xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001966283
007 cr mnu|||uuuuu
008 081020s2007 flu sbm 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0002226
Malaria, labor supply, and schooling in Sub-Sahara Africa
h [electronic resource] /
by Taiwo Abimbola.
[Tampa, Fla.] :
b University of South Florida,
ABSTRACT: The purpose of this study is to evaluate the causal effects of malaria and poor health in general on economic outcome in Sub-Saharan Africa. This study uses panel data from the Living Standard Measurement Survey (LSMS) for Tanzania from 1991 to 2004. Three main hypotheses are tested. First, the study evaluates the effect of malaria and other chronic illnesses on labor supply using the number of hours worked per week as a measure of outcome. Second, it determines the impact of poor health on human capital accumulation by measuring the number of weekly school hours lost to illness. The third objective deals with the question of whether changes in preconditioning factors such as income levels and healthcare accessibility have improved the disease environment in Sub-Saharan Africa over time. The study uses several identification strategies in the empirical estimation process. The first estimation strategy applies the standard Ordinary Least Squares (OLS) and Fixed Effects (FE) estimators to the schooling and labor supply models. In addition to OLS and FE, the preferred methods of estimating the causal effects of malaria on schooling and labor supply outcomes are Two Stage Least Squares (2SLS) and Limited Information Maximum Likelihood (LIML). Findings in this study suggest that malaria significantly increases school absenteeism. In particular, 2SLS and LIML estimates of the number of school hours lost to malaria suggests that children sick with malaria are absent from school for approximately 24 hours a week. However, the results show the effect of malaria on work hours is inconclusive. Furthermore, difference in difference estimates of the disease environment show slight improvements in the disease environment resulting from changes in income levels. The study finds no statistically significant improvements in the disease environment due to increases in the number of health facilities over time.
Dissertation (Ph.D.)--University of South Florida, 2007.
Includes bibliographical references.
Text (Electronic dissertation) in PDF format.
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
Title from PDF of title page.
Document formatted into pages; contains 128 pages.
Co-adviser: Gabriel A. Picone, Ph.D.
Co-adviser: Kwabena Gyimah Brempong, Ph.D.
Human capital accumulation.
t USF Electronic Theses and Dissertations.
Malaria, Labor Supply, and Schooling in Sub-Saharan Africa by Taiwo Abimbola A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy Department of Economics College of Business Administration University of South Florida Co-Major Professor: Gabriel A. Picone, Ph.D. Co-Major Professor: Kwabena Gyimah Brempong, Ph.D. Philip K. Porter, Ph.D. Christopher R. Thomas, Ph.D. Date of Approval: October 26, 2007 Keywords: Health, Human Capital Accumulation, Schoo l, Work Copyright 2007, Taiwo Abimbola
Dedication This dissertation is dedicated to my father whose wisdom has been my motivation. I also dedicate this to my husband, Ni cholas Michael Pallutto, my sister, Mojisola Abimbola, and my twin brother, Olatunji Ab imbola.
Acknowledgements I would like to thank my dissertation committee fo r their leadership and support. Especially, the co-chairs of the committee, Gabriel A. Picone and Kwabena GyimahBrempong who provided invaluable guidance throughou t the entire process. I am very grateful for their time. Furthermore, I also want t o thank the other committee members, Christopher R. Thomas and Philip K. Porter for thei r helpful comments and advice. I would have not been able to complete this dissertat ion without the mentorship of my entire committee.
i Table of Contents List of Tables..................................... ................................................... .............................iv List of Figures.................................... ................................................... .............................vi Abstract........................................... ................................................... ...............................vii Chapter 1: Introduction and Motivation 1.1 Motivation..................................... ................................................... ..................1 1.2 Malaria and its Impact on Sub-Saharan Africa... ...............................................2 1.3 Goals.......................................... ................................................... .....................4 Chapter 2: Theoretical Framework 2.1 Theoretical Framework of Health, Human Capital and Labor Supply..............8 2.2 Health, Labor Supply, and Human Capital Accumul ation Model...................11 2.2.1 Health and Human Capital Accumulation Model.. ...........................15 2.2.2 Health and Labor Supply Model................ .......................................17 220.127.116.11 Case 1..................................... ............................................19 18.104.22.168 Case 2..................................... ............................................20 Chapter 3: Literature Review 3.1 Malaria and Labor Supply in Sub-Saharan Africa. ..........................................19 3.1.1 Malaria and Agricultural Labor Supply in SubSaharan Africa.......20 3.1.2 Malaria and Schooling in Sub-Saharan Africa. ................................23 3.2 Health and Labor Supply and schooling in Sub-Sa haran Africa 3.2.1 HIV/AIDS and Labor Supply in Sub-Saharan Afri ca.......................24 3.2.2 HIV/AIDS and Schooling in Sub-Saharan Africa. ...........................25 3.2.2 Hookworm and Schooling in Sub-Saharan Africa. ...........................25 3.3 Additional Reviews of Health, Labor Supply, and Schooling 3.3.1 Health and Labor Supply: Developed Country Li terature................26 3.3.2 Health and Schooling: Developed Country Liter ature......................28 Chapter 4: Research Design 4.1 Objectives and Hypotheses...................... ................................................... .....29 4.2 Description of Data 4.2.1 The Kagera Region of Tanzania................ .......................................31 4.2.2 Current State of Malaria and Other Illnesses in the Kagera Region.33 4.2.3 LSMS Tanzania Kagera....................... ...........................................34 4.2.4 Attrition in the LSMS Tanzania Kagera...... ...................................35
ii 4.2.5 Sample Construction: School Enrollment....... ..................................35 4.2.6 Sample Construction: Labor Supply............ .....................................36 4.2.7 Definition: Prevalence of Illness............ ...........................................36 4.2.8 Definition: Malaria Prevalence in the LSMS Ta nzania Kagera.....37 4.3 Description of Variables 4.3.1 Outcome Variables............................ ................................................37 4.3.2 Explanatory Variables........................ ...............................................38 4.4 Econometric Strategy........................... ................................................... .........40 4.4.1 OLS, FE, 2SLS, and LIML Estimation Methods... ...........................41 4.4.2 2SLS Instruments............................. .................................................43 4.5 Difference in Difference Estimation Method..... .............................................44 4.6 Attrition and Estimation Problems.............. ................................................... ..45 Chapter 5: Research Results 5.1 Episodes of Malaria in the Study............... ................................................... ...46 5.2 School Enrollment in Kagera.................... ................................................... ....47 5.3 Summary Statistics: Kagera School Sample....... .............................................48 5.4 Summary Statistics: Kagera Work Sample......... .............................................49 5.4.1 Farm Employment Characteristics.............. ......................................51 5.4.2. Self Employment Characteristics............. ........................................54 5.4.3 Wage Employment Characteristics.............. .....................................56 5.5 Other Kagera Sample Characteristics............ ..................................................5 8 5.6 2SLS Estimation Results Â– Kagera School Sample 5.6.1 First Stage Results.......................... ................................................... 59 5.6.2 Second Stage 2SLS, OLS, and FE Results....... ................................61 5.6 2SLS Estimation Results Â– Kagera Work Sample 5.6.3 First Stage Results.......................... ................................................... 65 5.6.4 Second Stage 2SLS, OLS, and FE Results....... ................................67 5.7 LIML Estimation................................ ................................................... ..........73 5.7.1 LIML ResultsÂ– Kagera School Sample........... .................................74 5.7.1 LIML ResultsÂ– Kagera Work Sample............. .................................76 5.8 Analyzing Improvements in the Disease Environme nt...................................78 5.9 Other Findings................................. ................................................... .............79 Chapter 6: Conclusion 6.1 Summary of Findings............................ ................................................... ........81 6.2 Policy Implications............................ ................................................... ...........83 6.3 Limitations.................................... ................................................... ................84 6.4 Future Research................................ ................................................... ............85 References......................................... ................................................... ..............................86 Appendices Appendix A: Tables................................. ................................................... ...........92 Appendix B: Figures................................ ................................................... ...........99
iii Appendix C: Theoretical Framework.................. ................................................102 Appendix D: Average Treatment Effects Estimation (A TET)............................117 About the Author................................... ................................................... .............End Page
iv List of Tables Table 1: The Kagera Region Population.............. ................................................... ...........33 Table 2a: Malaria Diagnosis in the LSMS Tanzania Kagera..........................................37 Table 2b: Malaria Cases in the LSMS Tanzania Kage ra................................................. 47 Table 3: School enrollment in the LSMS Tanzania K agera............................................48 Table 4: Summary Statistics: Kagera School Sample.. ................................................... ...50 Table 5a: Summary Statistics: Kagera Farm Employmen t Sample...................................53 Table 5b: Summary Statistics: Kagera Self Employmen t Sample.....................................55 Table 5c: Summary Statistics: Kagera Wage Employmen t Sample..................................57 Table 6a: 2SLS First Stage Results: Kagera School S ample.............................................6 1 Table 6b: 2SLS Second Stage Results: Kagera School Sample........................................63 Table 7a: 2SLS First Stage Results: Kagera Work Sam ple...............................................6 6 Table 7b: 2SLS Second Stage Results: Kagera Work Sa mple..........................................69 Table 7c: 2SLS Comprehensive Second Stage Results: Kagera Work Sample................72 Table 8b: LIML: Kagera School Sample............... ................................................... .........75 Table 9a: LIML: Kagera Work Sample................. ................................................... .........77 Table 10: Improvements in the Disease Environment.. ................................................... ..79 Table A1: Number of Reported Cases of HIV, TB, and Malaria......................................92 Table A2: Per Capita Health Expenditures for Sub-Sa haran Africa Countries.................93 Table A3: Annual Reported Malaria Cases in Sub-Saha ran Africa Countries..................94
v Table A4: Construction of Price Variables.......... ................................................... ...........96 Table A5: Summary Statistics: Total Kagera Work Sam ple.............................................97 Table A6: Comprehensive First Stage Results: Kagera Work Sample..............................98 Table C1: ATET Comparison of Means Â– Kagera School Sample.................................108 Table C2: ATET Results Â– Kagera School Sample...... ................................................... 109 Table C3: ATET Comparison of Means Â– Kagera Work Sa mple...................................112 Table C4: ATET Results Â– Kagera Work Sample........ ................................................... 115
vi List of Figures Figure 1: Human Capital and Household Utility Maxim ization........................................16 Figure 2: Map of Kagera............................ ................................................... .....................32 Figure B1: Malaria Map............................. ................................................... .....................99 Figure B2: RosenfieldÂ’s Conceptual Framework for Tr opical Illness.............................100 Figure B3: Malaria Episodes and Rainfall Amounts in the Kagera Sample....................100 Figure C1: ATET Overlap Assumption................. ................................................... .......100
vii Malaria, Labor Supply, and Schooling in Sub-Saharan Africa Taiwo Abimbola Abstract The purpose of this study is to evaluate the causa l effects of malaria and poor health in general on economic outcome in Sub-Sahara n Africa. This study uses panel data from the Living Standard Measurement Survey (L SMS) for Tanzania from 1991 to 2004. Three main hypotheses are tested. First, th e study evaluates the effect of malaria and other chronic illnesses on labor supply using t he number of hours worked per week as a measure of outcome. Second, it determines the impact of poor health on human capital accumulation by measuring the number of wee kly school hours lost to illness. The third objective deals with the question of whet her changes in preconditioning factors such as income levels and healthcare accessibility have improved the disease environment in Sub-Saharan Africa over time. The study uses several identification strategies i n the empirical estimation process. The first estimation strategy applies the standard Ordinary Least Squares (OLS) and Fixed Effects (FE) estimators to the schooling and labor supply models. In addition to OLS and FE, the preferred methods of estimating the causal effects of malaria on schooling and labor supply outcomes are Two Stage L east Squares (2SLS) and Limited Information Maximum Likelihood (LIML). Findings in this study suggest that malaria significantly increases school absenteeism. In par ticular, 2SLS and LIML estimates of
viii the number of school hours lost to malaria suggests that children sick with malaria are absent from school for approximately 24 hours a wee k. However, the results show the effect of malaria on work hours is inconclusive. F urthermore, difference in difference estimates of the disease environment show slight im provements in the disease environment resulting from changes in income levels The study finds no statistically significant improvements in the disease environment due to increases in the number of health facilities over time.
1 Chapter 1 Introduction 1.1 Motivation Recently, the debate over the influence of disease on social welfare in Africa has inspired a collection of empirical papers on the ec onomic impact of illness on the continent. Most of these papers have focused on three major di seases: malaria, HIV/AIDS, and tuberculosis (TB) (Martine Audibert 1986; El Tahir Mohamed Nur 1993; Joseph K. WangÂ’Ombe and Germano M. Mwabu 1993; Martha Ainswor th, et al. 2005; Kathleen Beegle 2003; Anne Case, et al. 2002; Harsha Thirumu rthy, et al. 2006). More than any region in Africa, the Sub-Saharan belt (the region of Africa below the Sahara desert) has suffered the biggest impact of malaria, HIV/AIDS, a nd TB. Table A1 compares the global impact of malaria, HIV/AIDS and TB to the Sub-Sahar a burden of disease. Although the region is home to only 11% of the worl d population, it is also home to 70% of all people infected with HIV (the Global Fun d (GF)). Together, HIV and TB form a deadly force, as one-third of those infected with HIV will develop TB (GF Â– Social, Economic and Development Impact). On the other han d, malaria is prevalent in 105 countries, 45 of which are in the continent of Afri ca (GF Â– Facts and Figures on Malaria). Figure B1 shows a map of the estimated incidence of clinical malaria episodes by region. Sub-Saharan Africa also has the highest mortality r ates from malaria. A disturbing 90% of
2 malaria deaths occur in Sub-Saharan Africa every ye ar (World Health Organization (WHO) Roll Back Malaria Report (RBM) 2003). The Sub-Sahara region of Africa is unique because o f its disproportionate burden of disease and for its extremely high level of poverty ; a combination that fuels further spread of disease. In a region where total health expendi ture runs as low as $4 per capita, it is no surprise that poverty persists and high levels of m ortality and morbidity continue to plague Sub-Saharan Africa (Table A2). 1.2 Malaria and Its Impact on Sub-Saharan Africa Malaria is caused by four sub-species of a protozoo n of the plasmodium genus namely, falciparum vivax malariae and ovale The f alciparum sub-species cause the greatest illness and death in Africa. Human malari a is only transmitted through mosquito bites by the females of the genus Anopheles Anopheles mosquitoes prefer to feed on humans or animals and are more likely to transmit t he malaria parasites from one person to another (Center for Disease Control (CDC)). A p eriod of incubation takes place following the infective bite, which lasts up to 30 days. Shorter periods of incubation are observed with plasmodium falciparum compared to oth er sub-species. Upon transmission, people often experience fever, c hills, and flu-like illness. Severities of infection with malaria parasites rang e from mild symptoms to more severe disease and even death. The most severe manifestat ion of malaria in Sub-Saharan Africa is cerebral malaria, which is most common in childr en and persons without previous immunity. People without immunity to malaria or pe ople from regions where transmission is absent (i.e. persons living in regi ons of high altitudes or those from
3 countries where the disease is virtually extinct) w ho contract the disease are also more likely to die from malaria. Malaria prevention is usually implemented by vector control or by preventive therapy. Vector control involves the use of indoor residual spraying of long-lasting insecticides such as Dichloro-Diphenyl-Trichloroeth ane (DDT) and long-lasting insecticide-treated nets (ITNs). DDT and ITNs are cost effective and efficient methods of malaria vector control. Successful malaria cont rol efforts in the United States involved the use of DDT and anti-malarial treatment in the 1 950s. By the 1970s, environmental concerns over the use of DDT led to its ban in the United States. Since then, DDT has been shown to be safe for humans and the environmen t and the WHO has once again recommended using DDT alongside ITNs and malaria dr ugs. According to the WHO, there are currently no alternatives to insecticidal vector control methods and the development of new methods will be an expensive lon g-term effort. Trials of ITNs in the 1980s and 1990s showed a 20% reduction in deaths of young children (RBM). Unfortunately, malaria control by ITNs is very expensive for most households in Sub-Saharan Africa at risk of the dis ease. The cost of ITNs is not the only barrier to their effective use. Most people in the Sub-Sahara region are unfamiliar with ITNs and are uninformed regarding re-treatment, whi ch entails replenishing the nets with insecticides in order to maintain their effectivene ss. Malaria is often very costly and deadly, yet it is a preventable disease. The World Malaria Report 2005 (WMR 2005) lists malaria as the leading reason for doctor visits in Sub-Saharan Africa. According to the World Bank (2 003), malaria costs Africa more than
4 $12 billion annually. The lack of government reso urces for malaria prevention and treatment puts most of the financial burden of the disease on households in Sub-Saharan Africa. A study by M. Ettling, et al. (1994) of th e financial burden of malaria on households estimated that over 25% of income in low -income households in Malawi is spent on malaria treatments. In the past decade, reported cases of malaria have escalated significantly in SubSaharan Africa. Historical epidemiological data on reported malaria cases show an upward trend of the disease in many countries in Su b-Saharan Africa (Table A3). For instance, in Ghana the number of reported malaria c ases shows an upward trend from 1993 to 2003 (Table A2). Also in Ghana, malaria ac counts for 44% of outpatient visits and 22% of under-5 mortality (WMR 2005). The recen t estimate on reported episodes of malaria for Tanzania is comparable to the number of cases reported over a decade ago (Table A3). Established reasons for the observed t rends include malaria resistance to Chloroquine and Antifolate combination drugs and bi ological factors such as the human immunodeficiency virus (HIV), which may also affect response to treatment (Peter Bloland 2001). The global pattern of malaria suggests that the dis ease is centered in regions with the highest levels of poverty. As noted by Jeffery Sachs and Pia Malaney (2002), Â“where malaria prospers the most, human societies have pro spered the leastÂ” (pp. 681). How exactly does malaria limit economic prosperity in S ub-Saharan Africa? In the context of economic welfare, malaria can limit the number of h ealthy workdays for the adult population. For younger members of the society, ma laria can impede the development of human capital thereby promoting continued poverty i n economically deprived areas.
5 The prolonged devastation of malaria has prompted a n increased international effort to fund treatment and prevention mechanisms in less developed countries plagued by the disease. For example, from 1998 to 2005 the United States Agency for International Development (USAID) increased funding for malaria f rom $22 million to $89 million. Even though the financial resources for fighting ma laria are increasing, approximately 3,000 children in Africa continue to die from malar ia each day (United Nations ChildrenÂ’s Fund). Most research on health and the economics of develo pment have focused on the impact of chronic illness on economic prosperity wi th HIV/AIDS as the main subject of analysis. However, unlike HIV/AIDS, the quantitati ve effect of malaria on economic welfare of African countries has been largely overl ooked. Only a few studies relate malaria to economic outcomes in Sub-Saharan Africa (Audibert 1986; Nur 1993; WangÂ’Ombe and Mwabu 1993; Leighton and Foster 1993) These studies focus on the effects of malaria on labor supply to agricultural production and they tend to find no relationship between malaria and economic outcomes. None of these studies empirically measure the quantitative economic impact of malaria on the accumulation of human capital. 1.3 Goals The primary objective of this study is to evaluate the causal effects of malaria and poor health in general on labor supply and the accu mulation of human capital in SubSaharan Africa. This study also evaluates the exte nt to which improvements in the socioeconomic environment have altered the disease envir onment in the region. To arrive at
6 the first goal, the implicit cost of poor health in Sub Sahara Africa is measured using time lost to labor supply and the loss in human capital formation as measures of outcome. In the analysis of improvements in the disease environ ment, the study explores how changes in income levels and increased accessibility to hea lthcare have influenced the incidence of malaria in the region. Measuring the time lost in labor supply is generall y favored over the income or output approach of evaluating the effect of health on labor supply (Janet Currie and Brigitte Madrian 1999). Empirical evidence regardi ng the effects of health on wages and hours of work in the United States suggests that he alth has a greater effect on hours worked than on wages. These effects are even stron ger for developing countries where historical literature documents that improvements i n health are linked to bigger changes in standards of living over time (John Strauss and Duncan Thomas 1998). Considering this evidence, this study will not evaluate loss in income or output due to illness; instead, it focuses on measuring the number of work hours lo st due to illness. The causal effects of malaria on human capital accu mulation will be evaluated by examining child education. In this analysis, the a cquisition of human capital is treated as a household choice. The household decision to inve st in human capital depends on its future expectations about the returns to education. Therefore, malaria reduces the return to education, possibly reducing the incentive to ed ucate a child in Sub-Saharan Africa. Assessing the impact of illness on education is par ticularly important given the high prevalence of disease in young children in Afr ica. Although adults also suffer severely from malaria, infections in children are f ar more severe. Malaria can easily lead to seizures and coma in young children (GF). Repea ted episodes of infection in children
7 have also been linked to a reduction in social inte raction and educational opportunities, thereby contributing to poor development. In addit ion, children who have had malaria infections affecting the brain (cerebral malaria) s uffer from an increased likelihood of learning impairments and brain damage, including ep ilepsy (Sean Murphy and Joel Breman 2001). To evaluate the extent to which the disease environ ment has changed over time, this study employs the approach suggested by P.L. R osenfield (1984). The continuing increase in malaria rates demonstrates the importan ce of evaluating the changes in the disease environment in Sub-Saharan Africa. As ment ioned earlier, Table A3 shows an increase in the number of reported malaria cases fo r several Sub-Sahara countries including Tanzania. This new trend calls for an as sessment of how previous efforts to end the disease have benefited the population. To assess these benefits, the changes in income levels and changes in socio economic factors are related to the reported cases of malaria over time. The data used in this study is derived from the Liv ing Standards Measurement Study (LSMS) for the Kagera region of Tanzania. Th e panel design of the LSMS data controls for some of the limitations that hampered previous studies in the area. The data used in previous studies by Audibert (1986) and Lei ghton and Foster (1993) lacked sufficient information on the episodes of malaria a nd are based on unrealistic recollection periods1. Fortunately, LSMS Tanzania-Kagera contains five waves of health data capturing peak and off-peak prevalence of malaria. 1 The recall period for Audibert (1986) is 3 to 9 mo nths while the recall period of Leighton and Foster (1993) is as high as one year.
8 Previous studies on malaria and labor market outcom es (Audibert 1986; Nur 1993; WangÂ’Ombe and Mwabu 1993) limit their analysi s to farm employment. Audibert (1986) and WangÂ’Ombe and Mwabu (1993) rely on an ou tput measure of productivity. On the other hand, Nur (1993) examined the actual t ime lost in labor supply due to malaria. These studies all found no relationship b etween malaria and labor market outcomes. The output approach used in the existing literature attempts to capture reduced productivity by comparing output of healthy persons to diseased persons. This approach fails when an omitted variable bias arises. Specif ically, such bias arises when a study fails to capture the likelihood that sick workers m ay work longer hours with lower productivity. The lack of a significant relationship between mala ria and labor supply in the existing literature is also driven by the fact that labor substitution is far more likely in farm employment (especially for self-employed farme rs) than in other sectors of employment. It is reasonable to expect the inciden ce of malaria to vary by type of employment. Farmers in general tend to work outdoo rs, thus increasing the likelihood of contact with mosquitoes. Therefore, the likelihood of labor substitution in farm employment necessitates analyzing the causal link b etween malaria and non-farm employment where labor substitution is less likely. Unlike earlier studies, this paper will evaluate the causal effects of malaria on farm and non-farm labor supply in Sub-Saharan Africa. In the analysis of health and labor supply, labor s upply is measured as the number of hours worked per worker. The Â‘hours workedÂ’ app roach employed here controls for the above mentioned limitations in the output and incom e methods. The remaining sections of
9 the study are arranged in the following order. Cha pter two details the theoretical framework supporting the analysis. Chapter three r eviews the existing literature on the effects of malaria on labor supply and education. Chapter four explains the research design. Chapter five presents the main findings of the study and chapter six presents the conclusion.
