Insurance and propagation in village networks

Kinnan, Cynthia; Samphantharak, Krislert; Townsend, Robert M.; Vera Cossío, Diego Alejandro Working Paper Insurance and propagation in village networks IDB Working Paper Series, No. IDB-WP-1155 Provided in Cooperation with: Inter-American Development Bank (IDB), Washington, DC Suggested Citation: Kinnan, Cynthia; Samphantharak, Krislert; Townsend, Robert M.; Vera Cossío, Diego Alejandro (2020) : Insurance and propagation in village networks, IDB Working Paper Series, No. IDB-WP-1155, Inter-American Development Bank (IDB), Washington, DC, https://doi.org/10.18235/0001846 This Version is available at: https://hdl.handle.net/10419/237452 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/3.0/igo/legalcode Abstract* In village economies, insurance networks are key to smoothing shocks, while production networks can propagate them. The interplay of these networks is crucial. We show that a significant health expenditure shock to one household propagates to other linked households via supply-chain and labor networks. Imperfectly insured households adjust production decisions—cutting input spending and reducing labor hiring—affecting households with whom they trade inputs and labor. Household businesses proximate to shocked households in the supply chain network experience reduced local sales, and those proximate in the labor network experience a lower probability of working locally. As a result, indirectly shocked households’ earnings and consumption fall. These declines persist over several years because networks are rigid: households appear unable to form new linkages when existing links experience negative shocks. Propagation is a function of access to insurance networks: well-insured households do not cut spending when hit by shocks, leading to minimal propagation. A simple back-of-the-envelope exercise suggests that the total magnitude of indirect effects may be larger than the direct effects and that social (village-level) gains from expanding safety nets such as health insurance may be substantially higher than private (household-level) gains. JEL classifications: D13, D22, I15, O1, Q12 Keywords: Entrepreneurship, Risk sharing, Propagation, Production networks * Kinnan: Tufts University and NBER, [email protected]; Samphantharak: University of California San Diego and Puey Ungphakorn Institute for Economic Research (PIER), Bank of Thailand, [email protected]; Townsend: Massachusetts Institute of Technology and NBER, [email protected]; Vera-Cossio: Inter-American Development Bank, [email protected]. We thank numerous colleagues and seminar audiences for helpful suggestions. Townsend gratefully acknowledges research support from the University of the Thai Chamber of Commerce, the Thailand Research Fund, the Bank of Thailand, and the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD, grant number R01 HD027638). Opinions, findings, conclusions, and recommendations expressed here are those of the authors and do not necessarily reflect the views of the Bank of Thailand or the InterAmerican Development Bank. 1 Introduction Households in developing countries are both consumers and producers (Banerjee and Duflo 2007; Samphantharak and Townsend 2010). In addition to purchasing goods and services for consumption, they exchange productive inputs with each other (Braverman and Stiglitz, 1982). These inputs flow through local supply chain networks (networks of households buying and selling output, raw material or intermediate goods) and labor networks (networks of households providing and hiring labor). Households are also exposed to shocks, such as health, weather and business demand, affecting both consumption needs and production. These may be smoothed by transfers or loans from other households, via self-insurance, or not at all. Linkages between households mean that what happens to one household has the potential to propagate to others, too. This paper documents that this potential is empirically relevant. Local networks of gifts and loans (which we call “financial networks”) provide insurance, as previous work has shown (e.g., Udry 1994; Townsend 1994; Samphantharak and Townsend 2018). On the other hand, when financial networks fall short of full insurance, supply chain and labor networks propagate shocks. We leverage a unique dataset to examine network effects—both positive (insurance) and negative (propagation). The Townsend Thai data, constructed from 14 years of monthly panel surveys to households in rural and peri-urban Thai villages, allow us to identify idiosyncratic shocks to households’ budgets. Using detailed information regarding gifts, loans and transactions across family-operated businesses, we are able to observe labor, supply chain, and financial networks, and study their evolution over time and their role in mitigating and/or propagating shocks. Of course, shocks to household endowments are not typically exogenous, which makes it challenging to empirically identify their causal effects. To overcome this challenge, we exploit variation in the timing of episodes of sudden increases in health spending and construct counterfactuals for shocked households using households that experience similar shocks, but at different points in time (Fadlon and Nielsen, 2019). These sudden increases in health spending are idiosyncratic shocks to household endowments, which constitute exogenous and inescapable obligations that generate pressure on household budgets. We show that, conditional on ever experiencing a health spending shock, their timing is exogenous and uncorrelated across households. We also show that the shocks are severe—they are twice as large as average per capita food consumption and coincide with sharp increases in inpatient care. 1 Summary of Results We first analyze the direct effects of idiosyncratic shocks on household consumption, production, and financing decisions. We find that directly-shocked households are able to smooth the shock on the consumption side. We find neither significant nor substantial changes in non-health consumption for these households in the aftermath of a shock. This consumption-side smoothing is achieved, in part, through intra-village insurance. Shocked households were more likely to receive transfers from other households in the village, constituting a 29% increase in total incoming transfers relative to the pre-shock periods. This result highlights the importance of local financial networks in providing insurance against idiosyncratic shocks. However, while local financial networks provide insurance, this informal insurance is incomplete.1As a result, in order to fully smooth consumption, shocked entrepreneurs reduce business spending: in essence, financing the shock out of the business budget. They substantially reduce costs (23% decrease), and almost entirely cut their demand for external labor (79% decrease). This decrease in productive activities leads to an average 12% reduction in revenues, relative to the pre-period. Shocks to the health of prime-aged members could also affect labor endowments; however, we show that even shocks to children and elderly household members affect production. Thus, shocks to households’ consumption needs affect production-side decisions, in violation of the separation theorem (Benjamin, 1992). The impacts of health shocks are related to incompleteness in insurance and labor markets. Shocks to households with limited access to informal insurance (that is, with limited participation in gift and loan networks during the year preceding the shock) reduce costs and revenues by 34% and 27%, respectively. In contrast, for households with higher informal insurance participation, these decreases are almost fully attenuated. Thus, households that are not well-integrated into local informal insurance networks are most vulnerable and so, in turn, are their network connections. Additionally, the declines in production are larger when the shocks also affect household labor endowments and thus are unlikely to be fully insured by local financial networks. Namely, shocks related to the illness of prime-age household members (i.e., financial and labor endowment shocks) are associated with significant declines in hired labor and business costs, while shocks affecting elderly or children (i.e., mostly financial shocks) affect business revenues but do not significantly affect hired labor or costs. Thus there appears to be a missing market for individual-specific labor 1In the half-year of the shock, less than half of the spending need is met by transfers: see Figure 3. 2 inputs into household production, so naturally, the separation hypothesis fails. We next examine the indirect impact of these shocks: the effect on other local businesses and workers. Our empirical strategy relies on variation in the proximity of a given household to the shocked household through pre-period economic networks. We undertake a generalized differencein-difference analysis: comparing changes in outcomes before and after each shock, between moreexposed households (i.e., those that are closer to a shocked household in the pre-period network) and less-exposed households (those farther away). Closer households, with greater exposure, see larger falls in total transactions (a 21% decline for a unit change in closeness), and therefore a fall in income (12% decline for a unit change in closeness). They do not experience a change in incoming gifts. Given a fall in income and no offsetting change in gifts, consumption falls (4% decline for a unit change in closeness). These indirect effects are largely driven by shocks to underinsured households, suggesting that, by protecting directly hit households, local financial networks can also reduce or prevent propagation. Propagation occurs in a context of rigid networks: suppliers struggle to find new customers when their clients suffer a shock, and workers struggle to find new jobs when existing employers face shocks. We document that links are quite persistent: ceteris paribus, households that transacted at baseline are substantially more likely to transact ten years later, relative to households that did not transact at baseline. Kinship relationships are strong predictors of trade, highlighting the importance of time-invariant contract-enforcement barriers to trade across households (Ahlin and Townsend, 2007; Johnson et al., 2002). Frictions leading to rigid labor networks are particularly important in the Thai setting, as proximity to the shocked household via the labor market network is most strongly associated with indirect effects on income and consumption.2These indirect effects persist even two years after the shock, suggesting that, in a context of rigid networks, the recovery from indirect shocks can be sluggish. As expected, the magnitude of the indirect effect experienced by a single household is smaller than the magnitude of the effect felt by the directly-hit household (i.e., the one experiencing the health shock). The indirect effects, however, hit many more households. As a result, a simple back-of-the-envelope exercise suggests that the total village-level magnitude of the indirect effects of a given shock may be as large or larger than the direct effects of that shock, with a multiplier 2Evidence of frictional slack in goods and labor markets is also shown by Egger et al. (2019) in the context of rural Kenya and by Breza et al. (2020) for rural India. Chandrasekhar et al. (2020) present a model of network-based hiring arising from informational frictions. 3 (i.e., a ratio of the total indirect fall in consumption to the direct reduction in business spending) of approximately 1.5. Contributions to the Literature Previous studies have provided evidence of non-separability of household consumption (labor supply) and production (input or labor demand) decisions in developing countries.3We build on this literature by showing that idiosyncratic shocks to household spending can affect not only the production decisions of shocked households but also those of other (non-shocked) households. Our findings emphasize the dual role of networks in understanding non-separability and its consequences: risk-sharing networks provide insurance, but production networks increase the risk of propagation. In turn, we complement the empirical literature studying the firm-to-firm propagation of regional or sectoral shocks through production networks.4We leverage our unique data and the context of family-owned firms (micro and small enterprises) to show that, in line with recent macroeconomic models linking granular shocks to aggregate fluctuations (Gabaix, 2011; Acemoglu et al., 2012; Farhi and Baqaee, 2020), granular shocks to health spending affect family enterprises and propagate to other local firms. Our ability to focus on local and granular shocks is important, as a large share of firms across the world are small and family-operated (Beck et al., 2005; Banerjee and Duflo, 2007; La Porta et al., 1999; Bertrand et al., 2008), and thus exposed to shocks affecting family endowments. The fact that we uncover a multiplier (i.e., a ratio of indirect to direct effects stemming from a change in demand) that is similar to multiplier estimates for demand shocks in the United States (Nakamura and Steinsson, 2014; Su´arez Serrato and Wingender, 2016) and in Kenya (Egger et al., 2019) suggests that our results have applicability beyond the specific context of Thai villages, and that, in settings of imperfect insurance and rigid supply-chain and labor networks, idiosyncratic shocks can lead to multipliers similar to those generated by aggregate shocks. This paper also contributes to the literature studying local economic networks in developing countries (e.g., Bramoull´e et al. (2016); Chuang and Schechter (2015); Munshi (2014)). Previous studies have analyzed the ability of households to use local networks to buffer shocks (Townsend, 3See for example Benjamin 1992; Dillon and Barrett 2017; Dillon et al. 2019; Samphantharak and Townsend 2010; LaFave and Thomas 2016; Samphantharak and Townsend 2018, among others. 