Economic impacts of illness, coping strategies, health insurance and crowding-out effects : dynamic evidence from a financial diary survey in Nigeria

E

CONOMIC IMPACTS OF

I

LLNESS

,

C

OPING

STRATEGIES

,

H

EALTH INSURANCE AND

C

ROWDING

-

OUT EFFECTS

:

D

YNAMIC EVIDENCE FROM A FINANCIAL DIARY SURVEY IN

N

IGERIA

MASTER THESIS

AMSTERDAM,FEBRUARY 2016

ABSTRACT

Using a unique and detailed, high frequency diary dataset from Nigeria, this paper investigates the financial impacts of and responses to illness, comparing outcomes within individuals across states of health insurance enrollment. Dynamic fixed effects panel data models produce results indicating that the costs to illness are significant, with large increases in OOP expenditures and forgone earnings. To cope with these costs inter- and intra-household risk-sharing networks, saving accounts, and consumption out of stocks are extensively used and ranked in order of decreasing magnitudes. Significant intra-household cooperation is found, although the collective household model is rejected. Despite this, uninsured individuals are unable to smooth consumption across health shocks, resulting in detrimental welfare effects. Enrollment in health insurance vastly improves wellbeing by fully mitigating consumption volatility during illness. Surprisingly, insurance seems to significantly reduce OOPs only for females and never so for forgone earnings, although all estimates have the expected signs and are of considerable size, especially so for forgone earnings. Moreover, evidence is inconclusive, but suggestive, of inter-household transfers being partially crowded-out by formal insurance. Ultimately, the results suggest health insurance is a welcome addition to current risk coping strategies delivering substantial welfare gains, especially so for rural populations and females. The results are robust to heterogeneity in insurance uptake, state dependence of consumption and reverse causality issues.

JEL classification: D04, D14, I11, I13, I15, O12

Key words: Risk management, Health insurance, Health shocks, Coping mechanisms, Consumption

smoothening, Risk sharing, Nigeria

Author: Arno de Jager UvA ID: 6158021

Supervisor: Prof. M. Pradhan

B

R

De Jager, A. (2016). Economic impacts of Illness, Coping strategies, Health insurance and crowding-out effects: Dynamic evidence from a financial diary survey in Nigeria. University of Amsterdam, Graduate School of Economics, Master Thesis, p. 97.

S

O

This document is written by Student Arno de Jager who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

A

First of all, I would like to thank my thesis supervisor Professor Menno Pradhan who has always been readily available with insightful feedback when I got stuck and energized and motivated me throughout the sometimes lonesome process with various lengthy and lively discussions. Second, I would like to extend appreciation to Wendy Janssens and Marijn van der List in helping me thoroughly understand the Kwara state Health Insurance program and financial diary dataset. Without their help I would still be stuck analyzing the endless data files. Third, I would like to thank the AIID and the PharmAccess Foundation for trusting in my research capabilities and permitting me the opportunity to evaluate their insurance program and survey data.

On a more personal note, I would like to thank Bob, Nathan and Taco for numerous refreshing, enjoyable and longed-for coffee breaks. Further gratitude goes to Mandy, Jim and Frank for making the economics transition program so much more workable and fun. Last but not least, I am deeply grateful to my parents for their never-ending support and to Nina for upgrading my life in so many ways.

