APPLIED ECONOMICS WORKSHOP. Giacomo De Giorgi Stanford University. Wednesday, May 6, :30 to 2:50pm Location: HC 3B

Transcription

1 APPLIED ECONOMICS WORKSHOP Business Spring Quarter 2009 Giacomo De Giorgi Stanford University (Manuela Angelucci, Marcos A. Rangel, and Imran Rasul) "Insurance and Investment in Family Networks" Wednesday, May 6, :30 to 2:50pm Location: HC 3B For any other information regarding the Applied Economics Workshop, please contact Tamara Lingo (AEW Administrator) at , or stop by HC448.

2 insurance and investment in family networks manuela angelucci giacomo de giorgi marcos a.rangel imran rasul First Version: February, This version: May 3, 2009 PRELIMINARY AND INCOMPLETE PLEASE DO NOT QUOTE Abstract We study the role of extended family networks as providers of informal insurance and facilitators of investment in poor rural Mexican villages. Related households (or connected ) can achieve higher insurance and investment than unrelated households (or isolated ) if the family network eases commitment and information issues. Connected and isolated households look remarkably similar. However, the connected 1) share risk only within the extended family and are almost fully insured against idiosyncratic risk; 2) have a smoother consumption and a bigger long-term increase in consumption, income, and investment than the isolated; 3) disinvest less when hit by negative health shocks and invest more in human capital when offered a partial subsidy for schooling, compared with the isolated. The paper has been screened to ensure no confidential information is revealed and approved by an IRB approval of Stanford University. We thank Joe Altonji, Pat Bayer, Raj Chetty, Luigi Pistaferri, Nicola Pavoni, Mark Rosenzweig, Adam Szeidl, Chris Udry, and seminar participants at Duke, Yale, Stanford, Social Economics Workshop, UAB and UCSD, among others, for their comments. All errors remain our own. Department of Economics, University of Arizona [angelucm@eller.arizona.edu]. Department of Economics, Stanford University [degiorgi@stanford.edu]. Harris School of Public Policy, University of Chicago [rangelm@uchicago.edu]. Department of Economics, University College London [i.rasul@ucl.ac.uk]. 1

3 1 Introduction We study whether and to what extent members of extended family networks share resources, enabling them both to be insured against idiosyncratic income fluctuations and to increase investment. Poor households in developing countries face high income volatility but have limited access to formal insurance and credit markets, so resort to informal risk-sharing arrangements. Repeated interactions between small groups of agents may be an important mean to address the information and enforcement problems limiting the extent of informal insurance. These facts suggest the extended family might be an important channel to insure against risk: family members know each other well and are able to monitor and punish deviating behavior by imposing sanctions. Altruism and evolutionary biology (e.g. the need to transmit and protect one s genes) may contribute to the enforcement of insurance agreements. The presence of the extended family may favor investment through three different channels. First, by providing insurance, because investment is linked to consumption smoothing. Poor households may be willing to engage in costly activities to ensure their consumption never falls below subsistence levels, including reducing high-return investments. This need to disinvest to face short-term income fluctuations might have long-term negative consequences. A household that is forced to withdraw its children from school and deplete its assets to cope with a negative income shock is deemed to live in permanent poverty. Conversely, a household that smoothes consumption by sharing resources within the extended family may not need to disinvest as much when hit by a negative shock. On the other hand, uninsured households may invest more than insured households when they have a positive shock, for example because they don t have to share their resources with others. With decreasing marginal returns to the investment, insurance might have positive effects on aggregate investment. Second, the extended family may act as a shareholder if there are non-convexities in the return to the investment and a single household does not have enough resources. In this case multiple related households may pool resources to finance a given investment. Third, the availability of insurance may enable agents to undertake riskier, more profitable investment [Arrow 1971; Obstfeld 1994]. Since insurance affects both the smoothness of consumption and the investment prospects, 2