4There is a growing literature in international trade studying the propagation of shocks through production networks in the aftermath of natural disasters (Barrot and Sauvagnat, 2016; Carvalho et al., 2020), trade shocks (Tintelnot et al., 2018; Huneeus, 2019), and sectoral or regional shocks (Caliendo et al., 2017). 4 network (e.g., first-degree relatives), which is measured during the baseline survey in 1998.17 We include controls for distance with respect to demographic characteristics and a measure of distance between each pair of households based on baseline net worth (e.g., total assets net of liabilities).18 Finally, we also include household-fixed effects. The parameter of interest is ρ, which captures the persistence of the economic interactions between each pair of sample households. Table 2 presents estimates of equation (1). There is an important degree of persistence in the labor-market and supply chain networks, with raw auto-correlation coefficients of 0.46 and 0.42 (see column (1) in each sub-panel). These are substantially higher than that of the financial network (0.26). In the case of the labor market and the supply chain networks, having transacted during the previous period explains one-fifth of the overall variation in the current probability of trading. This pattern contrasts sharply with the case of the transactions in the financial markets (gifts and loans), as transactions in period t−1 only explain 7% of the overall variation in the probability of transacting at t. One explanation is that financial networks are less active, and they are probably responding to either unexpected business opportunities or shocks. We explore this premise in Section 4. Persistence remains substantial after controlling for village-year fixed effects, suggesting that economic linkages respond mostly to within-village variation (see column (2) in each sub-panel). In columns (3) and (4), we analyze whether persistence is related to kinship relationships, differences in demographic characteristics or differences in endowments (net worth). Although, in all three networks, controlling for baseline kinship links reduces the persistence coefficients, they are still high. Persistence does not seem to respond to including measures of differences in terms of demographic characteristics or initial wealth. In all cases, pairs that share kinship connections are 10 percentage points more likely to trade. The probability of trade in the supply chain and labor networks does not respond to differences in distance or wealth between the two households. In contrast, the probability of trading in the local financial network increases when households are different in terms of demographic characteristics, but it decreases when there are differences in baseline wealth in the pair. This pattern highlights two features of local financial networks. First, among those households with similar wealth, households that differ in demographic characteristics 17Two households share a link if they are first-degree relatives (including parents-in-law). 18Demographic distance is measured as the Euclidean norm of a vector of household attributes capturing household size, gender and age composition, as well as average age and education corresponding to members of the household at baseline. We then take logs of the resulting norm. Net worth distance is constructed by taking logs of the squared net-worth difference within each pair. 11 are more likely to transact, suggesting that one motive for trading is diversification, as shock type and occurrence may vary with demographics. Second, similarly wealthy households are more likely to trade, which suggests that, although diversification takes place, it is restricted to household pairs for whom insurance is more likely to be actuarially fair. In sum, the labor-market and supply chain networks exhibit a striking degree of persistence over time. One implication of this persistence is that the effects of shocks which propagate via these networks may be quite persistent, an implication that we test in Section 5. 3 Constructing Idiosyncratic Shocks Our goal is to examine how household production decisions respond to idiosyncratic shocks to household wealth and labor endowments, and whether these shocks propagate to other households through village economic networks. In this paper, we focus on idiosyncratic events associated with high levels of health spending to identify episodes of high financial stress.19 Analyzing these shocks is important for three reasons. First, it has been shown that health shocks affect household finance and labor supply (Gertler and Gruber, 2002; Genoni, 2012; Hendren et al., 2018). Second, because these shocks are uncorrelated across households, we are able to separate these idiosyncratic shocks from aggregate shocks that could directly affect economic activity through changes in the markets for final goods, intermediate inputs, and labor. Finally, we can analyze shocks that are, in principle, insurable through local networks, and thus understand whether individual responses to such shocks vary with access to local insurance networks. We identify the shocks as follows. On a monthly basis, we compute health spending as the sum of spending on medicines, transportation to medical facilities, and fees related to either inpatient or outpatient care. For each household, we identify the month with the highest amount of monthly health spending throughout the panel. We focus on the largest shocks, as we want to restrict the 19Thailand has a universal health insurance program, so these expenses are above and beyond those covered. The insurance program essentially covers expenses related to basic healthcare services. These services include medical visits at registered primary healthcare facilities (which must be located in the same area as each patient’s registered residential address), transferred patients from a primary facility to secondary or tertiary facilities for complicated cases, emergency cases at non-registered facilities, expenses for in-patients staying for less than 180 days for the same illness, and prescriptions of medicines as listed in the National List of Essential Drugs. For details, see Thailand’s National Health Security Office (NHSO), Administrative Manual, 2014 (in Thai). http://www.oic.go.th/FILEWEB/ CABINFOCENTER3/DRAWER091/GENERAL/DATA0000/00000367.PDF 12 analysis to shocks that pose a financial burden to the household.20 To facilitate measuring responses to the shocks by comparing households’ behavior before and after the episodes, we restrict the search to years 2-12 in the panel (out of 14 years of monthly data). This enables us to observe at least two years of both preand post-shock behavior for all households. Following this approach, we identify 505 episodes of sudden increases in monthly health spending, one per household.21 We exclude 32 events related to possible pregnancies/births from our analysis as the associated changes in health spending are likely to be anticipated.22 After possible pregnancies are excluded, there are 473 shocks. In addition, for a subset of 405 of these events, we were able to identify the health symptoms affecting household members at the time of the events, and when these symptoms were initially reported. Appendix Figure A1 shows that, prior to the sudden increase in health spending, the median number of consecutive months in which households report any health symptoms is three months. To account for potential anticipation effects, we define the beginning of each event by subtracting the number of months preceding the episode of high health spending during which household members reported health symptoms from the month corresponding to the episode.23 In the case of the 85 households for which we could not identify the beginning of the symptoms,24 we coded the beginning of the event as three months before the episode of high total health spending. We present robustness checks varying the beginning of the shock in Section 4.3.1. 3.1 Characteristics of the Shocks Relationship between health spending and health status. Appendix Figure A3 depicts total household health spending (left axis) and the probability of reporting health symptoms around the events of financial stress (right axis). The figure shows that health spending and self-reported symptoms co-move, confirming that the events are correlated with decreases in household health 20An alternative way of identifying shocks would be to identify households who report not having been able to work due to illness. Hendren et al. (2018) follow this approach using the same dataset. However, we follow a different approach as we are interested in extreme events that are related to severe financial needs. For instance, a worker could catch an infection and thus miss some weeks at work, but that may not necessarily imply large spending needs. 21In some cases, our approach identified more than one sudden increase per household—i.e., increases of the same magnitude. In such cases, we only focus on the first increase to avoid sample selection issues due to repeated shocks. 22As we do not observe pregnancies or births directly, we do so by excluding events that coincide with the inclusion of a new child in the household roster within 12 months of the sudden increase in health spending. 23For example, if the episode of high health spending was recorded in month 100 and the symptoms started being reported three months before, the beginning of the event is month 97. 24There were 19 households for which symptoms were repeatedly reported for two years or more, and 68 households for which we did not find information related to symptoms. 13 endowments. Appendix Table A1 reports the distribution of types of health symptoms reported by shocked households during the two years around the shock, during non-shock periods, and during all the sample periods. Relative to non-shock periods, there is a lower incidence of transitory symptoms such as headaches, colds, cough or influenza during shock periods. In contrast, during the periods related to the shocks, there is a higher incidence of other less common symptoms which might be more severe. Magnitude of the shock. Appendix Figure A3 also shows that episodes of high health spending represent a substantial financial burden for the households: on average, that increase in health spending is twice as large as the monthly average per-capita food expenditure, and represents 18% of average monthly household income. The figure also demonstrates that food consumption is smoothed even as health spending spikes, which we confirm via regression analysis below. Shocks to household budget or household labor supply? The shocks are related to a substantial increase in spending needs but also to substantial declines in health status. Appendix Figure A2 shows that 50% of the shocks affected family members that were 52 or older, and that 10% of the shocks affected children under the age of 18. Thus, the majority of shocks are related to illness of non-prime age household members. Indeed, Appendix Table A2 shows that in the month before experiencing the shocks, on average, affected individuals spent most of their days helping with household chores rather than working in family businesses. Moreover, the distribution of type of symptoms around the shock matches more closely that of older individuals (see Appendix Table A1). Thus, for the subsample of shocks affecting non-prime age household members, we interpret the shocks as financial shocks. In contrast, around 40% of the events relate to household members in prime-working age, and we interpret this subset of shocks as affecting both spending needs and labor endowments. Below we test whether “financial only” shocks have different effects than “financial plus labor” shocks. Are the shocks idiosyncratic? Our analysis requires that the events be idiosyncratic and that their occurrence be uncorrelated with trends in household behavior. The top panel of Appendix Figure A4 presents the distribution of months associated with the beginning of each event. It shows that the event start dates are spread through all the periods in the sample and suggests that the events are indeed idiosyncratic. Indeed, the bottom panel shows that, in over 87% of the cases, the shocks affected only one household per village at the same time.25 25We formally test whether village-level trends explain the occurrence of these events in Appendix Table A3. 