T

C

1. INTRODUCTION 1

2. RELATIONSHIP TO EXISTING LITERATURE 4

2.1. WELFARE EFFECTS OF ILLNESS: COSTS AND CONSUMPTION SMOOTHENING 4

2.2. GROUP-BASED RISK SHARING 6

2.2.1. INFORMAL RISK SHARING: INTER-HOUSEHOLD TRANSFERS 6

2.2.2. INFORMAL RISK SHARING: INTRA-HOUSEHOLD TRANSFERS 7

2.2.3. FORMAL RISK SHARING: MICRO HEALTH INSURANCE 8

2.3. SELF-COPING MECHANISMS 11

2.3.1. SAVINGS 11

2.3.2. ASSET SALES 12

3. THEORETICAL PREDICTIONS 13

3.1. EFFECTS ON OOPS,INCOME, AND CONSUMPTION SMOOTHENING (H1,H2,H3) 13

3.2. EFFECTS ON INTRA- AND INTER-HOUSEHOLD TRANSFERS (H4,H5) 14

3.3. EFFECTS ON SAVINGS AND ASSET SALES (H6,H7) 15

4. EMPIRICAL FRAMEWORK 16

4.1. EMPIRICAL SPECIFICATION 16

4.2. IMPROVEMENTS OVER PREVIOUS MODELS TESTING FOR RISK-SHARING 18

4.3. IDENTIFYING ASSUMPTIONS 18

5. RESEARCH SETTING AND DATA COLLECTION STRATEGY 21

5.1. RESEARCH SETTING:KWARA STATE,NIGERIA 21

5.2. KWARA STATE HEALTH INSURANCE PROGRAM 22

5.3. SAMPLING AND DATA COLLECTION STRATEGY 23

6. DESCRIPTIVE STATISTICS 25

6.1. DEMOGRAPHIC AND FINANCIAL VARIABLES 25

6.2. HEALTH SHOCKS AND THE BEHAVIORAL RESPONSE 28

6.3. DYNAMIC FINANCIAL FLOWS IN RESPONSE TO ILLNESS 32

7. RESULTS 34

7.1. COSTS OF HEALTH SHOCKS 34

7.1.1. OUT-OF-POCKET (OOP) HEALTH EXPENDITURES 34

7.1.2. FORGONE EARNINGS DUE TO ILLNESS 37

7.2. CONSUMPTION SMOOTHENING ABILITY 38

7.2.1. INDIVIDUAL SMOOTHENING 39

7.2.1. HOUSEHOLD WIDE SMOOTHENING 41

7.3. DISAGGREGATION OF COPING MECHANISMS 42

7.3.1. INFORMAL GROUP-BASED RISK SHARING ARRANGEMENTS: TRANSFERS 42

7.3.2. SELF-COPING MECHANISMS: SAVINGS AND ASSET SALES 45

7.4.1. GENDER DIFFERENCES 47

7.4.2. RURAL AND URBAN DISPARITIES 47

7.5. COPING WITH ILLNESS – AN ACCOUNTING OVERVIEW 50

7.5.1. COSTS AND COPING STRATEGIES VISUALIZATIONS 52

7.5.2. COSTS AND COPING STRATEGIES VISUALIZATIONS 53

7.6. ROBUSTNESS CHECKS 54

7.6.1. HEALTH INSURANCE AND ASSET SHOCKS 54

7.6.2. STATE DEPENDENCE AND REVERSE CAUSALITY 55

7.6.3. NUMBER OF OBSERVATIONS IDENTIFYING INSURANCE EFFECTS 57

8. CONCLUSION 59

8.1. DISCUSSION OF RESULTS 60

8.1.1. WHY DOES INSURANCE HELP SMOOTH CONSUMPTION WHEN IT FAILS TO SIGNIFICANTLY REDUCE

ILLNESS-RELATED COSTS? 62

8.2. POLICY IMPLICATIONS 63

9. REFERENCES 66

10. APPENDIX 70

10.1. APPENDIX A:THEORETICAL MODEL DERIVATION 70

10.1.1. THE STANDARD RISK SHARING MODEL AND THE IDENTIFICATION PROBLEMS IT FACES 70

10.1.2. TRANSFERS AS IDENTIFIER OF INFORMAL RISK-SHARING 73

10.1.3. EFFECTS OF HEALTH INSURANCE 74

10.2. APPENDIX B:COVERAGE LIST OF KSHI PLAN 75

10.3. APPENDIX C:DATA SEASONALITY 76

10.4. APPENDIX D:DYNAMIC FINANCIAL FLOW CHARTS 77

10.4.1. COSTS OF HEALTH SHOCK 77

10.4.2. SMOOTHENING ABILITY 78

10.4.3. COPING MECHANISMS 79

10.5. APPENDIX E:DYNAMIC FINANCIAL FLOW CHARTS FOR LARGE HEALTH SHOCKS 83

10.5.1. COSTS OF LARGE HEALTH SHOCK 83

10.5.2. SMOOTHENING ABILITY WITH LARGE SHOCKS 84

10.5.3. COPING MECHANISMS 85

10.6. APPENDIX F:HOUSEHOLD-WIDE SAVINGS REGRESSIONS 89

L

T

TABLE 1-HYPOTHESES FOR THE EFFECTS OF ILLNESS AND INSURANCE ON FINANCIAL OUTCOMES ... 13

TABLE 2-SAMPLE SIZES AND STRATIFICATION ... 23

TABLE 3–DEMOGRAPHICS (BASELINE DATA) ... 26

TABLE 4–FINANCIAL DESCRIPTIVES (WEEKLY AVERAGES) ... 27

TABLE 5-HEALTH SHOCK DESCRIPTIVES (AVERAGES OVER ENTIRE PERIOD) ... 28

TABLE 6–EFFECT OF HEALTH SHOCKS ON OOPHEALTH EXPENDITURES ... 35

TABLE 7–THE EFFECT OF ILLNESS AND INSURANCE ON EARNED INCOME ... 38

TABLE 8–TESTING CONSUMPTION SMOOTHENING ABILITY: THE EFFECT OF ILLNESS AND BEING INSURED ON CONSUMPTION AND PURCHASES ... 40

TABLE 9–THE EFECT OF ILLNESS ON TRANSFERS (TOTAL, INTRA- AND INTER-HOUSEHOLD)... 44

TABLE 10–THE EFFECT OF ILLNESS ON SAVINGS AND PRODUCTIVE ASSET SALES ... 46

TABLE 11-GENDER DISPARITIES: SEPARATE MALE AND FEMALE EFFECTS OF ILLNESS AND INSURANCE ON SELECTED OUTCOME VARIABLES ... 48

TABLE 12-RESILIENCE AND COPING DISPARITIES BETWEEN TOWNS AND VILLAGES... 49

TABLE 13-ABSOLUTE EFFECTS OF ILLNESS ON OOPS,INCOME,CONSUMPTION,INTRA- AND INTER -HOUSEHOLD TRANSFERS, AND SAVINGS (IN NAIRA AND AS % OF WEEKLY CONSUMPTION) ... 51

TABLE 14–THE EFFECT OF ASSET SHOCKS ON FINANCIAL OUTCOMES ... 55

TABLE 15–TESTING FOR STATE DEPENDENCE OF CONSUMPTION ... 56

L

F

FIGURE 2-HEALTH SHOCK OCCURENCE (AVERAGE OVER ENTIRE PERIOD) ... 28

FIGURE 3-HEALTHCARE PROVIDER CHOICE (AVERAGE OVER ENTIRE PERIOD) ... 30

FIGURE 4-MEDICAL EXPENSES PER PROVIDER ... 31

FIGURE 5–INDIVIDUAL COSTS AND COPING STRATEGIES OF UNINSURED ... 52

FIGURE 6–INDIVIDUAL COSTS AND COPING STRATEGIES OF INSURED _{... 52}

FIGURE 10-SEASONALITY IN HOUSEHOLD TRANSACTIONS (JANSSENS &KRAMER,2015) ... 76

FIGURE 11–EFFECT OF HEALTH SHOCK ON OOPHEALTH SPENDING (WEEKLY AVERAGES, IN NAIRA) ... 77

FIGURE 12–EFFECT OF HEALTH SHOCK ON EARNED INCOME (WEEKLY AVERAGES, IN NAIRA) ... 77

FIGURE 13–PURCHASE VOLATILITY IN WAKE OF ILLNESS (WEEKLY AVERAGES, IN NAIRA) ... 78

FIGURE 14–CONSUMPTION VOLATILITY IN WAKE OF ILLNESS (WEEKLY AVERAGES, IN NAIRA) ... 78

FIGURE 15–EFFECT OF HEALTH SHOCK ON INCOMING TRANSFERS (WEEKLY AVERAGES, IN NAIRA) ... 79

FIGURE 16–EFFECT OF HEALTH SHOCK ON OUTGOING TRANSFERS (WEEKLY AVERAGES, IN NAIRA) ... 79

FIGURE 17–EFFECT OF HEALTH SHOCK ON INCOMING INTRA-HOUSEHOLD TRANSFERS (WEEKLY AVERAGES, IN NAIRA) ... 80

FIGURE 18–EFFECT OF HEALTH SHOCK ON INCOMING INTER-HOUSEHOLD TRANSFERS (WEEKLY AVERAGES, IN NAIRA) ... 80

FIGURE 19–EFFECT OF HEALTH SHOCK ON SAVINGS WITHDRAWALS (WEEKLY AVERAGES, IN NAIRA) ... 81

FIGURE 20–EFFECT OF HEALTH SHOCK ON SAVINGS ACCUMULATION (WEEKLY AVERAGES, IN NAIRA) ... 81

FIGURE 21–EFFECT OF HEALTH SHOCK ON PRODUCTIVE ASSET SALES (WEEKLY AVERAGES, IN NAIRA) ... 82

FIGURE 22-EFFECT OF LARGE SHOCK ON OOPHEALTH EXPENSES (WEEKLY AVERAGES, IN NAIRA) ... 83

FIGURE 23-EFFECT OF LARGE SHOCK ON EARNED INCOME (WEEKLY AVERAGES, IN NAIRA) ... 83

FIGURE 24-PURCHASE VOLATILITY IN WAKE OF LARGE SHOCK (WEEKLY AVERAGES, IN NAIRA) ... 84

FIGURE 25-CONSUMPTION VOLATILITY IN WAKE OF LARGE SHOCK (WEEKLY AVERAGES, IN NAIRA) ... 84

FIGURE 26-EFFECT OF LARGE SHOCK ON INCOMING TRANSFERS (WEEKLY AVERAGES, IN NAIRA) ... 85

FIGURE 27-EFFECT OF LARGE SHOCK ON OUTGOING TRANSFERS (WEEKLY AVERAGES, IN NAIRA) ... 85

FIGURE 28-EFFECT OF LARGE SHOCK ON INTRA-HOUSEHOLD TRANSFERS (WEEKLY AVERAGES, IN NAIRA) 86 FIGURE 29-EFFECT OF LARGE SHOCK ON INTER-HOUSEHOLD TRANSFERS (WEEKLY AVERAGES, IN NAIRA) 86 FIGURE 30-EFFECT OF LARGE SHOCK ON SAVINGS WITHDRAWALS (WEEKLY AVERAGES, IN NAIRA) ... 87

FIGURE 31-EFFECT OF LARGE SHOCK ON SAVINGS ACCUMULATION (WEEKLY AVERAGES, IN NAIRA) ... 87

FIGURE 32-EFFECT OF LARGE SHOCK ON PRODUCTIVE ASSET SALES (WEEKLY AVERAGES, IN NAIRA) ... 88

FIGURE 30-HOUSEHOLD WIDE COSTS AND COPING STRATEGIES WHEN IDIOSYNCRATIC HEALTH SHOCK STRIKES UNINSURED HOUSEHOLD MEMBER... 91

FIGURE 32-HOUSEHOLD WIDE COSTS AND COPING STRATEGIES WHEN IDIOSYNCRATIC ILLNESS STRIKES INSURED HOUSEHOLD MEMBER ... 91

1. I

“One spoon of soup in need has more value than a pot of soup when we have an abundance of food” - African proverb

In developing countries individuals are subject to considerable risk due to large volatility of incomes, high prevalence of health and asset shocks, failing institutions and markets and/or natural disasters. Due to low average levels of wealth and liquid means exposure to risk can be detrimental. Reducing financial risk is welfare improving since most people, especially those living near or below poverty lines, are risk averse. Achieving informal or formal, full or partial, insurance against risk by paying a premium results in a utility gain for the risk-averse individuals equal to this reduction in risk, net of the premium (i.e. the knowledge that whether or not an adverse event occurs economic welfare is the same in both states). Especially in developing countries, demand for risk-hedging arrangements or product should therefore be large.