4 an important metric to understand the comparative importance of the family is not only the relative smoothness of consumption for households with and without an extended family, but also how different their investment strategies are both in terms of physical and human capital. Understanding whether and to what extent members of extended family networks can share resources and how this affects insurance and investment has important implications for the design of optimal policies. For example, it may help identify vulnerable households that are not well-insured against negative income shocks, or fine-tune investment programs for different household types. More broadly, it sheds light on the role of informal institutions as means to overcome market imperfections in developing countries. We model an economy where agents are able to smooth fluctuations in their endowments through transfers. However, transfers are costly. The transaction cost, born by the agent making the transfer, is a reduced form term for all information and enforcement problems that may limit the degree of insurance. The existence of transaction costs limits the amount of insurance between groups of agents. When the cost exceeds a certain threshold, the agents will not insure each other against idiosyncratic fluctuations. Since transaction costs are likely lower for related than for unrelated households, we expect related households, i.e. members of an extended family, to achieve a higher degree of insurance than unrelated households, at the very least, and only the former group to have any transfers at all in a more extreme case. We then extend the model allowing agents to invest. We show that the optimal aggregate investment is a positive function of insurance against idiosyncratic shocks when investments have decreasing marginal returns. The presence of the extended family may increase aggregate investment also when investment is lumpy. In this case the relatives pool resources to undertake an investment that is unaffordable or not optimal for a single household and share the returns. We look at a set of poor villages in rural Mexico, using the experimental data collected for the evaluation of Progresa, Mexico s flagship cash transfer program, as well as subsequent waves after the experiment ended. This is a population with a high need for insurance but no access to formal credit and insurance markets, as typical of developing countries. Our analysis exploits three unique features of the data. First, we have a census of 506 villages in rural Mexico. This is a panel of more than 20,000 households interviewed 8 times between 1997 and 2003, providing detailed information on consumption, income, health shocks, and informal loans and gifts, as well as other socio-demographic characteristics. Second, we 3

5 can identify the extended family in each village by matching household heads and spouses last names, exploiting the Hispanic convention of having two surnames, the first from the paternal lineage and the second from the maternal one. Third, between 1998 and 1999 Progresa is offered only to a random group of 320 villages, in which poor households are eligible for sizeable cash transfers, while no public transfers takes place in the control villages. Our empirical analysis begins by describing the characteristics of households with and without an extended family network in the village, which we call connected and isolated. At least 80% of the households in our villages are part of an extended family network. We show that connected and isolated households have similar observable characteristics in 1997, our first (pre-program) wave. The few differences seem related to life cycle and cohort effects, rather than to different needs and preferences for insurance. While similar at baseline, income, consumption, and investment change differently for connected and isolated households over time, increasing significantly more for the connected. This preliminary evidence is consistent with our hypothesis that the extended family, by providing insurance, enables its members to undertake more investment, increasing their income and consumption. Our next step is establishing whether households share risk only within the extended family or also with other villagers. Our experimental data enable us to test whether, when a given household has an exogenous income increase (the program transfer), consumption, transfers, and informal loans of its relatives and non-relatives increase. Connected households share risk within their family network but not with isolated households. When eligible households receive the Progresa grants, their ineligible relatives consume more and receive more informal transfers. At the same time, consumption and transfers to ineligibles who have no eligible relative (or no relative at all) in the village do not. We also show that insurance is related the connectedness of the households, rather than to other characteristics of the households. That is, insurance does not seem related to ethnicity or land ownership, and these variables are not associated with higher degrees of insurance over and above the one associated with connectedness. The relevant institution, therefore, is the extended family, rather than one s ethnicity or occupational group. Consistent with the model predictions, consumption for the ineligibles is a positive function of the size of the income shock per network members, i.e. the more money is injected in the network, the more its members will benefit, irrespective of which household receives the extra 4

6 income. On average, for each extra dollar received by an eligible individual, her non-durable consumption increases by 54 cents and the consumption of her ineligible relatives increases by 13 cents. The consumption of eligibles who have no ineligible relatives increases by 69 cents for each dollar received. There is a high degree of insurance among the extended family. Connected households achieve full risk sharing in 80% of the networks (at the 90% confidence level). As the isolated do not share risk within the village, they may be more vulnerable to idiosyncratic shocks. The longitudinal coefficients of variation of consumption is significantly smaller for connected than isolated households. However, this in itself is not necessarily evidence that the isolated are less insured than the connected, or that they engage in costlier activities to smooth consumption. Therefore, we compare the effect of idiosyncratic shocks on child labor supply, school enrollment, and livestock. When the household head falls ill, isolated households reduce school enrollment and increase child labor more than connected households. Therefore, consumption smoothing is costlier for the isolated, who, by reducing investment, are foregoing future wealth in order to stabilize current consumption. Conversely, when offered a government transfer, only connected households increase secondary school enrollment and attendance. Further, the connected are responsive to both negative and positive networks shock in terms of Progresa grant and illness of the network members while they are significantly less responsive to own negative shocks. The paper has the following structure: in Section 2 we sketch a simple risk-sharing model (with investment) to define the notation, show the standard full insurance results, and prove that full insurance cannot be achieved with positive transaction costs, which are probably higher between isolated than connected households. Sections 3 and 4 describe the data, the creation of family networks and their characteristics. The following three Sections show that risk-sharing occurs only within the extended family (Section 5), that consumption is smoother for the connected (Section 5.5), and that most families are fully insured against idiosyncratic risk (Section 5.4). Section 6 documents that isolated households engage in costlier activities to smooth consumption. Section 7 concludes. 5

