24 August 1912, p. 1). We find a break in AR on 16 January 1912 leading to a price decrease, while the price fall is lower than those of before. On 16 January 1912, Tercüman-ı Hakikat disseminated news for discussion on the end of the conflicts (Tercüman-ı Hakikat, 16 January 1912, p. 1). On the other hand, on 4 March and 11 May 1912, there are breaks in the volatility of DT’s stock return prior to the end of the Turco-Italian war, resulting in a longrun decrease in its price. The price increase becomes higher over time. On 4 March and 11 May 1912, Tercüman-ı Hakikat and Tanin argued that there severe conflicts and the Ottoman army got some support from ingenious people (Tercüman-ı Hakikat, 4 March 1912, p.1; Tanin, 11 Mayıs 1912, p.1). This finding suggests lower risk expectation for KM1, AR, and DT, as the end of the conflicts was approaching.

116


Table 2. Structural Break Dates and Corresponding Price Changes, 1910–14 OGI

Break dates

R Change over Break dates longer period

IOB Change over longer period

BM

Break dates

Change over longer period

Break dates


11.11.1910

0.87

27.05.1911

0.15

08.11.1910

-1.03

30.12.1910

2.90

17.11.1910

0.99

10.07.1911

0.45

06.12.1910

-0.81

16.01.1911

5.45

17.12.1910

0.92

26.07.1911

0.55

06.01.1911

-0.84

12.03.1911

18.08

26.09.1911

1.01

31.05.1911

-1.71

30.09.1911

1.08

01.07.1911

-1.85

25.02.1913

2.46

471

Obs

469

KM1

Break dates

469

434

KM2 Change over longer period

Break dates

EM

Change over longer Break dates period


06.10.1911

-35.55

19.12.1910

-32.47

20.12.1910

-9.54

12.10.1911

-35.18

14.04.1913

-35.83

03.02.1911

-7.98

25.03.1912

-29.97

02.06.1913

-36.25

22.04.1911

-7.30

21.06.1912

-29.94

19.07.1913

-35.38

13.06.1911

-6.34

23.08.1912

-29.72

09.01.1914

-36.91

22.09.1911

-1.74

01.03.1913

-31.28

02.03.1914

-37.21

Obs

460

460

463

Econometrics: Methods & Applications AR

Break dates

117

IF Change over longer Break dates period

DT Change over longer period

Break dates


14.09.1911

-1.30

22.04.1911

-1.21

15.04.1911

7.93

16.01.1912

-0.93

06.06.1911

-1.18

18.04.1911

7.89

01.02.1913

-0.27

18.02.1913

-0.30

10.07.1911

6.48

13.09.1913

-0.02

04.03.1912

5.57

14.11.1913

-0.02

11.05.1912

5.81

30.12.1913

-0.17

Obs

435 295 347 Note: The change observed in the stock price before and after the break times is defined as “change over longer period”.

We identify five break points prior to the end of The First Balkan war. By March 1913, there was a decreasing price in the long-run in R, AR, IF, as the price fall decreases. We find break points on 1, 18, and 25 February 1913 in returns of AR, IF, and R’s stocks. On 31 January 1913, Tercüman-ı Hakikat disseminated news on that the conflicts would began again after an armistice (Tercüman-ı Hakikat, 31 January 1913, p. 1). On 1 February 1913, Tercüman-ı Hakikat reported increasing prices, although there were uncertainties about the end of conflicts with Balkan states (Tercüman-ı Hakikat, 1 February 1913, p. 1). On 18 and 25 February 1913, Tercüman-ı Hakikat reported the ongoing conflicts in Rumelia (Tercüman-ı Hakikat, 18 February 1913, p.1; 25 February 1913). This result suggests a higher risk as the conflicts were ongoing, though the presence of the peace possibility. But it seems lower the risk level over time, which could imply that there could be an expectation for the end of the hostilities. However, after February 1913 just before the end of the First Balkan war, we identify breaks in KM1 and KM2, leading to lower price, as the fall increases. Surprisingly, 14 April 1913, Tercüman-ı Hakikat disseminated news that the conflicts would be located places which were close to capital city of the Ottoman Empire and came to end (Tercüman-ı Hakikat, 14 April 1913, p. 1) Overall, Table 2 suggests the absence of higher risk level due to the conflicts for many of companies. The wars could be responsible for different risk

118


perceptions for the various companies due to their characteristics. For the Turco-Italian war, there was a risk change perceived by investors prior to the outbreak of conflicts, as it was not persistent over time. This implies that the threats were seen as temporary by investors. For the First Balkan war, there was no initial sign of the outbreak of the war. The risk was high even if the end of the hostilities was approaching, although threat level decreased.

VII.

Discussion

If the stock investors believe that a conflict will have catastrophic consequences for their investments, then stock prices fall. Our findings indicate falls in stock prices for AR corresponding to volatility changes in their stock returns around the outbreak of the Turco-Italian war. In contrast, we find an increase of R’s stock price with higher volatility. In contrast to the investors of AR’s stocks, those of R’s stocks were received lower threats due to the war. We find breaks corresponding to lower prices and risks for KM1 during the course of TurcoItalian war. The findings also indicate decreasing risk for KM1, AR, and DT’s stocks as the Turco-Italian war was ending. Moreover, we identify break points prior to the end of the First Balkan war in R, AR, IF’s stocks, suggesting price fall and decreasing risk as the conflicts were ending. The AR was a railroad company established to construct the railways between İstanbul and Baghdad in 1888. It was owned by the Deutsche Bank and hired Italian engineers and workers. The hostilities with Italy during the Turco-Italian war created sanctions against Italian residents of the Ottoman Empire and Italian skilled workers were fired, leading to disruptions of construction (Akyıldız, 2011, pp. 112, 118; McMeekin, 2010, pp. 239–240). In addition, there were many attacks during the conflicts (Geyikdagi & Geyikdagi, 2011, p. 387). As resultantly, investors could have expected that the company’s activities would have been negatively affected due to the Turco-Italian war. The R was founded in 1883 with the cooperation of foreign investors and enjoyed a monopoly position in tobacco, salt, and alcohol. Even if we do not report here, there were 1 percent price falls during the break points corresponding to the Turco-Italian war. This shows the negative effects of the war during the beginning of the hostilities. Though the conflicts and oppositions against the renewal of its privileges in 1911, R had not lost much of income in 1911 and 1912, as it imposed a higher price mark-up and regained profitable privileges