10 Chapter 2 Theoretical Framework The first section of this chapter reviews the the oretical background of the relationship between human capital, labor supply an d health. The second section of this chapter introduces a two-period household utility m aximization model that links health directly to human capital and labor supply. Mathem atical solutions to the two-period household utility maximization model are presented in Appendix C. 2.1 Theoretical Framework of Health, Human Capital and Labor Supply There have been many empirical studies in the past three decades on health and economic welfare. At first, most of the research f ocused on documenting the health status of households in developed economies as a pa rt of a larger inventory of human capital components influencing labor supply and ear nings. More recently, this field of research has branched into evaluating the implicati ons of health on the economic welfare of households in developing countries, particularly in Africa. Given this recent interest in the economics of health in developing countries, a growing concern has been whether the traditional theoretical framework used in developed country literature applies to developing economies. A well-developed theory that explains the role of h ealth in labor productivity for developing countries, is presented in Strauss and T homas (1998). According to Strauss
11 and Thomas, there are several reasons why the relat ionship between health and labor market outcomes in developing economies should be o f special interest. First is the existing tradition of theoretical models of nutriti on-based efficiency wages in the development literature that established the link be tween nutrition of workers and productivity. A link that is highly relevant in th e developing countries context where the marginal productivity of health is likely to be hig her relative to developed economies. This coupled with the structure of employment in lo wer income economies, which often relies more on strength and endurance (i.e. good he alth) suggests that labor market consequences of poor health are more severe for lea st developed countries than for developed economies. To describe the theoretical link between health and labor supply, Strauss and Thomas begin with the traditional single-person hou sehold utility maximization model and household production function as presented in G ary Becker (1965) and the seminal work of Michael Grossman (1972). The household pro duction function presented in Strauss and Thomas assumes that health is increasin g in all inputs except for labor supply, which decreases the stock of health. Healt h impacts labor supply by influencing the decision to work, choice of work and hours of w ork. If health affects productivity, both the market wage and the marginal rate of subst itution (MRS) between consumption goods and leisure also depend on health outcomes. A structural labor supply function conditional on h ealth outcomes and earnings is also provided in Strauss and Thomas. Health affect s labor supply by influencing offered wages with resulting income and substitution effect s, and by affecting the MRS between consumption goods and leisure; a direct result of t he assumption that health directly
12 appears in the utility function. In addition to th e single-person household utility model, the paper mentions applications of the described in dividual labor supply behaviors to household choices. However, an extensive applicati on of the model at the household level was not presented in the paper. The existing theoretical background on health and s chooling only explains the causal link running from education to health. A we ll developed literature on the theory of human capital accumulation and health can be found in Becker (1964) and Grossman (1972). Other papers including Strauss and Thomas have noted that the individual or household utility function can be conditioned on sc hooling without providing the theoretical link between health and schooling. The purpose of this study is not to develop new the oretical insights, but to provide a conceptual framework to guide the empirical analy sis. This chapter borrows heavily from Becker (1964), Grossman (1972), Rosenfield (19 84), Paul Schultz and Aysit Tansel (1993), and Strauss and Thomas (1998) to develop th e theoretical framework for analyzing the causal effects of health on labor sup ply and the accumulation of human capital. This chapter introduces a two-period household mode l for evaluating the causal link between health and labor supply and health and human capital. The two-period model captures the household decision to invest in human capital and supply labor given its health production function. In period one, the household can choose to invest in human capital through educating its children and/or by sending them into the labor market. However, the entire household is expected to supply labor in period two after all
13 gains from investing in human capital have been rea lized. The labor supply model is formulated in period two when the entire household supplies labor. 2.2 A Model of Health, Education, and Labor Supply The study begins by proposing a two-period househol d utility function, which assumes that the household derives indirect utility from investments in health. In this framework, utility depends on the amount of non-hea lth goods, health goods, and labor supply per period. In period one, health is catego rized by two components Â– overall health of the household and health of the child wit hin the household. That is, it is assumed the household not only derives utility from its own personal health but also from the health of its children. Thus, our two-period u tility function can be expressed as2: ) , ( ) , (2 2 2 2 1 1 1 1 1L H C U L H H C U UH H K K H Hb+ =(3.2.1) Where ) (1 1 1 K K HL L L L-= and ) (2 2 HL L L = 1 K HL-and 1 KL are household less child labor supply in period one and child labor supply in period one, respectively. 2 HL is overall household labor supply in period two when all members of the household are assumed to pa rticipate in the labor market. As seen in this chapter, if the household makes the de cision to invest in human capital by educating its children in period one it is rewarded in next period in form of accumulated human capital. In period two, when all returns fro m the investment in human capital are realized, the entire household supplies the amount of labor denoted by2 HL 2 Appendix C provides a definition of variables in t he theoretical framework.
14 The householdÂ’s utility in period one depends on th e household consumption of non-health goods, 1 HC 1 K HHis the household health less the health of the chi ld in period one, 1KH is the health of the child in period one, and 1L is labor supply in period one. Notice that utility in period two only depend s on the consumption of health and non-health goods and labor supply for the entire ho usehold denoted by2 HC ,2 HH and 2 HL since all gains from investing in health must be re alized by the second period. The utility function differs from traditional two-period utilit y formulations only by the added health components. It is expected that1 HC ,1 K HH-, 1 KH 2 HC and 2 HH will have a positive impact on utility, while 1Land 2Lwill negatively impact utility. Unlike other consumption goods C, health as describ ed in equation (3.2.1) cannot be purchased in the market and instead has to be pr oduced. In addition, the individual also derives indirect utility from good health. A person who is in good health spends less time devoted to illness, which translates into more time to pursue activities that enhance his/her utility. With this in mind, the process th rough which the household produces health for itself,1 K HH-, for a child, 1KH and as a whole, 2 HH is represented by3: ) , ; (1 1 1 1 1 1 K H K H K H K H K H K He D A L HI H H-=m (3.2.2) ) , ; (1 1 1 1 1 1 K K K K K Ke D A L HI H Hm= (3.2.3) ) , , ; (2 2 1 1 2 2 2 2 H H K K H H H H He D A H H L HI H Hm-= (3.2.4) Here, health given by (3.2.2), (3.2.3), and, (3.2.4 ), is produced using a vector of health inputs, HI, and labor supply, L, which are controll ed by the individual and other 3 The household consumption of non-health goods can also be defined in terms of non-health consumption inputs, labor supply, and prices: ) , ; , (, H H H H H He D A P L CI C Cm=.
15 uncontrollable factors such as previous health hist ory. Health is increasing in inputs except labor supply, which consumes energy and in t his way taxes health. Although, schooling does not directly enter the utility funct ion specified in (3.2.1) it is an input into the health production function and contained in HI The levels of health produced in (3.2.2), (3.2.3), and, (3.2.4), are also likely to vary with socio-demographic characteristics, A, such as age and gender and the disease environment, D (Figure B2). There are two sources of unobserved heterogeneity t hat influence the health production formulation in (3.2.2), (3.2.3), and, (3 .2.4). The first deals with factors that are known to the individual but unobserved by the r esearcher such as the inherent healthiness of the individual, m The other unobservable is that which is unknown to the individual and the researcher, e, which includes me asurement error. The issue of unobserved heterogeneity is discussed in detail in chapter four. Now, suppose that the household can either make inv estment in the education of the child (1 KS ) in period one, and reap the fruit of this investm ent in period two in the form of human capital (KS ~ ) or the household can choose not to make this inve stment and send the child into the labor market (part-time or full-time) in period one. The level of accumulated human capital in period two can be defi ned as: ) , ; ( ~ ~ 1 1KS K K K K Ke SI A H S S Sm= (3.2.5) The level of education attained in period two is a function of previous years of schooling,1KS health status in period one, 1 KH and uncontrollable factors such as school
16 infrastructure and the quality of teachers, SI4. Schooling in period one (1KS ) is expected to positively influence the level of human capital in the future. Health status in the current period can either increase or decrease the value of KS ~ in the future. If the household in (3.2.1) earns labor income, w, and own s assets or non-labor income V, utility in (3.2.1) is bounded by the resource const raint: ) ~ ( ) ( ) ( ) ( ) (2 2 1 1 1 1 2 1 1 2 1 K H H K K K H K H K s H K K H HI H H cS L w L w V L w S P HI HI HI P C C P + + + = + + + + +(3.2.6) Equation (3.2.6) states that expenditures on consum ption goods, health inputs, and education cannot exceed total income. The right-ha nd side of this two-period resource constraint, is a result of the earlier assumption t hat the household can choose to fully invest in education, in which case the child would supply zero labor and 0 ) (1 1=K KL w or decide to send the child into the labor market in p eriod one part-time or full-time (0 ) (1 1K KL w ). If the household chooses to educated the child in period one and succeeds, then 2 Hw is expected to be higher since 2 Hw is a function of the human capital acquired, KS ~ Unlike the traditional labor supply models for deve loped countries, it is assumed that the household is better able to earn labor inc ome when it is healthy. Therefore, the earnings of every member of the household are a fun ction of health (, ,1 1 K K HH H-and2 HH ). With that in mind, the earnings functions for e ach member can be written as: ) , , ; (1 1K Hw K H K H K H K He IN E A L H w w--=a (3.2.7) 4 Note that (3.2.5) can also depend on the health st atus of the parent(s) within the household, 1 K HH-.
17 ) , , ; , (1 1 1 1 1 1Kw K K k K Ke IN E A S L H w wa= (3.2.8) ) , , ; ~ , (2 2 2 2 2Kw K H H H He IN E A S L H w wa= (3.2.9) The earnings function is also influenced by uncontr ollable factors such as local community infrastructure,IN, socio-demographic characteristics, A and education, E Notice that earnings in period two also depend on t he gains from investing in human capital in period one, KS ~ The term KS ~ is expected to positively impact overall earnings for the household given that it chose to invest in education in the previous period. Earnings will also be affected by unobservable fact ors, a such as ability, school quality, and random fluctuations in wages and measurement er ror captured by ew. The full extent of health on utility can be derive d by substituting (3.2.2), (3.2.3), and (3.2.4) into the utility function in (3.2.1). The utility function in (3.2.1) can then be rewritten as: ) , ; ), ( ( ) , ; ), ( ), ( (2 2 2 2 2 2 1 1 1 1 1 1 1 1 1x b xE A L L HI H C U E A L L HI H L HI H C U UH H H H H K K K K H K H K H H+ =(3.2.1 Â¢ ) The household can now choose1 HC ,1 K HHI-,1 KHI 1 K HL-, 1 KL 1 KS 2 HC 2 HHI and2 HL subject to the resource constraint in (3.2.6) (see Appendix C). 2.2.1 Utility Maximization: The Health and Human Ca pital Accumulation Model To determine the optimal investment in human capita l for the household represented in (3.2.1 Â¢ ), the first order necessary condition (FONC) is so lved for1KS This yields the demand equation for education in period one. ) , , ; (1 *KS K S K Ke SI A P H S Sm= (3.3.1)
18 The FONC for 1KS indicates that the household must weigh the price of schooling in the current period against the returns to schooling in the next period (i.e. the opportunity cost of schooling) in its decision to invest in human ca pital. Assuming that the second order sufficient condition (SOSC) for 1KS is satisfied, the demand curve for schooling in pe riod one is illustrated below, assuming a constant price for education in that period. Figure C1 illustrates that the household maximizes utility in period one by investing in KS amount of human capital. However, the optimal inve stment is determined by the childÂ’s health in period one, and other factors, which are out of the control of the household (as indicated by (3.3.1)). Good health/poor health in period one is expected to increase/decrease schooling (i.e. th rough the number of healthy school days and academic performance) and shift the demand curve rightwards/leftwards. Notice that health as a choice variable in the dema nd for schooling poses an estimation problem because 1 KH is endogenous in the human capital accumulation model. 1 KH is determined by immeasurable factors such as the inherent healthiness of the individual therefore; estimating (3.3.1) with o rdinary least squares (OLS) will produce biased estimates of the causal effect of he alth on schooling. The likelihood of endogeneity of 1 KH as well as the possibility of measurement error us ing instrumental variables can be mitigated. This can be done by us ing equation (3.2.3) as the first stage regression in our econometric analysis of the causa l effects of health on schooling, where PHI and D in (3.2.3) are valid instruments for1 KH A detailed description of these instruments is presented in chapter four and five.
19 2.2.2 Utility Maximization: The Health and Labor Su pply Model As mentioned earlier, the causal effect of health o n labor supply is to be determined in period two when the entire household supplies labor. In period two, all gains from investments in human capital are realize d and the household supplies an amount of labor equal to2 HL In this period, the study evaluates the househol dÂ’s labor supply decision and solves for the optimal amount o f labor supply. Two different cases are considered in solving for the optimal amount of2 HL In the first case, health is treated as an endogenous variable while in the second case health is considered to be exogenous. 22.214.171.124 Case 1: Health is Endogenous A maximization of (3.2.1 Â¢ ) with respect to C, HI, and L subject to (3.2.6) a nd (3.2.9) will produce the following reduced form equ ations for C, HI, and L (Appendix C). ) , , , , (a mIN D E A V P P C CHI C= (3.4.1) ) , , , , (a mIN D E A V P P HI HIHI C= (3.4.2) ) , , , , (*a mIN D E A V P P L LHI C H= (3.4.3) Although, (3.4.1) and (3.4.2) are informative, the problem with treating health as a choice variable is that 2 HH disappears from the reduced form equation for L in (3.4.3). The reduced form equation for labor supply, L in case o ne is only an assessment of the total effect of prices and the disease environment on lab or supply, which is not very helpful in understanding the causal effect of health on labor supply. To determine how health affects labor supply, a structural labor supply fun ction that is conditional on health is
20 needed. To derive this important link between heal th and labor supply, consider the following case. 126.96.36.199 Case 2: Health is Exogenous In this case, the individualÂ’s utility maximization problem remains the same except that 2 HH is now a conditioning variable. Case two yields t he structural equation for labor supply listed below which now depends on health, 2 HH (Appendix C). ) , , , ; (2 2 *a mIN E A V w P H L LH C H H= (3.4.4) Unlike (3.4.3), the structural equation for labor s upply (3.4.4) provides the opportunity to separately identify the causal effects of health on labor supply. First, once the model conditions on2 HH the prices of health care inputs, HIP and the disease environment, D, do not affect labor supply. Second, equation (3.2. 4) can be used as the first stage regression in our econometric analysis of labor sup ply, where HIP and D are valid instruments for health (2HH ). A detailed description of these instruments is presented in chapter four, section 4.4.2.
21 Chapter 3 Review of Relevant Literature Many empirical studies of the effects of health on economic welfare have been published over the past 3 decades. Initially, most research in this field sought to document health status of households in developed e conomies as a part of a larger inventory of human capital components influencing w ages, labor supply and earnings. This literature has branched in recent times into e valuating the implications of health on the economic welfare of households in developing co untries, particularly in Africa. The following review documents evidence on health a nd economic welfare based on developed and developing country experience. Th e first section analyzes relevant literature on the implications of disease on labor supply and human capital accumulation in Africa. The last section of this review is a br ief assessment of the empirical evidence of health on economic welfare in the developed coun try context. 3.1 Malaria, Labor Supply, and Schooling in Sub-Sah aran Africa This section reviews the existing literature on the role of malaria in the supply of labor and child education in Sub-Saharan Africa. V ery few empirical studies have analyzed the role of malaria in labor supply and hu man capital accumulation in Africa. Studies by Audibert (1986) and WangÂ’Ombe and Mwabu (1993) are the only known empirical studies on malaria and labor supply in Su b-Saharan Africa. Another study by
22 Nur (1993) also analyzes the role of malaria in the supply of labor in Sub-Saharan Africa but only through a descriptive analysis of the data Compared to the literature on the role of malaria i n labor supply in Sub-Saharan Africa, less work has been done on malaria and scho oling in the region. A study by Leighton and Foster (1993) attempted to evaluate th e economic impact of malaria on child schooling in Sub-Saharan Africa using a case study method. However, there is no real empirical study evaluating the causal effects of malaria on child schooling in SubSaharan Africa. 3.1.1 Malaria and Agricultural Labor Supply in SubSaharan Africa Most studies of health and labor supply in Sub-Saha ran Africa have focused predominantly on agricultural employment. Because most farmers (especially selfemployed farmers) enjoy the advantage of labor subs titution, the true cost of ill health on labor supply is often minimized. Labor substitutio n5 minimizes economic loss from illness by insuring workers against future losses i n income, which arises from inability to work due to illness. Therefore, when analyzing the impact of disease on the agricultural sector the observed cost of ill health may be lower than the true cost. More importantly, the prevalence of diseases such as malaria, HIV/AID S, and tuberculosis in Africa makes analyzing the impact of health on the labor supply patterns of working adults equally important for all sectors of employment, not just f arming. Studies evaluating the link between health and labo r supply have focused on patterns of agricultural labor supply in the event of a prime-age adult illness and/or death. 5Labor substitution in this context refers to the pr ocess whereby the hours lost in labor supply are compensated for by family members
23 Studies by Audibert (1986), Nur (1993), and WangÂ’Om be and Mwabu (1993) all address the issue of malaria and labor supply in terms of f arm occupation and generate similar results. Audibert (1986) was the first major longitudinal st udy to measure the effect of malaria on agricultural productivity in Sub-Saharan Africa. Audibert estimated the impact of a variety of illnesses, including malaria on agricultural non-wage peasant production using data on rice farmers in Cameroon. The study measured loss in productivity using a single production function wit h factors such as land properties and quality of labor as exogenous regressors affecting the production of rice. In contrast to Audibert, which uses an output approach, the presen t study evaluates the marginal impact of health status on labor using the number of hours of labor supplied as a measure of outcome. In addition, our goal is not to measure t he impact of health status on the marginal loss in output but instead to measure the impact of health status on the quantity of labor supplied. AudibertÂ’s study found malaria to be insignificant in explaining variations in rice output. Nur (1993) studied the effects of malaria on labor inputs into agriculture by examining the extent of family labor substitution i n the event of a male adult being incapacitated by malaria. In addition to family la bor substitution, he also examined the traditional system of mutual aid in which, other fa rmers provided assistance on a reciprocal basis when an adult male was ill. NurÂ’s sample was based on 250 randomly selected tenant farmers in Gezira, Sudan. Like Audibert, NurÂ’s main findings on the number of hours of labor supply lost due to illness were counter-intuitive. The results indicate that farmers in Sudan were
24 more likely to be incapacitated by malaria in the d ry season (low prevalence season), but less likely to be incapacitated in the peak prevale nce season. However, the study stated that high prevalence detected in the sample coincid es with the peak season (the land preparation and harvest seasons). The results in t he study revealed that the sum of reported hours lost in agriculture due to malaria i n the land preparation and harvest seasons exceed those lost in the low prevalence mon ths by almost 2200 hours. In terms of family labor substitution and mutual ai d, Nur determined that other family members compensated for all the hours lost i n agricultural labor supply due to malaria. The family members in the sample tend to contribute more than the actual number of hours lost. This is possibly because the time contributed by labor substitutes was not as productive as each hour lost. According to the study, women in the family are far more likely to be labor substitutes than childr en and non-family labor substitutes (mutual aids). Unlike Audibert, Nur failed to control for seasonal variation in labor use. Since the peak malaria season coincides with the harvest season, it is reasonable to expect that these months generate fewer hours lost due to illne ss than otherwise reported in the months of low prevalence (dry season). One reason is because labor substitution is less likely in the dry season than in the rainy or harve st season. Also, the opportunity cost of being sick is higher in the harvest season than it is in non-harvest seasons. These factors are not accounted for by NurÂ’s study since the resu lts are drawn from a simple review of frequency tables that are limited in interpretabili ty and lend no econometric legitimacy to the study.
25 WangÂ’Ombe and Mwabu (1993) also examined the direct effects of malaria on labor productivity. WangÂ’Ombe and Mwabu use a cros s-section of 302 households in Western Kenya. The study estimated a production fu nction for cassava farming to evaluate the extent that malaria affects productivi ty and household income, similar to equation (3.4.4). Because the production equation estimated in the study only controlled for family size and the number of malaria cases per household as exogenous regressors, the study found no statistically significant direct effects of malaria on productivity. The authors cite possible model misspecification for th e lack of significant results. The studies reviewed above all fail to control for factors that could have prolonged the duration of sickness and are not dire ctly related to malaria. For instance, the pre-existence of other chronic conditions such as HIV/AIDS and Sickle Cell Anemia was not accounted for. Failure to control for the existence of pre-existing conditions can create endogeneity in the estimation process. 3.1.2 Malaria and Schooling in Africa In the context of health and child schooling in Sub -Saharan Africa, much effort has been geared towards evaluating the impact of ch ronic illness and adult mortality and morbidity on schooling. One exception to the exist ing literature is the study by Leighton and Foster (1993) that attempted to measure the eco nomic impact of malaria on schoolchildren in Kenya and Nigeria. This study wa s based solely on focus group interviews and a spreadsheet method of estimation, which failed to control for microlevel factors that influence the decision to attend school. The study determined that between 4% and 40% percent of all school absences a nnually can be attributed to malaria
26 in Kenya and Nigeria, respectively. The findings i n the study are questionable for several reasons. First, the study is based on an unusually long recall period of one year. Second, it used a very small focus group sample. Third, as mentioned earlier it is based on a questionable estimation strategy. 3.2. Health and Labor Supply and Schooling in Sub-S aharan Africa This section reviews the existing literature on il lness and labor supply and schooling in Sub-Saharan Africa. Unlike the publis hed literature on malaria in the region, more studies have been conducted on the eff ects of chronic illness in Sub-Saharan Africa. These studies have mostly focused on the e ffect of HIV/AIDS on the Sub-Sahara population. This part of the review discusses stud ies on health and economic outcomes in Sub-Saharan Africa. 3.2.1 HIV/AIDS and Labor Supply in Sub-Saharan Afri ca Studies in the area of health and the labor supply in Sub-Saharan Africa have focused on the impact of adult morbidity and/or dea th on household decision to work. Beegle (2003) and Thirumurthy, et al. (2006) studie d the impact of HIV/AIDS on labor supply in Tanzania and Kenya, respectively. Beegle evaluated the impact of the HIV/AIDS epidemic on Sub-Saharan householdÂ’s farm l abor supply before and after the death of a prime-age adult. This study used data f rom the World Bank and the University of Dar Es Salaam for the Kagera region of Tanzania and studied over 800 households between 1991 and 1994. These data are now publishe d as part of the Living Standard Measurement Survey (LSMS).
27 Like other studies, Beegle ignored seasonal variati on in agricultural labor use and found insignificant changes in labor supply for ind ividuals in households that experienced a prime-age adult death from HIV/AIDS. Another study by Thirumurthy, et al. analyzed how antiretroviral therapy (ARV) influ ences the labor supply of treated patients and their family members. Results from Th irumurthy, et al. indicate that individuals just beginning ARV treatment are five t imes more likely to increase participation and four times more likely to increas e hours worked than those who have not undergone treatment. Like Beegle, Thirumurthy, et al. findings suggest that neither girls nor older boys experienced any significant sp illover effects in terms of increased labor participation rates when another household me mber suffered from or died due to HIV/AIDS. 3.2.2 HIV/AIDS and Schooling in Sub-Saharan Africa Martha Ainsworth, et al. (2005) and Anne Case, et a l. (2002) analyzed the impact of orphan status and the death of a prime-aged adul t in the household on school hours for children in Sub-Saharan Africa. Ainsworth, et al. estimated the number of school hours lost per week for orphaned girls to be between five and 13 hours conditional on school attendance. In addition, the study found the impac t of adult mortality to be greater on school attendance for younger children than for old er children. In particular, children in poor households with recent adult death had a 10 pe rcentage point lower attendance rate than poor children in households without an adult d eath. Case, et al. also examined the effect of orphan status on school enrollment and fo und results similar to Ainsworth, et al. for school enrollment rates of orphans in poorer ho useholds.