7 2 A model of partial insurance and investment The purpose of this section is to sketch a simple model of risk sharing and investment we can then take to the data. We are interested in understanding under what conditions members of an extended family achieve a higher level of insurance than non-members, and in which case the presence of the extended family favors investment. 2.1 Partial insurance We consider a simple exchange economy in which households smooth consumption by making transfers to each other. Transferring resources is costly and may reduce the amount of informal insurance (as in Schullofer-Wohl, 2008). This iceberg-type cost, modeled so that part of the transferred resources is lost in the transfer, captures all the potential reasons why full insurance may not be achieved - e.g. imperfect information or enforcement. We conjecture that members of an extended family face lower, possibly zero, costs of transacting among each others than with households outside their family ties. This is because family members are part of a small network (e.g. Genicot and Ray 2003) with repeated interactions and potentially high trust (Mobius and Szeidl 2007) and altruism (Altonji et al., 1992; Foster and Rosenzweig, 2001), features associated with stable networks achieving high levels of insurance. That is, the extended family may provide low-cost information about its members income and effort and reduce information and commitment problems, achieving a higher degree of insurance than unrelated households. Some of the latter households may have no informal insurance, if the transaction cost is sufficiently high. In our economy, risk-averse agents smooth consumption by making transfers to each other. Households receive an exogenous endowment, yh t (s t), in each of the finite states of the world, s t, and then choose how much to transfer to each other to maximize their discounted expected utility over a perishable composite consumption good, c. central planner as: The problem can be solved by a T max ω h E 0 β t 1 U ( c t c t h (st ) ) (1) h (st ) h t=1 s.t. yh t (st ) c t h (st ) + α h d ( yh t (st ), c t h (st ) ). (2) h h h 6

8 Where α h represents the cost to be paid for insuring household h, ω h is the Pareto weight of household h, and λ t is the lagrange multiplier. The transfer d(.,.) is a convex function with negative first cross-derivatives. The first-order condition of this problem is the following (dropping s t to simplify the notation). ( t 1 U(ct h ω h β ) c t = λ t 1 + α d ( yh t, )) ct h h h c t. (3) h If the transaction cost is zero α h = 0 we are in the standard full-insurance world, where household s consumption depends only from aggregate resources (and Pareto weights) while idiosyncratic fluctuations are smoothed out. However in general we will have partial insurance or even autarky when the cost of sharing is too high. To show that in this formulation household consumption depends on household s endowment we differentiate the first order condition above to show: c t h y t h = ω h β t 1 2 u(.) 2 c λ t α h 2 d(.,.) c y λ t α h 2 d(.,.) 2 c > 0. (4) In partial insurance, the household cannot completely smooth away fluctuations in endowment and the economy suffer a net welfare loss. Using the same argument as the above we can show that, in general, and for all j h: c t h yj t > 0, d t h < 0, a h d t h yj t > 0; (5) d t h < 0. (6) a h That is, when two households, h and j share risk, consumption of household h, and therefore the transfer this household receives, depend positively on household j s resources, as well as on h s resources. The amount of transfer that h receives in each state of the world, hence the degree of insurance, is smaller the higher the transaction costs, i.e. the more costly it is to 7

9 insure each other. In autarky, i.e. when α h is large : c t h = yt h ; (7) c t h yj t = 0; (8) d t h y t j = 0. (9) These above are some of the predictions we will take to the data. In partial insurance consumption a positive function of own endowment and aggregate resources, so that even households with the same pareto weights may consume different amounts depending on their endowment. This can be easily seen by taking the f.o.c above for two different households, their ratio of marginal utility is no longer equal to the inverse of their pareto weights, rather it involves the insurance costs. The household with a high realization makes a smaller transfer than in full insurance, because of the transaction cost. The household with the high endowment consumes more than the other and the longitudinal variance of consumption is greater than with full insurance. Suppose transaction costs are lower among related than unrelated households and unrelated households s costs are sufficiently high. In this case the households would engage in informal risk-sharing only with their relatives. We can test this hypothesis by looking at the implications for the level of consumption and transfers and the variance of consumption. The first set of tests requires exogenous income increases. As pointed out an increase in the endowment to a given household j would have a positive effect on all households with whom it shares resources, the precise impact would depend mainly on the relative pareto weights and transaction costs. A second set of tests compares consumption smoothness over time for connected and isolated households. Define the average longitudinal consumption variances for K and I households as V ar(c K ) and V ar(c I ). If our hypothesis is correct, V ar(c K ) < V ar(c I ). A variant of this test considers the relationship between the growth of household consumption and idiosyncratic income. Under the hypothesis that α = 0, taking the log of optimal consumption and then its first difference yields a well-known relationship between the growth 8