14 4 Direct Effects of Idiosyncratic Shocks 4.1 Research Design Estimating the effects of idiosyncratic shocks on household outcomes requires a valid comparison group. Ideally, we would like to compare changes in outcomes before and after the shock between shocked households and otherwise-similar households who, by chance, were not simultaneously exposed to such a shock. We follow Fadlon and Nielsen (2019) and exploit the plausibly random variation in the timing of severe health shocks to approximate this ideal scenario. That is, we compare households that suffer a shock in a given period with households who will suffer a similar shock, but later in time. We follow the approach of Fadlon and Nielsen (2015, 2019) to construct counterfactuals for shocked households. We compare the behavior of households that experienced the shock in period t(i.e., treated households) to the behavior of households from the same age group and village that did not experience the shock at time t, but experienced a similar shock later on, in period t+ ∆ (control households). We then undertake a difference-in-difference approach to estimate the effect of the shock by computing differential changes in outcomes, before and after the treated household’s shock, between treatment and control households. The underlying identification assumption is that, in the absence of the shock, the outcomes of households in the treatment and control group would have followed parallel trends. By comparing households in the same village and age group, we isolate contemporaneous village-specific shocks and potential differences in the trajectories of business and household-finance outcomes that could vary along the life cycle (Silva et al., 2019). We validate the identifying assumption using flexible difference-in-difference estimates that allow us to analyze whether there were systematic differences across treatment and control groups before the occurrence of the shocks. We operationalize this approach in three steps. First, we split households into two age groups— i.e., below and above the median household age at baseline (1997).26 Given our sample size, we choose two age group bins to ensure that we have multiple observations per bin in each village. Second, for each age group within each village, we split the panel into two equal-length sub-samples {θ1, θ2}by taking the midpoint between the months associated to the first and last shocks in each age group-village bin (∆), such that those households suffering a shock between periods tand 26One alternative way of assigning households into cohorts is by focusing on the age of the household head. However, that approach ignores the age structure of the household, as in several cases several families are part of the household. 15 tmed =t+∆ belong to the treatment group (θ1), and those experiencing the shock between periods tmed and ¯ tbelong to the control group (θ2).27 By construction, there is no overlap between the two groups. Third, we assign a placebo shock to each household in the control group ∆ periods before it experienced its actual shock. Thus, if a household in the control group experiences the actual shock in t00, its placebo shock is assigned to period t00 −∆. Because the timing of the shocks is evenly distributed over time (see Appendix Figure A4), the placebo shocks occur within the domain of the actual shocks. As 243 out of 473 shocked households experienced a shock in the earlier part of the panel, this process yields 243 households in the treatment group and 230 in the control group. By using households that experience a shock ∆ periods (approximately 5 years) in the future, this process ensures that none of the households in the control group experienced a shock themselves during the analysis period. This is potentially important, as Hendren et al. (2018) show that households that experience illness are more likely to experience other illness episodes in the future. This approach reduces the threat of biases arising from contemporaneous shocks affecting the control group, but it comes at the cost of precision since we do not exploit the occurrence of the actual shocks in the second part of the sample. To increase power, we also report estimates exploiting the variation associated with shocks to households in the second half of the sample for robustness. In this case, the comparison group consists of households that suffered the shock earlier on and their corresponding placebo shock occurs in period t0+ ∆; ∆ periods after their actual shock. Including this variation does not materially alter the point estimates, but it increases statistical power. Figure 2 plots means of health spending and the self-reported probability that at least one household member experienced health symptoms over time, for the treatment and control groups. It shows that the control group does not experience any change in health spending or health status around the placebo shock. In the case of the treatment group, the sharp increase in health spending coincides with sharp increases in spending on inpatient and outpatient care. The magnitude of the increase in health spending suggests that health shocks were quite severe. The figure also demonstrates that, prior to the shock, the treatment and control groups are on similar trajectories in terms of spending, symptoms, and probability of receiving care, supporting the parallel trends assumption. 27We define ∆ as ∆ = ¯ t−t 2for each age-group-village bin. On average, each sub-sample covers 66 months. We exclude shocks occurring during the first and last 24 survey waves to ensure that we observe pre and post outcomes for at least two years for all households—i.e., t >= 24 and ¯ t <= 148. 16 4.2 Estimation We estimate the following generalized difference-in-difference specification, following Fadlon and Nielsen (2019): yi,t = τ=3 X τ=−4,τ6=−1 βτI[t=τ]×Treatmenti+ τ=3 X τ=−4,τ6=−1 θτI[t=τ] (2) +γ+Xi,tκ+αi+δt+i,t where yi,t denotes the outcome of interest corresponding to household i, during period t. We control for time-invariant household characteristics and aggregate time-varying shocks by including household fixed effects (αi) and month fixed effects (δt). We denote T reatmentias an indicator of whether the observation belongs to the treatment (T reatmenti= 1) or control (Treatmenti= 0) group. Time to treatment is denoted by τi,t and is measured in half-years to increase precision. X is a vector of time-varying demographic characteristics including the number of male and female household members, age of the household head and maximum years of schooling in the household. The coefficients of interest are {βτ}τ=3 τ=−4, which compare differences in changes in outcomes with respect to the period preceding the shock (τ=−1) between households in the treatment and control group. We focus on a two-year (i.e., four-half year) time window before and after the shock as our panel is fully balanced during this period. We complement equation (2) with a more parsimonious difference-in-difference specification: yi,t =βP osti,t ×Treatmenti+θPosti,t +Xi,tκ+αi+δt+i,t (3) where Posti,t is an indicator that takes the value of 1 in periods following the shock, and 0 otherwise. The parameter of interest, β, compares differences in outcomes before and after the shock, between households in the treatment group and the comparison group. In both specifications, we cluster standard errors at the household level as our main source of variation comes from cross-household variation in the timing of events.28 4.3 Results Outflows of resources must equal inflows of resources plus changes in cash holdings. Thus, a shock to health spending, which entails large outflows of resources, can be financed through four types 28This also flexibly accounts for serial correlation within the household over time (Bertrand et al., 2004). 17 of (non-mutually exclusive) adjustments. First, the shock could crowd out non-health spending. Second, households may liquidate assets to finance spending needs. Third, households may receive other inflows of resources, either as gifts from other households, government transfers, or loans. Finally, the shock could affect household production decisions—reducing hired labor or business input spending to free up resources to meet the shock. Figure 3 reports flexible difference-in-difference estimates following the specification in equation (2). Panel (a) shows that, relative to the control households, the shocked households experience a sharp increase in total expenditure. Panel (b) shows that shocked households experience a sudden increase in incoming gifts from other households. This increase in gifts from other local households highlights the importance of local informal insurance networks; when idiosyncratic shocks occur, other households respond by providing gifts or transfers to the affected household. During the first six-month period after the shock, the increase in gifts and loans only covers around one-half of the increase in spending needs due to the shock.29 As households are only partially insured through gift/transfer networks, the shocks affect other household financial decisions, namely production-side decisions. Panel (c) shows that there is a decline in fixed assets after the shock. Panel (d) shows that, compared to households in the control group, labor usage declines for shocked households. Panel (e) shows that input spending falls after the shock. Finally, Panel (f) shows that the slowdown in input spending coincides with a slowdown in revenues after the shocks. Note that the sharp declines in input spending and revenues coincide with the sharp increase in spending induced by the shock. One implication is that, as fixed assets can be hard/costly to liquidate quickly, households may have to meet short-term liquidity needs by drawing down working capital. The dynamics of these effects suggest that households use gifts to (partly) finance immediate expenses, and they appear to rely on alternative sources for financing remaining expenses related to the shock. Thus households may follow a pecking order when it comes to financing or coping with adverse shocks; they first rely on gifts, which may be less costly, and then turn to resources meant to fund their family businesses, which could compromise future income. We explore this possibility in the next subsection. The graphical evidence suggests that despite the receipt of gifts and transfers, health-driven 29While gifts and loans are also received in later post-shock periods, we show below that households appear constrained in their ability to borrow and so, even if these future transfers are anticipated, they cannot be used to meet immediate spending needs. 18 shocks to household spending affect production decisions, inconsistent with the separation theorem. The graphical evidence also documents parallel pre-trends: for all six outcomes, there are no spuriously significant “effects” prior to the spending shock.30 To provide a quantitative assessment of the overall impact of the health shocks, in Table 3 we report difference-in-difference estimates of the effect of the shock on outcomes, corresponding to equation (3). Panel A examines household spending. Column 1 shows that the shock leads to a large increase in health spending. While this is by construction, the magnitude is notable, representing a roughly 350% increase relative to the baseline mean. Column 2 shows that during the two years following the shock, on average, total spending increases for shocked households, relative to control households, by an amount close to the effect on health spending, though the effect is not significant at conventional levels. Columns 3 to 5 shows that there are neither substantial nor significant effects on non-health spending (total, non-food, and food, respectively). Thus, in terms of non-health consumption spending, shocked households were able to buffer the shocks, despite strong pressures on household budgets. We next turn to analyze whether the shocked households draw down assets to finance their spending needs. Panel B of Table 3 shows that households did not rely significantly on either deposits in financial institutions or cash on hand to cover their health expenses. We also fail to detect significant changes in inventory or livestock, which are traditional proxies for buffer-stock savings. Similarly, we do not find significant changes in household fixed assets. While savings and the stock of fixed assets decrease, the decrease is not significant over the two-year post-shock period. Thus, incoming gifts allow households to (partly) cope with the shocks and, as we show below, the remainder of the shock is buffered by reducing business spending. Panel C of Table 3 shows that, although incoming gifts significantly increased, there were no detectable effects on borrowing. One interpretation is that obtaining credit from banks or community-based organizations is costly and/or entails a significant amount of delay. For instance, village funds do not meet often enough to evaluate loan applications quickly. Panel D of Table 3 shows that affected households significantly decrease spending on business inputs (column 1) and reduce the use of external labor (column 2). Households also appear to reduce the use of labor provided by household members (column 3), though the effect is not significant. There is no significant effect on business assets (column 4). As a result of reduced investment in inputs and labor (columns 1-3), there is a decrease in revenues from family enterprises, as seen in 30Appendix Figure A5 shows the same dynamics in the raw data. 19 column 5. (The effect on revenues has a p-value of 0.106.) Thus, households insure consumption against health shocks, but consumption smoothing comes at the cost of a decline in household business investment and revenues. 