119

from the Ottoman state in 1912 (Geyikdagi & Geyikdagi, 2011, p. 381; Parvus Efendi, 2014, pp. 145–75), which could have heralded a safe haven for investors during the war. DT was established in 1869 by a Greek-Ottoman entrepreneur and its activities were negatively affected by the wars, as the equipment was seized by the Ottoman state (Kayserilioğlu, 2003, pp. 38–39, 197). KM was a company established in 1893 by local entrepreneurs to operate copper, manganese mines in Rumelia (Akyıldız, 2011, p. 156; Yılmaz, 2011, pp. 321–325). Under the lack information, it could be argued that the mining activity was interrupted, as the Turco-Italian and Balkan led to difficulties in hiring workers as well as the closeness of mines in the combat zones. IF was the first joint-venture of the Ottoman Empire and established in 1851 by the support of the Ottoman state and the Balkan wars led to damages in its ships due to bombardments (Tutel, 1997, pp. 22–23, 161–162; Kazgan et al., 1999, p. 352), which could be responsible for the fall in the price increase, as the First Balkan war was ending. The absence of a statistically significant impact of war-related news on prices of many of the companies in the sample is in line with the strand of literature6 that finds no evidence for a negative and persistent effect of hostilities. The TurcoItalian war affected many ports in the Ottoman Empire. However, Hanedar (2016) shows that these disruptions on Ottoman foreign trade activities were not statistically significant, which could explain the temporary nature of stock price declines. There is another aspect supporting the findings, as our sample includes such industries and companies that may not be affected by conflicts, since as İkizlerli & Ülkü (2012) and Hobbs et al. (2016) showed, the negative effects of wars varies by industries. Investors, therefore, expected few losses in some industries due to the conflicts. Hanedar (2016) suggests that the Balkan wars only hurt Ottoman exports from the export-oriented farms of Rumelia. Geyikdagi & Geyikdagi (2011) state the absence of significant political risk increases for several joint-stock companies during the wars by 1919 due to home countries’ supports and firms’ privileges. 7 We can argue that investors might have believed that the war would not be that harmful for the non6

7

See Cutler et al. (1989), Corallo (2007), Franck & Krausz (2009), Brune et al. (2014), and Charles & Darne (2014). For instance, the Imperial Ottoman Bank operated as the state bank of the Ottoman Empire. Based on the data obtained from Hanedar (2015), the wars did not have significant effects on the Imperial Ottoman Bank Branches profits and foreign bank entry within the Ottoman Empire’s different povinces by 1914.

120


governmental economic and financial sectors, because the companies were either established or supported by foreign investors. Major European powers 8 protected the investments of their home countries economically and politically. The companies obtained revenue guarantees and privileges from the Ottoman state, making the investments secure. Major countries that invested in the Ottoman Empire were expecting its demise soon. Therefore, investors were likely to invest in the companies just for the sake of having territorial claim without much consideration of risk (Geyikdağı, 2011, pp. 54–55; Geyikdagi & Geyikdagi, 2011, pp. 395–398; Hanedar, 2013). In addition, as Amihud & Wohl (2004) imply, negative effects of wars depend on their durations and stock market investors might have expected a short duration of the Turco-Italian war, as the Ottoman army began to be defeated within a short period of time (Beehler, 1913, p. 20). The presence of immediate impacts of the Turco-Italian war is in line with Corallo (2007), Franck & Krausz (2009), and Kollias et al. (2013), suggesting short-run effects for different war-related events. Hanedar et al. (2015) showed a temporary reduction in the prices of Ottoman government bonds traded on the İstanbul bourse due to the outbreak of the Turco-Italian war.

VIII. Conlusion Because of data unavailability, the finance literature on the impacts of political risks on financial markets does not examine historical cases on the Middle Eastern countries to see the effects of wars on stock markets. Likewise, no previous empirical research focuses on the impact of increased likelihood of war outbreaks for the İstanbul bourse. We study the impacts of the Turco-Italian and the Balkan wars on the stocks traded at the İstanbul Bourse, using novel data which are manually collected from the daily Ottoman newspapers. We extend the previous literature by providing empirical evidence on the stock market, as the paper could be refined with future researches using additional data on other stocks, volume of trade, and investor profile, which seem to be non-existent at this time. Our empirical results indicate that many breaks in the volatility of stock returns were not correlated to war related events. There were falls in stock prices of 8

The UK, France, Germany, Italy, and Austria-Hungary.


121

export oriented monopoly, mining, and transportation companies only for a short time period during the Turco-Italian and the First Balkan wars. In contrast to the relatively timid responses of stock market investors to war-related risks, Hanedar et al. (2015) indicate higher responsiveness of government bond prices during the same period. So, it seems that government bondholders were more sensitive to the war-related risks, as the conflicts were highly related to government survival. In addition, the effects of wars depend on industry location, the ownership structure and networks of companies as well as privileges that were provided by goverment. Moreover, if the company has assets that are likely to be seized bu the government, the impact could not be negligible. To sum up, the risk of wars perceived by government bond investors seem to have been associated with higher uncertainty regarding the Ottoman state’s fiscal position rather than increasing costs of life within the whole country. During the nineteenth century, the Ottoman state had financial problems, leading to higher budget deficits and debt burden. Wars were important sources of the solvency problem (Kıray, 1995, pp. 213–221), which could explain the sensitivity of government bond prices to the conflicts studied here. Acknowledgements We have indebted to Ahmet Vecdi Can, Chris Colvin, Necla Geyikdağı, Oktay Güvemli, Nathan Marcus, Erdost Torun, and participants of ICAFR 2017 for their feedbacks and support as well as research grant (2017-03-18006) provided from the Sakarya University.