However, many financial markets in developing countries are missing, and those that do exist, often work imperfectly, making it difficult to insure against risks. Despite these missing formal credit and insurance markets, to reduce risks and smooth consumption, individuals in developing nations have developed a vast array of informal behavioral and institutional responses to fill the holes left by the formal markets (Murdoch, 1995). Many studies have found that despite large variances in household incomes, consumption patterns are remarkably smooth (e.g. Townsend, 1994; Fafchamps & Lund, 2003). However, these risk management mechanisms often come at high expense. For instance, precautionary saving of resources rarely yield interest, insuring in informal networks often comes at expense of large transaction costs (Jack & Suri, 2014), or forgoing upward income potential by not benefiting from rural-urban wage gaps to uphold informal risk-sharing networks in rural areas (Rosenzweig & Munshi, 2014).

An especially prevalent and costly source of risk is health. Many of the world’s poor live in areas plagued by disease, exposing them to considerable risk and hardship (Wagstaff & Lindelow, 2014; Heltberg & Lund, 2009). The economic costs associated with illness tend to be large since those costs have a dual source: (1) costs of medical care used to diagnose and treat the illness, and (2) costs of reduced income due to decreased labor supply or productivity. If health shocks are treated inadequately costs of the second category can explode easily, especially when this occurs in stages of early childhood development. The unpredictability and the size of the costs of most health shocks make it difficult for individuals and households to cope with illness episodes (Gertler & Gruber, 2002).

In recent years many developing and emerging countries have, in collaboration with NGO’s or the private sector, focused on improving healthcare coverage, especially so for the countries’ poorest inhabitants (Giedion et al., 2013). Such initiatives are sorely

needed to improve living standards but also hold the promise of seriously boosting economic growth through productivity enhancements (McKinsey Global Institute, 2014). Comprehensively understanding the impact of-, and behavioral and financial reactions to- health shocks is therefore vital to inform on the desirability, feasibility, and design of health care improvement programs. Despite the apparent value of such knowledge the body of research on this topic has been limited, ignores second order effects of insurance and neglects minor health shocks.

This paper aims to counter this knowledge gap by thoroughly studying the following in one coherent framework:

• What are illness-related costs in terms of medical spending and forgone income? • Is non-health consumption insured against health shocks?

• What coping mechanisms are utilized to cope with illness and to what extent? • What is the effect of being enrolled in health insurance on all of the above?

To do so a unique, high frequency financial and health diary dataset from town and village economies in a high risk but resource poor environment with limited access to formal financial services is employed. Weekly, individual data, during a full year, covering all financial transactions, consumption patterns, insurance statuses, health events and other shocks, enable insights and strong inference into health-related costs and risk coping mechanisms and provide pragmatic advice for future micro health insurance policy design. Due to the nature of the data strict focus is placed on financial responses to illness.

The results show that health shocks result in significant costs, through increasing out-of-pocket (OOP) health expenditures and forgone income. Individuals use several mechanisms to cope with illness: informal risk sharing networks through intra- and inter-household transfers, savings withdrawals and consumption out of stocks or self-cultivated produce. Specifically, comprehensive intra-household cooperation is found, although the collective household model is rejected. Nevertheless, Nigerian individuals are unable to smooth consumption across illness.

Being enrolled in health insurance fully mitigates this consumption volatility, and therefore reaps large welfare gains. No first order health costs reductions of insurance, through lowering OOPs or mitigating forgone earnings, are found. Evidence is presented that suggests insurance does lower illness-related costs but only significantly so when combined and identified through studying consumption patterns. Low utilization rates of insurance-covered facilities and increased health awareness and health demand further lower the financial impacts of insurance, albeit improving health status. Due to this increased health demand, forgone income is reduced much more through insurance than medical spending is. Second order crowding-out effects of inter-household risk sharing by insurance are not statistically proven, although the results are suggestive of its presence.

Endogeneity and measurement issues are circumvented by applying individual fixed effects models to short timespan, high frequency data. Moreover, all endogeneity in insurance uptake is captured in an insurance status dummy while all inference is conducted using the estimates on the exogenous insurance and health shock indicators

interaction term. Robustness checks show the results are not driven by or biased due to state dependence of consumption, reverse causality, remaining endogeneity in the interaction term, and limited variation in insurance status.

This paper adds to the current literature in five ways.

• Interesting results considering the timing of financial responses are found due to the high frequency nature of the data, and are novel to this literature.

• Weekly data points enabled insights into effects of more minor health shocks, which are often omitted in longer recall surveys. Insights on such shocks are of importance towards developing sustainable health insurance products.

• This work adds to the meager literature thoroughly evaluating first order effects of formal health insurance by introducing a novel empirical strategy to exogenously identify the impact of insurance passing several robustness checks, and by expanding micro health insurance evaluation literature geographically. • Moreover, to my knowledge, this is the first paper formally investigating second

order effects of formal insurance on a wide variety of coping mechanisms.

The rest of this paper is organized as follows. Section 2 present an overview of seminal works in related literature. Section 3 complements this empirical overview with theoretical dynamics and models to form testable hypotheses. Section 4 describes the empirical strategy, while section 5 provides some background information on the research setting and dataset. Section 6 informs on basic summary statistics and describes financial dynamics related to health shocks. Lastly, section 7 present the inferential results while section 8 concludes and discusses implications.

2. R

This paper aims to encompass several threads in the existing literature to create a thorough understanding of the financial impacts and employed coping strategies related to health shocks, and the interactions of formal insurance with these, in one coherent framework. In order to generate this holistic understanding the contemporary literature of the following subtopics will be studied in greater detail. The ordering is highly similar to the structure adopted in the subsequent empirical analysis. First, the current knowledge on welfare effects of illness is discussed. This comprises both the costs related to illness and the ability of household and individuals to smooth consumption across illness. The following two sections present the literature on adoption and effectiveness of coping mechanisms. First, the literature on group-based risk sharing is discussed. This literature can be sub-organized in papers focusing on informal risk sharing via transfers, either within or between households, and formal risk sharing via health insurance. Crowding-out theories between formal and informal risk sharing and other coping mechanisms are discussed as well. Lastly, the literature on self-coping mechanisms such as precautionary savings and asset sales is discussed.

2.1. Welfare effects of illness: costs and consumption smoothening

A necessary precondition for illness to have any financial effects is that adverse health events pose considerable costs. Basically there are two types of costs to illness. First, direct costs, consisting of expenditures made in direct relation to curing, dampening or preventing the illness. These consist out of out of pocket (OOP) medical expenses, transportation costs, or other costs related to caring for the ill. Second, more indirect costs constitute of forgone earnings caused by degradation of abilities and productivity due to illness. These latter costs can explode easily if illness causes persistent or permanent productivity decreases; something that happens much more frequently if improper healthcare is consulted, especially in early childhood development stages.

The above-described illness-related costs can translate into consumption reductions if the household or individual is not well insured. In poorer countries or families this can be a fruitful identification strategy for the direct observation of welfare effects and thus indirectly of the costs involved, through studying the causal effect of a change in health status on changes in consumption. Moreover, this strategy also allows for testing whether households or individuals can cope with these costs easily, i.e. is able to raise additional finance to cover OOP health expenses and income reductions and keep a constant consumption across such episodes. It is, however, not possible to definitely contribute a consumption reduction or lack thereof to high/low illness costs or inadequate/well-functioning insurance networks.