10 of individual and aggregate consumption for household h at time t: lnc t h = lnȳt We can then regress the growth in household consumption over the growth in aggregate consumption and household income, lnc t h = β 1 lny t + β 2 lny t h + ut h (10) where lnyh t is the endowment growth for household h at time t and ut h = ɛt h is an error term derived from assuming a multiplicative error in the measurement of consumption, c t h = ct h eɛt h. Under the null hypothesis that α = 0 consumption growth does not depend on the growth of the idiosyncratic component of the endowment; therefore β 2 = 0. We use equation (10) for two purposes. First, to show whether villages as a whole achieve full insurance. Second, to estimate for which share of the networks we cannot reject the null of α = 0 (β 2 = 0), i.e. a given extended family network is fully insured against idiosyncratic risk. 2.2 Resource sharing and investment The presence of an extended family network likely affects investment, through different channels. First, insurance and investment are related. That is, the existence of insurance stabilizes consumption, enabling households to invest more when hit by negative shock (but less when hit by a positive shock because they have to share it within their network). Second, connected households may invest more than if they had no family networks because their relatives may lend them money or become share-holders in the investment. Third, the availability of insurance may enable agents to undertake riskier, more profitable investment [Arrow 1971; Obstfeld 1994]. Understanding the link between income pooling and investment is important to evaluate the costs of being isolated and excluded from informal sharing networks. The lack of sharing mechanisms may limit investment for poor households even when its returns are certain. The goal of this section is to investigate in which cases aggregate investment is higher when agents can pool income and to study the partial effects of endowment changes on optimal investment in these cases. We study the relationship between transaction costs, resource sharing, and investment by allowing households to invest, as well as to consume their endowment. 9

11 Consider an investment, I, with gross returns, which are certain, given by the function f(i), with f (.) > 0, f < 0. 1 The central planner maximizes the following objective function: 2 2 max U β t 1 ln c t h s.t. h=1 t=1 c 1 1 = y1 1 + d 1 I 1 c 1 2 = y2 1 (1 + α)d 1 I 2 c 2 1 = y1 2 (1 + α)d 2 + f(i 1 ) c 2 2 = y2 2 + d 2 + f(i 2 ) d 1,2 0, I 1,2 (0, 1). The notation here is as before with the difference that we are limiting the model to a very simple 2 periods, 2 identical households model which allow us to discuss the main points. In particular, we model household 1 as having a low draw and receiving a transfer from household 2. The planner chooses the optimal transfers and investments profile given the gross returns, transaction costs and the initial endowments. With no lumpiness in the investment and decreasing marginal returns, that is for example f(i h ) = I γ h with γ (0, 1), there is a positive correlation between the amount invested and the degree of insurance in the aggregate. However depending on the main parameters of the model as discount rate, returns, level and correlation of the endowments there might be situations where the investment level is actually lower in full insurance than it is in autarky. The model has an analytical solution for the cases of full risk-sharing and autarky under further assumption regarding γ. Special cases have closed form solutions however in general it is not possible to solve the model analytically and we resort to numerical simulations. Simulation exercises show that the level of investment is higher when the endowments are not too different and aggregate resources are fixed. With lumpy investment, aggregate investment may also be higher the higher the level of in- 1 We assume the return is certain and observable for simplicity. In unreported results we considered the cases in which the investment is fully or partially hidden. The scope for investment is higher for insured agents when their investment is hidden and our point that insurance may favor aggregate investment does not change. 10