Sources Hüseyin Cahid (1911). Ne bekliyoruz?. Tanin, 6 October, p. 1. Hüseyin Cahid (1911). Siyaset-İtalyanların tardı. Tanin, 12 October, p. 1. Hüseyin Cahid (1911). Trablusgarb. Tanin, 22 September, p. 1. İkdam (1911). Borsa ahvali. 29 September, p. 1. İkdam (1911). Galata Borsası-Osmanlı eshamı. 25 September, p. 1. Tanin (1911). Borsa. 14 November 1911, p. 2. Tanin (1911). Borsa. 14 November, p. 2. Tanin (1911). İtalya aleyhinde ittihaz-ı husumet. 13 October, p. 1. Tanin (1911). Trablusgarb- İtalya. 12 September, p. 1.

122


Tanin (1911). Trablusgarb- İtalya. 19 September, p. 1. Tanin (1912). Kanakde hücumu. 11 May, p.1. Tanin (1912). Siyaset, Harp vergisi. 21 June, p. 1. Tanin (1912). Sulh şaiyatı. 24 August, p. 1. Tercüman-ı Hakikat (1911). İtalya Trablusgarb. 12 September, p.1. Tercüman-ı Hakikat (1911). Trablusgarb içün. 18 September, p. 1. Tercüman-ı Hakikat (1912), Caniler tuzağı. 25 March, p. 1. Tercüman-ı Hakikat (1912). Muharabe, 4 March, p.1. Tercüman-ı Hakikat (1912). Osmanlı İtalyan cengi. 16 January, p. 1. Tercüman-ı Hakikat (1913). Çatalca hattı harbinde. 14 April, p. 1. Tercüman-ı Hakikat (1913). Harp arifesindeyiz. 1 February, p. 1. Tercüman-ı Hakikat (1913). Muharebe devam ediyor. 18 February, p.1. Tercüman-ı Hakikat (1913). Muharebe devam ediyor. 25 February, p. 1. Tercüman-ı Hakikat (1913). Müzakere-i sulhiye kati edildi. 31 January, p. 1.

References Abramitzky, R. (2015). Economics and the modern economic historian. The Journal of Economic History, 75(4), pp. 1240–1251. Akyıldız, A. (2011). Osmanlı dönemi tahvil ve hisse senetleri. İstanbul: Türk Ekonomi Bankası. Al, H. and Akar, Ş. K. (2014). Osmanlıdan günümüze borsa: Dersaadet tahvilat borsası, 1874–1928. İstanbul: Borsa İstanbul. Amihud, Y. and Wohl, A. (2004). Political news and stock prices: The case of Saddam Hussein contracts. Journal of Banking and Finance, 28(5), pp. 1185–1200. Beehler, W. H. (1993). The history of the Italian-Turkish war, September 29, 1911 to October 18, 1912. US: Annapolis. Brune, A., Hens, T., Rieger, M. O., and Wang, M. (2015). The war puzzle: Contradictory effects of international conflicts on stock markets. International Review of Economics, 62(1), pp. 1–21.


123

Charles, A. and Darné, O. (2014). Large shocks in the volatility of the Dow Jones Industrial Average index: 1928–2013. Journal of Banking and Finance, 43, pp. 188–199. Childs, T. G. (2008). Trablusgarp savaşı ve Türk-İtalyan diplomatik ilişkileri. Deniz Berktay (Translated). İstanbul: İş Bankası kültür yayınları. Choudhry, T. (2010). World War II events and the Dow Jones industrial index. Journal of Banking and Finance, 34(5), pp. 1022–1031. Corallo, E. (2007). The effect of the war risk: A comparison of the consequences of the two Iraq wars. International Review of Economics, 54(3), pp. 371–382. Cutler, D.M., Poterba, J.M., and Summers, L. H. (1989). What moves the stock market?, Journal of Portfolio Management, 15(3), pp. 4–11. Dimic, N., Orlov, V., and Piljak, V. (2015). The political risk factor in emerging, frontier, and developed stock markets. Finance Research Letters, 15, pp. 239–245. Eldem, E. (1999). A History of the Ottoman Bank. İstanbul: Ottoman Bank Historical Research Centre. Erickson, E. J. (2001). Order to die: A history of the Ottoman Army in the First World War. London: Greenwood Press. Fertekligil, A. (2000). Türkiye’de Borsa’nın tarihçesi. İstanbul: İstanbul Menkul Kıymetler Borsası. Fogel, R. W. (1964). Railroads and American economic growth. Baltimore: Johns Hopkins Press. Franck, R. and Krausz, M. (2009). Institutional changes, wars and stock market risk in an emerging economy: Evidence from the Israeli stock exchange, 1945–1960. Cliometrica, 3, pp. 141–164. Geyikdaği, V. N. (2011). Foreign investment in the Ottoman Empire: International trade and relations, 1854–1914. London: I. B. Tauris and Co Ltd.. Geyikdagi, V. N. and Geyikdagi, M. Y. (2011). Foreign direct investment in the Ottoman Empire: Attitudes and political risk. Business History, 53, pp. 375– 400.

124


Giolitti, G. (2012). Dönemin İtalya başbakanının, Türk-İtalyan savaşına dair hatıraları, Trablusgarp’ı nasıl aldık. Tahsin Yıldırım (Translated). İstanbul: DBY yayınları. Guidolin, M. and La Ferrara, E. (2010). The economic effects of violent conflict: Evidence from asset market reactions. Journal of Peace Research, 47(6), pp. 671–684. Hall, R. C. (200). The Balkan Wars 1912–1913, prelude to the First World War. London: Routledge. Hanedar, A. O. (2013). Three Essays on the Economy of the Late Ottoman Empire. PhD thesis, Colchester: University of Essex. Hanedar, A. Ö. (2015). Foreign Bank Entry in the Late Ottoman Empire: The Case of the Imperial Ottoman Bank. Review of Middle East Economics and Finance, 11(3), pp. 207–223. Hanedar, A. Ö. (2016). Effects of wars and boycotts on international trade: Evidence from the late Ottoman Empire. The International Trade Journal, 30, pp. 59–79. Hanedar, A. Ö. Torun, E., and Hanedar, E. Y. (2015). War-related risks and the İstanbul bourse on the eve of the First World War. Borsa İstanbul Review, 15(3), pp. 205–212. Hanedar, A. Ö., Hanedar, E. Y., and Torun, E. (2016). The end of the Ottoman Empire as reflected in the İstanbul bourse. Historical Methods: A Journal of Quantitative and Interdisciplinary History, 49(3), pp. 145–156. Haupert M. (2006). History of Cliometrics, Diebolt C., Haupert M. (Eds.): Handbook of Cliometrics, Berlin: Springer, 2016, pp. 3–33. Hobbs, J., Schaupp, L. W., and Gingrich, J. (2016). Terrorism, militarism, and stock returns. Journal of Financial Crime, 23(1), pp. 70–86. Hudson, R. and Urquhart, A. (2015). War and stock markets: The effect of World War Two on the British stock market. International Review of Financial Analysis, 40, pp. 166–177. İkizlerli, D. and Ülkü, N. (2012). Political Risk and Foreigners' Trading: Evidence from an Emerging Stock Market. Emerging Markets Finance and Trade, 48, pp. 106–121.