28 3.2.3 Hookworm and Schooling in Sub-Saharan Africa Edward Miguel and Michael Kremer (2004) conducted a school-based mass randomized treatment experiment with deworming drug s in Kenya. The purpose of the field experiment was to identify the impacts of dew orming on education and health while controlling for the likelihood of treatment externa lities. In order to determine the extent of externalities resulting from treatment, seventyfive primary schools participating in the experiment were phased into deworming treatment in a randomized order. Miguel and Kremer found significant gains in school attendance due to treatment for both treatment and control groups. Findings in the study suggest that the program reduced overall school absenteeism by seven percent age points, a one-quarter reduction in total school absenteeism. In terms of externali ties, the study found that deworming creates positive externalities both within and acro ss schools. In particular, implementing treatment to a fraction of pupils in a school led t o a 6.2 percentage points gain in school participation for children in the control group at the same school. 3.3 A Brief Review of Developed Country literature This section provides a brief review of empirical evidence on health, labor supply and human capital accumulation in the United States 3.3.1 Health and Labor Supply: A Brief Review of De veloped Country literature Early works by Michael Grossman and L. Benham (1974 ), Ann Bartel and Paul Taubman (1979), and H.S. Luft (1975) emphasize the interrelationship between health
29 and labor force participation. They all conclude t hat poor health in preceding periods reduce labor supply in following periods. Grossman and Benham in particular, emphasize the importance of a model that views wage rates, hours of work and health as a set of interrelated household decisions. In the Grossman and Benham framework, work-time and wages are recognized as interdependent. Using this framework, Robert Havem an, et al. (1994) adopted a 3equation simultaneous equation model to capture the interrelationship between labor force participation and health. The study utilized annual data on white males with a history of significant labor force attachment from the Michigan Panel of Income Dynamics (PSID). The findings by Haveman, et al. s upport that of Grossman and Benham in concluding that prior health limitations have a significant negative effect on work time and wages. Another study examining the impact of health status on work hours is Susan Ettner, et al. (1997). This study examined the imp act of psychiatric disorders on employment and conditional work hours and income us ing the National Comorbidity Survey (NCS). Ettner, et al. found small and somet imes insignificant effects of health on work hours depending on whether two-stage instrumen tal variables (IV) estimation or ordinary least squares (OLS) was used. The slight reduction in conditional work hours was only observed for men. Ronald Kessler and Richard Frank (1997) also examin ed the relationship between psychiatric disorders and work impairment by using the same data as Ettner, et al. Kessler and Frank examined eight job conditions usi ng cluster analysis. Unlike Ettner, et al., Kessler and Frank found significant reduction in workdays. This result also varies
30 significantly depending on the constellation of dis orders and the occupation in which the worker is employed. Other works on the economic ef fects of poor health include Thomas Chirikos and Gilbert Nestel (1985), Jean Mitchell a nd Richard Burkhauser (1990) and Matthew Kahn (1998). The study by Chirikos and Nestel was based on a tob it regression of retrospective history of self-reported health appraisals on hours of work and wages over a ten-year period. The authors found enough evidence to suppo rt the significance of prior health status on labor hours and wages, but determined the strongest effect to be for hours worked, not wages. Mitchell and Burkhauser estimat ed the extent to which arthritis limits the ability of workers to fully function in the workplace. The study is also based on a simultaneous tobit model similar to Chirikos a nd Nestel. Mitchell and Burkhauser found that a history of ill health (in this case, A rthritis) has a greater effect on hours than on wages. Kahn (1998) investigated the effect of diabetic sta tus on health on a more general scale than Chirikos and Nestel and Mitchell and Bur khauser. The findings of this study suggest that the effects of diabetes duration on em ployment rates have lessened over time. The study determined that even though people with diabetes still earn less than their non-diabetic counterparts, diabetic income an d labor force participation rates are far higher than in earlier years. Overall, improvement s in labor market participation for diabetics can be attributed to changes in the Socia l Security Disability Insurance System and technological advances that improve the quality of life for diabetics.
31 3.3.2 Health and Schooling: A Brief Review of Devel oped Country literature Barbara Shapiro, et al. (1995) documents the effec ts of chronic illness on school attendance of children in the United States. The s tudy collected data on children and adolescents with sickle cell disease in order to ga in information on the natural history of pain and its impact on school attendance and sleep. The data was based on a home diary experiment from a self-report system used for resea rch on sleep and circadian rhythms. Shapiro, et al. found that sickle cell patients rep orted missing 41% of school days on which they reported pain versus an average of 9% on days without pain. On average, 2.7 consecutive school days over a period of 10 months were associated with clinic visits or other medical problems related to sickle cell disea se.
32 Chapter 4 Research Design This chapter describes the data and methods used i n this study. The first section explains the objectives and hypotheses underlying t his dissertation. The second section details the source of data used in the analysis. T he third section describes the variables used in the estimation of the models presented in t he study. The fourth and fifth sections focus on the methodologies and findings of the stud y. 4.1 Objectives and Hypotheses The objective of the study is to evaluate the caus al effects of malaria on labor supply and schooling outcomes. To arrive at this g oal, the results of the two-period household utility model developed in chapter two ar e applied to a sample of household in the Sub-Saharan Africa region. The first goal is t o estimate the equations for the demand for education (3.3.1) and labor supply (3.4.4). Assuming a linear specification for both schooling and labor supply, equations (3.3.1) and (3.4.4) can be written as: e SI B A B P H SS + + + + + + = m b b b 4 3 2 1 0 (4.1) a m b b b b b b b b + + + + + + + + + = IN E A w V P H LC 8 6 5 4 3 2 1 0 (4.2) In equation (4.1), the demand for schooling (S) is conditioned on health status ( H ), price of schooling (SP), socio-demographic characteristics ( A ), and school
33 infrastructure (SI). Health status, which is characterized in this s tudy as the incident of illness is defined as the presence of malaria. Hol ding constant all other factors that may influence the demand for human capital, malaria is expected to decrease the demand for schooling; therefore, a negative sign is expected o n1 b Improvements in school amenities and the quality of teachers denoted by SI are other factors influencing schooling outcomes in equation (4.1). These variab les are expected to positively impact the demand for schooling. On the other hand, incre ases in the level of price indicators such as the price of education,SP, will decrease the demand for school. In equation (4.2), labor supply is conditioned on h ealthH (), price of consumption goods (CP ), wages, non-labor income (V), socio-demographic characteristics ( A ), education ( E ), and community infrastructure (IN). Holding all other factors constant in equation (4.2), the partial effect of health on lab or supply (1 b ) is expected to yield a negative sign for those who are sick with malaria. On the contrary, increases in the price of non-health consumption goods, wages, the level o f education and community infrastructure are expected to increase the labor s upply in equation (4.2). Furthermore, an increase in non-labor income (V), is more likely to reduce labor supply. As indic ated in chapter two, current health status poses an estimat ion problem in the labor supply and schooling models. Current health status ( H ) in both models is determined by immeasurable factors such as the inherent healthine ss of the individual. The second goal of this study is to determine the e xtent to which improvements in preconditioning factors of the socio-economic and p hysical environment have impacted the incidence of malaria in Sub-Saharan Africa. As mentioned earlier in the chapter two,
34 the framework suggested by Rosenfield (1984) highli ghts the relevance of socio economic factors that act as baseline inputs into t he health production function (Figure B2). According to Rosenfield (1984), these factors predetermine health outcomes in tropical environments. In this study, RosenfieldÂ’s framework is used to evaluate the effect of changes in preconditioning factors of the socio-economic environment on health status over time. The goal is to determine how the disease environment has been altered by changes in baseline health production variables such as income and health care availability. The unit of analysis for both the labor supply and schooling models is the household. In the analysis of the disease environm ent, the unit of analysis is at the household and community levels. Income measures fo r the analysis of the disease environment are aggregated at the household level w hile data on health facilities are derived at the community level. 4.2 Description of Data The data used in this study are part of the Living Standard Measurement Surveys (LSMS). The World Bank began the surveys in the 19 80s for the purpose of developing new methods of monitoring levels of living, identif ying the effects of government policies, and advancing communications between thos e who collect and use data as well as policymakers around the world. To date, the LSM S has been conducted for the following five Sub-Saharan Africa countries Cote dÂ’Ivoire, Ghana, Malawi, South Africa and Tanzania.
35 The Cote dÂ’Ivoire LSMS is the oldest of the five li ving standard surveys, which began in 1985 and ended in 1988. The living standa rd surveys for Ghana, Malawi, South Africa, and Tanzania are more current. However, LS MS for Malawi and South Africa are each limited to one wave of information. Ghana and Tanzania contain the most recent and detailed panel information of the five c ountries surveyed in Sub-Saharan Africa. LSMS Ghana however, lacks detailed informa tion on household health compared to LSMS Tanzania. 4.2.1 The Kagera Region of Tanzania The analyses in this study are based on data from t he Tanzania Living Standard Measurement Survey. The dataset is arguably the mo st detailed of all five living standard surveys for measuring the effect of health status o f the Sub-Saharan Africa population. The living standard survey for Tanzania was conduct ed in the Kagera region of the country. Kagera is located on the western shore of Lake Victoria adjacent to Uganda, Rwanda, and Burundi. Kagera is the 15th largest region in Tanzania and its regional capital Bukoba Town is about 1,500 kilometers from the countryÂ’s capital, Dar Es Salaam. The region covers a total of 40,838 square kilometers and lies at 3,750 feet above sea level6. 6 Location information is derived from the Tanzania Chamber of Commerce Industry and Agriculture website at www.kagera.org.
36 Figure 2: Map of Kagera Tanzania Source: Tanzania Chamber of Commerce Industry and A griculture The Kagera region comprises of five districts name ly, Bukoba (the regional capital), Muleba, Karagwe, Biharamulo and Ngara. T he entire region has a population of approximately 2 million people (Table 1). The most populous region Bukoba, has a population of over 470,000. Total population for t he Kagera region is expected to increase by 400,000 for the year 2007. Table 1: Kagera Population Kagera Population District Population No. of Households Average Household Size Bukoba 476,351 64,510 4.3 Karagwe 425,476 89,047 4.8 Muleba 386,328 79,107 4.9 Biharamulo 410,794 67,131 6.1 Ngara 334,939 49,082 6.8 Total Population Kagera District 2,033,888 394,128 5.2 Projected population for 2007: 2,417,000
37 4.2.2 Current State of Malaria and Other Illnesses in Kagera Although malaria is generally prevalent throughout Tanzania, it is a big public health concern in Kagera where malaria is the leadi ng cause of death7. On July 10 2006, the Tanzania Red Cross National Society (TRCNS) rep orted a rise in the number of reported malaria cases in Kagera. The rise in repo rted cases had resulted in a drastic increase in malaria related mortality in the region The two districts most affected by the outbreak were Karagwe and Muleba. From January to May 2006, the number of deaths among children under the age of five rose to 3,944 in Karagwe and 3,542 in Muleba from about 300 deaths in January of 2006. Coupled with the threat of malaria in the region is the prevalence of HIV/AIDS in Tanzania. In 1983, the first reported cases of HIV /AIDS in Tanzania were from the Kagera region. The first cases of HIV/AIDS in the area were reportedly imported from neighboring countries as a result of Kagera having the largest common border with other countries in east Africa. Like most regions in Afr ica, there are too few hospitals in Kagera to sustain the extent of disease prevalence in the region. In 2006, the TRCNS estimated that there are 13 hospitals, 13 clinics, and 202 dispensaries providing health care serving a population of over 2 million in the Kagera region. The Living Standard Measurement Survey (LSMS) of th e Tanzanian region of Kagera began in 1991 with the goal of measuring the impact of adult mortality (predominantly due to AIDS) on households and evalu ating the effectiveness of policies geared toward preventing the disease in this Sub-Sa hara section. The desire for a LSMS 7 According to the Tanzania Chamber of Commerce Indu stry and Agriculture malaria is the leading cause of death in Kagera, Tanzania.
38 style survey in Kagera was prompted by the disturbi ng rate of HIV infection and AIDS death of the adult population in Kagera. 4.2.3 LSMS Tanzania Kagera The LSMS in Tanzania-Kagera now consists of five wa ves of household and community level data from 1991-2004. Over 800 hous eholds were surveyed in the first four rounds (1991-94). The final round (2004) cons ists of over 2700 households from the original baseline, which were re-contacted 10 to 14 years later. The sections of LSMS Tanzania-Kagera contain responses to questions on a vast array of topics. Of relevance to this study, are questions on current illness, sc hool and work hours, and other sociodemographic information. 4.2.4 Attrition in LSMS Tanzania-Kagera The Tanzania-Kagera LSMS had a low attrition rate o f 10% from wave one to four. The main reason for attrition in the LSMS of Tanzania-Kagera is death in the household, which led to the relocation of the house hold. However, attrition in the fifth wave is a bigger issue since wave five of the surve y was conducted a decade later. Section 4.6 presents the estimation problems posed by attrition in surveys such as the LSMS Tanzania Â– Kagera. 4.2.5 Sample Construction: Schooling In Kagera, it is expected that children be enrolled in school at the age of seven (Ainsworth, et al. 2005). School enrollment in thi s study is defined as students who
39 reported enrollment in formal schooling in all 5 wa ves of the survey. Children who were being home schooled as well as those only attending Koranic schools were excluded from the sample. The study also focuses on primary educ ation in Kagera as most children interviewed in the survey fall within this category Primary schools in the Kagera survey offer a maximu m of seven grade levels. The analysis in this study focuses on children aged seven to 20 who are enrolled in primary school at the time of survey. The mean age observed for the Kagera school sample is 12 years. Chapter five presents additiona l characteristics of the Kagera school sample. 4.2.6 Sample Construction: Labor Supply A sample of adults between the age of 18 and 65 was chosen as the likely work force for LSMS Tanzania Kagera. Majority of the Kagera work force reported working as a farmer from wave one to wave five. The second largest work category in Kagera at the time of the survey was those working for someon e other than himself or herself in form of formal employment (wage employment). Selfemployed individuals are the least of the Kagera work force. 4.2.7 Definition: Prevalence of Illness Estimates of malaria prevalence used in this study are based on self-reported measures. John Bound (1991) argued that self-repor ted measures of health may be more reliable than other objective measures of health. It is expected that a self-reported measure of health will be appropriate for identifyi ng malaria in Africa given the
40 commonality of the illness. In addition, the weakn ess of health systems in Sub-Saharan Africa may contribute self-treatment versus formal care for common ailments. Selftreatment for malaria is common in Sub-Saharan Afri ca where most people live on less than $1 a day and are unable to afford physician ca re (S.C. McCombie 1996). Studies that are more recent however, indicate a huge incre ase in formal treatment seeking in private health clinics and community health centers (Wakgari Deressa 2007). The analysis to follow utilizes both doctor diagnos ed and self-diagnosed selfreported episodes of malaria. A vast majority of r eported malaria cases in the LSMS for the Kagera region are self-diagnosed. Doctor diagn osed episodes only constitute a small fraction of the total (Table 2a). Self-diagnosed e pisodes of malaria in the Kagera sample, was lowest in wave four of the survey. Wave five o f the survey had the highest number of reported episodes (doctor and self-diagnosed)8. Table 2a Episodes of Illness by Diagnosis Type Malaria Diagnosis in LSMS Kagera Diagnosis Type Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 Self-Diagnosed 418 590 538 393 1969 Doctor-Diagnosed 190 229 163 165 998 4.2.8 Definition: Malaria The malaria status in the LSMS is based solely on s elf-reported episodes of doctor diagnosed and self-diagnosed cases. Respondents in the LSMS were asked to report all types of acute illness within the past four weeks a nd all types of chronic illness within the past 6 months. The malaria sample was then defined as inclusive of all individuals who 8 Wave five of the LSMS for Kagera was administered 10 years after the wave four survey.
41 reported a lack of school or work attendance due to the malaria illness in the past seven days. To minimize measurement error with this defi nition of malaria status, the estimation process also controls for pre-existing c onditions that can also result in low attendance. 4.3 Description of Variables The following sections describe the dependent and independent variables selected for the models in equation (4.1) and equation (4.2) for schooling and labor supply. 4.3.1 Dependent Variables The outcome variable in the labor force participati on equation (4.2) is the number of hours worked per week. This variable is defined as the number of hours worked within a seven-day workweek for adults aged 18-65. The LSMS labor force sample is made up of respondents between the ages of 18 and 6 5 who reported work hours in the past seven days as an employee, a farmer, or a self -employed businessman. The outcome variable in the schooling equation (4.1) is the number of school hours attended per week. This variable is defined as the number of hours pre sent in school during the school week for those between the ages of 7 and 20. 4.3.2 Explanatory Variables Factors that determine labor supply in equation (4. 2) are health status, (H), prices, (P), non-labor income, (V), socio-demographic indic ators, (A), education, (E), and community infrastructure, (IN).
42 Health Status, H In our conceptual framework, health status is appro ximated using self-reported measures of acute and chronic illness. This defini tion of health status differs from actual health by a measurement error, e: Actual Health = H + e The measurement error, e is sometimes treated as ra ndom and uncorrelated with other determinants of health. A problem with such treatm ent is that Â‘eÂ’ is unlikely to be random. For instance, in the labor supply case, in dividuals who have reduced their hours of work, are more likely to report poor health stat us, functional limitations, and other conditions (Chirikos and Nestel 1984). This can be attributed to the existence of factors that are known to the individual but unobserved by the researcher; described in the theoretical framework as m In this case, ignoring m would lead to biased estimates of the measure of health status. In addition, there are studies, which suggest that the concerns about non-random Â‘eÂ’ do not necessarily induce a bias on self-report ed measures of health. One case is when an instrumental variables approach is used to capture the other part of H that is not addressed by self reported H (Ettner 1997). Althou gh, self-reported health may bias the estimated coefficients downwards, the endogeneity o f self-reported health may also bias the estimated effect upwards. Therefore, the two e ffects may cancel out unlike in the case of more objective measures of health that are biased towards zero only. The determinants of health status (H) are evaluated as the existence of acute and chronic illness. Acute illness is measured as the incident of malaria. Acute illness in the
43 LSMS is defined as illness that limits the ability to work, attend school or perform regular chores. The incident of malaria in the survey is r ecorded by wave for individuals who reported being incapacitated by malaria within the survey period. Chronic conditions in the LSMS are addressed as pre-existing health condi tions (such as HIV, Asthma and Kwashiorkor, and Malnutrition) that prevent the ind ividual from working or attending school. Prices (P), Education (E), and Socio-demographic In dicators (A) The LSMS contains information on a collection of ho usehold expenditures on food and non-food items including the price of food pharmaceutical products, and school fees. The LSMS price questionnaire consisted of th irty food items, thirteen non-food items, and six pharmaceutical products collected at different markets in Kagera. With the exception of a few items, food items in the survey were weighted to the nearest 50 grams. Pharmaceutical products were not weighted in the su rvey but measured in tablets. Complete pharmaceutical price information is only a vailable for three of the six items in waves. A list of food and pharmaceutical items rel evant to this study can be found on Table A4. In addition to questions on household expenditures, the LSMS asks questions about the educational attainment of each respondent every wave. The survey also asks detailed questions on socio-demographic characteris tics such as age, sex, marital status, and income. Household income in the LSMS is define d as the sum of six components: employment income from self-employment in agricultu re; non-farm self-employment income; income from rent; transfer income from indi viduals and organizations; and other
44 non-labor income. Because no information on wages is available in the LSMS dataset, household income is used as a proxy for the wages i n the analyses. Community and School Infrastructure (SI and IN) In each wave, community leaders are interviewed by the LSMS. They are asked to identify the local schools and hospitals in the district. Representatives from each school and hospital facility are then re-interviewe d in each subsequent wave on the changes in the demand and supply of these facilitie s. Three community infrastructure variables were chosen for the labor supply equation s. These variables measure access to motorable roads, electricity and pipe borne water. For the schooling equation, the number of teachers per school was chosen as the mea sure of school infrastructure. 4.4 Econometric Strategy Equation (4.1) and (4.2) can be estimated using sta ndard econometric methods like OLS. OLS is unbiased in the absence of measur ement error and unobserved heterogeneity. However, as mentioned in the theore tical framework, 2 HH is endogenous to the labor supply model while, and 1 KH is endogenous to the human capital accumulation model. Given the possibility of endog eneity in both models and the likelihood of measurement error in the health statu s variable, OLS is biased. There are several solutions when dealing with unobs erved heterogeneity. This study considers three identification strategies usi ng the fixed effects, two stage least squares (2SLS), and the limited information maximum likelihood (LIML) estimators. The fixed effects (FE) estimator captures all unobs erved, time in-variant factors that
45 affect labor force participation and schooling. FE by no means solves the endogeneity problem and may also impose additional problems of attrition bias and insufficient within variation in panel data estimations. Given that th e endogeneity problem posed by 2 HH in the labor supply equation and,1 KH and in schooling equation will remain unsolved usi ng FE, a more appropriate method of estimating (4.1) a nd (4.2) is by instrumental variables. Another problem with the FE estimator is that it im poses a strict exogeneity assumption on the regressors within the model. Estimation by instrumental variables will involve estimation of equation (4.1) and equation (4.2) using factors such as the disease en vironment, D, and prices, as valid instruments for health. The instrumental variables regression can be implemented using the 2SLS and LIML methods. Having more than one in strument in the instrumental variables estimation also allows us to test overide ntifying restrictions. Consistency of the instrumental variables estimates rely on the validi ty of the instruments. Invalid instruments can produce 2SLS and LIML parameter est imates that much more inconsistent than OLS. 4.4.1 2SLS Estimation Methods: Two Stage Least Squa res (2SLS) Estimation of the effect of malaria on schooling a nd labor supply is based on the structural equation it it it itu X h y+ + + =d b a0 (1) where ity is the outcome (hours of school and work) in the sc hooling and labor supply equations for individual i at time t ; ith is a vector of malaria and pre-existing health
46 status indicators for an individual i at time t ;itX is a vector of socio-demographic, school and community level infrastructure, prices, and health status indictors. As discussed earlier, health status in (1) is correlat ed withitu Estimating the above structural model through OLS is biased. To mitigat e the endogeneity problem in (1), the analysis employs a 2SLS regression of the schooling and labor supply models where the reduced form equation is given by: it it it itv X Z h+ + + =2 1 0p p p (2) and itZ is a vector of indicators of the disease environmen t serving as additional instruments in the reduced form estimation, anditv is the reduced form error term. Assuming that itZ is uncorrelated with the structural itu and provided that the disease environment is sufficiently correlated with health status then, 0 ) (it itZ h Cov Finally, the study estimates a panel 2SLS model of schooling and labor supply where (2) is the first stage regression of health s tatus on a vector of instruments in (2) and (1) is the second stage regression. The second sta ge regression uses hours of school and hours of work as the two outcome measures as depict ed in equation (4.1) and equation (4.2) and a vector of health status indicators as w ell as a vector of socio-demographic, school and community level infrastructure, and pric e indicators as exogenous regressors. 4.4.2 Instruments As discussed earlier, the likelihood of endogeneity in both equations can be alleviated using instrumental variables. It was al so mentioned in chapter three that, PHI
47 and D, are valid instruments for mitigating the end ogeneity of ith in (1). Two instruments of the disease environment available in the LSMS data for the estimation of the reduced form model in (2) are a rainfall season indicator and a measure of rainfall amounts in the Kagera region. A good instrument satisfies the following assumptio ns Â– It must be uncorrelated with the error term and is correlated with the endo genous explanatory variable. As mentioned earlier, having more than one instrument in the 2SLS estimation facilitates the test for overidentification. The overidentificatio n test can be implemented by first estimating the structural equations in (1) to obtai n the 2SLS residuals1Âˆ m Then 1Âˆ m is regressed on all exogenous explanatory variables in (1) to obtain the R-squared,2 1R Under the null hypothesis that all instruments are uncorrelated with 1 m 2 2 1~qnRc where q is the number of instruments minus the total numb er of endogenous explanatory variables. If the null hypothesis is rejected 2 1nR > the critical value in the chi-Square distribution (2 qc), it is safe to conclude that at least one of the instruments is not exogenous. Considering the known correlation between malaria a nd high amounts of rainfall, monthly rainfall estimates and a rainfall season in dicator can be considered as measures of the disease environment (Chris Drakeley, et al. 2005)9. The LSMS keeps monthly record of the rainfall amounts in Tanzania-Kagera. In the LSMS, monthly rainfall records are measured in millimeters over a period o f 60 months. This record is only available for four of the five waves of the LSMS fo r Tanzania. To complete the rainfall 9 Drakeley et al. (2005) found that plasmodium falci parum prevalence showed a negative relationship wit h altitude and rainfall amounts in Tanzania
48 data, the LSMS rainfall estimates are merged with a dditional rainfall data from the NOAA National Data Centers for wave five. High prevalence for malaria usually falls in the ra iny season months, which varies between regions in the Sub-Sahara. The rainy seaso ns in the Kagera district of Tanzania are the short Â‘VuliÂ’ rainy season, which falls betw een October and January, and the long Â‘MsimuÂ’ rainy season from March to May. In additio n to instrumenting for the likely endogeneity of malaria, controlling for the rainfal l patterns using a season indicator ensures that seasonal variations in labor force par ticipation and schooling are reasonably captured. 4.5 Analyzing Improvements in the Disease Environme nt The panel structure of LSMS also allows for a direc t test of whether changes in pre-conditioning factors have translated into impro vements in the disease environment over time in Kagera. Factors preconditioning expos ure to malaria used in this study are access to community health facilities and changes i n household income levels over time. Using a difference-in-difference (DID) estimator, t he estimated change in a preconditioning factor from wave t-1 to wave t can be obtained as: 1 =t tg g d tgand 1-tg represent levels of measurable preconditioning fact ors over time and d is the difference between predicted values of these factor s over time. If the parameter d is significantly different from zero, it is safe to co nclude that the preconditioning factors have improved the disease environment in Kagera ove r time.