Some empirical work finds significant direct and indirect costs to illness (Rosenzweig & Wolpin, 1993; Genoni, 2012; Wagstaff & Lindelow, 2014; Wagstaff,

2007; Sepehri et al., 2006). Gertler and Gruber (2002) find that indirect costs through forgone earnings generally pose the largest share of costs.

However, most of the literature aims to evaluate the welfare effects of illness by studying the ability to smooth consumption, without studying costs of illness separately. Outcomes of these studies are mixed but many show that health shocks negatively affect consumption (e.g. Dercon & Krishnan, 2000; Gertler & Gruber, 2002; Wagstaff, 2007; Alkenbrack & Lindelow, 2013; Wagstaff & Lindelow, 2014). Some find no statistically significant effect (e.g. Genoni, 2012), but this can of course be due to effective utilization of coping mechanisms and insurance networks. Limited studies focus on the individual. Dercon & Krishnan (2003) show that in Ethiopia women from poor households bear the chief impact of total consumption volatility.

The empirical work on the welfare effects of illness faces four econometric challenges. First, it is difficult to measure changes in health reliably, consistently and exogenously. Over the years eight health measures have been commonly used of which self-reported health status, utilization of medical care and whether there are health limitations in the ability to work, are the most common (see Currie and Madrian (1999) or Mitra et al. (2015) for full overview). Reliable data that tracks health status over-time is very limitedly available. Even if quality measurements are available endogeneity could pose serious problems. Measurement errors on health status changes could very well be correlated with characteristics that are also determining economic wellbeing (e.g. wealth, education, and age). Indeed, there is some evidence that health shocks are differentially reported over wealth and income ranges, with the more affluent reporting illness more easily.

Second, unobserved heterogeneity affecting both economic welfare and health status can be present. This can be time invariant (e.g. early childhood malnutrition) or time varying (e.g. substance abuse or weather shocks).

Third, problems of reverse causality might bias the results. If feedback effects from economic welfare to health status exist, they are most likely to be positive (e.g. increased income enables improved nutrition leading to better health status), although negative effects could theoretically be possible as well (e.g. increased income enabling purchases of cigarettes, alcohol, motorbikes etc., all of which can lead to reductions in health status).

Lastly, state dependency could present an alternative interpretation of the estimates found in empirical work. If illness changes consumption preferences (e.g. lower demand for tobacco or alcohol) any change in observed consumption might not be due to binding budget constraints but due to welfare optimizing consumption choices. Gertler and Gruber (2002) rule this possibility out for their Indonesian dataset, but it is important to re-evaluate this concern in datasets that differ in location and timing.

To circumvent these challenges a variety of strategies have been used throughout the literature. Some have adopted more objective health measurements, looking for instance at systolic blood pressure, weight, height or BMI, to reduce measurement error in (past) health changes. Specifications including lagged health shocks aim to address issues of reverse causality (see e.g. Wagstaff, 2007). However, recent work finds that

this strategy does not fully avoid simultaneity bias (Reed, 2015). Using first differencing or fixed effects (FE) techniques the aim is to difference out time invariant unobserved and omitted variables and systematic measurement error (see e.g. Gertler & Gruber, 2002). However, since many studies investigate time periods of multiple years, it is likely that time varying factors still significantly bias the results. Recently, Genoni (2012) addressed time varying and time invariant unobserved heterogeneity and reverse causality by implementing an Instrumental Variables (IV) strategy using the changes in prices of health inputs as instruments. Although, the instruments pass the necessary test, OLS or FE results are not reported, making it difficult to assess the added value of controlling for time-varying effects.

In general, although econometric difficulties are abundant, a fairly large literature has found significant welfare effects of illness either through identifying related costs or consumption reductions. Although, many studies find a reduction in consumption during illness, this reduction is often much smaller than the total costs, suggesting other mechanisms are used to absorb these costs. The natural next question to ask is: how do individuals and households (partly) smooth consumption across illness? What coping mechanisms do they use? The following sections will elaborate on the contemporary knowledge addressing these questions. Section 2.2 will explore literature on group-based coping mechanisms (between and within household transfers, and formal health insurance) and section 2.3 will explore literature on self-coping mechanisms (savings and asset sales).

2.2. Group-based risk sharing

If smoothing consumption across time is infeasible due to missing or imperfect financial markets, smoothing consumption across entities is a natural second best option. Either through leveraging social networks or joining formal health insurance, if available, risk sharing can be achieved.

2.2.1. Informal risk sharing: inter-household transfers

Indeed, many studies have confirmed a large degree of risk sharing across households in developing economies. Most studies rely on the identifying assumption that under perfect risk sharing any idiosyncratic shock to a household should not influence its consumption pattern given the total village consumption. Empirical work therefore studies the relation between household health shock data and household consumption data within a predefined pool. Townsend’s (1994) seminal paper was one of the first to thoroughly investigate risk-sharing networks. Although, he finds that in high-risk Indian villages the perfect risk-sharing model is rejected, it does provide a surprisingly good fit with the data. The partial insurance that the village economies undertake is close to perfect, and results in the fact that household consumption is not much influenced by idiosyncratic shocks (Townsend, 1994). Many studies have found similar conclusion in different geographic areas (Deaton, 1992; Fafchamps & Lund; 2003; Udry, 1994). Dercon & Krishnan (2003) stress test these findings by studying the effect of positive food aid shocks instead, and find similar results.

Later studies aim to explain the rejection of the full risk-pooling hypothesis. Ligon, Thomas & Worral (2002) incorporate the impediment of limited commitment in risk sharing networks and conclude that this better fits observed data. Another explanation is found by changing the unit of analysis. The grounds for using the whole village as basis for full insurance can easily be contested, and several studies find that informal insurance networks within and across villages, consisting of, for instance friends and relatives, better describe observed risk sharing networks (De Weerdt & Dercon, 2006; Fafchamps & Lund; 2003).

So which mechanism do households adopt to informally share risk among each other? Available evidence suggests the following mechanisms are at least partly used to share risk between households: gifts and remittances, informal credit, trade credit, and lesser so labor sharing (Rosenzweig, 1988; Fafchamps & Lund, 2003; Fafchamps & Gubert, 2007; Udry, 1994). Informal saving and cooperative groups also help in risk sharing. Udry (1994) finds that credit transactions in rural Nigeria are effectively a form of informal insurance in itself: repayments are found to be state contingent, as likelihood of repayment depends on shocks affecting creditor and/or debtor. Fafchamps and Gubert (2007) further prove this feature of informal credit in the Philippines where loan repayment also turns out to be contingent on shocks affecting both involved parties. Thus gifts, remittances, informal credit and its repayments, labor sharing, and financial flows form informal saving groups or cooperatives can be classified as a form of inter-household risk sharing.

Inter-household risk sharing measures are potentially endogenous. Unobservable factors, such as for instance social capital or financial connectedness, can influence both economic wellbeing and the use of transfers. Studies aiming to identify direct effects of transfers only very limitedly control for this.

2.2.2. Informal risk sharing: intra-household transfers

The literature on intra-household transfers is much more limited, partly because required individual based data is scarce. However, in face of imperfect inter-household risk sharing, a natural place to share risks is within the household. Although shocks are often covariate within the household, making consumption smoothening difficult, it does reduce or overcome problems with asymmetric information and payment enforceability. Due to these reasons the literature on intra-household allocative efficiency has been growing in recent years. The seminal work by Chiappori and Browning (1998), who found evidence for efficiency in Canada, has sparked future research also covering developing countries. Most studies focus on whether (adult) household members are able to ex-post insure themselves in face of idiosyncratic income or expenditure shocks, thereby confirming to the collective household model. This strategy requires that the idiosyncratic shock does not alter any preferences or intra-household bargaining power. This limits the shock in both size of severity and duration, as too large or seemingly permanent shocks may change intra-household Pareto weights.