125

Inclan, C. and Tiao, G. (1994). Use of the cumulative sums of squares for retrospective detection of changes of variance. Journal of the American Statistical Association, 89, pp. 913–923. Karpat, K. (2002). Studies on Ottoman social and political history: Selected articles and essays. Leiden: Brill. Kayserilioğlu, R. S. (2003). Dersaadet’ten İstanbul’a Tramvay. 1. Cilt. İstanbul: FSF Matbaacılık. Kazgan, H. (1995). Tarih boyunca İstanbul Borsası. İstanbul: İstanbul Menkul Kıymetler Borsası. Kazgan, H., Ateş, T., Tekin, O., Koraltürk, O., Soyak, A., Eroğlu, N., and Kaban, Z. (1999). Turkish financial history from the Ottoman Empire to the present. Volume 1. İstanbul: İstanbul Stock Exchange. Kray, E. (1995). Osmanlı’da ekonomik yapı ve dış borçlar. İstanbul: İletişim yayınları. Kollias, C., Kyrtsou, C., and Papadamou, S. (2013). The effects of terrorism and war on the oil price–stock index relationship. Energy Economics, 40, pp. 743–752. Kollias, C., Papadamou, S., and Stagiannis, A. (2010). Armed conflicts and capital markets: The case of the Israeli military offensive in the Gaza strip. Defence and Peace Economics, 21, pp. 357–365. Le Bris, D. (2012). Wars, inflation and stock market returns in France, 1870– 1945. Financial History Review, 19(3), pp. 337–361. Mathy, G. P. (2010). Stock volatility, return jumps and uncertainty shocks during the Great Depression. Financial History Review, 23(2), pp. 165–192. McMeekin, S. (2010). The Berlin-Baghdad Express: The Ottoman Empire and Germany`s Bid for World Power, 1898-1918. London: Allen Lane. Önder, Z. and Şimga-Mugan, C. (2006). How do political and economic news affect emerging markets? Evidence from Argentina and Turkey. Emerging Markets Finance and Trade, 42, pp. 50–77. Özmucur, S., and Pamuk, Ş. (2002). Real wages and standards of living in the Ottoman Empire, 1489–1914. The Journal of Economic History, 62(2), pp. 293–321.

126


Quataert, D. (1996). The social history of labor in the Ottoman empire, 18001914. Ellis Jay Goldberg, The social history of labor in the Middle East (Boulder, Co., 1996), 23. Quataert, D. (2005). The Ottoman Empire, 1700-1922. Cambridge University Press. Pamuk, Ş. (2000). A monetary history of the Ottoman Empire. New York: Cambridge University Press. Rigobon, R. and Sack, B. (2005). The effects of war risk on US financial markets. Journal of Banking and Finance, 29(7), pp. 1769–1789. Schneider, G. and Troeger, V. E. (2006). War and the world economy stock market reactions to international conflicts. Journal of conflict resolution, 50(5), pp. 623–645. Tuncer, A. C. and Pamuk, Ş. (2014). Ottoman Empire: from 1830 to 1914, South-Eastern European monetary and economic statistics from the nineteenth century to world war. Athens, Sofia, Bucharest, Vienna: Bank of Greece, Bulgarian National Bank, National Bank of Romania, Oesterreichische National bank. Tutel, E. (1997). Şirket-i Hayriye. İstanbul: İletişim Yayınları. Urquhart, A. and Hudson, R. (2016). Investor sentiment and local bias in extreme circumstances: The case of the Blitz. Research in International Business and Finance, 36, pp. 340–350. Yılmaz, C. (2012). Osmanlı anonim şirketleri. İstanbul: Scala yayınları. Zussman, A., Zussman, N., and Nielsen, M. Ø. (2008). Asset market perspectives on the Israeli–Palestinian conflict. Economica, 75, pp. 84–115.

Chapter 9

Determinants of Poverty in Turkey Süreyya DAL 1

Abstract. Poverty can be regarded as a multidimentional social issue, where income poverty and human poverty constitute the two main dimensions. According to the Turkish Statistical Institute's (TUIK) 2010 Household Budget Survey, the poverty rate for Turkey is 3.66 per cent with US Dollar 4.3 (PPP) per capita daily limit. This paper investigates the determinants of poverty in Turkey by employing TUIK 2010 Household Budget Survey data. First, in order to measure the incidence or the spread of poverty the headcount ratio measures are calculated. Then factors contributing to poverty are investigated utilizing logistic regression analysis. Early empirical results indicate that in addition to economic factors, socio-demographic factors are also among determinants of poverty. The analysis suggest that there is a direct relationship between poverty and educational status and gender of household head. Poor people in Turkey generally live in extended families, have lower educational status, are likely to be women, lack social and health insurance, and work seasonally, implying that there is a significant association between social exclusion and poverty.

Keywords: Poverty, Poverty Measure, Logit Model, Categorical Data Analysis

1

Res. Ass. Trakya University, Department of Econometrics, Edirne,Turkey, [email protected]

128

Deteminants of Poverty in Turkey

I.