49 Using the first and last waves of the LSMS Kagera, improvements in the disease environments are determined using the following DID estimation of malaria on two sets of preconditioning factors. Two preconditioning fa ctors were chosen for the DID estimation: a count of health facilities in Kagera over time and changes in the levels of household income in Kagera over time. u lities healthfaci year income year lities healthfaci income year malaria + + + + + =+* *4 3 2 1 0 0b b b b d bIn the above DID equation, the parameters of intere st, 3 b and 4 b measure the effects of changes in income and changes in the number of heal th facilities on the incidence of malaria. The parameter 0dcaptures the changes in the reported cases of malar ia within the sample between 1991 and 2004. 4.6 Other Identification Strategies In addition to 2SLS and LIML, another identificati on strategy explored in the estimation of the labor supply and schooling equati ons is the Average Treatment Effects on the Treated (ATET) estimator. Unlike the 2SLS a nd LIML the ATET takes into consideration the different states of the health st atus indicator ( H ). A thorough evaluation of both equations using ATET is presente d in Appendix D.
50 Chapter 5 Research Results This chapter presents the estimation results for t he schooling and labor supply models. Section 5.1 presents the sample evidence o n episodes of malaria for the school and work samples. Section 5.2 documents the school enrollment characteristics of school age children in Kagera. Section 5.3 details the su mmary statistics of the Kagera school sample. Section 5.4 and 5.5 review the summary sta tistics of the Kagera labor force sample. Estimation results for the school and work sample using OLS, FE, 2SLS, and LIML are presented in section 5.6 and 5.7. Lastly, section 5.8 presents difference in difference estimates (DID) of changes in the diseas e environment in Kagera. 5.1 Episodes of Malaria in the Study The rates of reported malaria episodes by survey pe riod for the Kagera sample ranges between a low of 7% and a high of 23% for th e selected school and work samples. The percentage of reported malaria cases for childr en enrolled in school ranged between 7% and 13% from wave one to wave five of the LSMS ( Table 2b). Unlike the school sample, workers in Kagera reported a wider range of malaria episodes by wave. The malaria rate for self-employed farmers ranged betwe en 11% and 20% across waves. Non-farm self-employed workers and individuals in w age employment also reported a similar range of malaria cases (between 9% and 23%) across waves. Wave five had the
51 highest number of reported malaria cases in the LSM S for children enrolled in school and for most types of employment. Overall, fewer peopl e reported being sick with malaria in earlier waves than they did in more recent waves of the LSMS; a finding consistent with current trends in reported malaria cases for Tanzan ia (Table A3). Table 2b: Malaria Cases Reported Malaria Cases By W ave Sample Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 School Sample % Reporting a Malaria Episode 7% 11% 12% 11% 13% Work Sample % of Self Employed (Farm) Reporting a Malaria Episo de 11% 14% 13% 11% 20% % of Self Employed (Non-Farm) Reporting a Malaria Episode 13% 9% 11% 15% 23% % of Wage Employed Reporting a Malaria Episode 9% 1 7% 15% 10% 12% 5.2 School Enrollment in Kagera There were over 11,000 children between the age of 7 and 20 in the Kagera sample from 1991 to 2004. Majority of these childr en were introduced into the survey in 2004 (wave five). Approximately 6400 children of s chool age were enrolled in school in at least one wave of the Kagera survey. In wave on e, the school enrollment rate for children in the survey was 49%. Between wave two a nd wave three, the overall enrollment rate remained steady at 56%. By wave fi ve, school enrollment for children of school-going age had increased to 62%. In terms of school enrollment rates by gender compo sition, males were more likely to be to be enrolled in school than females. However, enrollment for both groups increased steadily in each wave with the exception of waves two and three when female
52 school enrollment stalled at 54%. In addition to t he overall low school enrollment in the sample, the school starting age was well over seven years. In fact, the lowest grade level, P1, consisted of students as old as age 16. This f inding is consistent with that of Ainsworth, et al. (2005), which found that school e nrollment was delayed for orphans in Tanzania and that children tend to drop out of scho ol due to orphan status or death of an adult in the family. Table 3: School Enrollment Rate Children Ages 7 to 20: School Enrollment Wave 1 Wave 2 Wave 3 Wave 4 Wave 5 Total 2232 2112 1996 1821 3114 Number Currently Enrolled in School 1104 1186 1125 1048 1936 Percentage of Total Currently Enrolled in School 49 % 56% 56% 58% 62% Percentage of Males Currently Enrolled in School 51 % 58% 59% 60% 67% Percentage of Females Currently Enrolled in School 48% 54% 54% 55% 58% 5.3 Summary Statistics Kagera School Sample Table 4 presents the summary statistics for the en tire Kagera school sample for all five waves of the LSMS data. The final primary sch ool sample for all five waves of the survey consisted of 4,189 students. This sample si ze was conditioned on having reported current school enrollment. Therefore, the school s ample excludes students who were at home on vacation or holiday at the time of the surv ey. The summary statistics for primary school students in the Kagera showed that school children in Kagera reported attending approx imately 28 hours of school per week between wave one and five of the study. The averag e school enrollment age in the schooling sample was 12 years. Overall, there were more students in lower grade levels
53 in the Kagera schools than there were in upper prim ary grade levels. In particular, there were 70% more students enrolled in the P1 (the lowe st grade level) than there were in P7 (the highest grade level). Again, this finding is consistent with that of Ainsworth, et al. (2005). The average annual household income for parents wi th children in primary school in the Kagera survey was well below $200. Most par ents with children in primary school in Kagera reported having less than an elementary s chool education. However, fathers in the sample were more likely to have completed prima ry school than mothers. In terms of malaria status, at least 11% of primary school chil dren reported an episode of malaria in all five waves of the survey. However, the rate of chronic illness in the sample was much lower than the malaria rate. Merely 3% of primary school children reported a chronic illness in all five waves. Other variables relevant to the Kagera school samp le include school level infrastructure variables and measures of the cost o f education. Approximately 10 teachers were staffed in a typical primary school i n Kagera during the survey period. However, each primary school in the Kagera sample w as staffed by a minimum of one teacher and a maximum of 29 teachers. In terms of costs, school students in Kagera live relatively close to home and on average commute abo ut three miles to school. Real expenditures on tuition for primary school in Kager a were negligible as the average household spent about $310 a year on school fees in the survey. 10 School fees in the analysis exclude expenditures s uch as money spent on books and supplies, school uniform and transportation as most of these variabl es contained missing observations in the sample.
54 Table 4: Summary Statistics Kagera School Characteristics (1991-2004) Variable Description Obs Mean Std. Dev. Min Max School Hours Hours School Hours (Past 7 Days) 4189 28.311 12.686 0.000 55.000 Grade Level p1 First Grade Level 4189 0.136 0.342 0.000 1.000 p2 Second Grade Level 4189 0.138 0.345 0.000 1.000 p3 Third Grade Level 4189 0.128 0.334 0.000 1.000 p4 Fourth Grade Level 4189 0.121 0.326 0.000 1.000 p5 Fifth Grade Level 4189 0.104 0.306 0.000 1.000 p6 Sixth Grade Level 4189 0.077 0.266 0.000 1.000 p7 Seventh Grade Level 4189 0.031 0.172 0.000 1.00 0 Socio-Demographic Characteristics Age Age in Years: Range 7-20 4189 12.604 2.940 7.00 0 20.000 Gender Percent Male: 1 = male; 0 female 4189 0.527 0.499 0.000 1.000 Mother Education Mother Education: 1 = primary education; 0 = no education 4189 0.136 0.343 0.000 1.000 Father Education Father Education: 1 = primary education; 0 = no education 4189 0.178 0.383 0.000 1.000 Household Income Annual family income in dollars 41 89 199.531 473.660 0.000 11880.640 School Characteristics Teachers Number of teachers at school 4189 9.671 4. 106 1.000 29.000 Distance Distance to school from home in miles 4189 3.276 26.109 0.000 500.000 Fees Annual school fees in dollars 4189 0.851 7.571 0.000 300 Health Status Malaria Malaria Cases 4189 0.113 0.317 0.000 1.000 Chronic Illness Cases of Chronic Illness 4189 0.033 0.180 0.000 1.000 Disease Environment Malaria Season Rainfall Season 4189 0.279 0.449 0.000 1.000 Rainfall Total Monthly Rainfall (mm)a 4189 388.826 287.313 2.700 872.300 aThis variable is constructed from the Living Standa rd Measurement Survey for Tanzania (1991-994) and N OAA National Data Centers (2004)
55 5.4 Summary Statistics Kagera Work Sample Three employment categories where analyzed in the labor supply estimation. These employment groups are self-employment in farm occupations, self-employment in non-farm occupations, and wage employment (Table A5 ). The largest category of the Kagera work sample was farm employment, which accou nted for 71% of the total workforce. Workers in wage employment made up 18% of the total work sample. The smallest employment category in the survey was the self-employment non-farm group, which accounted for 15% of the Kagera work sample. The typical worker in the Kagera sample was age 34 years old, female and married. Precisely, the gender composition of the Kagera work sample was 45% male and 55% female. In addition, 56% of these workers were married at the time of the survey and the average worker reported having at le ast a primary school education. Workers in the Kagera sample earned an average of $ 189.93 dollars a year and worked approximately 20 hours per week. In terms of community level indicators, access to community infrastructure was measured as the availability of to electric power, pipe borne water, and motorable roads in each community. With the exception of access to a motorable road, most Kagera workers reported living in communities without acce ss to either of the other two community infrastructures. Specifically, 80% of wo rkers in Kagera lived in areas without pipe borne water, while 59% resided in area s without electrical power. However, 96% of Kagera workers reported living near a motora ble road at the time of the survey.
56 5.4.1 Farm Self-employment Characteristics Two main types of farm activities were chosen for the definition of the farm employment work category. The first type of farm a ctivity refers to adults who reported work hours on a farm or garden belonging to themsel ves or their household at the time of the survey. The second farm activity pertains to t hose who served as caretaker for animals or transformed animal products belonging to themselves or their households during the survey period. These two types of farm activity make up the single definition of farm related self-employment used in the analysi s. Table 5a presents the summary statistics for the Ka gera farm sample. The average age for a typical Kagera farmer in the farm employment category was 35 years. Compared to other workers in non-farm self-employme nt, self-employed farmers are less likely to be educated. Most farmers in the survey had less than a secondary education. Majority of the farmers were female and married. S elf-employed farmers in the Kagera survey earned on average $182.54 annually during th e survey period. Farmers in Kagera were also less likely to have access to community l evel infrastructures such as motorable roads, pipe borne water and electric power supply c ompared to workers in non-farm self employment and wage employment. Ninety five percen t of self-employed farmers lived in communities near a motorable road, while 36% had access to electric power and only 17% had access to pipe borne water.
57 Table 5a: Summary Statistics Kagera Self-Employment (Farm) Characteristics (1991 -2004) Variable Description Obs Mean Std. Dev. Min Max Work Hours Hours Number of hours worked per week 7,822 24.585 15.576 0.000 130.000 Socio-demographic Characteristics Age Age in Years: Range 18 65 7,822 35.038 13.929 18.000 65.000 Gender Percent Male: 1 = Male; 0 Otherwise 7,822 0.412 0.492 0.000 1.000 Education Highest Level of Education: 1 = Primary; 2 = Secondary; 3 = College 7,822 1.893 1.300 0.000 6.000 Married Marital Status : 1 = Married; 0 Otherwise 7 ,822 0.580 0.494 0.000 1.000 Income Annual Income: Employment income in dollars 7,822 182.540 377.082 0.000 7401.588 Health Status Malaria Malaria Cases = 1 if reported malaria in at least 1 wave; 0 otherwise 7,822 0.154 0.361 0.000 1.000 Chronic Illness Chronic Illness = 1 if reported chr onic illness in at least 1 wave; 0 otherwise 7,822 0.107 0.309 0.000 1.000 Disease Environment Malaria Season Rainfall Season: 1 = Interviewed dur ing rainy season; 0 otherwise 7,822 0.885 0.319 0.0 00 1.000 Rainfall Total Monthly Rainfalla: Total Monthly Rainfall measured in millimeters 7,07 7 358.799 278.346 1.000 872.300 Community Infrastructure Electric Power Electric Power: 1= Access; 0 otherwi se 6,886 0.364 0.481 0.000 1.000 Pipe Water Pipe Borne Water: 1 = Access; 0 otherwis e 6,886 0.174 0.379 0.000 1.000 Road Motorable Road: 1 = Access; 0 otherwise 6,886 0.951 0.215 0.000 1.000 Pricesb Food Average food price in dollars 6,393 0.156 0.196 0.003 3.800 Pharmaceuticals Average pharmaceutical price in dol lars 5,445 0.032 0.106 0.001 0.700 aThis variable is constructed using estimates from t he Living Standard Measurement Survey for Tanzania (1991-994) and NOAA National Data Centers (2004); Food items are w eighted to the nearest 50 grams and pharmaceutical items are measured in tablets
58 In terms of weekly work hours, self-employed farmer s in the Kagera survey reported working on average 25 hours per week. The number of hours worked per week for self-employed farmers were closer to the work h ours of those in non-farm selfemployment but lower than hours reported by workers in wage employment. When asked about malaria status, 15% of self-employed fa rmers in the survey reported being incapacitated with malaria. This was the lowest pe rcentage of all types of employment in the Kagera work sample. Although, farmers reported lower malaria rates on average, they were also more likely to report having a chron ic illness. Eleven percent of surveyed farm workers in Kagera reported a chronic illness i n at least one wave of the survey. 5.4.2 Non-farm Self Employment Characteristics A non-farm self employed worker in the analysis is defined as an individual who owns his/her own business or is employed by his/her family in a non-farm sector. This is the smallest employment category in all five waves of the survey with (Table 5b). The average age for workers in non-farm self-employment is 34 years. On average, non-farm self-employed workers reported a higher income than those in the self-employment farm category. Self-employed workers in the Kagera samp le were mostly male and better educated than those in farm employment, but less ed ucated than workers in wage employment.
59 Table 5b: Summary Statistics Kagera Self-Employment (Non-farm) Characteristics ( 1991-2004) Variable Description Obs Mean Std. Dev. Min Max Work Hours Hours Number of hours worked per week 1,659 24.940 21.842 0.000 130.000 Socio-demographic Characteristics Age Age in Years: Range 18 65 1,659 33.474 12.015 18.000 65.000 Gender Percent Male: 1 = Male; 0 Otherwise 1,659 0. 592 0.492 0.000 1.000 Education Highest Level of Education: 1 = Primary; 2 = Secondary; 3 = College 1,659 2.001 1.151 0.000 5.000 Married Marital Status : 1 = Married; 0 Otherwise 1 ,659 0.586 0.493 0.000 1.000 Income Annual Income: Employment income in dollars 1,659 221.565 591.714 0.000 7401.588 Health Status Malaria Malaria Cases = 1 if reported malaria in at least 1 wave; 0 otherwise 1,659 0.165 0.371 0.000 1.000 Chronic Illness Chronic Illness = 1 if reported chronic illness in at least 1 wave; 0 otherwise 1,659 0.096 0.295 0.000 1.000 Disease Environment Malaria Season Rainfall Season: 1 = Interviewed dur ing rainy season; 0 otherwise 1,659 0.820 0.384 0.0 00 1.000 Rainfalla Total Monthly Rainfalla: Total Monthly Rainfall measured in millimeters 1,41 3 308.379 273.839 1.000 872.300 Community Infrastructure Electric Power Electric Power: 1= Access; 0 otherwi se 1,414 0.526 0.499 0.000 1.000 Pipe Water Pipe Borne Water: 1 = Access; 0 otherwis e 1,414 0.212 0.409 0.000 1.000 Road Motorable Road: 1 = Access; 0 otherwise 1,414 0.967 0.177 0.000 1.000 Pricesb Food Average food price in dollars 1,260 0.199 0.260 0.003 1.500 Pharmaceuticals Average pharmaceutical price in dol lars 1,031 0.044 0.124 0.002 0.700 aThis variable is constructed using estimates from t he Living Standard Measurement Survey for Tanzania (1991-994) and NOAA National Data Centers (2004); Food items are w eighted to the nearest 50 grams and pharmaceutical items are measured in tablets
60 In addition, work hours for non-farm self-employme nt were closer to those reported in farm employment. Specifically, the ave rage worker in non-self employment in the Kagera sample worked approximately 25 hours a week. These workers were more likely to reside in communities with access to moto rable roads, pipe borne water and electric power. Ninety seven percent of workers in non-farm self employment reported living in a community with access to a motorable ro ad; while 53% reported having access to electric power and 21% had access to pipe borne water. In terms of health status, non-farm self-employed workers were more likely to report episodes of malaria than workers in other em ployment categories. The malaria rate for workers in this group was 17% for all five waves of the survey. On the contrary, 10% of self-employed workers reported a chronic ill ness in all five waves of the survey. 5.4.3 Wage Employment Characteristics Wage employment in the analyses is defined as empl oyment by an employer who is not a member of the employeeÂ’s household (i.e. a firm or the government). The employment characteristics for this work group are presented below on Table 5c. Workers in the wage employment category reported wo rking an average of 40 work hours per week. This was the highest weekly work ho urs for all three work categories. A typical worker in this cohort was 32 years of ag e, male, and better educated than workers in other work categories. Unlike thos e in farm employment, workers in wage employment had better access to community leve l infrastructures such as electric power, pipe-water, and motorable roads. Approximat ely, 96% of wage-employed
61 workers had access to a motorable road, 25% had acc ess to pipe borne water, while 50% had electric power supply at the time of the survey Workers in wage employment also earned more income than workers in the other two ca tegories of employment over the sample period. Particularly, a worker in wage empl oyment had twice the earning power of a farm worker and earned 39% more income than a self-employed individual in Kagera. In terms of health status, 16% of workers in wage employment reported being sick with malaria during the survey period. This r ate of malaria infection is high relative to that reported in farm employment but slightly lo wer than the malaria rate for those in self-employed non-farm occupations. On the contrar y, wage workers were significantly less likely to report a chronic illness when compar ed to their counterparts. Only 7% of workers in wage employment suffered from a chronic condition in at least one wave of the survey. 5.5 Other Kagera Sample Characteristics Food and pharmaceutical price data were used in ad dition to community infrastructure as indicators of the Kagera work env ironment. These variables are derived directly from the LSMS dataset. The calculation of food prices in the dataset is based on an average of a 30 food items priced at different p oints in time and weighted to the nearest 50 grams. Pharmaceutical prices are measure d in tablets in the survey. Table A4 provides a list of the food and pharmaceutical item s relevant to this study. The prices were collected for each item at the cluster level, where one cluster consists of 16 households. Food and pharmaceutical prices were on ly used in the labor supply
62 estimation. Over the survey period for the Kagera work sample, food prices for all 30 items averaged at of 0.2 cents and peaked at 3.80 d ollars. Pharmaceutical prices for selected items averaged at 0.04 and peaked at 0.70 cents. In addition, the estimated labor supply and school ing equations also include an indicator of rainfall amounts and a measure of seas onal rainfall variations. Rainfall amounts in the LSMS were measured in millimeters ov er a 60-month survey period from wave one to four. Rainfall estimates for wave five was derived from the NOAA national data centers. Monthly rainfall estimates ranged be tween one and 872.30 millimeters over the sample period (Table 4). For the school sample monthly rainfall amounts averaged at 384.706 millimeters over the sample period (Tabl e 4). Average rainfall amounts for the total work sample was 341.27 millimeters for al l waves. A malaria season indicator was also included in th e analyses. The malaria season indicator is a dummy variable capturing rainy and n on-rainy months in Kagera. Additional coverage of seasonal rainfall patterns i n Kagera can be found in chapter four.
63 Table 5c: Summary Statistics Kagera Wage Employment Characteristics (1991-2004) Variable Description Obs Mean Std. Dev. Min Max Work Hours Hours Number of hours worked per week 1,987 40.650 21.095 0.000 130.000 Socio-demographic Characteristics Age Age in Years: Range 18 65 1,987 32.389 11.637 18.000 65.000 Gender Percent Male: 1 = Male; 0 Otherwise 1,987 0.721 0.449 0.000 1.000 Education Highest Level of Education: 1 = Primary; 2 = Secondary; 3 = College 1,987 2.121 1.280 0.000 6.000 Married Marital Status : 1 = Married; 0 Otherwise 1,987 0.508 0.500 0.000 1.000 Income Annual Income: Employment income in dollars 1,987 365.769 567.481 0.002 7300.002 Health Status Malaria Malaria Cases = 1 if reported malaria in at least 1 wave; 0 otherwise 1,987 0.161 0.368 0.000 1.000 Chronic Illness Chronic Illness = 1 if reported chr onic illness in at least 1 wave; 0 otherwise 1,987 0.074 0.262 0.000 1.000 Disease Environment Malaria Season Rainfall Season: 1 = Interviewed dur ing rainy season; 0 otherwise 1,987 0.831 0.374 0.0 00 1.000 Rainfalla Total Monthly Rainfalla: Total Monthly Rainfall measured in millimeters 1,71 8 335.444 281.118 1.000 872.300 Community Infrastructure Electric Power Electric Power: 1= Access; 0 otherwi se 1,679 0.496 0.500 0.000 1.000 Pipe Water Pipe Borne Water: 1 = Access; 0 otherwis e 1,679 0.245 0.430 0.000 1.000 Road Motorable Road: 1 = Access; 0 otherwise 1,679 0.958 0.200 0.000 1.000 Pricesb Food Average food price in dollars 1,458 0.183 0.204 0.005 1.500 Pharmaceuticals Average pharmaceutical price in dol lars 1,188 0.042 0.122 0.002 0.700 aThis variable is constructed using estimates from t he Living Standard Measurement Survey for Tanzania (1991-994) and NOAA National Data Centers (2004); Food items are weight ed to the nearest 50 grams and pharmaceutical items are measured in tablets
64 5.6 Main Estimation Results The estimation results of the causal effects of ma laria on school and work hours are presented in this section. Section 5.6.1 summa rizes the estimated results for the school sample beginning with the first stage 2SLS e stimates and the overidentification test. Section 5.6.2 presents the 2SLS, OLS and FE results for the entire school sample. As indicated earlier, the Kagera labor supply sampl e is categorized into three main types of employment (namely, self-employed farmers, selfemployed non-farm workers and wage employed workers) in order to separately ident ify the impact of malaria on each group. A separate binary indicator is included in the analyses identify each work group. Section 5.6.3 and 5.6.4 presents the first and seco nd stage estimation results for the three employment groups in the sample. 5.6.1 First Stage IV Estimation Results Kagera Sc hool Sample First stage estimation results for the Kagera samp le are presented in Table 6a. The three instruments used in the first stage as pr oxies for the disease environment are one binary indicator of the seasonal variations in rainfall and a level and squared indicator of monthly rainfall amounts11. The first stage results show a significant correlation between malaria and seasonal variations in rainfall for the Kagera region. The coefficient on the monthly rainfall indicator is sm all magnitude. Even so, the first stage results show a positive correlation of the three in struments with the incidence of malaria. A joint significance test of the instruments indica tes the statistical significance of the 11 Analysis of the reported cases of malaria in the o verall Kagera data showed a non-linear relationship between malaria episodes and rainfall amounts (Figu re A2).