Rainfall shocks often qualify these requirements and are used extensively. Duflo & Udry (2003) acknowledge that in much of Africa husbands and wives farm different crops on separate plots and rainfall shocks differentially affect those crops. They find that the composition of purchases is sensitive to the gender of the recipient of a rainfall shock. Dubois & Ligon (2009) have a similar set-up, but allow productivity to depend on nutrition, this can explain some but not all of the response of consumption to the income shock. Both studies thereby reject the hypothesis of perfect intra-household insurance.

Health shocks can also serve as valid identifiers. Dercon & Krishnan (2000) use antromorphic indicators on adult nutrition to test if unanticipated illness shocks are insured within the household. While they cannot reject the collective household model for richer households, poor households fail to insure perfectly, especially if women are experiencing health shocks. Finally Goldstein (2004) also rejects perfect intra-household insurance. Although individual consumption in intra-households is not affected by idiosyncratic agricultural or health shocks, this smoothening is attributable to risk sharing networks from outside of the household. Goldstein (2004) thereby shows that studying either inter- or intra-household risk sharing in isolation can provide falls conclusions.

Robinson (2012) stress tests these findings by providing positive income shocks to randomly selected families in Kenya. Intra-household allocative efficiency is rejected since men increase private consumption while women do not.

Synthesizing this information, it is clear that risks are not shared perfectly within the household. Nevertheless, all studies find at least some risk-sharing making this potential important coping mechanism.

Lastly, it is important to stress that in traditional economies, especially in Africa, households maintain discrete accounts in which money from different activities and/or earned by different household members is kept separate (Duflo & Udry, 2004). For instance, money earned by the household head or earnings from cash crops are stored in different (mental) accounts. Fungibility of income in this sense is often limited. If households do engage in transactions from one account to another the transactions are made explicitly and awareness of those transactions across household members is high. Data collectors can exploit this feat and capture intra-household transfers in detail. Karanja-Diejomaoh (1978) confirms this situation holds in Nigerian as well.

2.2.3. Formal risk sharing: micro health insurance

Risk sharing arrangements can also be arranged on larger and more formal scale, resulting in formal micro health insurance products. In light of the available evidence on imperfect intra- or inter-household informal insurance, formal insurance can possibly fill the void. In recent years various insurance schemes with a focus on the poor have been established, including rainfall index insurance and health insurance. Health insurance schemes hold the promise of lowering detrimental OOP medical expenses, thereby enabling smoother consumption overtime (e.g. Sepehri et al., 2006). Moreover, since coverage effectively lowers healthcare prices for insured individuals, it

encourages ex-post moral hazard, increasing demand for healthcare. This increased utilization leads to improved health outcomes. Moreover, since demand for healthcare is price inelastic (Wagstaff & Lindelow, 2008), and coverage for certain services is often full, in total health insurance will lead to a reduction in OOPs. All this is welfare improving for the risk-averse individual, especially in lesser-developed countries where healthcare utilization is low.

However, impacts of health insurance schemes in developing countries have resulted in mixed results. To my knowledge, no earlier work has formally studied the effects of insurance on consumption smoothening through illness. Alkenbrack and Lindelow (2015) did compare the percentage of individuals who bought less or cheaper food, and spent less on children’s needs, of which only the last category showed a significant difference over uninsured and insured groups. Many papers studying direct costs and health effects report positive impacts, such as a reduction in OOPs (Wagstaff & Pradhan, 2005; Sepehri et al., 2006; Alkenbrack & Lindelow, 2015), and an increase in utilization (Alkenbrack & Lindelow, 2015; Van der Gaag et al., 2015; Wagstaff & Lindelow, 2008). Results often differ over rural or urban settings (Wagstaff, 2007) or over income quintiles where Alkenbrack and Lindelow (2015) find that insurance is inequality promoting. Other work reports mixed impacts (e.g. Wagstaff, Lindelow, Jun, & Juncheng, 2009; Alkenbrack & Lindelow, 2013). There are two theoretical lines of reasoning why the impacts of health insurance are not positive across the board: first, insurance can crowd-out informal group-based coping mechanisms, and second, insurance can potentially increase financial risk.

To elaborate on the first line of reasoning, enrollment into formal insurance potentially interacts with existing informal insurance networks in two ways. First, formal insurance can crowd-out informal coping mechanisms if and only if contracts are not perfectly enforceable. This limited commitment requires for each contract to uphold that individual expected future offs of risk sharing should be larger than the pay-offs of reverting to autarky at all times and states of the world. In this case insurance coverage increases the value of deviating from any informal risk-sharing contract relative to staying in contract, and hence will reduce the to overall degree of informal risk sharing (Dercon & Krishnan, 2003; ligon et al., 2002; De Weerdt & Fafchamps, 2011). These can lead to partial or even excess crowding out of informal insurance if the dismantling of informal risk sharing arrangements is severe enough. Second, existing informal risk sharing arrangements can crowd-out formal health insurance through free riding problems. Janssens and Kramer (2012) show in a framed public good experiment in Tanzania that in jointly liable groups, where risks are shared between members, insurance is effectively a public good. This leads to free riding by less risk averse individuals and ultimately to a sub-optimal rate of insurance within the risk-sharing group. Jowett (2013) finds further evidence for the theory that informal risk sharing arrangements can crowd-out, or prevent adoption of formal insurance. She concludes that demand for formal insurance is lowest in highly cohesive communities and those where individuals rely more heavily on informal financial arrangements. This suggests formal and informal risk-sharing arrangements are in fact substitutes if coverage of

either is adequate enough, and that market imperfections or inefficiencies in either will influence demand of the other.

The welfare effects ultimately depend on the relative efficiencies of formal and informal insurance arrangements. Although it seems clear that in order to fully comprehend and evaluate the welfare effects of a formal insurance program informal risk sharing arrangements have to be taken into account, the literature on this topic has developed only very marginally. Rosenzweig (1988) shows that formal loans and informal inter-household transfers can crowd each other out. While Dercon & Krishnan (2002) show that informal risk sharing can also be crowded out by positive food aid shocks. Alkenbrack and Lindelow (2015) report significant crowding-out by simple means of usage comparisons. They show insured individuals are half as likely to receive help from friends, relatives or the village in case of illness. By far most studies, however, do no not consider potential crowding-out effects when evaluating insurance schemes, although the relevance is apparent (e.g. Wagstaff & Lindelow, 2008; Sehehri et al., 2006).

To elaborate on the second line of reasoning, insurance can potentially also increase exposure to financial risk if healthcare providers can engage in supplier induced-demand (SID). Providers can exploit information asymmetries to over-treat insured patients, resulting in increased revenues if payouts adhere to a fee-for-service system. In effect, providers hereby shift the demand curve of insured individuals for healthcare to the right, which ultimately can result in higher OOPs if insured. Wagstaff and Lindelow (2008) find evidence for this exact line of reasoning in China. They conclude that health insurance “increases the risk of high and catastrophic spending” and that this is mainly due to insurance encouraging people to seek care more often when ill and to seek care from higher-quality providers. However, these results rely heavily on institutional factors such as means of payment, as several geographically and institutionally distinctive studies find that OOPs do not increase with insurance (e.g. Sepehri et al., 2006; Alkenbrack & Lindelow, 2015).

Lastly, to fully comprehend the financial protection that insurance provides it is important to study its effects on self-coping mechanism described below. Aggerwal (2010) finds that insurance lowers the use of savings and asset sales at times of illness, while Alkenbrack and Lindelow (2015) find significant effect for asset sales only.