Introduction

Poverty can be regarded as a multidimensional social issue. Income poverty and human poverty constitute the two main dimensions. Income poverty is the lack of income necessary to satisfy basic needs whereas human poverty is the lack of human capabilities, for example poor life expectancy, poor maternal health, illiteracy, poor nutritional levels, poor access to safe drinking water and perceptions of well-being (UNDP, 2003). However the lack of economic resources is not the only determinant of destitution, in that the social context of the issue should also be taken into account. In this regard, the definition of poverty is increasingly being framed in terms of capacity to participate in the society in which a person lives. Thus poverty in this way is directly linked to social inclusion (Atkinson and Marlier, 2010). Poverty is defined as the situation in which one has a limited capacity to intervene in his own circumstances and the society he lives in, which means he is socially excluded in various aspects of life such as lack of money and power, ethnic, linguistic, racial and cultural isolation, physical, mental and health disabilities. The limited capacity of the individual, in turn, contributes to his failure to integrate into the labour market, which restrains him from pursuing his individual welfare as an active, responsible agent. Barry (1998) states that social exclusion is a violation of the demands of social justice in two ways: It conflicts with equality of opportunity, as social exclusion leads to unequal educational and occupational opportunities. Furthermore social exclusion is associated with an inability to participate effectively in politics, resulting in unequal opportunity in relation to politics. Thus poverty can be represented as mainly the outcome of the lack of human capital endowments. In addition to the low initial level endowments, reflected in low schooling and skills and poor health, relative abundance of labour and the presence of an informal economy prevent poor participating in labour market. Accordingly, in order to eradicate poverty, public policies should be implemented, aiming to improve the human capital potential of the poor and integrate them to the market. These policies include provision of basic social (health, education, nutrition) and infrastructure services to build their human capital capacities. There has been a continuing effort towards the alleviation of poverty and achieving greater social cohesion and social inclusion. In 2001 Turkey suffered the most severe economic crisis the country had known in its modern history.

Econometrics : Methods & Applications

129

As a rapid response to financial crisis, the Social Risk Mitigation Project (SRMP) has been initiated in 2001, which was financially supported by the World Bank. The SRMP was designed to empower and expand the available social safety net programs, aiming to alleviate the impact of the recent economic crisis on poor households. In order to achieve these objectives the SRMP initiated Conditional Cash Transfer Programs (CCT). The CCTs provide continuous cash transfers conditional on positive behavioral change, in order to induce the demand for education and health services among the poor. These programs represent a shift in government’s approach of focusing on the supplyside to a demand driven approach. The CCTs target disadvantaged people and aim to enhance future human capital, as well as providing immediate poverty relief. CCT programs gained worldwide popularity, as they achieved success in reaching excluded groups, notably the extreme poor living outside the reach of social protection programs tied with informal sector employment. Especially, targeting of households with children and making monetary transfers to women are important features of CCT programs contributing to efforts towards greater policy inclusion. The aim of this study is to investigate key determinants of poverty in Turkey for targeting households in the poverty alleviation policies. For this purpose, firstly, we review previous poverty measurement studies in Turkey. Secondly, we discuss the equivalence scales. Next, the estimation results will be elaborated upon. The estimation results are consequently subdivided into an investigation of effects of TurkStat equivalence scale on measures of poverty and determinants of poverty. Finally, some concluding remarks are made.

II.

Previous Poverty Measurement Studies in Turkey

Poverty is a social problem that cannot be ignored in Turkey as well as all over the world. To solve this problem, First of all, the concept of the poverty should be better understood. Then the causes of poverty should be put forward in a quantitative manner. In Turkey, most of the available studies are descriptive and focus mainly on poverty alleviation policies. However, the econometric methods used in poverty studies are quite limited. The poverty reports suggest that there is a direct relationship between poverty and educational status and gender. There are twice as many women as men who were illiterate in the poor group. Additionally people who live in an extended

130


family in rural parts of Turkey have a higher likelihood of being in poverty compared to those who live in urban areas as a nuclear family. Poor people in Turkey have lower educational status, are likely to be women, lack social and health insurance, and work informally in the agricultural sector as family workers (Dansuk, 1997; Saatci & Akpinar, 2007). Hence there is a significant association between social exclusion and poverty in Turkey. People who live in rural areas, in extended families and usually work as family labor force have limited or no access to education and health facilities, which in turn contributes to their poverty prevailing generations. Hentschel et al. (2010) argue that wealth and measured circumstances are closely related in Turkey. They state that 85 per cent of women in the least wealthy households are born in rural areas whereas this holds for only 20 per cent of women who live in a wealthy household. Moreover there is a direct relationship between parental education and wealth: as the level of parental education increases so does wealth. Additionally regional differences of birth place also affect the wealth status. People born in urban and Western Turkey are more likely to be wealthier than those born in rural and Eastern Turkey. Overall, Hentschel et al. (2010) find that at least one third of the wealth disparity in Turkey can be explained by inequalities in opportunity. Likewise, Ferreira et al. (2011) explore the opportunity profile for Turkey, constructed by ranking household types by chosen valuation of their opportunity sets: mean imputed consumption. Empirical analysis reveals that the bottom 10% of the distribution is 88% rural and 96% Eastern by birth. Moreover, around 70 per cent of them have six or more siblings, in addition to low parental education levels. Aktan (2002) identified the following as causes of poverty and income inequality: rapid population growth, unfair tax system, tax evasion, high interest rate policy, climate, natural conditions and natural disasters, unemployment, inflation, inheritance, government grants, monopolization of the market, differences in the ability of individual and being unable to work. Kızılgöl and Demir (2010) estimate a logit model using pooled data of household budget survey between 2002 and 2006. The results of the logit model show that age of the household head, education of the household head and household size are important determinants of poverty. The probability of being poor decreases by rise of age and education level. Furthermore, they find a negative relationship between household size and the probability of being poor. Kabaş (2009) examines Turkey’s poverty profile, it is seen that poverty changes directly