65 three instruments in the first stage. Table 6a sho ws that the partial F-test on the three instruments is statistically significant at the 1% level. The n*R-Sq test was used to test overidentifying r estrictions in the first stage IV estimation process. The overidentification test is based on the null hypothesis that the instruments are uncorrelated with the error term in the first stage IV regression. The n*R-Sq estimate in the first stage estimation is 1. 028 with a p-value of 0.5982. Therefore, the null hypothesis is rejected and it i s safe to conclude that at least one of the IVs is not exogenous in the first stage estimation. Although the overidentification test and n*R-Sq test confirm the validity of the instrum ents in the first stage estimates, they do not guarantee the exogeneity of the health statu s variables in the second stage. In addition to the instruments evaluated at the fi rst stage, the first stage estimates also provide a meaningful insight into the determin ants of malaria in the Kagera school sample. Two variables found to be significantly co rrelated with the incidence of malaria in the school sample are age and the household spen ding on tuition. As presented in the first stage results on Table 4a, age decreases the probability of reporting an incidence of malaria in the Kagera school sample by 0.003%. On the other hand, a unit increase in household spending on school fees increases the pro bability of reporting an incidence of malaria by 0.002%. An explanation of the significa nt positive correlation between the probability of reporting an episode of malaria and household spending on education lies in basic economic theory. The positive correlation indicates a likely tradeoff between household spending on education and household spend ing on health. In other words, the more the household allocates resources to education the less it has left to allocate to health.
66 Table 6a: First Stage Results Kagera School Sample Dependent Variable: Malaria (N = 4189) 1st Stage Explanatory Variables Coefficient Standar d Error t P>t Age -0.003917* 0.002121 -1.850 0.065 Grade 1 -0.005497 0.016421 -0.330 0.738 Grade 2 0.026833 0.016353 1.640 0.101 Grade 3 0.008775 0.017239 0.510 0.611 Grade 4 0.026348 0.018167 1.450 0.147 Grade 5 0.036911* 0.019930 1.850 0.064 Grade 6 0.033975 0.022608 1.500 0.133 Gender -0.014960 0.009957 -1.500 0.133 Mother Education 0.003289 0.015023 0.220 0.827 Father Education 0.006277 0.011492 0.550 0.585 household Income 0.000002 0.000011 0.180 0.858 Distance to School 0.000229 0.000205 1.120 0.264 School Fees 0.002353*** 0.000744 3.160 0.002 Number of Teachers -0.001640 0.001219 -1.350 0.179 Chronic Illness -0.014631 0.028712 -0.510 0.61 Instruments Rainfall Season 0.036441*** 0.013810 2.630 0.008 Total Rainfall -0.000132 0.000084 -1.590 0.112 Total Rainfall Squared 1.10E-07 0.0000001 1.170 0.2 43 F-Statistic (P-value) 2.73 (0.000) Joint Significance IVs Only (P-value) 7.11 (0.000) Overidentification Test N*R-Sq (P-value) 1.028 (0 .5982) ***1% Significance Level **5% Significance Level *1 0% Significance Level Overidentification Test (N*R-Sq): Reject H0; at least some IVs are not exogenous The malaria season indicator is positively related to malaria incidence and statistically significant even though the magnitude of the coefficient is small. The coefficients on the rainfall amount variables are n ot individually statistically significant. However, an overall test reveals the joint statisti cal significance of all the instruments in the schooling estimation at the 1% level.
67 5.6.2 2SLS, OLS and FE Estimation Results Kagera School Sample Table 6b presents the 2SLS, OLS, and FE estimation results for the dependent variable school hours for the Kagera school sample. The table shows a statistically significant causal relationship between malaria and school hours. The incidence of malaria in the Kagera school sample is associated w ith a 23.67 hour decrease in weekly school attendance. This estimate is statistically significant at the 1% level. OLS and FE estimates of the relationship between malaria and s chool hours also show a significant causal effect although the magnitudes of the coeffi cients for both regressors are smaller than the estimated 2SLS effect. Unlike 2SLS malari a estimates, the OLS and FE results are likely to be biased downwards due to the endoge neity of malaria. Beside the malaria variable, another health status indicator in the school sample estimations is chronic illness. This variable is i nsignificant in explaining school hours for the Kagera school sample. This lack of significanc e can be attributed to the relatively small number of students in the Kagera sample repor ting a known form of chronic illness (see Table 4). Another indicator of the number of school hours attended per week is age. 2SLS estimates show that being a year older increas es weekly school attendance by 0.82 hours. Grade level was also a strong indicator of school attendance. The omitted grade level in the analysis grade seven. With the except ion of students in the 1st grade, hourly school attendance per week is positively correlated with grade level for all grades. In particular, students who had completed the 4th grade have the highest hourly school attendance rate compared to those in grade seven at the time of survey.
68 In addition to age in the schooling equation gende r also proved significant in predicting school hours. Males in the Kagera schoo l sample are less likely to attend school on a weekly basis compared to females. In p articular, male students in the Kagera sample attended 1.076 fewer weekly school hours com pared to female students. Furthermore, 2SLS, OLS, and FE estimates all show t hat family income increases weekly school attendance by 0.002 hours a week. However, parental education proved insignificant in explaining school attendance. In terms of the indicators of school infrastructur e and the cost of schooling, school fees and the number of teachers in a school are bot h significant predictors of school attendance. School fees, an indicator of the cost of schooling is significant in both 2SLS and FE estimations at the 10% level or better. 2SL S estimates show that a one unit increase in school fees increases weekly school att endance by 0.065 hours, while OLS and FE estimates predict a much lower effect. The predicted positive relationship between weekly school attendance and school fees is counterintuitive.
69 Table 6b: Second Stage 2SLS Results Kagera School Sample Dependent Variable: Hours of School Attended Per We ek (N = 4,189) Two Stage Least Squares Ordinary Least Squares Fixe d Effects Variables Coefficient Std. Error Coefficient Std. Error Coefficient Std. Error Malaria -23.674*** 9.018 -2.918*** 0.534 -2.064*** 0.802 Age 0.820*** 0.097 0.900*** 0.067 0.682*** 0.198 Grade 1 -0.563 0.681 -0.735 0.541 -0.031 0.983 Grade 2 8.902*** 0.732 7.698*** 0.540 8.829*** 1.231 Grade 3 9.012*** 0.723 8.573*** 0.569 10.883*** 1.410 Grade 4 9.424*** 0.795 8.573*** 0.607 9.903*** 1.573 Grade 5 6.441*** 0.907 5.881*** 0.673 6.289*** 1.740 Grade 6 8.365*** 1.009 7.694*** 0.753 7.512*** 1.937 Gender -1.076** 0.433 -0.881*** 0.335 Mother Education -0.407 0.624 -0.689 0.490 -0.145 0.913 Father Education 0.107 0.484 0.067 0.362 -1.755*** 0.681 Income 0.002*** 0.000 0.002*** 0.000 0.002*** 0.001 Distance to School -0.012 0.009 -0.016** 0.007 -0.011 0.008 School Fees 0.065* 0.039 0.024*** 0.006 0.013 0.022 Number of Teachers 0.057 0.052 0.080** 0.041 -0.149 0.193 Chronic Illness -1.462 1.205 -1.413 1.010 2.453 1.442 Constant 15.675*** 1.958 12.304*** 0.934 16.177*** 2.886 Hausman Statistic (p-value) 31.42 (0.009) ***1% Significance Level **5% Significance Level *1 0% Significance Level Note: Malaria = 1 if reported at least one episode of malaria in all 5 waves; 0 otherwise. Age = age i n years; Grade 1-7 = Completed Primary School Grade Levels; the omitted group is grade 7. Gender = 1 if male, 0 otherwise; Mother/Father Education = 1 if mother/father has at least a prima ry school education; 0 otherwise. Household Income = income in dollars; Chronic illness = 1 if reported at least one chronic illness; 0 oth erwise Distance to School = Estimated distance trav eled from home to school in miles. Number of teache rs (this is a community level variable) = Number of teachers per school.
70 Higher school fees might reflect higher school qua lity, which is not controlled for in the estimation. In addition, parents who spend more on school related expenses may be inclined to encourage higher school attendance c ompared to those who spend less. The positive relationship between school attendance and tuition can also be attributed to school quality and perhaps a public versus private school quality phenomenon since the majority of students in the survey were enrolled in public schools. In fact, there are no records of child enrollment in private schools for the first 4 waves of the survey and the fraction of reported enrollments in public school i n wave five is small. In addition to school quality, other unobserved factors such as fa mily characteristics or family values regarding education could be driving the observed e ffect. As described in chapter four, the main proxy for s chool infrastructure in the estimation is number of teachers per school. 2SLS and FE estimates of the relationship between the number of teachers per school and schoo l attendance are statistically insignificant. However, the OLS coefficient for sc hool infrastructure is statistically significant and suggests that an additional teacher staffed in a Kagera school moderately increases school attendance by 0.08 hours per week. 5.6.3 First Stage IV Estimation Results Kagera Wo rk Sample The first stage results for the Kagera work sample are presented in Table 7a. The instruments in the first stage 2SLS regression for the work sample are identical to those used in section 5.6.1 for the school sample. All t hree instruments for the disease environment for the labor supply sample are jointly statistically significant in the first stage. The n*R-Sq test was done to ensure the endo geneity of the instruments in the first
71 stage. With a 2c statistic of 3.479 and a p-value of 0.1756, it is safe to conclude that at least one of the three instruments is not exogenous in the first stage. As indicated earlier, studies evaluating the causa l effects of malaria on labor supply have thus far focused on farm employment. T he motivation behind disaggregating the Kagera work sample by employment type was to determine the extent to which type of employment influences the incidenc e of malaria, with the expectation that farm employment increases the likelihood of co ntracting malaria. To clarify this, the analyses begin with evaluating the employment type indicators in the first stage 2SLS estimation. According the first stage results, workers in wage employment have a higher probability of reporting an episode malaria compare d to those in self-employment. There is no statistically significant difference in the i ncidence of malaria between workers in farm employment and those in self-employment. The e ffect of malaria on workers by type of employment is further investigated in the s econd stage through interacting malaria with employment type variables. As seen on Table 7a, age is a significant indicato r of malaria incidence for the work sample. However, the magnitude of the age eff ect on the adult work sample is rather small. Chronic illness and food prices also proved significant in the first stage work sample estimation. Chronic illness for the en tire work sample increases the incidence of malaria for working adults by 2 percen tage points. On the other hand, food prices have the strongest effect on malaria inciden ce in terms of magnitude. With the exception of the age, chronic illness and food pric es, other regressors in the labor supply estimation proved insignificant in explaining the i ncidence of malaria; a finding similar to
72 that obtained the first stage estimation for the Ka gera school sample. An overall significance of the malaria first stage regression was evaluated using a joint F-test. The joint F-test confirmed the statistical significance of the first stage labor supply 2SLS estimates at the 1% level. Table 7a: First Stage Results (without interaction terms) Kagera Work Sample Dependent Variable: Malaria (N = 5,876) 1st Stage Explanatory Variables Coefficient Std. Er r. t P>t Farm Employment -0.016 0.011 -1.510 0.131 Wage Employment -0.0009*** 0.013 -0.670 0.502 Age in Years 0.001 0.000 2.820 0.005 Gender 0.003 0.009 0.370 0.715 Education -0.001 0.004 -0.200 0.845 Married -0.008 0.009 -0.830 0.408 Income -8.60E-06 0.000 -0.800 0.422 Chronic Ill 0.021* 0.013 1.670 0.096 Electric Power -0.006 0.013 -0.450 0.653 Pipe Water -0.010 0.014 -0.710 0.480 Motorable Road -0.027 0.029 -0.950 0.344 Pharmaceutical Price 0.099 0.094 1.050 0.295 Food Price 0.051* 0.029 1.760 0.079 Instruments Malaria Season -0.110** 0.044 -2.510 0.012 Total Rainfall -0.0002*** 0.000 -3.940 0.000 Total Rainfall Squared 2.86E-07*** 7.08E-08 5.110 0 .000 F-Statistic (P-value) 3.26 (0.000) Joint Significance IVs Only (P-value) 9.07 (0.000) Overidentification Test N*R-Sq (P-value) 3.479 (0 .1756) Overidentification Test (N*R-Sq): Reject H0; at least some IVs are not exogenous
73 5.6.4 2SLS, OLS and FE Estimation Results Kagera Work Sample This section presents the main estimation results for the Kagera work sample. First, the 2SLS12 estimates for the work sample are presented on Tab le 7b without interaction terms. In Table 7c, the complete work sample results are presented with interaction terms to determine the extent to which malaria limits work hours for each employment type. 2SLS estimates of work hours without interaction o f malaria with type of employment (Table 7b) show that a typical Kagera wo rker who reported being sick with malaria lost 13.9 hours of work per week over the s ample period. This estimate is however insignificant at the 10% level. OLS and FE estimates of the causal effects of malaria on labor outcomes also indicate a loss in w eekly work hours. Although, the coefficients derived by the OLS and FE estimators a re significantly smaller in magnitude than the 2SLS result, the OLS and coefficients on m alaria are statistically significant at the 5% and 10% levels. Table 7b also shows a negative casual effect of ch ronic illness on the Kagera work sample. 2SLS estimates of the effect of chron ic illness for all types of employment indicates that being chronically ill decreases the number of work hours by approximately 2.5 hours a week. OLS and FE effects of chronic il lness on work hours are similar to those obtained by 2SLS and range between -2.68 and -1.45 hours per week. Estimates of the casual effect of chronic illness on weekly work hours are significant at the 1% level for 2SLS and OLS and at the 5% level for FE. 12 OLS and FE results are presented solely for the pu rpose of comparison. Given the endogeneity of malar ia status in the estimation of the schooling and labor supply models, OLS and FE estimates of the causal effect are biased.
74 Table 7b also shows the effect of occupation choic e on work hours. 2SLS estimates suggest that workers in farm employment a nd those in wage employment work more hours on average than those in self-employment All three estimators confirm that workers in farm employment worked 15 hours more per week than those in selfemployment; whereas, those in wage employment suppl ied 16-21 more hours of work per week compared to those in self-employment. All est imates on work hours by employment type are significant on the 1% level. Other statistically significant indicators of labo r supply for the Kagera adult sample are, age and gender, marital status, income, and community infrastructure. In terms of age and gender, 2SLS estimates show that b eing a year older increases work hours by approximately 0.040 hours per week, while being male increases work hours by 4.46 hours per week. 2SLS estimates also show that married workers work on average 1.77 hours more than their unmarried counterparts. Also, household income slightly increases the number of weekly hours by 0.006 hours in the 2SLS estimation. In terms of community infrastructure, access to a motorable road, electric power, and pipe borne water are all statistically signific ant indicators of work hours using 2SLS. Access to a motorable road increases the number of hours worked per week by 4 hours while, electric power supply increases work hours b y 1.5 hours per week. Accessibility of pipe borne water also increases weekly work hour s by 3.3 hours. Pharmaceutical and food prices proved insignificant in predicting work hours in the 2SLS, OLS, and FE estimates. In addition, education had no statistic ally significant effect on the number of hours worked per week for the Kagera work sample.
75 Table 7b: Second Stage 2SLS Results (without intera ction terms) Kagera Work Sample Dependent Variable: Number of Hours Worked Per Week (N = 5,876) Two Stage Least Squares Ordinary Least Squares Fixe d Effects Variables Coefficient Standard Error Coefficient S tandard Error Coefficient Standard Error Malaria -13.914 8.730 -1.368** 0.544 -1.184* 0.619 Farm Employment 15.306*** 0.502 15.737*** 0. 443 15.893*** 0.530 Wage Employment 20.783*** 0.576 21.093*** 0. 534 16.572*** 0.723 Age in Years 0.040** 0.017 0.024* 0.014 -0.215 ** 0.087 Gender 4.459*** 0.432 4.573*** 0.401 Education 0.238 0.170 0.217 0.156 -0.592* 0.349 Marital Status 1.786*** 0.416 1.848*** 0.386 2.558** 1.048 Income 0.006*** 0.000 0.006*** 0.000 0.00 4*** 0.001 Chronic Illness -2.468*** 0.608 -2.688*** 0.55 3 -1.452** 0.705 Electrical Power 1.529*** 0.572 1.043** 0.497 Pipe Borne Water 3.312*** 0.648 3.868*** 0.5 60 Motorable Road 4.114*** 1.310 4.723*** 1.234 Pharmaceutical Price 1.680 4.411 -1.199 3.170 -7.09 8 7.048 Food Price -1.025 1.410 -1.028 1.186 1.742 1.456 Constant -1.770 1.906 -3.767 1.452 14.276 3.406 Hausman Statistic (p-value) 24.43 (0.0274) ***1% Significance Level **5% Significance Level *1 0% Significance Level Note: Malaria = 1 if reported at least one episode of malaria in all 5 waves; 0 otherwise Farm employm ent = 1 if worker is a self employed farmer; 0 othe rwise. Wage Employment = 1 if worker is employed by a firm or government; 0 other wise (base group is self employment non-farm). Age = age in years; Gender = 1 if male, 0 otherwise; Ed ucation = 1 if worker has at least a primary school education; 0 otherwise; Marital st atus =1 if worker is married; 0 otherwise. Income = income in dollars; Chronic illness = 1 if reported at least one chronic illness; 0 otherwise Electrical power (this is a household lev el indicator) = 1 if worker resides in a residence with electrical power; 0 otherwise. Pipe Borne wate r (this is a community level indicator) = 1 if worker lives in area with access to pipe borne water; 0 otherwise. Motorable road (this is a comm unity level indicator) = 1 if worker lives in area with motorable road; 0 otherwise. Pharmaceutical prices = average price of selected p harmaceutical products priced at the community leve l (see chapter four for further description). Food prices = average price of selected food items priced at the community level (see chapt er four for further description)
76 Table 7c presents a comprehensive evaluation of th e causal relationship between the malaria incidence and work hours by type of emp loyment. This particular estimation differs from the parsimonious 2SLS estimation prese nted in Table 7b as interaction terms between malaria and the employment type indicators are introduced. Interacting malaria with employment variables facilitates the assessmen t of the causal effect of malaria from one sector to the other. The first stage estimates of this model are presented in the appendix on Table A7. The comparison group in this estimation is the self-employment group. The overidentification test for the comprehensive labor supply model is presented on Table A7. The 2c statistic and a p-value for the N*R-Sq test are 11 .168 and 0.0833, respectively. Therefore, the comprehensive model f ails the test overidentification test suggesting that all the instruments in the first st age are exogenous. The previous work sample results presented on Tabl e 7b suggested a statistically insignificant causal effect of malaria on weekly wo rk hours. Once malaria incidence is defined as a function of employment type, the casua l effect of malaria on labor supply (in terms of work hours) is now centered on farm employ ment. In essence, associating malaria with type of employment concentrates the ma laria effect into the farm employment sector where farmers lose approximately 35 hours of work per week compared to workers in self-employment. This estim ate is statistically significant at the 5% level in the 2SLS regression. In terms of the e ffect of malaria on wage employment, the comprehensive model produced statistically insi gnificant results indicating no decipherable causal relationship between malaria an d work hours for workers in wage employment compared to those in self-employment.
77 Although the 2SLS results in the comprehensive mod el, confirm a negative causal relationship between malaria incidence and work hou rs, the concern over the validity of the rainfall instruments in the first stage remains The failed overidentification test on Table A7 arises from the estimation of 2SLS in inst rumental variable settings with weak instruments. This particular problem was worsened by the inclusion of additional instruments to in the comprehensive model13. Recently in empirical economic literature, concern s have been raised over the reliability of inference based on conventional inst rumental variables settings using 2SLS when the instruments are only weakly correlated wit h the endogenous regressors. Older studies including that of John Bound, et al. (1995) emphasize the potential pitfalls of estimating 2SLS in the case of weak instruments. T he most citied example in these studies is that of Joshua Angrist and Alan Krueger (1991) (AK-91), which estimated wage equations using quarter of birth as an instrum ent for educational attainment. Paul Bekker (1994), Bound, et al. (1995), Gary Chamberla in and Guido Imbens (2004) and Jeffery Wooldridge and Guido Imbens (2007), to name a few, contend that the results obtained by AK-91 are influenced by weak instrument s. Although, the standard errors reported in AK-91 are reasonable and the specified model passed the overidentification test, the coverage properties are still very poor s uggesting that 2SLS estimates may be misleading especially in the case of many weak inst ruments Wooldridge and Imbens (2007). 13 Introducing interaction terms to the labor supply model to create a comprehensive model (Table A7) necessitated instrumenting for the additional endog enous regressors (i.e. the interaction terms betwee n malaria and employment type).
78 Table 7c: Second Stage 2SLS Results (with interacti on terms) Kagera Work Sample Dependent Variable: Number of Hours Worked Per Week (N = 5,876) Two Stage Least Squares Ordinary Least Squares Fixe d Effects Variables Coefficient Std. Err. Coefficient Std. Er r. Coefficient Std. Err. Malaria 4.909 11.851 -1.035 1.105 -2.184* 1.236 Malaria Farm Employment -35.100** 16.302 -0 .207 1.237 1.727 1.369 Malaria Wage Employment 14.946 21.015 -1.218 1. 536 -2.135 1.738 Farm Employment 19.961*** 2.200 15.773* ** 0.478 15.647*** 0.569 Wage Employment 19.263*** 2.624 21.24 8*** 0.568 16.802*** 0.749 Age in Years 0.041** 0.019 0.024* 0.014 -0 .215** 0.087 Gender 4.384*** 0.450 4.579*** 0.401 Education 0.181 0.193 0.214 0.156 -0.572* 0.349 Marital Status 1.760*** 0.483 1.854*** 0.3 86 2.586** 1.048 Income 0.006*** 0.001 0.006*** 0.000 0. 004*** 0.001 Chronic Illness -2.165*** 0.699 -2.680*** 0. 553 -1.449** 0.705 Electrical Power 1.753*** 0.642 1.043** 0.498 Pipe Borne Water 3.090*** 0.731 3.875*** 0. 560 Motorable Road 3.978*** 1.453 4.714*** 1.234 Pharmaceutical Price -5.343 5.304 -1.155 3.171 -6.9 99 7.047 Food Price -1.099 1.560 -1.030 1.186 1.728 1.456 Constant -4.125 2.298 -3.818 1.458 14.379 3.408 Hausman Statistic (p-value) 29.30 (0.0147) ***1% Significance Level **5% Significance Level *1 0% Significance Level Malaria*Farm = malaria interacted with farm employm ent indicator. Farm = 1 if worker is a self-employe d farmer; 0 otherwise. Malaria*Wage = malaria inter acted with wage employment indicator. Wage = 1 if worker is employe d by a firm or the government; 0 otherwise. Non-far m self employment is the omitted group.
79 5.7 Limited Information Maximum Likelihood (LIML) E stimation The main concern with estimating 2SLS in cases wit h weak or irrelevant instruments involves the construction of confidence intervals with good coverage properties. Wooldridge and Imbens (Summer, 2007) s tate that although, conventional methods such as 2SLS and LIML are rarely misleading when there are many weak instruments, LIML is generally much better than 2SL S. In particular, LIML has better properties compared to 2SLS when proportional adjus tments to the LIML standard errors are made. A simple adjustment to the LIML inferenc e statistics for the endogenous explanatory variable can be made as proposed in Mar celo Moreira (2001). Moreira developed reliable statistical inference tests with better coverage for structural parameters based on instrumental variable settings with weak i nstruments. Moreira showed that confidence regions based on the likelihood ratio (L R) statistics have coverage probabilities close to their nominal levels no matt er how weak the instruments. Others, such as Jean-Marie Dufour (1997), have also showed that LR-type tests behave more smoothly in the presence of identification problems compared to Wald-type tests14. The instruments in the labor supply estimation are weak as indicated by the overidentification test (Table A7). The result of estimation under weak instruments is that the obtained 2SLS inference statistics are pot entially misleading. A viable alternative to 2SLS under weak instruments is LIML. However, conventional LIML standard errors are invalid, but proportional adjus tments can be made. The labor supply and schooling equations are re-estimated in section 5.7.1 and 5.7.2 using the LIML estimator based on the conditional likelihood ratio (CLR) approach developed in Moreira. 14 Dufour (1997) shows that LR-type tests do not requ ire the indentifiability; a property not shared by Wald-type statistics.