4.3.1 Robustness Robustness to using shocks occurring in the second half of the panel. Our main analysis uses households that experienced the shock in later periods as a comparison group for households that experienced the shock earlier on. To increase power, we also report results using households that experienced the shock in the earlier periods as a comparison group for households that suffered the shock in later periods. Appendix Table A4 replicates the results from the previous section and shows results that are quantitatively similar, but estimated with higher precision since we now use 473 shocks as opposed to only 243, as in Table 3. By adding more shocks, we are able to detect significant declines in fixed assets, household labor, and revenues. Alternative definitions of shock onset. Throughout our analysis, we use the first consecutive month in which households reported experiencing health symptoms within the six months preceding the peak in health spending in order to account for potential anticipation effects that could bias the results. Appendix Table A5 reports results from two alternative definitions of the beginning of the shock. Panel A reports estimates of the effects of the health shocks assuming that the beginning of the event coincided with the peak in health spending. This approach provides a lower bound, since households may have adjusted their behavior before the peak. Panel B reports estimates assuming that the event started six months before the observed peak. In both cases, the estimates are qualitatively similar to those from our main specifications. Alternative definitions of comparison groups. We report two robustness checks that rely on different comparison groups for our analysis. Our main specification assigns placebo shocks ∆ periods away from the actual shocks, within village-age groups bins. An alternative approach would be to randomly allocate the placebo event within each village bin. The main difference between these approaches is that our main specification ensures that the placebo group does not suffer a shock during the two-year comparison window. In contrast, the random assignment of the placebo event could coincide with other shocks. Panel A in Appendix Table A6 reports results using the random placebo assignment, based on a uniform distribution between the months of the first and last shock in each village. The results are qualitatively similar to those from our main specifications. 20 no evidence of differential pre-trends. Figure 5 shows an analogous result for total consumption expenditure, which declines in the post-shock period (and exhibits no differential trend in the pre-period). The effects on transactions, income and spending are evident in all three half-year periods following the shock and do not exhibit evidence of shrinking in magnitude over time. Thus the effects are quite persistent. In theory, indirectly-hit households might attenuate these effects over time by finding new local trading partners. However, the evidence on the rigidity of local networks shown above (section 2.4) demonstrates that such reorganization of local ties is very difficult, at least over the span of 1-2 years. Panel A of Table 5 shows difference-in-difference estimates corresponding to (6). It documents significant post-shock declines in the number of monthly transactions in the supply-chain (column 1) and labor-market networks (column 2), and in total transactions (column 3). These effects are large, representing declines of 20%, 24% and 21% relative to the pre-period means, respectively. Column 4 shows that these changes, in turn, reduce income: a one-unit increase in Closeness is associated with a fall in income of THB 1,236, or 11.7% of the pre-period mean. As seen in column 5, incoming gifts and loans do not change in response to indirect exposure to shocks. With income falling and gifts and loans not filling the gap, consumption spending falls by THB 294, or 4.1% of its pre-period mean (column 6).42 The fall in consumption is smaller than the fall in income, suggesting that indirectly shocked households were able to partly, though not completely, smooth out the health shocks affecting their networks.43 The reason the (non-health) spending of directly-shocked households does not fall in response to a health shock (see Section 4), while the consumption of indirectly-shocked households does fall, relates to the differing response of incoming gifts. Directly shocked households see economically and statistically significant increases in transfers, while indirectly shocked households do not. (It is worth noting that, in addition to receiving transfers, directly shocked households take other, costly, steps to buffer consumption, namely scaling back on business activities.) Two factors, not mutually exclusive, may explain the divergence in transfer behavior. First, the direct shocks are large increases in health spending, associated with changes in health symptoms. These shocks 42Recall that these are the effects associated with moving from Closeness = 0 0, i.e., being unconnected to the directly shocked household, to Closeness = 1, i.e., being directly connected to the shocked household. The mean level of Closeness = 0.42, so that the average indirect effect is 42% of the coefficient. 43In Appendix Table A8, we re-estimate equation 5, including village-by-month fixed-effects (υv,t) to control for potential village-and-time-specific shocks. The results are quite similar to those from the main specification. 27 are salient and relatively unlikely to raise concerns of effort (moral hazard), verifiability (hidden income), etc. The indirect shocks, on the other hand, arise from reductions in supply and demand facing household businesses. Such shocks are likely less salient and potentially more subject to concerns of effort and verifiability, and hence potentially less insurable. Second, because the indirect shock, by its nature, affects many interlinked households, the shock becomes de facto aggregate, which makes the potential for insurance via gifts from other villagers more limited. 5.2.1 Propagation of Health Shocks through Supply Chain vs. Labor Networks In Panel B of Table 5, we examine whether the effect of exposure through the supply chain network has different effects than exposure through the labor market network. Because the two networks are correlated, we analyze the effect of exposure to one controlling for the effect of the other.44 Column 1 shows that, conditional on proximity in the labor market network, a one-unit increase in proximity in the supply chain network is associated with a significant fall in input/output transactions of 0.228. There is no effect on input/output transactions associated with proximity through the labor network. Analogously, column 2 shows that proximity through the labor market network has a negative and significant effect (-0.21) on transactions involving paid labor, while there is no effect seen via the supply chain network. The fact that proximity through the supply chain (labor) network is associated with changes in input/output (hired labor) transactions, and not vice versa, is supportive of the identification assumption, as confounds (e.g., differential trends between closer vs. more distant households) would likely manifest in both sets of outcomes. In column 3, proximity via the supply chain network and the labor market network both have negative and significant effects on the total number of transactions (-0.206 and -0.244, respectively). In column 4, proximity via the labor market network is associated with a large and significant drop in income, while the effect of proximity via the supply chain network is small and insignificant. Column 5 shows that neither networks’ proximity is associated with a large or significant effect on gifts/transfers. Finally, in column 6, proximity via the labor market network is associated with a large and significant drop in consumption spending, while the effect of proximity via the supply chain network is small and insignificant. We return to the interpretation of this difference in the 44On average, 41% of households share a direct or indirect link to the shocked households through both, supplychain and labor-market network, 16% are directly or indirectly linked to the shocked household only through the supply-chain network, 13% are directly or indirectly connected to the shocked households only through the labor network and 30% of households are neither connected to the shocked households through the supply-chain nor labor network. 28 impacts of exposure via the supply chain vs. labor markets in Section 6.2. Transactions in vs. transactions out Motivated by the fact that directly shocked households reduce their demand for inputs and labor, we have focused our discussion on “upstream” propagation, to households selling goods or labor to the directly shocked household. However, directly shocked households also see falls in revenues (Table 3, Panel D, col 5), so there are likely to be negative indirect “downstream” effects on indirectly shocked households’ ability to purchase inputs and labor. We investigate these effects in Appendix Table A9, which separately presents effects on outgoing/upstream transactions (columns 1 to 6) and incoming/downstream transactions (columns 7 to 12). Panel A shows that, using the measure of network proximity which combines the supply chain and the labor market networks, the fall in outgoing transactions documented above is matched by a fall in incoming transactions (input purchases and labor hiring, as well as total incoming transactions) which is similar in magnitude; both sets of effects are statistically significant. In Panel B, we distinguish closeness through the supply chain network and the labor market network. Controlling for closeness in the labor market network, closeness in the supply chain network is associated will falls in both outgoing purchases and incoming sales. Controlling for closeness in the supply chain network, closeness in the labor market network is associated with falls in both outgoing labor provision and incoming labor hiring. Both networks independently predict falls in total transactions. In sum, the health shocks we study generate indirect effects both upstream and downstream, as the costly adjustments taken by the directly shocked household reverberate through the local economic networks. 6 Propagation Mechanisms 6.1 Access to Insurance and the Propagation of Shocks Section 4 presented suggestive evidence that idiosyncratic shocks were more likely to trigger declines in business activities when the shocked household had limited access to informal insurance. This suggests that shocks to uninsured households may propagate more to other households. To test this hypothesis, we estimate the following difference-in-difference model: 29 yi,t,j =β1Postt,j ×Closenessi,j +β2P ostt,j ×Closenessi,j ×Accessj(7) +β3Postt,j ×Accessj+β4Closenessi,j ×Accessj+γClosenessi,j (8) +Xi,t,jκ+αi+ωj+δt+θτ+i,t,j where Accessjis an indicator of whether directly shocked household jhad above-median pre-period access to informal insurance networks, measured by the number of transfers and loans exchanged with other households in the village during the year preceding the shock. As above, Closenessi,j denotes inverse distance between household iand directly shocked household jduring the year preceding the shock. For simplicity and to maximize power, we focus on overall closeness. The coefficient β1measures the change in outcomes after the shock associated with a one-unit change in proximity to the shocked household when that shocked household has below-median access to informal insurance, and β2captures the differential indirect effects when the shocked household had above-median access to informal insurance networks. The total indirect effect associated with a oneunit change in proximity to a shocked household with above-median access to informal insurance networks is β1+β2. Table 6 provides evidence that shocks to less-insured households propagate differently, compared with those to more-insured households. Column 1 shows that input/output transactions by indirectly affected households fall by 0.26 (from a base of 0.999) when the shocked household had low access to insurance in the pre-period. When the shocked household had high access to insurance, however, the fall in sales is much smaller (the differential effect is significant at the 1% level) and the net effect is no longer significant. For hired labor (column 2), the fall for indirectly affected households is 0.107 (from a base of 0.470) when the shocked household had low access to insurance in the pre-period; in this case, the fall when the shocked household had high access to insurance is similar. Column 3 shows that the fall in total transactions (summing input/out sales and hired labor), is 0.367 (from a base of 1.47) when the shocked household had low access to insurance in the pre-period; when the shocked household had high access to insurance, the fall in sales is reduced by 0.09 percentage points; although this difference is not precisely estimated it accounts for 25% of the decline in total transactions in the case of shocks to uninsured households. Column 4 shows that income displays the same pattern: when the shocked household had low access to insurance in the pre-period, the fall in income associated with one-unit greater Closeness is 1,654 baht. When the shocked household had high access to insurance, the fall in income is 30 lower by 1,031 baht (the differential effect is significant at the 1% level) and the net effect is not significant. Column 5 shows that incoming gifts to indirectly shocked households do not respond, regardless of the baseline insurance status of the directly shocked household. Finally, column 6 shows that the consumption spending of indirectly affected households falls by 321 baht, or roughly 5%, when the shocked household had low access to insurance in the pre-period (the p-value of this effect is 0.134). When the shocked household had high access to insurance, however, the fall in consumption spending is reduced by 94 baht (the differential effect is significant at the 1% level). 