Econometrically testing the impact of insurance is difficult due to the highly endogenous nature of voluntary insurance. Since insurance up-take is a free choice selection biases are prominent. This can hamper inference, since for instance, those who enroll in insurance are also potentially wealthier, more educated, more financial literate, etc. Moreover, many of those characteristics that correlate with insurance also help to smooth risk. To mitigate these problems the literature has used different strategies. Sepehri et al. (2006) adopt fixed and random effects models to control for time invariant unobserved heterogeneity. They conclude that failure to recognize the

endogeneity of insurance results in estimates that are biased upwards1_{. However, time}

varying heterogeneity is uncontrolled for in their set-up, although, this is likely considering the five year data timespan. Wagstaff and Lindelow (2008) also adopt a FE approach, however, they recognize time varying heterogeneity may be a concern in their dataset spanning over 12 years. To investigate and potentially mitigate this bias they report results from an IV set-up next to FE and non-FE models2_{. This shows there}

is large endogeneity in non-FE models, which is mitigated through using instruments. IV and FE estimates, however, do not differ significantly suggesting both do an adequate job at making insurance exogenous. Lastly, Alkenbrack and Lindelow (2015), and Aggerwal (2010) adopt a propensity score matching (PSM) approach to account for heterogeneity. Although in the former, results are robust to an alternative approach, combining PSM with a weighted regression, PSM strategies only remove biases arising from observables. Any unobserved heterogeneity could still bias estimates, and therefore this technique is considered of secondary quality.

So the limited literature evaluating health insurance programs in developing countries suggests that positive impacts can be large but need to be evaluated with care. Studying interactions of insurance with existing risk sharing arrangements is critical in order to understand welfare effects fully, however, to my knowledge only one paper has done this thoroughly. It employs, however, econometric techniques of secondary quality (Alkenbrack & Lindelow, 2015).

2.3. Self-coping mechanisms

Individuals and households can also employ several techniques to cope with illness shocks themselves. Two especially common mechanisms are build-up of precautionary savings and build-up of assets, both to be used or sold-off in times of illness. Both coping mechanisms, however, bring considerable costs.

2.3.1. Savings

Precautionary savings bring utility costs especially in markets such as Nigeria, where the cost of holding capital is generally high, borrowing is very limitedly possible or prohibitively expensive and risk aversion is often high (Caballero, 2011)3_{. Empirical}

work anonymously agrees savings are being employed to cope with illness. For instance, Fafchamps & Lund (2003) find that savings are significantly used in the Philippines to smooth health shocks, and Aggerwal (2010) finds similar results in India.

1 _{Sepehri et al. (2006) find no effect of insurance on OOPs when not adopting fixed effects, while insurance}

significantly dampens OOPs in models including fixed effects.

2 _{Reported models are poisson/probit for non-FE and IV, and possion/logit for FE, to deal with large incidence of zero}

values in OOPs. Adopted instruments vary over different insurance schemes but include: indicators for being a governement official, the head of household, a working member of the household, and perceived quality of the local township health center.

3 _{Experiments have shown risk aversion responds to income references that guarantee a minimally healthy diet, with}

risk aversion being large if just above and low if below this level (Caballero, 2011). This means precautionary savings are used excessively when above subsistence level, limiting further income growth.

2.3.2. Asset sales

Sales of productive assets can help to reach necessary calorie intakes in short-run but caps future income by reducing productivity (see e.g. Rosenzweig & Wolpin, 1993). It can therefore in certain cases be welfare improving to increase consumption volatility in order to keep asset stocks at sufficient levels. Empirical work struggles to find clear effects of such asset sales, since many of the world’s poor operate shops or have farms meaning that asset sales are their main source of income. Therefore, when ill two contrasting effects will materialize. First, due to illness the individual is less able to operate the shop and thus earn money by selling asset stocks. Second, being ill can impose considerable costs, which could institute the need for finance, which in turn could be obtained through selling more assets. Indeed, Fafchamps & Lund (2003) find no evidence of asset sales, in the form of livestock and crop sales, being used to cope with illness in the Philipines. Aggerwal (2010) and Genoni (2012) fail to find significant effects in India and Indonesia, respectively. Alkenbrack and Lindelow (2015) do report that 9% of individuals sell off assets or livestock in case illness, but do not test its significance.

3. T

This part will present theoretical dynamics describing how the financial variables of interest will react to health shocks and whether and how insurance will dampen or boosts these effects. Combined with the empirical literature overview this section will present seven testable hypotheses, each describing the effect of illness and insurance on a financial outcome variable. These are summarized in Table 1.

TABLE 1-HYPOTHESES FOR THE EFFECTS OF ILLNESS AND INSURANCE ON FINANCIAL OUTCOMES

Outcome variable Previous

Empirical Results Hypothesis

H1 OOPs Illness ++ ++/+ 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 − − H2 Income Illness −− −−/− 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 N/A − H3 Consumption Illness − − 𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 N/A +

H4 Intra-household transfers Illness ++ ++

𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 N/A No effect

H5 Inter-household transfers Illness ++ ++

𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 − −

H6 Savings Illness ++ ++

𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 − −

H7 Asset sales Illness No effect No effect / +

𝐼𝑛𝑠𝑢𝑟𝑎𝑛𝑐𝑒 𝑥 𝑖𝑙𝑙𝑛𝑒𝑠𝑠 − −

Notes: ‘++’ refers to strong positive relative effect, ‘+’ to weak positive relative effect. ‘--‘ and ‘-‘ have similar meaning

but negative. N/A refers to non-existent in previous empirical literature. The hypothesis that the collective household model will be rejected is excluded from this table but will be formally tested in the remaining of the paper.

3.1. Effects on OOPs, Income, and consumption smoothening (H1, H2, H3)

The effect of illness on OOPs and earned income is relatively straightforward. It is expected that OOPs increase while earned income decreases due to reduced productivity and/or inability to work. The size of the effects is expected to be smaller than previous literature has found, since this dataset will register more illness episodes, including minor health shocks.

The theoretical effect of insurance on income and OOPs is ambiguous. Most insurance programs are designed to lower OOPs and increase health status. By covering (part of) the medical expenses, OOPs are lowered. By increasing health awareness and lowering healthcare costs demand for healthcare is increased, affecting OOPs upwardly. Since most empirical work finds a significant decrease in OOPs, and those papers that find otherwise have significant institutional differences, I expect insurance to lower OOPs slightly.

For earned income, insurance could significantly reduce the negative impact of a health shock as better quality healthcare can lead to a reduced productivity impact of illness and faster recovery. On the other hand, the increased health awareness that

insurance brings can result in more frequent hospital visits and increased illness time perceptions. No previous empirical work has studied the individual effects of insurance on income during illness in great detail. Although I expect that both positive and negative effects of insurance on income during illness are small, if any effect is to be found I expect the overall effect to be reducing the negative impact of illness.

To ultimately assess the welfare effects of illness I test for consumption smoothening ability. This test shows the consumption outcome after all illness-related costs and coping mechanisms have been incorporated, essentially testing the ability of an individual or household to insure him/her/itself. This test has the underlying assumption that individuals and households maximize welfare by smoothening consumption 4 _{. Taking this assumption for granted, analyzing consumption}

smoothening is an overarching test for studying the welfare effects of illness and the effectiveness of all coping mechanisms combined, i.e. risk-coping ability, including risk sharing and self-coping mechanisms. Previous literature has falsely identified this as a test for risk sharing, although this need not necessarily be the case5_{. Below I will show}

how to adjust this analysis to correctly test for risk sharing in a more direct way. Since the sample used is primarily poor and options for formal risk smoothening are limited, I expect illnesses to be imperfectly insured and thus to have a negative impact on consumption. This negative impact is expected to be smaller than the illness-related costs due to coping mechanisms covering part of the costs. By testing both individual and household wide consumption reactions to illness the collective household model can be tested, which I expect to reject based on previous literature.