131

(positively) with the household size, indirectly (negatively) with the education level. The handicapped people, children and retired people are among the most vulnerable and fragile groups in Turkey. There is a large gender gap in education and employment sectors, central and local decission taking processes and social environments. Cafri (2009) estimates Tobit and probit model for the province of Adana. According to the results of the analysis, Key determinants of probability and depth of poverty are: the age of the household head, female population of household, level of education and total expenditure. Çağatay et al (2009) estimate linear regression model using data of the 2003 Household Budget Survey for identifying the determinants of poverty in the AntalyaBurdur-Isparta region. They used poverty gap as the dependent variable and a limited of individual characteristics (e.g. employment and education status, gender, etc.) as explanatory variables. The estimation results show that the poverty gap decrease by the number of women in the household who work. Dayıoğlu (2007) investigate the determinants of child labour in urban Turkey with a special reference to low household income or poverty as one of its root causes. The data from urban Turkey indicate that children from poorer families stand at a higher risk of employment. This ﬁnding is conﬁrmed using various measures of household material well-being. Simulation results have further pointed out that current interventions are not likely to produce a sizeable impact on the child labour problem. Şengül and Tuncer (2005) examine poverty levels and the food demand of poor and extremely poor households in Turkey by using the Household Consumption Expenditure Survey of 1994. First, a least-cost food poverty line was determined. Then, some aggregate poverty measures, namely the Head-Count Ratio, the Poverty Gap Ratio, the Sen Index and the Foster, Greer and Thorbecke Index, were calculated and employed for assessment of poverty levels in Turkey as a whole and in both urban and rural areas. These indices indicate that approximately 46.8% of the households are poor, and 7% of the households are extremely poor in Turkey. Akder (1999) build poverty map for Turkey. The population size, life expectancy, literacy rate and per capita income in dollar terms are taken into account to calculate "Human Development Index" according to their location in Turkey. To sum up, Household Budget Survey is used in the applied studies to identify determinants of poverty. According to the literature of poverty, age, gender and education of household head, household size and residence are key determinants of poverty.

132


III.

Estimation Results

The estimation results are consequently subdivided into an investigation of effects of TurkStat and EuroStat equivalence scales on measures of poverty and identification of key determinants of poverty in Turkey. First, Different measures of poverty are constructed for TurkStat and EuroStat equivalence scales. Then, a logit model has been estimated to identify key determinants of poverty in Turkey. Turkish Household Budget Survey 2010 was used for estimation results. Poor households are determined according to US Dollar 4.3 (PPP) per capita daily limit. If a daily consumption of a household is under US Dollar 4.3 (PPP), we take this household as a poor and we denote it by 1. Otherwise, we denote households by 0. Measures of Poverty in Turkey The most widely used index is headcount index in poverty studies. This index is simple to construct and easy to understand. However this index does not take the intensity of poverty into account. This index assumes that degree of the poverty is same among poor. The poverty rate for Turkey is 8.48 per cent by using TurkStat equivalent scales. It means that 8.48% of the population suffers from poverty. Poverty gap index, which can be thought as a cost of eliminating poverty, has been calculated 2.19%. Squared Poverty Gap Index which takes into account inequality among poor has been calculated 4.78%. According to TurkStat equivalence scale, poverty measures are found higher than EuroStat equivalence Scale. Sensitivity analyses suggest that while the level and, in particular, the composition of income poverty are affected by the use of different equivalence scales, trends over time and rankings across countries are much less affected (Burniaux et al., 1998). Table 1. Poverty Index 2010 with US Dollar 4.3 (PPP) per capita daily limit. Poverty Index 2010:Turkey(%) Headcount Ratio, P 0 Poverty Gap Index, P 1 Squared Poverty Gap Index, P 2

TurkStat Equivalence Scale 8.48 2.19

EUROSTAT Equivalence Scale 1.09 0.22

4.78

0.05


133

Logit Model Results Logit model has been estimated by using a backward stepwise selection method. To identify key determinants of poverty we first computed a dichotomous variable indicating whether the household is poor or not. That is, P= 1, if household is poor 0, otherwise where P denotes poverty. For this purpose, we calculated daily consumption per person with respect to TurkStat equivalence scale. The captivity parameters may be explained by what characterizes deprivation, vulnerability (high risk and low capacity to cope), and powerlessness (Lipton and Ravallion, 1995; Sen, 1999). These characteristics weaken people's sense of wellbeing. In this context, the explanatory variables include: number of room per person, log monthly income per person, number of children, age (64+years old), level of education (illiteral, primary school and high school), marital status (single and divorced), green card (health insurance for poor people) ownership, being housewife, employment status (seasonal worker and civil servant), female household headed, residential characteristics (number of internet connection, number of mobile phone, number of LCD television, number of washing machine and toilet ownership, housing challenges (difficulty of access to public transport), vehicle ownership (number of car), real property ownership (field crop ownership), household type (nuclear family and couple without children) and fuel type (wood, cool and cow dung) in the final model that was fit to the Household Budget Survey 2010 (see Table 2). Table 2. Definitions of variables used in the Logit Model Explanatory variables Age 64+ Illiterate Primary School attained High School attained Marital Status

=1 if 64+ years old, 0 otherwise =1 if individual can read and write, 0 otherwise =1 if in primary school degree, 0 otherwise =1 if in high school degree, 0 otherwise =1 if married, 0 otherwise =1 if divorced, 0 otherwise Green Card (Health insurance for poor =1 if yes and 0 otherwise people) ownership Main occupation of individual =1 if housewife, 0 otherwise =1 if seasonal worker, 0 otherwise =1 if civil servant, 0 otherwise Female household headed =1 if female household headed, 0 otherwise Toilet ownership =1 if yes and 0 otherwise Field crop ownership =1 if yes and 0 otherwise Difficulty of access to public transport =1 very difficult, 2 difficult, 3 very easy, 4 easy

134 Household Type Fuel Type

Deteminants of Poverty in Turkey =1 if in nuclear family, 0 otherwise =1 if in couple without children, 0 otherwise =1 if wood and 0 otherwise =1 if coal, 0 otherwise =1 if dried cow dung, 0 otherwise