80 In addition, the Anderson-Rubin statistic, the Lagr ange Multiplier score, and the Wald statistic are reported alongside the Moreira CLR st atistic for comparison. 5.7.1 Limited Information Maximum Likelihood (LIML) Estimation Results Kagera School Sample Table 8a provides the LIML estimates for the Kager a school sample. The objective of the LIML estimation is to provide pote ntially better inference tests given the possibility of weak instruments. Although the weak instrument problem is a bigger concern in the labor supply estimates, the LIML res ults provide a sensitivity check for the schooling estimates in section 5.6.2. The causal effect of malaria on the Kagera school sample obtained using the LIML estimator is similar to that obtained by the 2 SLS estimator. According to the LIML estimates, malaria is associated with a 24 hou r loss in school attendance per week. The CLR statistic for malaria indicates that the ma laria coefficient is statistically different from zero at the 1% level. The Anderson-Rubin stat istic, the Wald statistic, the Lagrange Multiplier score also indicate the statistical sign ificance of the causal effect. Table 8a also displays the estimated LIML coeffici ents for the other variables in the Kagera schooling model. Standard errors are inv alid for the LIML estimation therefore, additional inference on other variables in the model is invalid. Nonetheless, the estimated LIML coefficients are similar in magn itudes to those obtained in the 2SLS estimation of the Kagera schooling model. The dire ction of the malaria effect on the Kagera school sample generated by the LIML estimato r is also comparable that obtained via OLS and FE.
81 Table 8a: LIML Results Main Estimation Results: Kagera School Sample Dependent Variable: Hours of School Attended Per We ek (N = 4,189) LIML 2SLS Variables Coefficient Std. Error Coefficient Std. Error Malaria -24.064 8.951 -23.674*** 9.018 Age 0.8501 0.098 0.820*** 0.097 Grade 1 -0.562 0.683 -0.563 0.681 Grade 2 8.925 0.734 8.902*** 0.732 Grade 3 9.019 0.725 9.012*** 0.723 Grade 4 9.433 0.797 9.423*** 0.795 Grade 5 6.45 0.908 6.441*** 0.907 Grade 6 8.385 1.01 8.364*** 1.009 Gender -1.079 0.434 -1.076** 0.433 Mother Education -0.68 0.726 -0.407 0.624 Father Education -0.102 0.667 0.107 0.484 Income 0.002 0.0004 0.002*** 0.000 Distance to School -0.011 0.009 -0.012 0.009 School Fees 0.067 0.039 0.065 0.039 Number of Teachers 0.059 0.052 0.057 0.052 Chronic Illness -1.578 1.173 -1.462 1.205 Constant 15.675*** 1.958 Hausman Statistic (p-value) 31.42 (0.009) ***1% Significance Level **5% Significance Level *1 0% Significance Level CLR Test: H0: b[malaria] = 0.0000 Likelihood Ratio Statistic: 9.7279; Critical Value : 3.8415; Reject H0 Anderson-Rubin Statistic: 10.7741; 95% Critical Val ue: 7.8147; Reject H0 Wald Statistic: 7.1322; Critical Value: 3.8415; Rej ect H0 Lagrange Multiplier Score: 9.2363; 95% Critical Val ue: 3.8415; Reject H0
82 5.7.2 Limited Information Maximum Likelihood (LIML) Estimation Results Kagera Work Sample Table 8b presents the LIML estimation results for the Kagera work sample. The LIML estimates on Table 8b account for the weak ins trument problem. As indicated earlier, LIML estimates have improved coverage prop erties over 2SLS in settings with weak instruments. Therefore, a gain in statistical inference power is expected with LIML over 2SLS. Table 8a presents LIML results alongsid e 2SLS, OLS, and FE estimates for comparison. The LIML estimates of the causal effect of malari a on work hours are similar to those obtained using 2SLS with the exception of an improved significance of the coefficient. The 2SLS results described in section 5.6.4 revealed a negative casual effect of malaria on work hours, which was statistically s ignificant at the 15% level. LIML estimates of the same specification using rainfall amounts and a rainfall season indicator as instruments produced similar results. In partic ular, the LIML coefficient on malaria suggests that workers in Kagera lose approximately 15 hours to malaria per week. Table 8b also presents the test CLR, Anderson-Rubi n, Wald, and LM statistics for the work sample. The CLR test verifies the extent of the weak instruments problem and confirms that the malaria effect for the Kagera wor k sample is not different from zero. Anderson-Rubin, Wald, and LM tests also suggest tha t the malaria effect is not different from zero. In all, it is reasonable to conclude th at the rainfall instruments do not properly mitigate the endogeneity of malaria in the labor su pply equation.
83 Table 8b: LIML Results Main Estimation Results: Kagera Work Sample Dependent Variable: Number of Hours Worked Per Week (N = 5,876) LIML 2SLS Variables Coefficient Standard Error Coefficient S tandard Error Malaria -15.746 9.434 -13.914 8.730 Farm Employment 15.275 0.512 15.306*** 0.502 Wage Employment 20.771 0.585 20.783*** 0.576 Age in Years 0.041 0.017 0.040** 0.017 Gender 4.452 0.436 4.459*** 0.432 Education 0.234 0.172 0.238 0.170 Marital Status 1.777 0.42 1.786*** 0.416 Income 0.006 0.001 0.006*** 0.000 Chronic Illness -2.431 0.618 -2.468*** 0.608 Electrical Power 1.539 0.579 1.529*** 0.572 Pipe Borne Water 3.292 0.655 3.312*** 0.648 Motorable Road 4.082 1.325 4.114*** 1.310 Pharmaceutical Price -1.443 4.478 1.680 4.411 Food Price -0.923 1.437 -1.025 1.410 Constant -1.54 1.973 -1.770 1.906 Hausman Statistic (p-value) 24.43 (0.0274) ***1% Significance Level **5% Significance Level *1 0% Significance Level CLR Test: H0: b[malaria] = 0.0000 Likelihood Ratio Statistic: 3.4720; Critical Value : 3.8415; Fail to Reject H0 Anderson-Rubin Statistic: 6.9141; 95% Critical Valu e: 7.8147; Fail to Reject H0 Wald Statistic: 2.9395; Critical Value: 3.8415; Fai l to Reject H0 Lagrange Multiplier Score: 3.0182; 95% Critical Val ue: 3.8415; Fail to Reject H0
84 5.8 Difference in Difference Estimates of Changes i n the Disease Environment The results of the DID estimation of changes in th e disease environment due to increases in income levels and the number of health facilities in Kagera are presented below on Table 10. The Â‘yearÂ’ variable captures ch anges in malaria prevalence over time in Kagera from 1991-2004. As indicated below on Ta ble 10, an increase in reported malaria cases occurred between 1991 and 2004. Prec isely, the DID estimates show that malaria prevalence in the region increased by 43% p ercent over the sample period and is statistically significant on at the 1% level. In terms of improvements in the disease environmen t over time due to rising income levels, the derived DID estimates are counte rintuitive. The coefficient on the income interaction term is positive suggesting that malaria prevalence has increased over time with changes in income levels. Although, the magnitude of the coefficient on the income interaction is negligible, it is statistical ly significant at the 10% level. One reason for the observed relationship is that even though i ncome levels have risen over time, the disease environment may have worsened at a faster p ace. On the other hand, the effect of a rise in the num ber of health facilities is also suspect. The model failed to confirm a statistical ly significant improvement in the disease environment due to changes in the number of health facilities. The coefficient on the health facilities interaction with time is none theless negative although statistically insignificant.
85 Table 10: Improvements in the Disease Environment DID Estimates of Improvements in the Disease Enviro nment Dependent Variable: malaria (N = 14,647) Variable Coefficient Std. Err. z P>|z| income -0.0094 0.0056 -1.69 0.090 health facility -0.0254 0.1089 -0.23 0.816 year 0.4348 0.0361 11.41 0.000 year income 0.0098 0.0056 1.75 0.081 yr health facility -0.1394 0.1132 -1.23 0.218 _cons -1.3159 0.0299 -43.89 0.000 Chi-Sq Statistic = 197.15 P>Chi-Sq = 0.000 A joint significance test of the estimates in Tabl e 10 reveals the statistical significance of the income and health facilities in dicators in predicting changes in the disease environment over time. A consistent conclu sion about the disease environment cannot be reached from Table 10. However, the resu lts generated by the DID estimator make intuitive sense when looking at the reported c ases of malaria in Tanzania over time and the current state of the disease in the Kagera region. The upward trend in reported cases of malaria for Tanzania on Table A3 suggests an increased persistence of the disease irrespectiv e of the changes in the socio-economic environment. In addition, the recent report by the TRCNS of a rise in the number of malaria cases in Kagera for the year 2006 lends som e validity to the results obtained by the DID estimator. A 43% increase in malaria preva lence in the LSMS sample over time and the recent report by the TRNS in Kagera suggest a possible adaptation of the malaria virus over time to the disease environment. Eviden ce of such adaptation of the malaria virus to treatment in Sub-Saharan Africa has been d ocumented in recent epidemiological literature (Bloland 2001).
86 5.9 Other Findings This study also made additional attempts to better identify the labor supply equation using Average Treatment Effects on the Tre ated (ATET). The Average Treatment effects estimator has been mostly utilize d in the labor supply literature to measure the impact of specific interventions on an outcome of interest. In health economics literature, studies using the ATET estima tor analyze the effect of specific health interventions on health related outcomes (an example is Gabriel Picone, et al. 2006). As an extension to the analyses in section 5.6, the ATET estimator was also applied to both the labor supply and schooling mode ls. A description of the ATET estimator and how it was applied to the data as wel l as the results derived from the analyses are presented in Appendix D. As part of the identification strategy, a separate analysis carried out in the study involves substituting household income for a measur e of the household wage rate in the labor supply model. The structural labor supply mo del presented in chapter two defines labor supply as a function of household health, pri ces, wages, non-labor income, sociodemographics, and community level infrastructure. However, the analyses so far have relied on household income as a proxy for the wage rate. As an extension to the labor supply estimations in section 5.6, the model is reestimated using a proxy for the wage rate, which was defined in terms of the number of h ours worked per member of the household. This particular specification yielded c ausal effects identical to those obtained in previous regressions, but the proxy wage variabl e was statistically insignificant. A formal table of the results for the wage rate regre ssion will not be presented in this study.
87 Chapter 6 Conclusion This chapter summarizes the main findings of the e ffects of malaria on schooling and labor supply. Section 6.2 discusses the policy implications of the findings in the study. Additionally, the limitations of this resea rch and opportunities for future research are discussed in section 6.3 and 6.4, respectively. 6.1 Summary of Findings This dissertation is the first attempt to empirica lly measure the causal effect of malaria on schooling in Sub-Saharan Africa. It is also the first attempt to investigate the effect of malaria on different sectors of employmen t in the region. The causal effects of malaria on economic outcomes are examined for the y oung and adult segments of the population in the Kagera region of Tanzania using p anel data from 1991-2004. The main findings drawn from this study are summarized below Health status as described in chapter four is main ly measured as the incidence of malaria. Given that malaria status is endogenous i n equations (4.1) and (4.2), a seasonal indicator of the variations in rainfall patterns in Tanzania and a rainfall amount variable are used as instruments to mitigate the endogeneity problem, while also controlling in the estimation for the existence of chronic illness.
88 This dissertation finds a negative causal relation ship between malaria and school attendance for children of schooling age in Kagera. Following the same definition of health status, the effect of malaria on adults of w orking age in Kagera is inconclusive. In terms of the school sample, this dissertation fi nds that children of schooling age in Kagera may lose up to 24 hours of school per week d ue to malaria. The largest causal effects generated for the school sample was derived using the 2SLS and LIML estimator at 23-24 hours lost in school attendance per week. Standard regression methods such as OLS and FE generate modest estimates suggesting a 2 -3 hour decrease in weekly school attendance attributable to malaria. The attempt to mitigate the endogeneity of malaria in order to obtain unbiased estimates of the causal effect of malaria on labor supply proved insufficient. Although similar instruments were employed in the labor supp ly estimations and the results point to a negative causal effect, identification problems w ithin the model could not be resolved. Therefore, inference on the malaria effect is inval id and the overall evidence on the causal effect of malaria on the Kagera work sample is inconclusive. Even though this study finds that the effect of ma laria on healthy work days is inconclusive; it is important to note that this stu dy does not measure the impact of malaria on labor productivity. Malaria may very we ll impact labor productivity more than it influences labor supply. In addition, the short-term effects of missing work may also be mitigated by labor substitution; a phenomen on not studied in this dissertation. It is likely that the extent of economic loss from ill ness may be minimized through family labor substitution, even if the time contributed by a labor substitute is not as productive as
89 the time lost. Therefore, in some cases the tradeo ff is between the number of hours worked and labor productivity. Another element of health status in equation (4.1) and (4.2) is chronic illness. This study finds a negative effect of chronic illne ss on labor supply. In particular, chronic illness can be related to a loss of 1-2 hou rs in weekly school attendance for the school sample and similarly for the labor supply sa mple. The effect of chronic illness on school attendance is statistically insignificant. 6.2 Policy Implications This dissertation has shown that malaria limits th e number of healthy school days in Sub-Saharan Africa. The adverse effects of mala ria on schooling are likely to exceed school absenteeism, as absenteeism may increase fai lure and school dropout rates for children. Therefore, policies geared towards malar ia eradication in Sub-Saharan Africa will not only benefit the health of the population, but also enhance the quality of human capital in the region. Malaria eradication in SubSaharan Africa is feasible. However, some methods of eradication of the malaria virus ha ve been strongly opposed by policy makers. Successful efforts to eradicate malaria in the ind ustrialized world involved house spraying of residual insecticides (such as DDT) and anti-malarial drug treatment (CDC Â– Eradication of Malaria in the U.S.). While eradica tion was a success for industrialized countries, developing countries face a bigger obsta cle in their efforts to wipe out the disease. DDT spraying is by far the cheapest metho d of eradicating malaria. However, environmental concerns have led to its ban in most countries. Since many Sub-Saharan
90 African countries depend on international aid to fu nd many disease prevention and treatment programs, they face the likelihood of don or sanctions if DDT is used for eradicating malaria. Beyond eradication, other costs of malaria treatmen t and prevention include the costs of insecticide treated nests for households i n Sub-Saharan Africa and the emergence of drug resistant plasmodium falciparum. The direc t costs of malaria prevention and treatment also eat into the meager incomes of house holds in Sub-Saharan Africa. Guyatt Brooker, et al. (2000) showed that the actual costs of prevention and treatment for households in Sub-Saharan Africa are unrealisticall y expensive in a region where most people live on less than $1 a day. Ettling, et al. (1994) also evaluated the financial burden of malaria on households in Sub-Saharan Africa and found that over 25% of earnings in low-income households in Malawi is spent on malaria treatments. The findings in this dissertation suggest that mala ria does limit the number of healthy school days for children in Sub-Saharan Afr ica. Cumulative effects of absenteeism can indeed translate in to indirect costs to societ y in terms of human capital accumulation. Effective malaria control policies may salvage such losses and improve overall economic welfare in the region. 6.3 Limitations A major limitation of this research is the lack of strong instruments for mitigating the endogeneity of malaria, especially for the labo r supply model. Finding a better measure of the disease environment that is correlat ed with malaria incidence should improve the indentifiability of the causal effect f or the labor supply sample. Other
91 potential instruments for malaria in the labor supp ly and schooling models are the cost of anti-malaria drugs and geographic elevation. Altho ugh, the LSMS collected data on prices for Chloroquine, a malaria treatment drug, n o information is available in the survey on anti-malaria drugs. Drakeley, et al. (20 05) determined that malaria is negatively correlated with altitude in Tanzania. A lthough data on geographic elevation lacks significant variation, this problem can be ov ercome by interacting geographic elevation with other variables such as regional ide ntifiers in the model. In addition to the problem of weak instruments, an other problem faced in this study is the unavailability of panel type data for other countries in Sub-Saharan Africa. There are relatively few panel data collected on ho useholds in Sub-Saharan Africa. Although LSMS datasets are available for 5 countrie s in Sub-Saharan Africa to date, most of these datasets are limited to one wave whil e others lack detailed information on household health. 6.4 Future Research Future research should consider the cumulative imp act of malaria on labor supply and schooling. This can be done by estimating the probability of repeating a grade due to repeated episodes of malaria for the children of sc hool age. The cumulative impact of malaria on labor supply can be analyzed through est imating the extent of job loss stemming from multiple episodes of malaria. Future research should also address the costs and benefits of malaria eradication and antimalaria treatment initiatives possibly through field experiments.
92 References Ainsworth, Martha; Beegle, Kathleen and Koda, Godli ke. 2005. "The Impact of Adult Mortality and Parental Deaths on Primary Schooling in North-Western Tanzania." The Journal of Development Studies 41:3, pp. 412-39. Amexo, Mark; Tolhurst, Rachel and Barnish, Guy, et al. 2004. "Malaria Misdiagnosis: Effects on the Poor and Vulnerable." Lancet 364, pp. 1896-98. Audibert, Martine. 1986. "Agricultural Non-Wage Pro duction and Health Status." Journal of Development Economics 24, pp. 275-91. Angrist, Joshua D. and Krueger Alan B. 1991. "Does Compulsory School Attendance Affect Schooling and Earnings?" The Quarterly Journal of Economics CVI: 4, pp. 9791014 Bartel, Ann and Taubman, Paul. 1979. "Health and La bor Market Success: The Role of Various Diseases." The Review of Economics and Statistics 61:1, pp. 1-8. Becker, Gary S. 1965. "A Theory of the Allocation o f Time." The Economic Journal 75, pp. 493-517. Becker, Sascha O. and Ichino, Andrea. 2002. "Estima tion of Average Treatment Effects Based on Propensity Scores." The Stata Journal 2:4, pp. 358-77. Beegle, Kathleen. 2003. "Labor Effects of Adult Mor tality in Tanzanian Households." World Bank Policy Research Working Paper 3062 Bekker, Paul A. 1994. "Alternative Approximations o f the Distributions of Instrumental Variable Estimators." Econometrica 62:3, pp. 657-81. Bloland, Peter B. 2001. "Drug Resistance in Malaria ." World Health Organization Communicable Disease Surveillance & Response (CSR) 4. Bound, J. 1991. "Self-reported versus objective mea sures of health in retirement models." Journal of Human Resources 26, pp. 106-38. Bound, John; David, Jaeger A. and Baker, Regina M. 1995. "Problems with Instrumental Variables Estimation When the Correlation between t he Instruments and the
93 Endogeneous Explanatory Variable is Weak." Journal of the American Statistical Association 90:430, pp. 443-50. Brooker, S; Guyatt, J. and Omumbo, R., et al. 2000. "Situation Analysis of Malaria in School-aged Children in Kenya What Can Be Done?" Parasitology Today 16:5, pp. 183-86. Case, Anne; Paxson, Christina and Ableidinger, Jose ph. 2002. "Orphans in Africa." NBER Working Paper 9213 Center for Disease Control and Prevention. 2006. Ma laria. http://www.cdc.gov/malaria/disease.htm (accessed September 22, 2006) Chamberlain, Gary and Imbens, Guido. 2004. Random e ffects estimators with many instrumental variables.Â” Econometrica 72:1, pp. 295-306. Chirikos, Thomas N. and Nestel, Gilbert. 1984. "Eco nomic determinants and consequences of self-reported work disability." Journal of Health Economics 3, pp. 11736. ______. 1985. "Further Evidence on the Economic Eff ects of Poor Health." The Review of Economics and Statistics 67:1, pp. 61-69. Currie, Janet and Madrian, Brigitte C. 1999. "Heal th, Health Insurance and the Labor market." Handbook of Labor Economics 3, pp. 3309-415. Dehejia, Rajeev H. and Wahba, Sadek. 2002. "Propens ity Score-Matching Methods for Nonexperimental Causal Studies." The Review of Economics and Statistics 84:1, pp. 151-61. Deressa, Wakgari. 2007. "Treatment-seeking behavior for febrile illness in an area of seasonal malaria transmission in rural Ethiopia." Malaria Journal 6:49. Drakeley, Chris; Carneiro, Ilona and Reyburn, Hugh, et al. 2005. "Altitude Dependent and -Independent Variations in Plasmodium falciparum Prevalence in Northeastern Tanzania." The Journal of Infectious Diseases 191, pp. 1589-98. Dufour, Jean-Marie. 1997. "Some Impossibility Theor ems in Econometrics With Applications to Structural and Dynamic Models." Econometrica 65:6, pp. 1365-87. Ettling, M, McFarland, DA and Schultz, LJ, et al. 1 994. "Economic Impact of Malaria in Malawian Households." Tropical Medicine and Parasitology 45:1, pp. 74-9.
94 Ettner, Susan L.; Frank, Richard G. and Kessler, Ro nald C. 1997. "The Impacts of Psychiatric Disorders on Labor Market Outcomes." Industrial and Labor Relations Review 51:1, pp. 64-81. Fox, Matthew P.; Rosen, Sydney and MacLeod, William B., et al. 2004. "The Impact of HIV/AIDS on Labor Productivity in Kenya." Tropical Medicine and International Health 9:3, pp. 318-24. Grossman, Michael. 1973. "The Correlation Between H ealth and Schooling." NBER Working Paper 22 Grossman, Michael and Benham, L. 1974. Â“Health, hou rs, and wages. In The Economics of Health and Medical Care (Mark Perlman, ed.). London. Haveman, Robert; Wolfe, Barbara and Kreider, Brent, et al. 1994. "Market work, wages, and men's health." Journal of Health Economics 13:1994, pp. 163-82. Imbens, Guido and Wooldridge, Jeffrey. Summer 2007. "Estimation of Average Treatment Effects Under Unconfoundedness." NBER Lecture Notes _____. 2007. "Weak Instruments and Many Instruments ." NBER Lecture Notes Kahn, Matthew E. 1998. "Health and Labor Market Per formance: The Case of Diabetes." Labor Economics 16:4, pp. 878-99. Kessler, Ronald C. and Frank, Richard G. 1997. "The imapact of psychiatric disorders on work loss days." Psychological Medicine 27, pp. 861-73. Leighton, Charlotte and Foster, Rebecca. 1993. "Eco nomic Impacts of Malaria in Kenya and Nigeria." A Health Financing and Sustainability Project (Abt Associates). Lloyd, Cynthia B. and Blanc, Ann K. 1996. "Children 's Schooling in Sub-Saharan Africa: The Role of Fathers, Mothers, and Others." Population and Development Review 22:2, pp. 265-98. Luft, H.S. 1975. Â“The Impact of Poor Health on Earn ings.Â” Review of Economics and Statistics 57 (February), pp. 43-57. McCombie, S.C. 1996. "Treatment Seeking for Malaria : A Review of Recent Research." Social Science and Medicine 43:6, pp. 933-45. Miguel, Edward and Kremer, Michael. 2004. Â“Worms: I dentifying impacts on education and health in the presence of treatment externaliti es.Â” Econometrica 72:1, pp. 159-217.