6.2 Total Spillovers and Access to Insurance Our results highlight that the extent to which shocks propagate is a function of whether the directlyshocked household has access to informal insurance, with shocks to well-insured households being buffered more successfully and hence spreading less through networks. To assess how the total indirect impacts of shocks relate to informal insurance, we calculate how much shocks to morevs. less-insured households propagate. Households with low access to insurance have a mean (median) of 21.42 (17) contacts and mean (median) closeness of 0.39 (0.38). The figures for those with high (i.e., above median) pre-shock access to insurance are similar, with a mean (median) of 21.14 (18) contacts and mean (and median) closeness of 0.46. Using the figures from Table 6, column 6, the consumption effect associated with a one-unit increase in Closeness to a poorly-insured shocked household is a fall of 1,654 baht (significant at 1%). The implied total indirect effect from poorly-insured households using mean values is therefore −1,654 ×0.39 ×21.42 = −13,817 baht per month. Using median values instead gives an implied total indirect effect of -10,685. For well-insured households, the effect associated with a one-unit increase is Closeness is a fall of 622.2 baht (not significantly different from zero). The implied total indirect effect from wellinsured households using mean values is -6051 baht per month. Using median values instead gives an implied total indirect effect of -5,152. Thus, the aggregate “spillover” effect is nearly twice as large when the shock hits a poorly-insured household. Taken together, these results suggest that when households facing health spending shocks have better ex-ante access to informal insurance networks, the propagation of idiosyncratic shocks to other households is meaningfully reduced. This, in turn, suggests that policies that improve informal insurance—e.g., by reducing frictions due to moral hazard, limited commitment, hidden income or asymmetric information—will have spillover benefits even to those whose direct access to risk31 sharing does not improve. The cash vs. labor costs of health shocks. The direct-exposure results in table 4 show that, while input spending and revenues of the shocked household decline less when that household has access to insurance, hired labor is affected even for high-insurance access households (column 6). Likewise, the propagation results in Table 6 show no significant differences in labor-market transactions (column 2) as a function of the insurance status of the directly-shocked household. These results suggest that, while incoming transfers help shocked households with expenses, they are unable to compensate for the time required to take care of the ill household member. As illnesses to prime-aged members induce a decline in hired labor provided by household members (Table 4, Panel B, col 6), this suggests that household labor and hired labor are complements, namely that household labor is required to supervise or monitor hired labor. This is further evidence of labormarket imperfections driving propagation, and moreover, such imperfections blunting the ability of gifts and transfers to smooth the impact of shocks. A related finding appears in Table 5, Panel B, namely that the indirect-effect declines in income and consumption associated with exposure via the labor market network are large and highly significant, while those associated with exposure via the supply chain network are indistinguishable from zero. This suggests that households connected to the shocked household via the labor market network may suffer a double impact, namely, reduced labor demand via an income effect as the directly hit household scales back, as well as a further hit due to the complementarity between household and hired labor. Propagation via Insurance? It is also possible that households connected to shocked households could have provided transfers in order to help them attenuate the consequences of the shocks. Thus, the indirect effect on consumption (Table 5, column 6) could be a consequence of a decline in cash on hand/liquidity arising from helping the directly shocked household. However, Appendix Table A10, shows that neither transfers nor loans given by the indirectly shocked household to other households increase following the shock. If anything, the point estimates suggest that the provision of transfers to other households declined among higher-exposure households, perhaps because they were experiencing negative indirect effects themselves. Moreover, the effect on consumption, while meaningful, is less in both absolute and percentage terms than the effect on income (column 4), suggesting that the fall in consumption can be fully explained by the fall in income stemming from reduced local sales in the input/output and labor networks. 32 7 Putting the Findings in Context In this section, we offer two exercises to help benchmark the magnitude of our results. First, we show how the propagation of household-level shocks compares with that of a sector-wide shock. Second, we perform a simple back-of-the-envelope exercise to estimate the total magnitude of indirect vs. direct effects on revenues. 7.1 Aggregate vs. Idiosyncratic Shocks To illustrate the idea that local networks may be less able to provide insurance against aggregate shocks, Figure 6 provides a graphical comparison of the direct responses to idiosyncratic vs. aggregate shocks. We consider the timing of the European Union ban on Thai shrimp imports, which was announced in May 2002 and directly affected over 30% of the households in Chachoengsao, the shrimp-producing province in our dataset.45 The “sectoral” line in the figure depicts differences in changes in incoming gifts, before and after the EU ban, between shrimp farmers and non-shrimp farmers.46 The “idiosyncratic” line depicts changes in incoming gifts before and after health shocks, relative to a placebo group as in equation (3). There is only a small, insignificant increase in gifts within a year of the shrimp ban, in marked contrast to the sudden increase in gift inflows in the aftermath of the idiosyncratic shocks. This simple comparison suggests that the effectiveness of local risk-sharing networks depends on the nature of the shock and echoes our finding that, while households directly hit by idiosyncratic health shocks see an increase in gifts (as seen in the figure), indirectly-shocked households do not, as the indirect effects are quasi-aggregate, hitting many households in the network. 7.2 The Multiplier Effect of Idiosyncratic Shocks As documented above, idiosyncratic health shocks have both direct costs (analyzed in section 4) and indirect costs (analyzed in Sections 5 and 6). The former are larger on a per-household basis, but the latter can potentially affect many more households. In order to compare their overall magnitude, and so obtain an estimate of the overall “multiplier effect” of the fall in spending 45Giannone and Banternghansa (2018) show that the EU ban led to significant declines in revenues, to spillovers to non-shrimp households, and to reallocation of resources towards non-shrimp businesses. 46In this case, we estimated a regression of incoming gifts on time-to-treatment fixed effects, the interaction of time-to-treatment fixed effects with an indicator of whether the households operated shrimp farms at baseline, and household fixed effects. We plot the coefficient associated with the interactions. 33 associated with the shock, we perform a simple calculation of the total indirect fall in consumption for each baht of reduced business spending by directly affected households. The indirect effect on consumption associated with a 1-unit change in Closeness, from Table 5, Panel A, column 6 is a fall of 308.9 baht (significant at 10%). The mean (median) level of Closeness in the village network is 0.42 (0.43) and the mean (median) number of indirectly exposed households (i.e., households who are connected to the shocked household via the network) is 21.23 (16).47 The implied total indirect effect using mean values is therefore −308.9×0.42 ×21.23 = −2,754 baht per month. Using median values instead gives an implied total indirect effect of -2,125. From Table 3, Panel D, column 1, the fall in business costs for a directly affected household is 1,783.4 baht, so the indirect effects using mean and median closeness represent multiplier effects of 1.54 and 1.19, respectively. For comparison, Egger et al. (2019) estimate a consumption-expenditure multiplier of 1.7 from cash transfers in Kenya, and in the United States, Nakamura and Steinsson (2014) estimate an “open economy relative multiplier” of 1.5 and Su´arez Serrato and Wingender (2016) estimate a local income multiplier of government spending of 1.7 to 2. These estimates are quite similar to ours despite very different data and methods. While our multiplier estimates are admittedly back-of-the-envelope, they demonstrate that, because the indirect effects are economically meaningful and affect many households for each directly affected household, the total indirect effects are of a similar order of magnitude, and perhaps larger than, the direct effect itself. 8 Concluding Remarks We study the dual role of networks—providing insurance and propagating idiosyncratic shocks. Such shocks are widespread in both highand low-income countries. Moreover, the novel coronavirus (COVID-19), while to a large extent an aggregate shock, has significant idiosyncratic aspects due to variation in household infection risks and realizations (Jordan et al., 2020), the ability to work from home (Dingel and Neiman, 2020; Angelucci et al., 2020), and the extent to which different workers and sectors of the economy are affected by shutdowns and social distancing (Daly et al., 2020). Thus the importance of understanding the ramifications of such shocks has, if anything, increased in the COVID-19 era. 47We report medians as well as means since the median is less sensitive to networks with a high number of connections or many distant (low-Closeness) connections, where the linear specification for Closeness may be less appropriate. 34 If markets were complete, idiosyncratic shocks would be fully insured. Shocks would not affect production and hence no propagation would take place through production networks. However, in the absence of complete markets, spending-side shocks affect production-side decisions and, in turn, ripple out to other households. In such a situation, indirectly-affected households will see falls in local transactions and income and so are forced to cut consumption. The impact of propagation through supply chain or labor networks could be long-lasting, as suppliers and customers may eventually switch to other partners or activities. This is potentially costly, as there may be fixed costs of each partner (e.g., workers may need to be specifically trained for each employer). These costly adjustments could have been avoided if there were full insurance. Empirically, we use variation in the timing of severe shocks on health spending experienced by households in Thai villages. We document consumption smoothing, i.e., no impacts on non-health consumption expenditure for directly shocked households. This smoothing is partially achieved through local gift and loan networks. However, insurance is partial and, as a result, shocked households need to adjust their production decisions–drawing down working capital, cutting input spending and reducing labor hiring. This propagates the shock to other households. Businesses close to the shocked household in the supply chain network experience reduced local sales. Workers closer to the shocked household in the labor network experience declines in the probability of working locally and reduced earnings. As a result, their consumption falls. Our findings suggest that two sets of interventions might be beneficial: i) policies that prevent granular shocks from propagating and causing aggregate fluctuations, and ii) policies that mitigate the indirect, propagated shocks, if they cannot be prevented. First, improved safety nets can help prevent granular shocks from amplifying to become aggregate. Given that the ability to share idiosyncratic shocks increases with the number of households participating in the insurance network, local networks alone may not be enough to diversify this idiosyncratic risk, such that the shared risk remains covariate. Formal commercial insurance contracts (through companies operating nationally or beyond) or social insurance (through central governments) could allow better risk-coping and better prevention of propagation through production and labor networks. Second, in order to mitigate the impacts of propagation through supply chain and labor networks, one can consider broadening the extent of product and factor markets beyond the local village market via regional or economy-wide platforms, to take advantage of more systematic multilateral matching. 