The theoretical effect of insurance on consumption levels through illness episodes is ambiguous since it essentially combines the effects of insurance on all illness costs and risk coping levers. Mechanically, I expect the effect of insurance on consumption to be positive since it should lower illness costs and I do not expect that these costs reductions will be fully offset through crowding-out effects in the risk-coping levers. Empirically there have been no formal tests of insurance effects on consumption smoothing while ill.

3.2. Effects on intra- and inter-household transfers (H₄, H₅)

The effect of illness on intra-household transfers is theoretically unambiguous if we assume at least some form of cooperation within the household. Theoretically, it can be shown that in a closed exchange economy without storage possibilities it is welfare maximizing to share risk perfectly within the household. This means that idiosyncratic shocks should influence individual consumption levels of household members identically6_{. To share these risks within the household intra-household transfers need}

to be employed. In reaction to a health shock it is therefore expected that

4 _{Although, this assumption seems plausible there could arise situations in which intertemporal utility is actually}

maximized by forgoing consumption in the illness period to, for instance, prevent selling off productive assets responsible for generating income in future periods. In effect, consumption volatility can therefore even be seen as an additional coping mechanism.

5 _{There are many other mechanisms, beside risk-sharing, that can potentially result in smooth consumption in times}

of illness. For instance, neglibile costs to illness, using savings or selling assets.

household transfers increase. In well cooperating households crowding-out effects of insurance are unlikely, since often decisions on whether to insure one or more household members are made in a mutually coordinated fashion. If a household, however, does not cooperate perfectly the observable benefits of insurance could partly crowd-out intra-household transfers. Empirical evidence on this issue is so far nonexistent.

The effect of a health shock on inter-household transfers in theory can be ambiguous. If people engage in informal risk-sharing networks between households, transfers will increase if any member of the network experiences an adverse shock, like an illness. It could, however, be welfare optimizing for a risk-sharing network to remove stricken households, if shocks are persistent or permanent, effectively reducing aggregate inter-household transfers. Considering the high frequency dataset I expect there to be enough minor illnesses such that the overall effect will be positive. Other empirical work has shown this result as well.

The theoretical model in appendix A derives that insurance benefits should perfectly crowd-out transfers, resulting in no net welfare effects. This, however, assumes risk sharing is perfect, a situation that is rarely found empirically. Where risk sharing is imperfect scope exists for formal insurance products to increase welfare by patching the holes in the informal safety net. Moreover, if shocks are covariate informal risk sharing easily breaks-down increasing the welfare impact of formal insurance.

Contrary, as outlined in the literature overview, insurance can also be welfare decreasing if it sufficiently crowds-out informal risk-sharing group arrangements. A detrimental domino effect could arise if a few people exiting the network due to insurance enrollment induce more people to exit, ultimately leading to the entire collapse of the risk-sharing network. The empirical work on this issue is very marginal but suggest some crowding-out to be present.

3.3. Effects on savings and asset sales (H₆, H₇)

Savings and assets are often accumulated as precautionary measures for harsh times. If a health shock strikes savings or assets can be used to finance the illness-related costs. Therefore in reaction to illness, savings withdrawals and asset sales are likely to rise in order to cope with these adverse effects. Empirical evidence on savings reinforces this assertion while no effects are found for asset sales.

Insurance theoretically should lower the use of savings and asset sales, and a very limited empirical literature recognizes this. My hypotheses will follow along these lines.

4. E

The aim of this paper is to identify the costs to illness, the ability to smooth consumption across illness, the coping mechanisms utilized to cope with illness, and study the effects of enrollment in micro health insurance on all of the previous. The coming section will elaborate on the empirical specification designed to accomplish these research goals while simultaneously mitigating five widespread econometric problems in this literature.

4.1. Empirical specification

To validate the research objectives and exploit the high frequency dataset a dynamic, autoregressive, fixed effects panel model will be estimated. The following specifications are based on the theoretical model developed in Appendix A and loosely follow insights from Gertler & Gruber (2002) and Jack & Suri (2014). The general model will be as follows: 𝑙𝑛�𝑌𝑗𝑡+ 1� = 𝛼𝑖 + � 𝛽𝑠𝐻𝑆ℎ𝑜𝑐𝑘𝑖,𝑡−𝑠 1 𝑠=0 + 𝛿1𝐼𝑛𝑠𝑢𝑟𝑖𝑡+ 𝛿2𝐼𝑛𝑠𝑢𝑟𝑖𝑡× 𝐻𝑆ℎ𝑜𝑐𝑘𝑖𝑡 + 𝜑1𝑙𝑛�𝑌�𝑗,𝑇−1+ 1� + 𝜑2𝑋𝑖𝑗𝑡+ 𝑉𝑡𝑐+ 𝜇𝑖𝑡 (1) 𝑤𝑖𝑡ℎ 𝑌𝑗𝑡∈ �𝑂𝑂𝑃𝑗𝑡, 𝐼𝑗𝑡, 𝑃𝑗𝑡, 𝐶𝑗𝑡, 𝜏𝑗𝑡, 𝜏𝑗𝑡𝑤, 𝜏𝑗𝑡𝑜, 𝑆𝑗𝑡, 𝐴𝑗𝑡� & 𝑗 ∈ {𝑗 = 𝑖, 𝑗 ≠ 𝑖}

where 𝑌𝑗𝑡 is one of the outcome variables of interest for individual 𝑖, if 𝑗 = 𝑖, or for

household 𝑗, if 𝑗 ≠ 𝑖, in week 𝑡. Natural logarithm transformations are adopted to dampen outliers and generate near normal distributed data. Square root transformations have been used to stress test the outcomes and have shown very similar patterns7_{. To investigate the economic welfare effects of illness the outcome}

variables used are 𝑂𝑂𝑃𝑗𝑡, out-of-pocket health expenditures including transportation

costs, 𝐼𝑗𝑡, earned income, 𝑃𝑗𝑡, purchases net of health-related expenditures, and 𝐶𝑗𝑡,

consumption net of health-related expenditures. To investigate the use of copings mechanisms the dependent variables used are 𝜏𝑗𝑡, total transfers received, 𝜏𝑗𝑡𝑤, transfers

received from within the household, 𝜏𝑗𝑡𝑜, transfers received from outside of the

household, 𝑆𝑗𝑡, saving withdrawals, and 𝐴𝑗𝑡, productive asset sales. In particular, 𝑃𝑗𝑡

includes all registered purchases while 𝐶𝑗𝑡 adds the monetary value of consumption out

of self-produced goods and stocks8_{. 𝜏𝑗𝑡 aggregates incoming remittances, gifts, informal}

loans, informal loan repayments, advance payments, and saving club/cooperative

7 _{Results available upon request.}

8 _{Data on consumed quantities have been multiplied by local monthly price levels, using Kormawa & Ogundapo}

(2004) for metric conversions, and averaged out over the number of adult and infant household members to form an individual consumption metric.

payouts. All variables are aggregated from individual transaction information and cover all available transaction means (cash, credit/bank, in-kind).

The taste shifts 𝑧𝑖𝑠𝑡ℎ and transitory income shifts 𝛿𝑖𝑠𝑡ℎ form the theoretical model in

Appendix A are captured by 𝐻𝑆ℎ𝑜𝑐𝑘𝑖𝑡, which is a dummy variable equaling 1 if

individual 𝑖 incurred a health shock during week 𝑡, and is 0 otherwise. 𝐼𝑛𝑠𝑢𝑟𝑖𝑡 is a

dummy variable which equals 1 if individual 𝑖 is insured in month 𝑡. The interaction term 𝐼𝑛𝑠𝑢𝑟𝑖𝑡× 𝐻𝑆ℎ𝑜𝑐𝑘𝑖𝑡 effectively captures 𝑣𝑖𝑠𝑡 , and therefore allows for the

identification of the financial effects of having insurance. 𝑋𝑖𝑗𝑡 is a vector of controls, in

particular household demographics, household head, 1st _{wife, and female headed}

household dummies, illnesses by children in the household, and days being unable to perform daily activities not due to illness. In OLS regressions more extensive controls are added (e.g. wealth, education, financial product use, etc.). These are unnecessary in the FE specification above, since 𝑎𝑖 represents individual fixed effects, effectively

controlling for these time-invariant differences. Adopting individual fixed effects most importantly means that the effect of insurance is identified by comparing within individuals the impacts of illness when insured with times when not insured. 𝑙𝑛�𝑌�𝑗,𝑇−1+ 1� is a term that captures the weekly average value of 𝑌 taken over month 𝑇.