Table 3 presents Logit model estimation results for identifying key determinants of poverty in Turkey. According to the estimation results, Number of room per person is estimated to have a negative effect on the probability of being poor. Since shared space indicates household living standards. Throughout the world it has been found that the probability of finding employment rises with higher levels of education, and that earnings are higher for people with higher levels of education. The results for Turkey support that a better educated household is less likely to be poor. Being illiterate is positively associated with poverty, while having primary and high school degree are negatively associated with poverty. The coefficient estimates for illiterate and primary and high school degree are statistically significant at less than the 10% level. Individuals living in female headed households face higher probabilities of being poor, and lower living standards, compared to those in male headed households. Total income is chronically low in many female-headed households due to the presence of at most only one (female) wage earner and women’s lower earnings in the labor market (Snyder et al., 2006). Main occupation of households is widely used in poverty studies. In the model, as would be expected, working as a seasonal worker is statistically significant at less than 1% level and it is positively related into poverty status. It means that seasonal workers who have not regular income have a positive effect on the probability of being poor. However, civil servants in Turkey are less likely to be poor. Public transportation access appears to have a negative effect on household welfare. Poor areas are poorly served by roads and transport services. More broadly, poor people can be classified and monitored according to their living conditions. In the model, the coefficients which are related to housing facilities are statistically significant at less than the 5% level. With respect to the expectations, they have a negative effect on the probability of being poor. It means that households who have higher living conditions are less likely to be poor. Number of children in the household is positively related to poverty status. Similarly, the estimations showed that nuclear family and couple without children are less likely to be poor.


135

Table 3. Estimation results for the best fitting Logit model Dependent Variable: 1, if household is poor; 0, otherwise Pseudo R2: 0.49 Wald Chi2(27)=4078.57 (0.00) Rob. Rob. Variable Coef. Std. z p>|z| Variable Coef. Std. z Err. Err. Housing Cons 11,19 0,49 22,78* 0,00 Challenges Difficulty of Number of room access to -0,14 0,03 -5,41a -1,28 0,14 -9,21* 0,00 public per person transport Log (monthly Vehicle income per -5,21 0,19 -28,09* 0,00 Ownership person) Number of Number of 0,04 0,02 2,29** 0,02 -0,36 0,09 -4,15* children car Real Personal Property Characteristics Ownership Field crop Age 64+ 0,36 0,10 3,58* 0,00 -0,25 0,06 -4,28* ownership Household Education Type Nuclear -0,62 0,05 -11,49* Illiterate 0,14 0,07 2,04** 0,04 family Couple -0,35 0,19 -1,82^ Primary -0,23 0,09 -2,59** 0,01 without children High School -0,27 0,15 -1,85^ 0,06 Fuel Type Marital Status Wood 1,17 0,18 6,64* Single 0,27 0,08 3,37* 0,00 Coal 1,11 0,18 6,09* Dried Cow Divorced 0,69 0,22 3,11* 0,00 1,14 0,20 5,76* Dung Housewife -0,20 0,07 -2,69* 0,00 Green Card(Health 0,47 0,06 8,51* 0,00 insurance for poor people) ownership Seasonal worker 0,72 0,24 2,99* 0,00 Civil servant -0,58 0,33 -1,75^ 0,07 Female household 0,33 0,14 2,27** 0,02 headed Housing Facilities Number of internet -1,31 0,18 -7,36* 0,00 connection

p>|z|

0,00

0,00

0,00

0,00 0,07

0,00 0,00 0,00

136


Number of -0,14 0,03 -5,41* 0,00 mobile phone Number of LCD -0,55 0,22 -2,53** 0,01 television Number of washing -0,67 0,06 -4,28* 0,00 machine Toilet ownership -0,55 0,06 -9,73* 0,00 *, **, ^ significant at 1, 5 and 10 per cent level.

Performance Evaluation of Logit Model Table 4 summarizes sensitivity, specificity and accuracy of Logit Model. The table rows give the results of the Logit Model, as either positive for the poverty status or negative for the poverty status. The columns indicate the true poverty status, as either poor or not poor. Sensitivity is the probability of a positive model result (that is, the model indicates the presence of poverty) for a poor household. It means that sensitivity is the probability of a being poor among poor households. Specificity, on the other hand, is the probability of a negative model result (that is, the model does not indicate the presence of poverty) for a poor household. Specificity is the probability of a being not poor among nonpoor households. Overall accuracy of Logit model has been found 93.68%. The findings show that the model is considerably strong to distinguish poor households. Table 4. Classification table for the Logit model Classified + Total

True Poor (D) 1495 1743 3238

Classified + if predicted Pr(D)>=.5 True D defined as poor !=0 Sensitivity Specifity Positive Predictive Value Negative Predictive Value False + rate for true ~D False + rate for true ~D False + rate for classified + False - rate for classified Correctly classified

Not Poor (~D) 673 34295 34968 Pr (+|D) Pr (-|~D) Pr (D|+) Pr (~D|-) Pr (+|~D) Pr (- |D) Pr (~D|+) Pr (D|-) Pr (+|D)

46.17% 98.08% 68.96% 95.16% 1.92% 53.83% 31.04% 4.84% 46.17% 93.68%

Total 2168 36038 38206


137

ROC curves are constructed from sensitivity and specificity to evaluate performance of the Logit model. An ROC curve begins at the (0, 0) coordinate, corresponding to the strictest decision threshold whereby all model results are negative for poverty status. The ROC curve ends at the (1, 1) coordinate, corresponding to the most lenient decision threshold whereby all model results are positive for poverty status. Figure 1 illustrates an empiric ROC curve constructed from Logit model estimations in Table4. The area under the ROC curve shows the ability of the model to correctly classify poor and non-poor. According to findings, the area under the ROC curve is 0.95. It would be considered to be “excellent” at distinguishing poor households from non-poor households.

0.00

0.25

Sensitivity 0.50

0.75

1.00

Figure 1. ROC curve for the Logit model

0.00

0.25

0.50 1 - Specificity

0.75

1.00

Area under ROC curve = 0.9512

IV.