35 References Aaronson, D., R. W. Bostic, P. Huck, and R. Townsend (2004). Supplier relationships and small business use of trade credit. Journal of Urban Economics 55(1), 46 – 67. Acemoglu, D., V. M. Carvalho, A. Ozdaglar, and A. Tahbaz-Salehi (2012). The network origins of aggregate fluctuations. Econometrica 80 (5), 1977–2016. Ahlin, C. and R. M. Townsend (2007). Using repayment data to test across models of joint liability lending. The Economic Journal 117(517), F11–F51. Aiyagari, S. R. (1994). Uninsured idiosyncratic risk and aggregate saving. The Quarterly Journal of Economics 109(3), 659–684. Angelucci, M., M. Angrisani, D. M. Bennett, A. Kapteyn, and S. G. Schaner (2020). Remote work and the heterogeneous impact of covid-19 on employment and health. Technical report, National Bureau of Economic Research. Angelucci, M. and G. De Giorgi (2009). Indirect effects of an aid program: how do cash transfers affect ineligibles’ consumption? American Economic Review 99 (1), 486–508. Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson (2013). The diffusion of microfinance. Science 341 (6144). Banerjee, A., A. G. Chandrasekhar, E. Duflo, and M. O. Jackson (2019, 02). Using Gossips to Spread Information: Theory and Evidence from Two Randomized Controlled Trials. The Review of Economic Studies. Banerjee, A. V. and E. Duflo (2007). The economic lives of the poor. Journal of economic perspectives 21(1), 141–168. Barrot, J.-N. and J. Sauvagnat (2016). Input Specificity and the Propagation of Idiosyncratic Shocks in Production Networks. The Quarterly Journal of Economics 131(3), 1543–1592. Beaman, L., A. BenYishay, J. Magruder, and A. M. Mobarak (2018, August). Can network theorybased targeting increase technology adoption? Working Paper 24912, National Bureau of Economic Research. 36 −1000 0 1000 2000 3000 THB −4 −3 −2 −1 0 1 2 3 Time to event (half year) Coef 90% CI Total consumption spending (a) Total consumption spending −1000 −500 0 500 1000 1500 THB −4 −3 −2 −1 0 1 2 3 Time to event (half year) Coef 90% CI Gifts/transfers (b) Incoming gifts/transfers −20000 −10000 0 10000 20000 THB −4 −3 −2 −1 0 1 2 3 Time to event (half year) Coef 90% CI Fixed Assets (c) Fixed assets −40 −20 0 20 Hours −4 −3 −2 −1 0 1 2 3 Time to event (half year) Coef 90% CI Hired labor (hours/month) (d) Hired labor (hrs/month) −3000 −2000 −1000 0 1000 2000 THB −4 −3 −2 −1 0 1 2 3 Time to event (half year) Coef 90% CI Business spending (e) Input spending −4000 −2000 0 2000 4000 THB −4 −3 −2 −1 0 1 2 3 Time to event (half year) Coef 90% CI Revenues (f) Revenues Figure 3: Changes in household outcomes before and after the shock Note: Each dot represents differences between treatment and placebo households in changes in outcomes relative to the period preceding the beginning of the shock (τ=−1).The estimating sample includes 2 years before and after the shock divided in half-year bins. All specifications control for household time-variant demographic characteristics, as well as household and month fixed effects. 90% confidence intervals are computed using standard errors clustered at the household level. Costs and revenues exclude costs and earnings associated with the provision of labor to other households or firms. 43 (a) Total transactions (b) Sales network transactions (c) Labor network transactions (d) Total income Figure 4: Indirect effects on transactions and income Note: The Figure presents flexible difference-in-difference estimates of the indirect effects of idiosyncratic shocks on local businesses, following equation (5). All regressions include household fixed effects, event fixed effects, month fixed effects, villageand year-fixed effects, and household size, household average age and education, and the number of adult males and females in each household. Each dot captures differences in changes in outcomes with respect to the half-year preceding the shock (-1) between moreand less-exposed households. Standard errors are two-way clustered at the household (i) and shock level (j). 44 Figure 5: Indirect effects on total consumption spending Note: The Figure presents flexible difference-in-difference estimates of the indirect effects of idiosyncratic shocks on local businesses, following equation (5). All regressions include household fixed effects, event fixed effects, month fixed effects, villageand year-fixed effects, and household size, household average age and education, and the number of adult males and females in each household. Each dot captures differences in changes in outcomes with respect to the half-year preceding the shock (-1) between moreand less-exposed households; that is, the coefficient on I[τj,t =k]×Closenessi,j , where Closenessi,j denotes inverse distance to the shocked household during the year preceding the shock to j. Standard errors are two-way clustered at the household (i) and shock (j) level. 45 −1000 0 1000 2000 3000 −4 −3 −2 0 1 2 3 Idiosyncratic Sectoral By shock origin Effects of shock on incoming gifts Figure 6: Effects on gift reception by type of shock Note: The figures depict flexible difference-in-difference estimates of the effects of idiosyncratic shocks on incoming gifts (in blue) and of the EU shrimp ban on incoming gifts (in red). The figures focus on households in the 4 villages corresponding to the Chachoengsao province, where most of the shrimp activity takes place. The European Union import ban on Thai shrimp was announced in May 2002. The the effects of idiosyncratic shocks are estimated using the specification detailed in equation (3). The effects of the shrimp-ban shock are estimated using a regression of incoming gifts on month fixed effects, normalized with respect to May of 2002, and interactions of month fixed effects with an indicator that takes the value of 1 if the household had a shrimp farm before the shock. In this case, the effects of the shock are captured by the plotted interaction coefficients. In both cases, standard errors are clustered at the households level and are used to plot 90% confidence intervals. 46 Tables Table 1: Summary statistics Panel A: Household baseline characteristics N Mean S.D. 10th %ile 90th%ile Number of household members 509 4.54 1.87 2 7 Number of adults 509 2.87 1.38 1 5 Household head age 507 51.95 13.45 35 70 Average age 509 34.14 12.11 21 52 Household head is a male 507 0.77 0.42 0 1 Years of schooling: Household head 504 4.49 2.59 3 7 Years of schooling: Household maximum achievement 509 8.19 3.64 4 14 Years of schooling: Household average 509 5.09 2.17 3 8 Panel B: Household finance (annual data) N Mean S.D. 10th %ile 90th%ile Net Income in THB: Farm 7635 134389 1378506 -150 316500 Off-farm family business 7635 19095 115540 0 40700 Labor 7635 52816 108492 0 152222 Total from operations (farm+off-farm + labor) 7635 173327 618277 4974.07 410723 Gift/transfers 7635 24107 183826 -11613 75706 Total net income (Operations+Gifts/Transfers) 7635 197434 644150 16241 446693 Consumption in THB Food 7635 32952 21915 11931 60559 Total consumption 7635 98149 99486 24330 204512 Household Assets and Debt Total Assets (THB) 7635 2448596 7431394 194277 4817110 Fixed Assets/ Total Assets (%) 7635 53 27 13 88 Total debt/Total assets (%) 7635 12 21 0 27 Households with outstanding loans (%) 7635 83 38 0 100 Households with outstanding loans from institutions (%) 7635 48 50 0 100 Households with outstanding loans from personal lenders (%) 7635 30 46 0 100 Panel C: Village networks N Mean S.D. 10th %ile 90th%ile Baseline kinship networks: Degree (Number of links) 8344 2.36 2.19 0 6 Baseline kinship networks: Access ( any link) 8344 0.77 0.42 0 1 Financial networks: Degree 8344 0.65 1.36 0 2 Financial networks: Access 8344 0.35 0.48 0 1 Sales networks: Degree 8344 1.26 2.64 0 3 Sales networks: Access 8344 0.48 0.50 0 1 Labor-market network: Degree 8344 3.07 4.42 0 9 Labor-market network: Access 8344 0.62 0.49 0 1 Note: Panel A reports summary statistics about demographic characteristics, measured at baseline. Panel B reports household financial characteristics based on annual averages using a balanced panel of 509 households. Farm income includes income from agriculture, livestock, fishing and shrimping. Off-farm income excludes earnings from labor provision. In both cases income is net of operation costs. Gifts and transfers include transactions from both households inside and outside the village, as well as the receipt of government transfers. Consumption includes spending but also consumption of home production. In Panel C, all networks are unvalued and undirected; all links have equal weight and the direction of the transaction is not considered. Kinship networks are measured at baseline, while transaction networks are measured on an annual basis. Financial networks are constructed based on gifts and loans between households in the same village. Supply chain networks include transactions of raw material and intermediate goods between businesses operated by households in the same village. Labor networks include relationships through paid and unpaid labor between households in the same village. Degree: Number of households with whom each household transacted in each year. Access: Takes the value of 1 if the household has participated in the network during a given year and 0 otherwise. 47 Table 2: Persistence in transaction networks, by network type Probability of a direct link Supply chain Labor Market Gift or Loans VARIABLES (1) (2) (3) (4) (1) (2) (3) (4) (1) (2) (3) (4) Lagged Prob. of link (ρ) 0.469*** 0.460*** 0.379*** 0.379*** 0.426*** 0.401*** 0.333*** 0.333*** 0.260*** 0.258*** 0.209*** 0.209*** (0.015) (0.014) (0.011) (0.011) (0.012) (0.013) (0.011) (0.011) (0.015) (0.015) (0.013) (0.013) Kinship connection 0.099*** 0.099*** 0.109*** 0.109*** 0.091*** 0.091*** (0.006) (0.006) (0.007) (0.007) (0.006) (0.006) Demographic ( log distance) -0.014 -0.107 0.141** (0.120) (0.131) (0.071) Net-worth distance (log of squared differences) -0.037 -0.006 -0.035** (0.027) (0.031) (0.017) Village fixed effects No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes Year fixed effects No Yes Yes Yes No Yes Yes Yes No Yes Yes Yes Household fixed effects No No Yes Yes No No Yes Yes No No Yes Yes Observations 233,240 233,240 233,240 233,240 233,240 233,240 233,240 233,240 233,240 233,240 233,240 233,240 R-squared 0.221 0.227 0.268 0.268 0.189 0.207 0.241 0.241 0.067 0.069 0.102 0.102 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The table presents regression coefficients following the specification in equation (1). We model the probability that a pair of households {i, j} trades in year tas a function of whether the couple traded in period t−1, by type of transaction. Column (1) presents raw correlations, Column (2) includes village-year fixed effects. Columns (3) and (4) control for kinship first-degree connections, differences in baseline demographic characteristics, differences in baseline wealth (e.g., assets net of liabilities), and household fixed effects. The coefficients of Demographic and Net-worth distance are re-scaled by 100. All regressions are estimated over a sample of dyads of households included in the survey sample that responded in all 172 monthly waves of the survey. Standard errors are two-way clustered at the household iand jlevels, and are presented in parentheses. 