This term adds significant explanatory power since all dependent variables, except for 𝑂𝑂𝑃𝑗𝑡, are autocorrelated. The weekly average of the previous month is taken to

minimize previous health shocks influencing this control.

Community-by-time fixed effects, 𝑉𝑡𝑐, are powerful controls capturing all internal

and external differences between communities and allows for differential time fixed effects per community. The unobserved aggregate consumption and taste shift variables are thus captured by 𝑉𝑡𝑐 (Fafchamps & Lund, 2003). It also controls for differences

across communities and time in wealth, economic activity, disease susceptibility, demand specificities, prices levels or inflation, insurance up-take seasonality, culture etc. Since there is a sharp divide in communities being villages or towns, this set-up also controls for differences between rural or urban areas. These dummies are especially important in this research since there is very high seasonality of economic activity and illness incidence mainly caused by distinct wet and dry seasons9_.

The entire set-up above will also be evaluated with the dependent variables being aggregated at household level (the case where 𝑗 ≠ 𝑖). This allows for further testing of intra-household cooperation that potentially is not covered by the intra-household transfer variable, 𝜏_𝑖𝑡𝑤_{, due to imperfect reporting. Moreover, for some variables of}

interest the effects are more long-term and often spread over several weeks. To increase the precision of estimates and combine the effects of consecutive weeks for these variables models aggregating all data over months are also presented.

Interpretation of the above-described model goes as follows. In specification 𝛽0

shows the direct effect of illness, while 𝛽1 shows the lagged effect of illness by either a

week or month depending on the specification. Significant difference from 0 of 𝛿2

implies the financial effects of insurance on 𝑌𝑗𝑡.

To show this in more detail consider the case where 𝑌𝑗𝑡= 𝐶𝑗𝑡 and 𝑗 = 𝑖 (individual

set-up). In equation (1) if 𝛽0 < 0 consumption is not smoothened across illness.

Enrollment in insurance reduces illness-related consumption volatility if −𝛽0> 𝛿2 > 0,

and completely removes it if 𝛿2 > −𝛽0 > 0.

To better understand the model interpretations for copings strategies, consider the case where 𝑌𝑗𝑡 = 𝜏𝑗𝑡𝑜 = outside household transfers. In equation (1) if 𝛽0 > 0 direct

informal inter-household risk sharing is present, since illness will increase received transfers from non-household members. If 𝛽1 > 0 this means that inter-household

transfer endure to or only materialize one week after illness strikes. Formal health insurance crowds-out informal inter-household risk-sharing arrangements if 𝛿2 < 0. In

the more unlikely case that |𝛿2 | > 𝛽0 or 𝛿2+ 𝛽0 < 0 insured individuals actually

receive less inter-household informal help in case of an health event, meaning excess crowding out is present. Insurance endogeneity, e.g. those who are insured have larger social networks, is captured in 𝛿1 and therefore no deductions will be made based on

this estimate. In the case where 𝑌𝑗𝑡= 𝜏𝑗𝑡𝑤 = within household risk sharing, similar logic

applies. If 𝛽0 > 0 intra-household risk sharing is present, since illness increases

received transfers from household members. If 𝛿2 < 0, insurance crowds out

intra-household transfers, which is expected to be more unlikely in cooperating intra-households.

4.2. Improvements over previous models testing for risk-sharing

This empirical specification provides clear advantages for identifying informal risk-sharing arrangements (𝜏𝑗𝑡). Most of the early empirical literature relies on consumption

smoothening tests to identify risk sharing. This strategy, however, is flawed since consumption can be smoothened without the use of risk sharing arrangements in three ways. First, intertemporal accumulation and reduction of savings or assets makes it possible for households to smooth consumption without informally sharing risks among their network. Second, fluctuations in local financial worth of assets and consumption varying with aggregate shocks can cause co-movement between individual and aggregate consumption even in the absence of risk sharing (Fafchamps, 2008; see Appendix A for more thorough explanation). Third, intra- and inter-household risk sharing arrangements can have opposing effects possibly leading to failures to reject hypotheses (see e.g. Goldstein, 2004). To circumvent these problems and enable correct inference the departures of the original model have been made. These are: (1) taking the individual instead of the household as unit of analysis throughout most models, (2) using transfers instead of consumption as identifier for risk sharing and (3) incorporation of crowding-out effects by formal health insurance. In appendix A this model is formally derived.

4.3. Identifying assumptions

Five econometric problems complicate finding consistent and unbiased estimates and need to be resolved to be able to validate the research objectives. These are: (1) non-random measurement error in health status, (2) unobserved heterogeneity in characteristics affecting both health status and outcome variables, (3) selection bias in

insurance status (4) reverse causality between health status and outcome variables, and (5) state dependency of consumption. This section will describe how each is dealt with in this paper.

First, it is possible that unreported illnesses are the ones that caused limited financial reactions, leading to upward biases in the effects. However, due to the weekly data significantly more minor illnesses are captured when compared to ordinary research, minimizing recall biases. It is difficult to control for biases in self-assessment of illnesses. A possibility is that illness assessments change over wealth or other variables. This bias is ultimately taken away by taking objective health measures. The adopted dataset unfortunately does not have any such measurements.

Second, to mitigate unobserved heterogeneity in characteristics that affect both health status and outcome variables individual fixed effects are included. This controls for all time-invariant heterogeneity and since the data span only one year, it is highly unlikely that significant time-varying heterogeneity is present.

Third, insurance enrollment is endogenous due to selective uptake or renewal associated with observables such as solvability, liquidity, and residential area, as shown in Janssens & Kramer (2015), and potentially others such as education, wealth, or unobservables10_{. Two strategies aim to mitigate this bias. (i) Due to the 52 consecutive}

weeks of data it is possible to purge the regressions from all factors that do not vary over this 52-week period but are causing endogeneity of insurance, by using individual fixed effects. Note that this is a powerful tool as it effectively controls for a wide variety of individual characteristics that may determine baseline levels and do not alter overtime, including wealth levels, trust, ability, risk and time preferences etc. It is credible to assume these attributes are time-invariant considering the fairly short time span of the data. Adopting individual fixed effects does imply that differences in financial impacts of illness over insurance status are studied within individuals, meaning that only those individuals that switch insurance status and experience at least one health shock when insured and uninsured help to identify the insurance effects. The high rate of switching in insurance status allows for significant variances in the insurance dummy such that the individual fixed effects do not capture enrollment status for all11_{. Moreover, in section 7.6.3 it is concluded that sufficient observations help}

identify the insurance effects in this empirical set-up by comparing standard errors of OLS and FE models. (ii) The second strategy to mitigate selective uptake bias argues in similar vein to Jack & Suri (2014) that all the remaining, time-varying, endogeneity of insurance is absorbed in the main effect of being insured (𝛿1), an estimate that is not the

interest of this study. Therefore 𝛿2 should be exogenous and able to capture the effect of

insurance on the dependent variables. This identification strategy fails, however, if there are unobservables that correlate with insurance status and help household smooth risk better in wake of shocks. In the robustness checks section (7.6.1) it will be shown that the results are not driven by unobservables that correlate with insurance

10 _{Will be tested at descriptive statistics part}