95 Mitchell, Jean M. and Burkhauser, Richard V. 1990. "Disentangling the effect of arthritis on earnings: a simultaneous estimate of wage rates and hours worked." Applied Economics 22, pp. 1291-309. Moreira, Marcelo J. 2001. "Tests With Correct Size When Instruments Can be Arbitrarily Weak." Center for Labor Economics Working Paper Series, 37 UC Berkeley Murphy, Sean C. and Breman Joel G. 2001. "The Child hood Malaria Burden in Africa: Cerebral Malaria, Neurological Sequelae, Anemia, Re spiratory Distress, Hypoglycemia, and Complications of Pregnancy." American Journal of Tropical Medicine and Hygiene 64:(1,2)S, pp. 57-67. Nicholson, Walter. 2002. Microeconomic Theory: Basi c Principles and Extensions. Mason, Ohio: Southern-Western (Thomas Learning) Nur, El Tahir Mohamed. 1993. "The Impact of Malaria on Labor Use and Efficiency in the Sudan." Social Science and Medicine 37:9, pp. 1115-19. Picone, Gabriel; Khwaja, Ahmed and Salm, Martin. 20 06. "A Comparison of Treatment Effect Estimators Using Severity of Illness Informa tion from Hospital Charts." Forthcoming, Journal of Applied Econometrics Roll Back Malaria (RBM). 2006. Malaria in Africa. h ttp://www.rollbackmalaria.org/ (accessed August 25, 2006). Rosenfield, P.L; Golladay, F. and Davidson, R.K. 19 84. "The Economics of Parasitic Diseases: Research Priorities." Social Science and Medicine 19:10, pp. 1117-26. Sachs, Jeffery and Malaney, Pia. 2002. "The economi c and social burden of malaria." Nature 415, pp. 680-85. Sascha, Becker O. and Ichino, Andrea. 2002. "Estima tion of average treatment effects based on propensity scores." The Stata Journal 2:4, pp. 358-77. Shapiro, Barbara S.; Dinges, David F. and Orne, Emi ly C., et al. 1995. "Home management of sickle cell-related pain in Children and adolescents: natural history and impact on school attendance." Pain 61, pp. 139-44. Shchultz, Paul T. and Tansel, Aysit. 1983. "Measure ment of Returns to Adult Health: Morbidity Effects on Wage Rates in Cote d'Ivoire an d Ghana." Living Standards Measurements Study Working Paper No. 95 Societies, International Federation of Red Cross an d Red Crescent. 27 February 2007. "Tanzania: Malaria Outbreak in Kagera." DRF Bulletin
96 Strauss, John and Thomas, Duncan. 1998. "Health, Nu trition, and Economic Development." Economic Literature 36:2, pp. 766-817. The Global Fund. 2006. Fighting Malaria. http://www.theglobalfund.org/en/about/malaria/ (accessed August 25, 2006). _____. 2006. Fighting AIDS. http://www.theglobalfun d.org/en/about/aids/ (accessed September 22, 2006). _____. 2006. Fighting Tuberculosis. http://www.theglobalfund.org/en/about/tuberculosis/ (accessed September 22, 2006). Thirumurthy, Harsha; Zivin, Joshua G. and Goldstei n, Markus. 2006. "The economic Impact of AIDS Treatment: Labor Supply in Western K enya." Job Market Paper United Nations ChildrenÂ’s Fund (UNICEF). 2006. Heal th Â– Malaria. http://www.unicef.org/health/index_malaria.html. (accessed September 22, 2006). United States Agency for International Development (USAID). 2006. Accelerating the Fight Against Malaria. http://www.usaid.gov/press/f actsheets/2005/fs050915.html (accessed August 2, 2006). Wang'Ombe, Joseph K and Mwabu, Germano M. 1993. "Ag ricultural Land Use Patterns and Malaria Conditions in Kenya." Social Science and Medicine 37:9, pp. 1121-30. Wooldridge, Jeffery M. 2003. Introductory Econometr ics: A Modern Approach. Mason, Ohio: South-Western (Thomson Learning). World Health Organization (WHO) Health System Stati stics. 2006. http://www.who.int/whosis/database/core/core_select _process.cfm?countries=all&indicat ors=nha (accessed August 25, 2006). _____. 2006. International Travel and Health. http: //www.who.int/ith/en/ (accessed June 23, 2006). World Health Organization (WHO) Â– Roll Back Malaria (RBM). 2006. The Africa Malaria Report 2003. http://www.rbm.who.int/amd2003/amr2003/amr_toc.htm (accessed November 13, 2006). _____. 2006. World Malaria Report 2005. http://www.rbm.who.int/wmr2005/html/toc.htm (accessed June 23, 2006).
97 Appendix A: Tables Table A1: Number of reported cases of HIV, TB and M alaria Country HIV Prevalence ( 2005) TB Prevalence ( 2004) Malari a Prevalence ( 2002) Global 38 600 000 14 602 353 408,388,001 Sub-Saharan Africa 24 500 000 3 843 167 43 614 238 Angola 320 000 48,006 1409328 Benin 87 000 9,779 Â… Botswana 270 000 46,815 28858 Burkina Faso 150 000 41,049 1451125 Burundi 150 000 36,473 1808588 Cameroon 510 000 1,556 Â… Central African Republic 250 000 21,882 Â… Chad 180 000 53,448 Â… Comoros 500 740 Â… Congo 120 000 18,035 2640168 Cte d'Ivoire 750 000 116,349 Â… Democratic Republic of Congo Â… 307,554 Â… Djibouti 15 000 1,585 5021 Equatorial Guinea 8900 18,498 Â… Eritrea 59 000 403,098 75386 Ethiopia Â… 4,619 427831 Gabon 60 000 4,858 Â… Gambia 20 000 81,480 Â… Ghana 320 000 37,739 2830784 Guinea 85 000 4,713 Â… Guinea-Bissau 32 000 297,330 194976 Kenya 1 300 000 9,782 124197 Lesotho 270 000 14,493 Â… Liberia Â… 63,606 Â… Madagascar 49 000 63,159 1543130 Malawi 940 000 75,845 2853317 Mali 130 000 14,975 723077 Mauritania 12 000 1,670 167423 Mauritius 4100 123,360 22 Mozambique 1 800 000 11,767 4458589 Namibia 230 000 38,850 442527 Niger 79 000 683,847 681707 Nigeria 2 900 000 58,658 2605381 Rwanda 190 000 387 Â… Senegal 61 000 51,383 Â… Sierra Leone 48 000 66 Â… Somalia 44 000 45,215 96922 South Africa 5 500 000 316,260 15649 Sudan 350 000 131,543 3056400 Swaziland 220 000 11,580 14863 Togo 110 000 43,012 Â… Uganda 1 000 000 179,843 7216411 United Republic of Tanzania 1 400 000 180,069 74898 90 Zambia 1 100 000 81,187 Â… Zimbabwe 1 700 000 87,006 1252668 North Africa 440 000 78 978 424 Algeria 19 000 17,432 307 Egypt 5300 25,364 10 Libyan Arab Jamahiriya Â… 1,146 107 Morocco 19 000 32,640 Â… Tunisia 8700 2,396 Â… WHO Estimates 2005 WHO Estimates 2004 WHO Estimat es 2002
98 Table A2: Per Capita Health Expenditures in Sub-Sah aran Africa Per capita total expenditure on health at average e xchange rate (US$) Country 2000 2001 2002 2003 2004 Angola 15.5 21.7 18 23.9 25.5 Benin 15.1 16.7 16.8 20.8 24.2 Botswana 129.8 129.8 143.9 231.6 328.6 Burkina Faso 11.1 11.3 13.7 18.9 24.2 Burundi 3.4 3.1 3 2.7 3 Cameroon 30.7 31.3 34.8 44.3 50.7 Cape Verde 55.5 61.3 66.2 77.7 97.8 Central African Republic 9.8 9.7 10.5 12.1 13.2 Chad 10.7 11.8 12.3 16.3 19.6 Comoros 8 7 9.6 12.7 13.2 Congo 19.8 19 20.1 24.1 27.6 Cte d'Ivoire 29.5 24.1 25.6 30.8 33 Democratic Republic of the Congo 9.8 4.3 3.3 4.1 4. 7 Equatorial Guinea 42.9 64.9 166.6 111.2 168.2 Eritrea 8.9 8.4 7.4 8.7 9.9 Ethiopia 5.1 5 4.9 5 5.6 Gabon 164.4 190.7 190.3 206.3 231.3 Gambia 21 20 16.9 17.4 18.5 Ghana 17.4 18.6 18.8 24.2 27.2 Guinea 17.2 18 20 22.1 21.8 Guinea-Bissau 7.4 6.2 7 7.6 8.7 Kenya 18.1 17.7 18.6 19.9 20.1 Lesotho 28 30.8 25.8 39.7 49.4 Liberia 5.8 6.1 5.1 5.4 8.6 Madagascar 6 10 10.3 10.9 7.3 Malawi 9.2 11.3 16 18.3 19.3 Mali 14.3 16.1 16.3 20.7 23.8 Mauritania 11.3 9.1 13.3 13.9 14.5 Mauritius 142.2 142.2 167 180.3 222.3 Mozambique 11.3 9.3 10.3 10.6 12.3 Namibia 126.1 109.9 97.4 147.4 189.8 Niger 5.3 6.2 6.4 7.6 8.6 Nigeria 18 19.1 18.8 20.8 23 Rwanda 9 8 8.3 13.4 15.5 Sao Tome and Principe 20.8 30.4 31.4 45.2 47.8 Senegal 19 20.3 24.2 31.6 39.4 Seychelles 403.5 406.5 463.1 515.4 534.4 Sierra Leone 5.4 5.8 6.8 6.8 6.6 South Africa 235.6 216 198.4 300.7 390.2 Swaziland 82.7 67.3 63 111.4 145.8 Togo 11 12.1 12.9 15.6 17.9 Uganda 15.6 16.9 17.7 17.5 19 United Republic of Tanzania 10.8 10.9 10.4 10.6 12 Zambia 17.1 18.7 21.2 23.9 29.6 Zimbabwe 44.4 65.1 150.5 39.9 27.2 Source: World Health Statistics
99 Table A3: Annual Reported Malaria Cases Â– Sub-Sahar a Countries (WHO Estimates) Annual Reported Malaria Cases Sub-Sahara Countrie s Country 2003 2002 2001 2000 1999 1998 1997 1996 199 5 1994 1993 Angola .. 1,409,328 1,385,597 1,635,884 1,471,993 1 ,169,028 893,232 .. 156,603 667,376 722,981 Benin .. .. 779,041 707,408 709,348 650,025 670,857 623,396 579,300 546,827 403,327 Botswana 22,418 28,858 48,237 71,403 72,640 59,696 101,887 80,004 17,599 29,591 55,331 Burkina Faso .. 1,451,125 1,203,640 1,032,886 867,8 66 721,480 672,752 582,658 501,020 472,355 502,275 Burundi .. 1,808,588 2,855,868 3,057,239 1,936,584 687,301 670,857 974,226 932,794 831,481 828,429 Cameroon .. .. .. .. .. 664,413 787,796 931,311 784 ,321 189,066 478,693 Cote d'Ivoire .. .. 400,402 1,491,943 .. .. 983,089 1,109,011 755,812 .. 421,043 Central African Rep 95,644 .. 140,742 89,614 127,96 4 105,664 99,718 95,259 100,962 82,057 82,072 Chad .. .. 386,197 369,263 392,815 395,205 343,186 278,048 293,564 278,225 234,869 Comoros .. .. 3,718 9,618 9,793 3,844 .. 15,509 15, 707 13,860 12,012 Congo .. .. .. .. .. 17,122 9,491 14,000 28,008 35, 957 15,504 Djibouti 5,036 5,021 4,312 4,667 6,140 5,920 4,314 6,105 5,982 6,140 4,166 Equatorial Guinea .. .. .. .. .. .. .. .. 12,530 14 ,827 17,867 Eritrea 72,023 75,386 125,746 119,155 147,062 255,1 50 .. 129,908 81,183 .. .. Ethiopia 565,273 427,831 400,371 383,382 647,919 60 4,960 509,804 478,411 412,609 358,469 305,616 Gabon .. .. .. .. .. 80,247 57,450 74,310 54,849 82 ,245 70,928 Gambia .. .. .. .. 127,899 .. 325,555 266,189 135,9 09 299,824 .. Ghana 3,552,869 2,830,784 3,383,025 3,349,528 2,895 ,079 1,745,214 2,227,762 2,189,860 1,928,316 1,672, 709 1,697,109 Guinea .. .. .. 889,089 807,895 817,949 802,210 772 ,731 600,317 607,560 .. Guinea-Bissau .. 194,976 202,379 246,316 197,454 2, 113 10,632 6,457 197,386 .. 158,748 Kenya .. 124,197 132,590 74,194 122,792 80,718 .. 3 ,777,022 4,343,190 6,103,447 .. Liberia .. .. .. .. .. 777,754 826,151 239,998 .. . .. Madagascar 2,114,400 1,543,130 1,429,491 1,383,239 1,141,474 .. .. .. 196,358 .. .. Malawi .. 2,853,317 2,955,627 3,774,982 4,193,145 2 ,985,659 2,761,269 6,183,290 .. 4,736,974 4,686,201 Mali 809,428 723,077 612,895 546,634 530,197 12,234 384,907 29,818 95,357 263,100 295,737 Mauritania .. 167,423 243,942 259,093 253,513 168,1 31 189,571 181,204 214,478 156,080 43,892 Mauritius .. 22 61 62 73 52 65 82 46 65 54 Mozambique 5,087,865 4,458,589 3,978,397 3,278,525 2,336,640 194,024 .. 12,794 .. .. .. Namibia 444,081 442,527 537,115 519,113 429,571 353 ,110 390,601 345,177 275,442 401,519 380,530 Niger .. 681,707 606,802 646,757 815,895 872,925 97 8,855 1,162,824 778,175 806,204 726,666 Nigeria 2,608,479 2,605,381 2,253,519 2,476,608 1,9 65,486 2,122,663 1,148,542 1,149,435 1,133,926 1,17 5,004 981,943
100 Table A3: Annual Reported Malaria Cases Â– Sub-Sahar a Countries (WHO Estimates) Annual Reported Malaria Cases Sub-Sahara Countrie s (Contd.) Country 2003 2002 2001 2000 1999 1998 1997 1996 199 5 1994 1993 Rwanda 856,233 .. .. 915,916 906,552 1,279,581 1,33 1,494 1,145,759 1,391,931 371,550 733,203 Senegal .. .. .. 1,120,094 1,145,112 948,823 861,27 6 .. 628,773 450,071 .. Sierra Leone .. .. .. .. 409,670 249,744 209,312 7, 192 .. .. .. Somalia 23,349 96,922 10,364 10,364 9,055 .. .. .. .. .. 3,049 South Africa 13,446 15,649 26,506 64,622 51,444 26, 445 23,121 27,035 8,750 10,289 13,285 Sudan 3,084,320 3,056,400 3,985,702 4,332,827 4,215 ,308 5,062,000 4,065,460 4,595,092 6,347,143 8,562, 205 9,867,778 Swaziland 36,664 14,863 19,799 45,581 30,420 4,410 23,754 38,875 .. .. .. Tanzania 10,712,526 7,489,890 .. .. 423,967 30,504, 654 1,131,655 4,969,273 2,438,040 7,976,590 8,777,3 40 Togo .. .. 431,826 398,103 412,619 368,472 366,672 352,334 .. 328,488 561,328 Uganda 12,343,411 7,216,411 5,622,934 3,552,859 3,0 70,800 2,845,811 2,317,840 .. 1,431,068 2,191,277 1 ,470,662 Zambia .. .. 2,010,185 1,139,489 2,992,203 3,399,63 0 .. 3,215,866 2,742,118 3,514,000 3,514,000 Zimbabwe .. 1,252,668 1,609,296 1,533,960 1,804,479 1,719,960 1,849,383 1,696,192 761,791 324,188 877, 734 Source: WHO Reported Malaria Cases
101 Table A4: Construction of Price Variables Food Items Measurement Pharmaceutical Items Measure ment Raw Cassava Grams Aspirin Tablets Dry Cassava Grams Paracetamol/Panadol Tablets Cassava Flour Grams Nivaquine/Cloroquine Tablets Maize Flour Grams Sorghum Grams Finger Millet Grams White Bread Grams Rice Grams Sugar Grams Sweet Potato Grams Irish Potato Grams Kidney Beans Grams Salt Grams Groundnuts Grams Tomato Grams Cooking Banana Grams Sweet Banana Grams Orange Grams Cooking Oil Grams Local Brew Grams Tea Leaves Grams Onions Grams Chicken Eggs Grams Chicken Grams Beef Grams Goat's Meat Grams Fish Grams Peas Grams Fresh Milk Grams Milk Powder Grams
102 Table A5: Summary Statistics Â– Total Work Sample Kagera Work Characteristics (1991-2004) Variable Description Obs Mean Std. Dev. Min Max Work Hours Hours Number of hours worked per week 11,955 19.89 19.14 0 130.00 Employment Type Farm Employed Self Employment in the Agricultural S ector 11,050 0.71 0.45 0 1.00 Self Employed Self Employment in Non-Agricultural Sector 11,040 0.15 0.36 0 1.00 Wage Employed Wage Employment in Private or Governm ent Sector 11,042 0.18 0.38 0 1.00 Socio-demographic Characteristics Age Age in Years: Range 18 65 11,955 33.94 13.52 18 65.00 Gender Percent Male: 1 = Male; 0 Otherwise 11,955 0.45 0.50 0 1.00 Education Highest Level of Education: 1 = Primary; 2 = Secondary; 3 = College 11,955 1.84 1.29 0 6.00 Married Marital Status : 1 = Married; 0 Otherwise 11,955 0.56 0.50 0 1.00 Income Annual Income: Employment income in dollars 11,955 189.93 409.61 0 7401.59 Health Status Malaria Malaria Cases = 1 if reported malaria in at least 1 wave; 0 otherwise 11,955 0.15 0.36 0 1.00 Chronic Illness Cases of Chronic Illness = 1 if reported chronic il lness in at least 1 wave; 0 otherwise 11,955 0.10 0.30 0 1.00 Disease Environment Malaria Season Rainfall Season: 1 = Interviewed dur ing rainy season; 0 otherwise 11,955 0.81 0.39 0 1. 00 Rainfall Total Monthly Rainfalla: Total Monthly Rainfall measured in millimeters 10,2 04 341.27 280.68 1 872.30 Community Infrastructure Electric Power Electric Power: 1= Access; 0 otherwi se 12,009 0.41 0.49 0 1.00 Pipe Water Pipe Borne Water: 1 = Access; 0 otherwis e 12,009 0.20 0.40 0 1.00 Road Motorable Road: 1 = Access; 0 otherwise 12,009 0.96 0.21 0 1.00 Pricesb Food Average Food Price in dollars 9,394 0.17 0.21 0.002 3.80 Pharmaceuticals Average Pharmaceutical Price in dol lars 7,707 0.04 0.12 0.001 0.70 aThis variable is constructed using estimates from t he Living Standard Measurement Survey for Tanzania (1991-994) and NOAA National Data Centers (2004); Food items are weight ed to the nearest 50 grams and pharmaceutical items are measured in tablets
103 Table A6 First Stage 2SLS Results (with interaction terms) Â– Kagera Work Sample2SLS First Stage Dependent Variable: Malaria Dependent Variable: Malaria in Farm Employment Dependent Variable: Malaria in Wage Employment 1st Stage Explanatory Variables Coefficient Std. Er r. Coefficient Std. Err. Coefficient Std. Err. Farm Employment -0.0796 0.0872 0.2497*** 0.0751 0.0 067 0.0334 Wage Employment 0.0015 0.1099 0.0325 0.0947 0.2946* ** 0.0421 Age in Years 0.0009*** 0.0003 0.0006** 0.0003 0.000 2 0.0001 Gender -0.0034 0.0095 -0.0027 0.0082 0.0032 0.0036 Education -0.0007 0.0037 -0.0032 0.0032 -0.0017 0.0 014 Married -0.0077 0.0091 -0.0029 0.0078 0.0035 0.0035 Income -9.00E-06 0.0000 -5.10E-06 9.32E-06 6.88E-08 4.14E-16 Chronic Ill 0.0214* 0.0128 0.0228** 0.0110 0.0069 0 .0049 Electric Power 0.0053 0.0127 0.0111 0.0110 0.0017 0 .0049 Pipe Water -0.0094 0.0143 -0.0121 0.0123 0.0025 0.0 055 Motorable Road -0.0264 0.0288 -0.0255 0.0248 -0.010 5 0.0110 Pharmaceutical Price 0.0940 0.0948 0.0073 0.0816 0. 0960*** 0.0363 Food Price 0.0503* 0.0291 0.0281 0.0211 -0.0017 0.0 112 Instruments Rainfall Season -0.1169 0.0741 0.0199 0.0638 0.0112 0.0284 Total Rainfall -0.0395*** 0.0138 0.0011 0.0119 0.00 22 0.0053 Total Rainfall Squared 4.28E-07*** 1.58E-07 -3.46E08 1.36E-07 -6.66E-08 6.05E-08 Season*Farm Employment 0.0352 0.0909 -0.1015 0.0783 -0.0113 0.0348 Season*Wage Employment -0.2905 0.1139 -0.0452 0.981 0 -0.1548*** 0.0436 Total Rainfall*Farm Employment 0.0173 0.0150 -0.022 0* 0.0130 -0.0028 0.0058 Total Rainfall*Wage Employment 0.0089 0.0174 -0.002 0 0.0149 -0.0213 0.0066 Total Rainfall Sq*Farm Employment -1.71E-07 1.71E-0 7 2.69E-07 1.47E-07 8.47E-07 6.56E-08 Total Rainfall Sq*Wage Employment -7.46E-08 1.94E-0 7 1.01E-07 1.67E-07 2.69E-07*** 7.45E-08 Constant 0.3046 .0767 -0.0099 0.0661 -0.0075 0.0294 F-Statistic (P-value) 2.49 (0.0001) 11.06 (0.00 00) 34.97 (0.0000) Joint Significance IVs Only (P-value) 3.29 (0.00 05) 2.38 (0.0109) 5.40 (0.0000) Overidentification Test N*R-Sq (P-value) 11.168 ( 0.0833) Omitted group in first stage is Non-farm Self Emplo yment group; Overidentification Test (N*R-Sq): Reje ct H0; IVs are exogenous
104 Appendix B: Figures Figure B1: Estimated incidence of clinical malaria episodes Â– caused by any species Â– resulting from l ocal transmission, country level Averages, (2004) Figure B1 is derived from: WHO Â– World Malaria Repo rt 2005
105 Figure B2: RosenfieldÂ’s Conceptual Framework for Tr opical Illness Figure B2 is derived from: Rosenfield (1984)
106 Figure B3: Relationship between Episodes of Malaria and Rainfall Amounts MALARIA AND RAIN AMOUNTS0 20 40 60 80 100 120 140 160 02004006008001000 RAIN (MM)# OF MALARIA CASES BY WAVE # of Malaria Cases Figure B3 shows monthly rainfall estimates plotte d against malaria cases per month for the year 1991 -1994 and 2004.