48 Table 3: Effects on spending, assets, transfers, and family businesses Panel A: Effects on Spending (1) (2) (3) (4) (5) Health Total Non-health Total Non-Food Food Post X Treatment 534.8*** 585.1 50.36 9.039 41.32 (91.18) (367.4) (344.4) (314.0) (69.65) Baseline mean (DV) 152.6 5451.0 5298.3 2903.7 2394.7 Observations 22709 22709 22709 22709 22709 Number of households 451 451 451 451 451 R-Squared 0.0489 0.154 0.146 0.0886 0.676 Panel B: Effects on household savings and assets (1) (2) (3) (4) (5) Savings Cash in hand Livestock Inventories Fixed Assets Post X Treatment -1211.1 -16081.6 -1035.3 -7771.7 -5094.7 (1847.5) (22697.4) (2467.5) (7114.0) (5471.4) Baseline mean (DV) 5278.6 370556.0 36674.4 101918.1 88184.4 Observations 22709 22709 22709 22709 22709 Number of households 451 451 451 451 451 R-Squared 0.0627 0.885 0.926 0.885 0.854 Panel C: Effects on gifts, transfers and debt (1) (2) (3) (4) (5) Gifts from village hhs Gifts/Transfers Borrowing Gifts+Loans Prob. count Post X Treatment 0.0103 0.0140 570.4*** 113.0 668.1* (0.006) (0.009) (220.5) (258.8) (398.7) Baseline mean (DV) 0.0192 0.0249 1957.3 234.4 2736.5 Observations 22709 22709 22709 22709 22709 Number of households 451 451 451 451 451 R-Squared 0.140 0.0695 0.159 0.0105 0.0399 Panel D: Effects on family businesses (1) (2) (3) (4) (5) Costs Hired labor HH Labor Biz. Assets Revenues Post X Treatment -1783.4** -14.33* -10.89 1510.0 -1744.4 (842.6) (7.504) (8.809) (2101.4) (1075.5) Baseline mean (DV) 7610.2 18.11 154.1 33376.9 14939.0 Observations 22709 22708 22708 22709 22709 Number of households 451 451 451 451 451 R-Squared 0.781 0.578 0.712 0.911 0.620 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table reports estimates of βfrom equation (3) for different outcomes. Each column reports differences between treatment and placebo households in changes in outcomes before and after the shock. All regressions control for household demographic characteristics, household and month fixed effects. Standard errors are clustered at the household level. Costs, labor, assets and revenues are aggregated across all businesses operated by household members, and exclude revenues and costs of wage labor provision to other businesses or households. Hired labor and labor provided by household members are measured in hours/month. 49 Table 4: Heterogeneity in the effects of health shocks Panel A: Effects of the shocks by pre-period access to informal insurance (1) (2) (3) (4) (5) (6) (7) (8) (9) Health Spending Total spending Prob. Gift (in village) # of Gifts (in village) Gifts Hired labor HH Labor Costs Revenues Post X Treatment (β1) 402.8*** 628.9 0.00574 0.00494 458.5** -8.895 -18.65* -2596.8*** -4070.7*** (65.57) (520.5) (0.00427) (0.00556) (206.7) (6.467) (9.830) (824.1) (1099.7) Post X Treatment X High Access (β2) 82.76 124.2 0.00594 0.0169 1.537 -5.949 7.040 2053.1* 3947.1** (108.9) (630.5) (0.0110) (0.0156) (351.9) (10.57) (14.69) (1231.9) (1647.3) Effect: High Acceess (β1+β2) 485.6 753.1 0.0117 0.0218 460.0 -14.84 -11.61 -543.8 -123.6 P-val: High Access 0.000 0.0650 0.259 0.138 0.109 0.0952 0.294 0.555 0.919 Baseline mean (DV) 148.8 6055.5 0.0203 0.0258 2248.3 16.93 144.2 7524.5 15228.3 Observations 37325 37325 37325 37325 37325 37325 37325 37325 37325 Adj. R-Squared 0.0435 0.101 0.113 0.0605 0.138 0.677 0.649 0.743 0.542 Panel B: Effects of the shocks by age of ill household member ( prime age 18-60) (1) (2) (3) (4) (5) (6) (7) (8) (9) Health Spending Total spending Prob. Gift (in village) # of Gifts (in village) Gifts Hired labor HH Labor Costs Revenues Post X Treatment (β1) 563.6*** 1064.4*** 0.00896 0.00883 533.7* -1.136 -5.741 -1468.8 -1962.8* (112.1) (403.6) (0.00920) (0.0126) (292.5) (2.999) (12.56) (894.9) (1134.8) Post X Treatment X Prime-working age (β2) -255.8** -935.0 -0.00244 0.00165 -325.8 -17.03* -4.344 -746.8 -726.2 (123.7) (574.3) (0.0119) (0.0159) (363.2) (9.626) (15.63) (1293.4) (1642.1) Effect: Prime-working age (β1+β2) 307.8 129.4 0.00652 0.0105 207.8 -18.17 -10.09 -2215.6 -2689.0 P-val: Prime-working age 0.000 0.754 0.444 0.335 0.339 0.0785 0.280 0.0186 0.0297 Baseline mean (DV) 135.7 5632.1 0.0194 0.0246 2134.2 18.50 142.9 7040.2 14449.8 Observations 29478 29478 29478 29478 29478 29478 29478 29478 29478 Adj. R-Squared 0.0376 0.168 0.0738 0.0399 0.116 0.706 0.659 0.735 0.525 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table reports estimates of β1and β2from equation (4) for different outcomes by access to informal insurance networks, in Panel A, and by whether the shocks relate to illness of non-prime-age or prime-age family members, in Panel B. Each column reports differences between treatment and placebo households in changes in outcomes before and after the shock, and interactions of these differences with the relevant variable capturing heterogeneity. High Access equals to 1 if the number of transfers or loans given/received to/from other households in the village during the year preceding the shock are above the sample median. Prime working age equals to 1 if the shock relates to the illness of a family member of age 18 to 60. Hired and household labor are measured in hours per month. All regressions control for household demographic characteristics, household and village-month fixed effects as well as flexible time-to-treatment trends by access to informal insurance. Standard errors are clustered at the household level. 50 Table 5: Propagation through village networks Panel A: Propagation through village networks (1) (2) (3) (4) (5) (6) Input/Output Hired labor All transactions Income Incoming gifts Consumption Post X closeness (village network) -0.199*** -0.115** -0.314*** -1,236.121*** -107.381 -294.199* (0.061) (0.044) (0.077) (450.332) (125.041) (160.936) Observations 410,578 410,578 410,578 410,578 410,578 410,578 R-squared 0.440 0.231 0.374 0.197 0.140 0.620 Pre-period Mean 0.999 0.470 1.469 10486 2339 7265 Number of events 391 391 391 391 391 391 Panel B: Propagation through supply-chain and labor-market networks (1) (2) (3) (4) (5) (6) Input/Output Hired labor All transactions Income Incoming gifts Consumption Post X closeness (supply-chain network) -0.228*** 0.022 -0.206** -73.483 -251.662 65.746 (0.065) (0.040) (0.081) (489.309) (160.727) (174.830) Post X closeness (labor-market network) -0.034 -0.210*** -0.244*** -1,323.731*** 177.292 -485.050*** (0.066) (0.043) (0.083) (447.973) (138.280) (159.340) Observations 410,578 410,578 410,578 410,578 410,578 410,578 R-squared 0.535 0.274 0.483 0.316 0.254 0.819 Pre-period Mean 0.999 0.470 1.469 10486 2339 7265 Number of events 391 391 391 391 391 391 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table presents estimates of βfrom equation (6). Closenessi,j denotes inverse distance to the shocked household during the year preceding the shock to j. Each coefficient captures differences in changes in outcomes before and after the shock between moreand less-exposed households, through village networks in Panel A, and through supply-chain and labor-market networks in Panel B. Each regression includes household (i), event j, and month fixed effects, as well as demographic characteristics such as household size, average age, education and number of male and female adults. Standard errors are two-way clustered at the household (i) and event (j) level. 51 Table 6: Propagation of health shocks and access to informal insurance of shocked households (1) (2) (3) (4) (5) (6) VARIABLES Input/Output Hired labor All transactions Income Incoming gifts Consumption Post X Closeness (village network) -0.260*** -0.107** -0.367*** -1,654.198*** -40.953 -321.272 (0.066) (0.048) (0.085) (565.595) (150.804) (213.899) Post X Closeness X Insurance (shocked household) 0.116*** -0.026 0.090 1,031.979*** -90.126 93.541*** (0.030) (0.078) (0.130) (64.348) (59.874) (10.928) Effect: High Access (β1+β2) -0.144 -0.133 -0.277 -622.2 -131.1 -227.7 P-val: High Access 0.442 0.233 0.376 0.198 0.141 0.622 Observations 405,073 405,073 405,073 405,073 405,073 405,073 R-squared 0.0498 0.0482 0.0150 0.267 0.371 0.287 Pre-period Mean 0.999 0.470 1.469 10486 2339 7265 Number of events 386 386 386 386 386 386 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table reports estimates of β1and β2from the pooled difference-in-difference equation (7) for different outcomes by access to informal insurance networks of the shocked household. Access to informal insurance : Number of transfers or loans given/received to/from other households in the village (during the year preceding the shock) above the sample median. Each regression controls for household, event, and month fixed effects as well as time-varying demographic characteristics. The estimating sample excludes households who suffered a direct health shock during any of the 24 months following the shocks to other households in their village. Standard errors are two-way clustered at the household (i) and event (j) level. 52 Table A3: Timing of health shocks and village and household characteristics (1) (2) (3) VARIABLES ∆ P(event) ∆ P(event) ∆ P(event) Lagged ∆ P(event) -0.500*** -0.501*** -0.5010*** (0.001) (0.001) (0.0012) Lagged ∆ Total net operating income 0.0002 (0.0009) Lagged ∆ Consumption spending -0.0022 (0.0017) Lagged ∆ Consumption of household production 0.0082 (0.0913) Lagged ∆ Borrowing -0.0008 (0.0010) Lagged ∆ Lending -0.0070* (0.0039) Lagged ∆ Inflows (transfers) 0.0008 (0.0010) Lagged ∆ Outflows (transfers) 0.0003 (0.0004) Lagged ∆ Livestock value 0.0002 (0.0005) Lagged ∆ Cash in hand 0.0005 (0.0004) Lagged ∆ Fixed assets - excludes land 0.0006 (0.0005) Lagged ∆ Land value 0.0003 (0.0004) Observations 80,750 77,163 77,163 R-squared 0.252 0.275 0.2754 Month FE Yes Yes Yes Village FE No Yes Yes Number of households 475 475 475 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The table reports OLS coefficients from changes in the the probability of suffering a shock on period ton lagged changes and village fixed-effects in columns 1 and 2. The bottom panel reports an F-test for the joint significance of the village fixed effects. Column 3 reports similar coefficients including lagged firstdifferences of household-finance variables. Standard errors are clustered at the household level to control for serial correlation. 59 0 1000 2000 3000 THB −4 −2 0 2 4 Time to shock (half years) Treatment Placebo Total consumption (a) Total consumption −500 0 500 1000 THB −4 −2 0 2 4 Time to shock (half years) Treatment Placebo Gifts (b) Assets −5000 0 5000 10000 15000 THB −4 −2 0 2 4 Time to shock (half years) Treatment Placebo Fixed Assets (c) Incoming gifts/transfers −10 −5 0 5 10 hours −4 −2 0 2 4 Time to shock (half years) Treatment Placebo Hired Labor (d) Hired labor (hrs/month) −1000 0 1000 2000 THB −4 −2 0 2 4 Time to shock (half years) Treatment Placebo Costs/Input spending (e) Input spending −1000 0 1000 2000 3000 THB −4 −2 0 2 4 Time to shock (half years) Treatment Placebo Revenues (f) Revenues Figure A5: Changes in household outcomes before and after the shock Note: The Figure plots means of average monthly consumption, savings, cash holdings, and incoming gifts for the four quarters preceding and following the shock. All variables are normalized with respect to the pre-shock mean. Period τ=−1 denotes the half year preceding the sharp increase in health spending. Total consumption spending includes health spending. Revenues include income streams from all household enterprises and exclude earnings from providing wage labor to other households. 60 Table A4: Robustness to including shocks occurring in the second half of the sample. Panel A: Effects on Spending (1) (2) (3) (4) (5) Health Total Non-health Total Non-Food Food Post X Treatment 431.6*** 711.3** 279.7 249.4 30.30 (53.09) (350.6) (346.0) (325.5) (61.10) Baseline mean (DV) 147.3 6025.9 5878.5 3217.0 2661.5 Observations 37881 37881 37881 37881 37881 Number of households 471 471 471 471 471 R-Squared 0.0503 0.102 0.0932 0.0524 0.673 Panel B: Effects on household savings and assets (1) (2) (3) (4) (5) Savings Cash in hand Livestock Inventories Fixed Assets Post X Treatment -1287.6 -6886.9 -1179.9 -3086.1 -10366.0* (1253.1) (16517.8) (1982.7) (5062.9) (6056.6) Baseline mean (DV) 6287.7 434066.6 29655.4 123317.6 95742.2 Observations 37881 37881 37881 37881 37881 Number of households 471 471 471 471 471 R-Squared 0.0813 0.849 0.775 0.834 0.759 Panel C: Effects on gifts, transfers and debt (1) (2) (3) (4) (5) Gifts from village hhs Gifts/Transfers Borrowing Gifts+Loans Prob. count Post X Treatment 0.00873* 0.0131* 479.8*** 229.5 752.4** (0.00522) (0.00724) (171.9) (266.7) (347.1) Baseline mean (DV) 0.0201 0.0255 2247.8 -30.78 2796.6 Observations 37881 37881 37881 37881 37881 Number of households 471 471 471 471 471 R-Squared 0.110 0.0585 0.136 0.0207 0.0382 Panel D: Effects on family businesses (1) (2) (3) (4) (5) Costs Hired labor HH Labor Biz. Assets Revenues Post X Treatment -1666.6*** -11.26** -16.38** -767.4 -2271.6*** (607.2) (5.427) (7.310) (1934.4) (808.3) Baseline mean (DV) 7447.5 16.80 143.6 33048.1 15110.3 Observations 37881 37880 37880 37881 37881 Number of households 471 471 471 471 471 R-Squared 0.743 0.677 0.645 0.880 0.541 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table reports OLS estimates of βfrom equation (3) for different outcomes. The estimating sample includes results 471 households in the treatment and control group. Each column reports differences between treatment and placebo households in changes in outcomes before and after the shock. All regressions control for household demographic characteristics, household and month fixed effects. Standard errors are clustered at the household level. Costs, labor, assets and revenues are aggregated across all businesses operated by household members, and exclude revenues and costs of wage labor provision to other businesses or households. Hired labor and labor provided by household members are measured in hours/month. 