Conclusion

In this paper an attempt has been made to explore effects of TurkStat and EuroStat equivalence scales on measures of poverty and to identify the key determinants of poverty in Turkey. First, Headcount ratio, poverty gap index and squared poverty gap index are constructed for TurkStat and EuroStat equivalence scales. The results indicate that income poverty are affected by the use of different equivalence scales. 8.48% of the Turkey’s lives in poverty with respect to TurkStat equivalence scale. Then, We have employed binomial logit model using the 2010 Household Budget Survey data. For this purpose, in the

138


analysis we use TurkStat equivalence scale which represent better characteristics of Turkey. From the logit model analysis we obtain a set of findings that are broadly consistent with the poverty literature. Education coefficients have been found statistically significant at three levels of education. Households who can not read or write are more likely to be poor, while primary and high school degree are negatively associated with poverty. Educational investments can be powerful instruments for long-term poverty reduction. As the results indicate, individuals living in female headed households face higher probabilities of being poor. Women’s lower average earnings compared to men, less access to remunerative jobs, and productive resources such as land and capital contribute to the economic vulnerability of female-headed households. Gender related policies with which we can eliminate gender-bias against women can be applied to alleviate poverty in Turkey. More broadly, poor people can be classified and monitored according to their living conditions. According to the results, households who have higher living conditions face lower probabilities of being poor.As would be expected, for the poor, labour is often the only asset they can use to improve their well-being. The results support this with a significant coefficients of employment status at three level. Hence It is crucial to provide suitable jobs that both secure income and empowerment for the poor, especially women and younger people. Overall accuracy of Logit model has been found 93.68%. According to the findings, the area under the ROC curve is 0.95. It would be considered to be “excellent” at distinguishing poor households from non-poor households. The findings show that the model is considerably strong to identify key determinants of poor households. İdentifying determinants of poverty may well be the answer for the poverty alleviation policies. This paper has attempted to identify key determinants of the poor households in Turkey. In this manner we can better fight against poverty.

References Akder, H. (1999). Dimensions of Rural Poverty in Turkey. Turkey Economic Reforms, Living Standards and Social Welfare Study, Vol. II. World Bank. Aktan, C., C. (2002). Yoksulluk Sorununun Nedenleri ve Yoksullukla Mücadele Stratejileri, Yoksullukla Mücadele Stratejileri, Hak-İş Konfederasyonu Yayını (pp. 20). Ankara.


139

Atkinson, A., B., & Marlier, E. (2010). Analysing and Measuring Social Inclusion in a Global Context. UN Department of Economic and Social Affairs, ST/ESA/325. Barry, B. (1998). Social exclusion, social isolation and the distribution of income. CASE Paper 12. London, Centre for Analysis of Social Exclusion, London School of Economics. Burniaux, J., M., Dang, T.,T., Fore, D., Förster, M., F., Mira d'Ercole, M. & Oxley, H. (1998). Income Distribution and Poverty in Selected OECD Countries. OECD Economics Department Working Paper, 189. Cafri, R. (2009). Analysis of Poverty in Adana: A Research With Limited Dependent Variables. Unpublished MA Thesis. Department of Economics, University of Çukurova. Çağatay, S., Zanbak, M., & Duman, K. (2009). TR61 Bölgesinde Yoksulluk Profili ve Hane İçi Kaynak Dağılımının Yoksulluk Üzerindeki Etkisi. EconAnadolu 2009: Anadolu International Conference in Economics. Eskişehir, Turkey. Dansuk, E. (1997). Measuring poverty in Turkey and the relationship with socioeconomic structures. State Planning Organization Dissertation in Turkish. Dayıoğlu, M. (2006). The impact of household income on child labour in urban Turkey. The Journal of Development Studies, 42(6), 939-956. Hentschel, J., Aran, M., Can, R., Ferreira, F.H.G., Gignoux, J., &Uraz, A. (2010). Life Changes in Turkey: Expanding Opportunities for the Next Generation. World Bank Publication. Kabaş, T. (2009). Causes of Poverty and Poverty Reduction Strategiesin Developing Countries. Unpublished PhD Thesis. Department of Economics, University of Çukurova. Kızılgöl, Ö.& Demir, Ç. (2010) Türkiye’de Yoksulluğun Boyutuna İlişkin Ekonometrik Analizler. İşletme ve Ekonomi Araştırmaları Dergisi,1(1), 2132. Lipton, M., & Ravallion, M. (1995). Poverty and policy. In H. Chenery & T.N. Srinivasan (Eds.), Handbook of Development Economics (pp. 2551-2657), California, CA: Elsevier. Saatci, E., & Akpinar, E. (2007). Assessing poverty and related factors in Turkey. Croatian Medical Journal, 48(5), 628–635. Sen, A.K. (1999). Development As Freedom. Oxford: Oxford University Press.

140


Snyder, A., R., McLaughlin, D., K.& Findeis, J.(2006). Household Composition and Poverty among Female-Headed Households with Children: Differences by Race and Residence. Rural Sociology, 71(4), 597-624. Şengül, S., & Tuncer, İ. (2005) Poverty levels and Food Demand of the Poor in Turkey. Agribusiness. 21(3), 289-311. TÜİK. (2008). Tüketim Harcamaları, Yoksulluk ve Gelir Dağılımı: Sorularla Resmi İ statistikler Dizisi-6. http://www.tuik.gov.tr. UNDP. (2003). Human Development Report 2003. New York.

Editor: M. Kenan TERZİOĞLU Co-Editor: Süreyya DAL


ISBN 978-605-344-671-2

9 786053 446712


econometrics: methods & applications

econometrics: methods & applications

Suggest Documents

Applied Econometrics with Eviews Applications

Applied Econometrics with Eviews Applications

Applied Econometrics with Eviews Applications

Applied Econometrics with Eviews Applications

Quantitative Methods in High-Frequency Financial Econometrics ...

Econometrics

Geophysical Methods & Applications

Principles, Methods and Applications

Bootstrap Methods and Applications

Bootstrap Methods and Applications

Advancing Methods and Applications

Methods and Applications

METHODS AND APPLICATIONS

econometrics

ECONOMETRICS

Econometrics

econometrics

Econometrics

Applied Econometrics Applied Econometrics - IES

Developing Hypermedia Applications with Methods

(English) Quantitative Methods and Applications

MRI Relaxometry: Methods and Applications

MRI Relaxometry: Methods and Applications

Driver Measurement: Methods and Applications