107 Appendix C: Two Period Household Health, Human Capi tal, and labor Supply Model Health and Education Model ) ), ( ( ) ), ( ), ( (2 2 2 2 2 2 1 1 1 1 1 1 1 1 1 H H H H H K K K K K H K H K H HL L HI H C U L L HI H L HI H C Ub+(3.2.1Â¢) ) ~ ( ) ( ) ( ) ( ) (2 2 1 1 1 1 2 1 1 2 1 K H H K K K H K H K s H K K H HI H H cS L w L w V L w S P HI HI HI P C C P + + + = + + + + +(3.2.6) ) , ; (1 1 1 1 1 1 K H K H K H K H K H K He D A L HI H H-=m (3.2.2) ) , ; (1 1 1 1 1 1 K K K K K Ke D A L HI H Hm= (3.2.3) ) , , ; (2 2 1 1 2 2 2 2 H H K K H H H H He D A H H L HI H Hm-= (3.2.4) ) , ; ( ~ ~ 1 1KS K K K K Ke SI A H S S Sm= (3.2.5) ) , , ; (1 1K Hw K H K H K H K He IN E A L H w w--=a (3.2.7) ) , , ; , (1 1 1 1 1 1Kw K K k K Ke IN E A S L H w wa= (3.2.8) ) , , ; ~ , (2 2 2 2 2Kw K H H H He IN E A S L H w wa= (3.2.9) Utility Maximization Problem: Period 1 ) , ; ), ( ( ) , ; ), ( ), ( ( max2 2 2 2 2 2 1 1 1 1 1 1 1 1 1x b xE A L L HI H C U E A L L HI H L HI H C U U LH H H H H K K K K H K H K H H+ = =-] ) ( ) ( ) ~ ( ) ( ) ( [1 2 1 1 2 1 2 2 1 1 1 1 K s H K K H HI H H c K H H K K K H K HS P HI HI HI P C C P S L w L w V L w + + + + + + +-l 0 :1 1 1= C H HP C U Cl 0 :1 1 1 1 1= HI K H K H K H K HP HI H H U HIl 0 ] ~ ~ :1 2 1 1 1 1 1= + K K K H HI K K K KHI S S w P HI H H U HIl
108 0 ~ ~ :1 2 1= + -K K K K S KS S S w P Sl l Rewritten, S K K K KP S S S w = 1 2 ~ ~ Demand for Education in period one ) , , ; (1 1KS K S K Ke SI A P H S Sm= (3.3.1) Figure C1 SPSP 1 KS ) , , ; (1KS S Ke SI A P H Sm KS
109 Health and Labor Productivity Model Case 1: H* is endogenous; Individual chooses C, HI, and L ) , ; ), ( ( ) , ; ), ( ), ( ( max2 2 2 2 2 2 1 1 1 1 1 1 1 1 1x b xE A L L HI H C U E A L L HI H L HI H C U U LH H H H H K K K K H K H K H H+ = =-] ) ( ) ( ) ~ ( ) ( ) ( [1 2 1 1 2 1 2 2 1 1 1 1 K s H K K H HI H H c K H H K K K H K HS P HI HI HI P C C P S L w L w V L w + + + + + + + +-l 0 :2 2 2= C H HP C U Cl 0 :2 2 2 2 2= HI H H H HP HI H H U HIl 0 :2 2 2 2 2 2= + H H H H Hw L H L U Ll Reduced Form Equations for C, HI, and L ) , , , , (a mIN D E A V P P C CHI C= (3.4.1) ) , , , , (a mIN D E A V P P HI HIHI C= (3.4.2) ) , , , , (2a mIN D E A V P P L LHI C H= (3.4.3) Case 2: H is given; Individual chooses C, and L 0 :2 2 2= C H HP C U Cl 0 :2 2 2 2 2= + H H H HH w L U Ll
110 Structural Equation for L ) , , , ; (2 2 2a mIN E A V P w H L LC H H H= (3.4.4) Figure C2 Hw Hw 2 HL HL ) , , , 2 ; 2 (a mIN E A V c P H w H H L
111 Dictionary of Variables in Theoretical Framework A = Socio-demographic Characteristics (age, gender, and marital status) 1 HC = Household Consumption of Non-Health Goods in Per iod One 2 HC = Household Consumption of Non-Health Goods in Per iod Two D = Disease Environment E = Education 1 K HH-= Household Health less that of the Child in Period One 1 KH = Child Health in Period One 2 HH = Overall Household Health in Period Two HI = Health Inputs 1 K HHI-= Household Health Inputs less that of the Child in Period 1 1 KHI = Child Health Inputs in Period 1 2 HHI = Overall Household Health Inputs in Period 2 IN = Community Infrastructure 1 L = Labor Supply in Period One 2 L = Labor Supply in Period Two 1 K HL= Labor Supply (Household minus Child) in period o ne 1 KL = Labor Supply (Child) in period one 2 HL = Household Labor Supply in period two HIP = Price of Health Inputs SP = School Fees
112 1KS = Child Schooling in Period 1 KS ~ = Accumulated Human Capital by Period 2 SI= School Infrastructure V = Non-Labor Income 1 K Hw-= earnings (Household minus Child) period one 1 Kw = earnings (Child) period one 2 Hw = Household earnings period two a, x, , e = error terms
113 Appendix D: Analysis of Malaria, Labor Supply and S chooling using Average Treatment Effects on the Treated (ATET) OLS, FE and 2SLS, and LIML estimates of the causal effect of health status on schooling and labor supply outcomes were presented in the previous sections. This section explores a different estimation procedure f or further identification of the parameters in (1) of chapter four. The approach ch osen for further identification of the schooling and labor supply equations is based on th e method of average treatment effects on the treated (ATET) with instrumental variables. Unlike the estimation procedure detailed in section 4.4.1 of chapter four, which ig nores the different states of treatment within the sample, estimating ATET considers differ ent states of the health status indicator. In general, estimating equation (1) of chapter fou r by 2SLS assumes that the predicted person specific gain from treatment is ze ro given a vector of observables. This is equivalent to the ATET case of estimating equati on (3) below: ) ( ) | (0 0X g h X h y E+ + =a m (3) Where h is the endogenous binary treatment indicator with 1 = h denoting treatment and 0 = h otherwise, X denotes a vector of observable variables, and ) (0X g is the expected value of the person specific effects of treatment c onditional on the observables. In previous chapters, it was assumed0 ) | (0 1= -X v v E ; where v1 and v0 are the person specific effects of treatment. If this assumption holds true, the gain from treatment is the same for everyone in the treatment group in (3) and estimating the effect of health status on the schooling and labor supply models by the sta ndard 2SLS procedure is sufficient.
114 If the0 ) | (0 1 -X v v E the effect of malaria is not the same for everyon e in the sample and 2SLS is insufficient. Clearly, making s uch a restrictive assumption as 0 ) | (0 1= -X v v E is highly unrealistic in this case, as one may exp ect that the effect of malaria on school and work hours vary from one indi vidual to another. In other words, to allow for varying effects of malaria on the school and work outcomes, the assumption of zero person-specific gains from treatment is relaxe d to derive the ATET estimator: )] ( ) ( [ ) | (0 1 0X g X g h h X h y E+ + =a m (4) Where ) ( ) (1 0X g X gmeasures the average gain of treatment to a parti cipant with characteristics X. To identify the parameters in (4), two assumptions are required about treatment assignment. The first assumption i s referred to as conditional independence or selection on observables or unconfo undedness. Assumption (1a): Unconfoundedness or Conditional In dependence X h y y | ) (1 0 ^ Conditional independence states that conditional on X the outcomes are independent of treatment15. A weaker conditional independence assumption can be specified if one is only interested in the treatment effect on the trea ted. Assumption (1b): Weak Conditional Independence X h y |0 ^ Assumption (1b) implies that there is no omitted variable bias once X is controlled for. Alternatively, this implies that m alaria status in our sample is exogenous. Assumption (1b) is not sufficient in our case knowi ng the likely endogeneity of malaria 151y and 0yare the outcomes for the treated and non-treated gr oups, respectively.
115 status; therefore, the analysis relies on assumptio n (1a). Assumption (1a) is not directly testable however; there are indirect ways to assess it. The test of conditional independence relies on estimating the causal effect that is known to equal zero. If the null hypothesis is rejected, then the likelihood of conditional independence is less plausible (Imbens and Wooldridge 2007). In additio n to (1a), equation (4) also relies on the overlap or matching assumption. Assumption (2): Overlap 1 ] |1 Pr[ 0 < = < X h Under assumption (2), there exist treated and nontreated cases for each value of X in the sense that there is overlap between the tr eated and untreated sub-samples. Although, this assumption is not required to identi fy the treatment parameters for the treated group, it is needed for identifying the tre atment parameter for a randomly selected individual within the sample. The lack of overlap can be detected by plotting the distribution of the covariates by treatment groups. Because this becomes cumbersome in cases with more than two covariates, a direct metho d of determining overlap is to examine the distribution of propensity score in bot h treatment groups. ATET Method: Propensity Score Matching Methods Matching by propensity score is recommended as a w ay to control for confounding factors that may bias the effect of tre atment in observational studies (Sascha Becker and Andrea Ichino 2002)16. Propensity score matching estimates are superior to 16 According to Becker and Ichino, Â“propensity score matching method is a way to correct the estimation of treatment effects controlling for the existence of these conf ounding factors based on the idea that the bias is reduced when the
116 other matching estimators when the dimensionality o f observable characteristics ( X ) is high (Rajeev Dehejia and Sadek Wahba 2002). When X is of a small dimension, matching is straightforward and non-propensity scor e matching methods are sufficient for deriving unbiased estimates of the treatment effect This study, explores 4 matching methods namely, stratification matching, nearest-ne ighbor matching, radius matching, and kernel matching methods, all of which can be de fined in terms of propensity scores. The propensity score ) ( X p is the probability of an individual receiving trea tment conditional on observables, X. If 1 ) ( 0 < < X p and ) ( X p is a consistent estimate of) ( X p, then a consistent estimator of the ATET can be ob tained using a parametric probit or logit model as: n =r r= = N i i i i i N i ix p Y x p h N h N ATET1 1 1 1 1)) ( Âˆ 1 /( )] ( Âˆ [ In the analyses to follow, consistent ATET estimates are obtained using 4 propensity score matching methods with the logit model. Matching on propensity score is based on pairing c omparable treated and control units in terms of observable characteristics, which are assumed independent of treatment. Due to the unobserved heterogeneity problem with th e health status indicator, malaria, an instrumental variables approach to ATET will be use d to mitigate this problem. This is arrived at in the estimation process by adding the instruments to the list of covariates when estimating the propensity scores which will be used in calculating the treatment effects. comparison of outcomes is performed using treated a nd control subjects who are as similar as possibleÂ” (Becker and Ichino 2002).
117 The stratification, nearest neighbor, radius, and kernel matching methods are the four most widely used propensity score matching met hods. The reason for using all four methods in this study is mainly for comparison. Ea ch one of these propensity score matching methods differ by the process of selection of the comparison units (matches) and selection of one method over another boils down to a tradeoff between the quality and quantity of matches (Becker and Ichino 2002). The stratification matching method provides ATET estimates which are based on dividing the range of variation of the propensity s cores into intervals (blocks) such that within each block, treated and untreated units have on average the same propensity score. After matching, the ATET estimates are then obtaine d as an average of the ATET of each block with weights given by the distribution of the treated units across blocks. One disadvantage of this method is that it ignores obse rvations in blocks where either treated or untreated units are missing. On the other hand, the nearest neighbor method makes up for this deficiency by taking each treated unit and locating the untreated unit with the closest propensity score as a match (i.e. the neare st neighbor to each treated unit). After matching is complete, the ATET estimates are derive d as the average of the difference between the outcome of the treated units and that o f the control/untreated units. While this method ensures that all treated units find a m atch, it does not necessarily improve the quality of the matches for those, which would have been disregarded using the stratification method. The radius method offers a solution to the above d eficiency. The radius matching method pairs the treated units only with the untrea ted units whose propensity score fall in a predefined radius (neighborhood) of those of the treated units. The downside to this
118 method lies in the specified radius in that, if the radius is set too small, it is possible that no matches exits in the neighborhood because the n eighborhood contains no untreated units. Alternatively, specifying a small radius al so improves the quality of the matches. Another matching approach is the kernel method, whi ch is based on matching treatment units using a weighted average of all control/untre ated units with weights that are proportional to the distance between the propensity scores of the treated and untreated units. All ATET estimates for the schooling and la bor supply equations are presented below. ATET Assumptions Although the ATET assumption of unconfoundedness i s not directly testable, Imbens and Wooldridge (2007) propose an indirect te st (described under assumption (1)). When the indirect test is applied to the labor supp ly and schooling model, neither model passes the conditional mean independence test. In addition to the test of conditional independence, Imbens and Wooldridge (2007) suggest a direct way of assessing overlap in ATET estimations (described under assumption (2) ). Two histogram plots of the propensity scores for the labor supply and schoolin g models are used to access the overlap assumption (Figure D1). The histogram plot s show considerable evidence of overlap in the distribution of the propensity score s for both samples.
119 ATET Estimation Results Â– Comparison of Means (Scho ol Sample) Table D1 presents the sample means for the two com parison groups of the Kagera schooling sample. Of the 4,951 observations from t he Kagera school sample, 549 fall into the treatment group for the ATET analyses. A comparison of means by the number of school hours attended per week suggests that tho se without treatment attend on average 3 hours more school per week than those wit h treatment (i.e. children reporting an episode of malaria). However, children in the t reated group were of similar age to those in the control group. In terms of the gender composition of both groups, females were just as likely to be in the treated group as m ales. Surprisingly, parents of children in the treatment category are more likely to be educat ed than those without treatment even though those in treatment are more likely to come f rom households with lower income. Furthermore, the likelihood of falling into treatme nt does not seem to vary by the preexistence of chronic illness since fewer children i n the treated group have a chronic illness. Table D1: ATET Comparison of Means Â– Kagera School Sample Sample Means of Characteristics for the Kagera Scho ol Sample (N = 4,189) N Age Gender Mother Education Father Education Household Income Chronic Illness Malaria Season Total Rainfall School Hours Treated Group Mean 549 12.51 0.5 0.27 0.29 157.48 0.02 0.32 349.6 6 25.58 Std. Dev. 12.51 0.50 0.44 0.45 355.13 0.14 0.47 294.79 13.97 Control Group Mean 4402 12.52 0.53 0.21 0.26 171.31 0.03 0.22 381 .6 28.26 Std. Dev. 3.00 0.50 0.41 0.44 452.02 0.17 0 .42 287.74 12.76 Note: School hours = number of hours of school atte nded in a week; Age = age in years; Gender = 1 if m ale, 0 otherwise; Mother/Father education = if mother/father has at l east a primary school education, 0 otherwise; House hold Income = income in dollars; Chronic illness = 1 if reported at least o ne chronic illness, 0 otherwise; Malaria season = 1 if month of interview falls within malaria season, 0 otherwise; Total Rainfall = rain amounts in mm.
120 ATET Estimation Results Â– Treatment Effects (School Sample) Four types of matching estimators in the estimatio n of the treatment effects were explored in the ATET analyses. All estimates of th e treatment effect use rainfall amounts and a rainfall season indicator as instruments. Th e results of all four matching estimators are presented on Table D2 for the purpose of compar ison. The radius matching method by definition matches treated units only with contr ol units whose propensity score fall in a predefined radius of those of the treated units. The analyses using radius matching first compares neighbors within a (0.001) radius; then, t he radius is gradually expanded for a more realistic comparison of both groups given the tight range of the estimated propensity scores (Table D2)17. Radius matching estimates of the treatment effect range between 2.397 and 3.865 school hours per week lost due to malaria illness. Using the nearest neighbor, stratification, and kernel matching methods all gen erate the same treated effects as the radius method. All estimates of the treatment effe ct are statistically significant but far smaller in magnitude than the 2SLS estimates derive d in section 5.6.2. The estimates produced by the propensity score matching methods a re however very close to those derived with the OLS and FE estimators. Findings o f treatment effects similar to OLS are not uncommon in practice. Picone, et al. (2006 ) estimated treatment effects on 17 Note that the tradeoff of expanding the radius to match pairs with very close propensity scores is a decrease in the number of possible matches. As show n in Table D2, the number of close matches falls drastically when a smaller radius is defined.
121 individuals with Acute Myocardial Infraction (AMI) using Medicare claim data and found no advantages over using a standard least squ ares regression. Table D2: ATET Results Â– Kagera School Sample Malaria Impact: Estimates of treatment effect Propensity Score Range [.04736965, .40351615] Matching Method Number Treated Number in Control AT ET Standard Error Radius r = 0.001 524 4171 -2.397 0.650a r = 0.0001 507 3140 2.793 0.673a r = 0.00001 331 609 -3.865 0.932a Nearest Neighbor 525 485 -2.811 0.866a Stratification 461 3715 -2.706 0.692a Kernel 525 4231 -2.639 0.617a a Analytical standard b Bootstrapped standard errors with 250 replication s 5.7.4 ATET Estimation Results Â– Comparison of Means (Work Sample) A comparison of means by work category is presente d in Table D3. Comparison of means for the farm employment category yielded 1 208 treated units and 6637 control units. In terms of non-farm self-employment, 275 w orkers fall into the treated group while, 1369 make up the control group. On the othe r hand, wage employment generated 316 treated units and 1668 control units. Although non-farm self-employment had the least number of workers in both treated and control groups this employment category also had the highest percentage of workers in treatment as a percentage of total workers in each group. Seventeen percent of non-farm self-emp loyed workers reported having malaria compared to 15% of self-employed farmers an d 16% of workers in wage employment.
122 As shown on Table D3, the largest difference in ho urs worked per week between workers in the treated group and those in the contr ol group is also for the non-farm self employment category. Workers in the treated group for non-farm self employment worked 10 fewer hours per week than those in the tr eated farm employment group and 23 hours less per week than those in treated wage empl oyment category. In terms of wages, treated workers in self-employment earn less income than their other treated counterparts. In particular, treated units in wage employment ear n on average 65% more income than treated units in self-employment while, treated uni ts in farm employment earn 7% more than treated units in self-employment. Also, a majority of treated units in farm employme nt are more likely to be female and less educated than the treatment units in other employment classifications. In terms of pre-existing health conditions, treated and untr eated farmers are more likely to have an existing chronic condition than workers in other gr oups. In addition, the treated units are in every work classification on average more likely to be chronically ill compared to workers in the control units. On the contrary, comparisons of means for the rain fall and malaria season variables yielded counter-intuitive results of both treated and control units for each work sample. For all three classification of employment rainfall amounts for the survey period are greater for workers in the treated units compar ed to those in the control groups. Likewise, those in treatment are less likely to rep ort being in malaria season compared to those in the control group. This calls to question the validity of the two instruments in the ATET estimations to follow. However, as shown earlier in figure B3, the overall data
123 suggests a non-linear relationship between episodes of malaria by wave and rainfall amounts, which may in part explain the observation on Table D3. ATET Estimation Results Â– Treatment Effects (Work S ample) Panels A D of Table D4 present the treatment eff ects for the dependent variable Â‘work hoursÂ’ using the radius, nearest neighbor, st ratification, and kernel matching methods. As indicated earlier, the farm employment group is the largest of the 3 work categories. The analysis begins with the treatment effect with the farm employment work group (Panel A; Table D4). According to the ATET estimates using 4 types of m atching methods, self employed farmers in Kagera who reported being sick during the sample period lost between 1.712 and 2.640 hours of weekly work hours. The lowest treatment effect for this work group was derived by the 0.001 radius mat ching method. Although, the 0.00001 radius generated the largest treatment effe ct (-2.640 hours), it also paired the least number of workers in both treatment and contr ol groups (Panel A; Table D4). All estimates of the treatment effects for the farm emp loyment sample are statistically significant at the 1% level.
124 Table D3: ATET Comparison of Means Â– Kagera Work Sa mple Sample Means of Characteristics for the Kagera Work Sample Age Sex Education Marital Status Income Chronic Illness Electric Power Pipe Borne Water Mo torable Road Drug Prices Food Prices Malaria Season Total Rainfall (mm) Work Hours Self Employment (Farm) Treated Group (N = 1208) Mean 35.52 0.38 1.64 0.63 157.27 0.12 0 .41 0.18 0.94 0.03 0.19 0.79 322.85 18.60 Std. Dev. 13.74 0.49 1.23 0.48 335.45 0.32 0.49 0.38 0.2 3 0.10 0.26 0.41 284.36 12.02 Control Group (N = 6637) Mean 34.93 0.42 1.93 0.57 187.54 0.10 0. 37 0.17 0.95 0.03 0.16 0.90 364.53 19.79 Std. Dev. 13.97 0.49 1.30 0.49 384.93 0.31 0.48 0.38 0. 21 0.08 0.21 0.30 276.68 12.27 Self Employment (Non-Farm) Treated Group (N = 275) Mean 34.01 0.57 1.80 0.61 146.78 0.11 0. 61 0.24 0.96 0.04 0.23 0.70 261.06 8.30 Std. Dev. 12.38 0.50 1.08 0.49 419.07 0.32 0.49 0.43 0. 20 0.12 0.31 0.46 269.47 14.24 Control Group (N = 1369) Mean 33.37 0.60 2.06 0.58 244.70 0.09 0. 62 0.20 0.97 0.04 0.19 0.84 319.68 12.49 Std. Dev. 11.93 0.49 1.16 0.49 634.92 0.29 0.49 0.40 0. 18 0.11 0.27 0.37 275.04 18.10 Wage Employment Treated Group (N = 316) Mean 32.66 0.68 1.81 0.55 423.78 0.09 0. 57 0.27 0.95 0.04 0.21 0.72 290.57 31.25 Std. Dev. 11.20 0.47 1.21 0.50 610.39 0.28 0.50 0.44 0. 22 0.10 0.24 0.45 279.74 23.01 Control Group ( N = 1668) Mean 32.36 0.73 2.18 0.50 360.78 0.07 0 .56 0.26 0.96 0.03 0.19 0.86 343.11 33.01 Std. Dev. 11.72 0.44 1.28 0.50 568.36 0.26 0.50 0.44 0. 19 0.10 0.27 0.35 280.68 22.68 Note: work hours = number of hours worked per week. Age = age in years; Gender = 1 if male, 0 otherwis e; Education = 1 if worker has at least a primary s chool education; 0 otherwise; Marital status =1 if worker is married; 0 otherwise. Income = income in dollars; C hronic illness = 1 if reported at least one chronic illness; 0 otherwise Electrical power (this is a h ousehold level indicator) = 1 if worker resides in a residence with electrical power; 0 otherwise. Pipe Borne water (th is is a community level indicator) = 1 if worker li ves in area with access to pipe borne water; 0 othe rwise. Motorable road (this is a community level in dicator) = 1 if worker lives in area with motorable road; 0 otherwi se. Pharmaceutical prices = average price of select ed pharmaceutical products priced at the community level Food prices = average price of selected food items priced at the community level Malaria season = 1 if month of interview falls within malaria season, 0 otherwi se; Total Rainfall = rain amounts in mm.
125 Panel C details the results of the ATET estimation s for the wage employment sample. The results of in Table D4 show that worke rs in wage employment lost -11.05 hours weekly work hours to malaria during the surve y period. This estimate is significant at the 1% level using the 0.0001 radius matching me thod. No other treatment estimates of radius matching method or other matching algorit hms are significant for the wage employment sample and will not be discussed. Results for the total sample using all four matchi ng methods are presented in panel D. For the entire work sample, a 0.001 radiu s matching and kernel matching methods produce statistically significant estimates of the ATET at the 5% level between 1.052 and -1.537 hours. Similar results were gener ated using a 0.0001 radius matching, which was significant at the 10% level. As seen previously with the ATET results for the s chool sample, propensity score estimates may produce results parallel to OLS. Wit h the exception of the ATET results derived for the self-employment and wage employment samples the estimated treatment effects mirror the causal effects generated using s tandard OLS. In addition, propensity score methods appear to mirror OLS estimates the cl oser the range of the propensity score estimates (Table D4).
126 Table D4: ATET Results Â– Kagera Work Sample Panel A Self Employment (Farm) Malaria Impact: Es timates of treatment effect Propensity Score Range [.06469756, .44094733] Matching Method Number Treated Number in Control ATET Standard Error Radius r = 0.001 561 3495 -1.712 0.549a r = 0.0001 536 2468 -2.029 0.586a r = 0.00001 310 442 -2.640 0.882a Nearest Neighbor 564 505 -2.246 0.73a Stratification 510 3445 -1.860 0.562a Kernel 564 3558 -1.785 0.532b Panel B Self Employment (Non-Farm) Malaria Impact : Estimates of treatment effect Propensity Score Range [.0134235, .78899154] Matching Method Number Treated Number in Control ATET Standard Error Radius r = 0.001 58 142 -6.178 2.724a r = 0.0001 15 15 -8.767 8.034a r = 0.00001 2 2 11.75 11.750a Nearest Neighbor 78 65 -4.513 3.485a Stratification 70 318 -3.925 2.180a Kernel 78 347 -4.814 2.119b Panel CWage Employment Malaria Impact: Estimates of treatment effect Propensity Score Range [.05082005, .56096773] Matching Method Number Treated Number in Control ATET Standard Error Radius r = 0.001 86 397 -2.763 2.739a r = 0.0001 38 53 -11.051 3.709a r = 0.00001 5 5 -6.400 10.668a Nearest Neighbor 98 84 1.250 3.793a Stratification 79 529 -4.029 2.823a Kernel 98 580 -1.131 2.543b Panel D Malaria Impact (Total): Estimates of trea tment effect Propensity Score Range [.06072613, .22507454] Matching Method Number Treated Number in Control ATET Standard Error Radius r = 0.001 842 5360 -1.537 0.694a r = 0.0001 821 4554 -1.318 0.713a r = 0.00001 641 1240 0.650 0.923a Nearest Neighbor 843 757 0.723 0.944a Stratification 843 5428 -1.052 0.682a Kernel 843 5428 -1.425 0.630b a Analytical standard b Bootstrapped standard errors with 250 replication s
127 Figure D1: ATET Assumptions Â– Overlap Test of Overlap Â– Total Kagera School Sample Test o f Overlap Â– Total Kagera Work Sample 0 10 20 30 .1 .2 .3 .4 .1 .2 .3 .4 0 1DensityEstimated propensity scoreGraphs by malaria 0 10 20 30 .1 .15 .2 .25 .3 .1 .15 .2 .25 .3 0 1DensityEstimated propensity scoreGraphs by malaria
128 About the Author Born in Nigeria, Taiwo Abimbola relocated to the U nited States of America in 1998 to obtain her undergraduate degree from Lincol n University of Pennsylvania. She graduated summa cum laude with a B.A. in Economics from Lincoln University in 2002. In 2004, she earned an M.A. in Economics at the Uni versity of South Florida (USF). Thereafter, Taiwo was accepted into the PhD program at the USF College of Business, where her main areas of research were Health Econom ics and Development Economics. While attending USF, Ms. Abimbola was an instructor for both Economics of Health Care and Principles of Microeconomics courses.