61 Table A5: Robustness to alternative definitions of the onset of the shocks Panel A: Beginning of event coincides with the observed peak in health spending (1) (2) (3) (4) (5) (6) (7) (8) (9) Health Total Prob. Gift (in village) # of Gifts (in village) Gifts/Transfers Costs Hired labor (Hrs/Month) HH Labor (Hrs/Month) Revenues Post X Treatment 494.9*** 741.4** 0.0104* 0.0130 550.3*** -1649.0** -10.70* -8.833 -1233.8 (82.52) (326.2) (0.00612) (0.00870) (208.2) (838.9) (5.637) (9.079) (1102.4) Baseline mean (DV) 147.9 5397.2 0.0201 0.0265 1977.3 7561.6 19.01 155.5 14643.0 Observations 22643 22643 22643 22643 22643 22643 22642 22642 22643 Number of households 451 451 451 451 451 451 451 451 451 R-Squared 0.0810 0.196 0.163 0.0929 0.175 0.784 0.570 0.718 0.629 Panel B: Beginning of event starts 6 months before the observed peak in health spending (1) (2) (3) (4) (5) (6) (7) (8) (9) Health Total Prob. Gift (in village) # of Gifts (in village) Gifts/Transfers Costs Hired labor (Hrs/Month) HH Labor (Hrs/Month) Revenues Post X Treatment 280.8*** 9.094 0.00802 0.0104 470.1** -1762.6** -15.23* -11.15 -1637.2 (89.16) (404.9) (0.00664) (0.00972) (211.3) (863.6) (8.310) (8.506) (1088.1) Baseline mean (DV) 214.5 5525.7 0.0178 0.0236 2030.9 7618.8 17.17 151.6 14977.5 Observations 22742 22742 22742 22742 22742 22742 22741 22741 22742 Number of households 451 451 451 451 451 451 451 451 451 R-Squared 0.0762 0.182 0.173 0.101 0.182 0.790 0.596 0.725 0.631 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table reports OLS estimates of βfrom equation (3) for different outcomes. Each column reports differences between treatment and placebo households in changes in outcomes before and after the shock. All regressions include a vector of demographic characteristics as well as household and month fixed effects. Standard errors are clustered at the household level. 62 Table A6: Robustness to alternative placebo groups and specifications Panel A: Randomly assigned placebo shocks (1) (2) (3) (4) (5) (6) (7) (8) (9) Spending Gifts (in village) Gifts (total) Production Health Total Prob. Count Costs Hired Labor HH Labor Revenues Post X Treatment 450.3*** 712.5** 0.00811* 0.00534 271.7* -975.3** -6.471 -14.57** -1671.2*** (64.81) (324.2) (0.00447) (0.00705) (147.7) (474.3) (5.619) (6.006) (633.7) Baseline mean (DV) 198.1 6186.2 0.0226 0.0275 2375.8 7707.3 17.13 142.1 15193.4 Observations 43194 43194 43194 43194 43194 43194 43193 43193 43194 Number of households 472 472 472 472 472 472 472 472 472 R-Squared 0.0620 0.119 0.154 0.0698 0.156 0.779 0.735 0.699 0.572 Panel B: Panel regression with household and month fixed effects (Gertler and Gruber, 2002) (1) (2) (3) (4) (5) (6) (7) (8) (9) Spending Gifts (in village) Gifts (total) Production Health Total Prob. Count Costs Hired Labor HH Labor Revenues Post 522.6*** 94.54 -0.00506* -0.00369 235.8* -472.2* -1.936 -5.541* -986.0** (88.57) (334.3) (0.00288) (0.00480) (130.7) (246.8) (1.349) (3.212) (430.4) Baseline mean (DV) 161.9 6306.1 0.0198 0.0257 2387.3 7938.9 18.92 142.9 15869.2 Observations 21362 21362 21362 21362 21362 21362 21361 21361 21362 Number of households 469 469 469 469 469 469 469 469 469 R-Squared 0.0248 0.0764 0.158 0.0836 0.184 0.781 0.764 0.762 0.582 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The table reports difference-in-difference estimates corresponding to equation (3). Panel A report estimates in which the placebo shocks are allocated randomly. Panel B reports estimates excluding households who suffered the shock in the second half of the survey and their respective placebo group. Standard errors are clustered at the household level. All regressions include a vector of demographic characteristics as well as household and village-month fixed effects. 63 Table A7: Heterogeneity in the effects of health shocks Panel A: Effects of the shocks by pre-period access to informal insurance (1) (2) (3) (4) (5) (6) (7) (8) (9) Health Spending Total spending Prob. Gift (in village) # of Gifts (in village) Gifts Hired labor HH Labor Costs Revenues Post X Treatment (β1) 458.3*** 569.7 0.0135* 0.0120 523.4* -14.12 -15.44 -2703.7** -3712.4** (98.17) (553.0) (0.00706) (0.00885) (271.1) (11.38) (12.45) (1261.5) (1603.3) Post X Treatment X High Access (β2) 172.7 -65.72 -0.00740 0.00237 35.81 -1.011 12.38 1828.8 4037.4* (180.2) (815.9) (0.0130) (0.0195) (452.2) (18.01) (19.40) (1709.7) (2075.9) Effect: High Access (β1+β2) 631.0 504.0 0.00612 0.0143 559.2 -15.13 -3.060 -874.9 325.0 P-val: High Access 0.0000616 0.358 0.577 0.404 0.121 0.217 0.827 0.451 0.816 Baseline mean (DV) 154.7 5476.4 0.0193 0.0252 1952.0 18.25 154.6 7696.3 15039.3 Observations 22289 22289 22289 22289 22289 22289 22289 22289 22289 Adj. R-Squared 0.0495 0.154 0.141 0.0702 0.163 0.578 0.716 0.783 0.624 Panel B: Effects of the shocks by age of ill household member ( prime age 18-60) (1) (2) (3) (4) (5) (6) (7) (8) (9) Health Spending Total spending Prob. Gift (in village) # of Gifts (in village) Gifts Hired labor HH Labor Costs Revenues Post X Treatment (β1) 698.3*** 1209.4* 0.00562 0.00491 584.4 3.194 -6.243 -1714.7* -2869.6** (195.4) (618.9) (0.0112) (0.0181) (379.3) (2.699) (14.53) (901.8) (1292.8) Post X Treatment X Prime-working age (β2) -344.8* -1081.4 0.00924 0.00979 -286.3 -31.88* 10.46 -674.4 1578.2 (200.3) (772.2) (0.0149) (0.0213) (450.1) (17.48) (19.28) (1729.6) (2229.9) Effect: Prime-working age (β1+β2) 353.5 128.0 0.0149 0.0147 298.1 -28.69 4.222 -2389.1 -1291.4 P-val: Prime-working age 0.00 0.78 0.15 0.24 0.22 0.09 0.74 0.13 0.51 Baseline mean (DV) 138.2 5064.4 0.0171 0.0223 1838.1 19.34 157.6 6891.0 14090.5 Observations 17579 17579 17579 17579 17579 17579 17579 17579 17579 Adj. R-Squared 0.0336 0.155 0.0747 0.0380 0.118 0.603 0.722 0.746 0.587 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table reports estimates of β1and β2from equation (4) for different outcomes by access to informal insurance networks, in Panel A, and by whether the shocks relate to illness of non-prime-age or prime-age family members, in Panel B. Each column reports differences between treatment and placebo households in changes in outcomes before and after the shock, and interactions of these differences with the relevant variable capturing heterogeneity. High Access equals to 1 if the number of transfers or loans given/received to/from other households in the village during the year preceding the shock are above the sample median. Prime working age equals to 1 if the shock relates to the illness of a family member of age 18 to 60. Hired and household labor are measured in hours per month. All regressions control for household demographic characteristics, household and village-month fixed effects as well as flexible time-to-treatment trends by access to informal insurance. Standard errors are clustered at the household level. 64 Table A8: Propagation through village networks (village X month FE) Panel A: Propagation through village networks (1) (2) (3) (4) (5) (6) VARIABLES Input/Output Sales Hired labor provision All transactions Income Incomig gifts Consumption Post X closeness (village network) -0.184*** -0.083*** -0.268*** -811.808* -167.341 -65.344 (0.045) (0.027) (0.055) (456.340) (123.281) (151.918) Observations 410,578 410,578 410,578 410,578 410,578 410,578 R-squared 0.499 0.319 0.448 0.253 0.187 0.636 Pre-period Mean 0.999 0.470 1.469 10486 2339 7265 Number of events 391 391 391 391 391 391 Panel B: Propagation through supply-chain and labor-market networks (1) (2) (3) (4) (5) (6) VARIABLES Input/Output Sales Hired labor provision All transactions Income Incomig gifts Consumption Post X closeness (supply-chain network) -0.190*** -0.014 -0.204*** -273.328 -278.366* -19.238 (0.049) (0.026) (0.057) (418.882) (151.786) (134.974) Post X closeness (labor-market network) -0.051 -0.144*** -0.195*** -585.115* 80.119 -29.148 (0.048) (0.026) (0.059) (342.531) (113.183) (126.717) Observations 410,578 410,578 410,578 410,578 410,578 410,578 R-squared 0.584 0.357 0.544 0.363 0.295 0.826 Pre-period Mean 0.999 0.470 1.469 10486 2339 7265 Number of events 391 391 391 391 391 391 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table presents estimates of βfrom equation (6). Closenessi,j denotes inverse distance to the shocked household during the year preceding the shock to j. Each coefficient captures differences in changes in outcomes before and after the shock between moreand less-exposed households, through village networks in Panel A, and through supply-chain and labor-market networks in Panel B. Each regression includes household (i), event j, and village-month fixed effects, as well as demographic characteristics such as household size, average age, education and number of male and female adults. Standard errors are two-way clustered at the household (i) and event (j) level. 65 Table A9: Propagation effects on downstream and upstream transactions Panel A: Effects on downstream and upstream transactions Outgoing transactions (Upstream) Incoming transactions (Downstream) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) VARIABLES Input/output sales Labor provision Outgoing transactions Input/output purchases Labor hiring Incoming transactions Post X closeness (village network) -0.078* -0.084** -0.087*** -0.048*** -0.165*** -0.132*** -0.121*** -0.100*** -0.028 -0.036* -0.150*** -0.136*** (0.044) (0.034) (0.025) (0.015) (0.050) (0.038) (0.032) (0.024) (0.027) (0.018) (0.044) (0.032) Observations 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 R-squared 0.574 0.597 0.181 0.271 0.496 0.534 0.369 0.424 0.243 0.285 0.341 0.390 Pre-period Mean 0.497 0.497 0.182 0.182 0.679 0.679 0.501 0.501 0.288 0.288 0.790 0.790 Number of events 391 391 391 391 391 391 391 391 391 391 391 391 Panel B: Effects on downstream and upstream transactions by exposure in the supply-chain and labor-market network Outgoing transactions (Upstream) Incoming transactions (Downstream) (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) VARIABLES Input/output sales Labor provision Outgoing transactions Input/output purchases Labor hiring Incoming transactions Post X closeness (supply-chain network) -0.086* -0.072* -0.017 -0.016 -0.103** -0.088** -0.142*** -0.118*** 0.039 0.002 -0.104* -0.116*** (0.047) (0.040) (0.016) (0.011) (0.048) (0.040) (0.040) (0.030) (0.031) (0.021) (0.054) (0.038) Post X closeness (labor-market network) 0.002 -0.016 -0.109*** -0.059*** -0.107* -0.075* -0.036 -0.035 -0.101*** -0.085*** -0.136*** -0.119*** (0.054) (0.041) (0.027) (0.016) (0.061) (0.045) (0.031) (0.026) (0.027) (0.018) (0.041) (0.031) Observations 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 410,578 R-squared 0.575 0.597 0.182 0.271 0.496 0.535 0.369 0.424 0.243 0.286 0.341 0.390 Pre-period Mean 0.497 0.497 0.182 0.182 0.679 0.679 0.501 0.501 0.288 0.288 0.790 0.790 Number of events 391 391 391 391 391 391 391 391 391 391 391 391 Village-Month FE No Yes No Yes No Yes No Yes No Yes No Yes ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table presents estimates of βfrom equation (6). Closenessi,j denotes inverse distance to the shocked household during the year preceding the shock to j. Each coefficient captures differences in changes in outcomes before and after the shock between moreand less-exposed households, through village networks in Panel A, and through supply-chain and labor-market networks in Panel B. Each regression includes household (i), event j, month fixed effects (odd columns), and village-month (even columns), as well as demographic characteristics such as household size, average age, education and number of male and female adults. Standard errors are two-way clustered at the household (i) and event (j) level. 66 Table A10: Indirect effects of health shocks on gift/transfers to other households (1) (2) (3) VARIABLES Gifts provided Total gifts Gifts+Loans Post X Closeness -0.007 -45.810 -67.610 (0.007) (57.812) (68.728) Observations 410,578 410,578 410,578 R-squared 0.066 0.293 0.216 Pre-period Mean 0.0283 903.9 1035 Number of events 391 391 391 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table presents estimates of the indirect effect of the idiosyncratic health shocks on gifts and transfers provided to other households in the village. The Table presents estimates of βfrom equation (6). Closenessi,j denotes inverse distance to the shocked household during the year preceding the shock to j. Each coefficient captures differences in changes in outcomes before and after the shock between moreand less-exposed households, through village networks. Each regression includes household (i), event j, month fixed effects (odd columns), and village-month (even columns), as well as demographic characteristics such as household size, average age, education and number of male and female adults. Standard errors are two-way clustered at the household (i) and event (j) level. 67 Table A11: Effects of indirect shocks on Savings and Assets (1) (2) (3) (4) (5) VARIABLES Savings Cash in Hand Livestock Inventories Fixed assets Post X Closeness 417.250 -18,216.659 -2,201.766 -11,883.668** -3,579.940 (990.391) (19,842.938) (2,081.598) (4,958.028) (4,087.105) Observations 410,578 410,578 410,578 410,578 410,578 R-squared 0.070 0.818 0.802 0.864 0.777 Pre-period Mean 5612 400531 29204 125374 92866 Number of events 391 391 391 391 391 ∗∗∗p < 0.01,∗ ∗ p < 0.05,∗p < 0.1 Note: The Table presents estimates of βfrom equation (6). Closenessi,j denotes inverse distance to the shocked household during the year preceding the shock to j. Each coefficient captures differences in changes in outcomes before and after the shock between moreand less-exposed households through village networks. Each regression includes household (i), event j, month fixed effects, as well as demographic characteristics such as household size, average age, education and number of male and female adults. Standard errors are two-way clustered at the household (i) and event (j) level. 68