12 surance. Lumpiness is common in cases of physical and human capital investment, the common types of investment in developing countries. The two investment models, with and without non-convexities, we present here are quite different. The first one is an income pooling model in which related households share their endowments in order to maximize the returns to the joint investments, as well as smooth idiosyncratic risk. The model mechanism behind the working of the model with lumpy investment departs from the classical insurance models. Here one of the two related households is acting as a shareholder, providing part of the capital to undertake the investment and then receiving part of the profit. In this case, both households are sacrificing a large part of current consumption in favor of future consumption. Note the following implication: both agents would like to borrow against the future returns to the investment, if they could. Therefore, there is scope for government intervention to complete the financial markets even in this case, in which, in principle, households can entirely smooth idiosyncratic risk. An alternative for the investing household would be to form a link with a household with a similar income process (in our case one with a high initial income draw, followed by a lower one). In this sense, there is a trade off between the insurance and the investment motive to form resource-sharing networks. We believe the former prevails for poor households in developing countries, for which the cost of a negative shock is potentially very high An example Suppose we take the following extreme example: a 2-by-2 model solved in full insurance and autarky. Let s assume γ =.5, then we can analytically solve the previous model, and in full sharing we know that the two investment levels chosen by the planner are going to be identical. The optimal allocation is: 11

13 I F I β 2 (Y 1 ) 2 = ( ( Y 2 + 2β (2 β) Y 1 + (Y 2 2) 2) ).5 2 ; (11) c 1F I = Y 1 2 β 2 (Y 1 ) 2 ( ( Y 2 + 2β (2 β) Y 1 + (Y 2 ) 2) ).5 2 ; (12) c 2F I = Y βy 1 Y 2 + 2β (2 β) Y 1 + (Y 2 ) 2. (13) where Y 1, Y 2 are the aggregate endowments in period 1, 2. comparative statics: We can also look at some I F I y 1 h c 1F I y t h > 0, > 0; I F I y 2 h c 2F I y t h < 0; (14) (15) > 0. (16) In autarky we have the following optimal allocation: 12

14 I AUT 1 = I AUT 2 = c 1AUT 1 = y 1 1 c 1AUT 2 = y 1 2 c 2AUT 1 = y 2 1 c 2AUT 2 = y 2 2 β 2 (y1 1 (y )2 ( 21 + β (2 + β) y1 1 + ( y1 2 β 2 (y1 2 (y )2 ( 22 + β (2 + β) y2 1 + ( y2 2 β 2 (y1 1 (y )2 ( 21 + β (2 + β) y1 1 + ( y1 2 β 2 (y1 2 (y )2 ( 22 + β (2 + β) y2 1 + ( y2 2 βy 1 1 ) 2 ).5 ) 2 ; (17) ) 2 ).5 ) 2 ; (18) ) 2 ).5 ) 2 ; (19) ) 2 ).5 ) 2 ; (20) y1 (β 2 + (2 + β) y1 1 + ( ) y1 2 2 ); (21) βy 2 1 y2 (β 2 + (2 + β) y2 1 + ( ) y2 2 2 ); (22) (23) We can also look at some comparative statics: Ih 1AUT yh 1 Ih 1AUT yh 2 c taut h yh t > 0, < 0, > 0, Ih AUT yj 1 Ih AUT yj 2 c taut h yj t = 0; (24) = 0; (25) = 0; (26) 2 I F I ( I A 1 UT + I A 2 UT ) or < 0. (27) An important point to notice is that in this model it is possible to have a lower aggregate level of investment under full insurance than in autarky. However, that occurrence is rare in simulation exercises and disappears at higher aggregate level of endowment. It is further 13

15 possible to show that: I F I h y t h < IAUT h yh t. (28) That is the response to own shocks is lower (on average) in fully insured households than in autarkic ones. The reason being that shocks are shared in full insurance while they are suffered in full in autarky Testable hypotheses We summarize our testable hypotheses, as follows. First, if two households share resources, consumption is a positive function of the current and future endowments of all households (within the sharing network) and investment is a positive function of the current endowments (and a negative function of the future endowments) of both households. Conversely, for autarkic households consumption and investment depend only on own endowments. [HP 1a] : [HP 1b] : c tf h I yh t > 0, c taut h yh t > 0, Ih F I yh 1 > 0, c taut h yj t = 0 Ih F I yh 2 Ih AUT yh 1 < 0; h (29) > 0, I F I h y 2 h < 0. (30) (31) Second, the idiosyncratic variance of consumption is smaller for resource-sharing households than for autarkic households, if both sets of households have the same endowment process. Third, investment is higher for resource-sharing than for autarkic households when the investment has decreasing marginal returns, or when there are some constraints in the minimal investment that bind for the autarkic households but not for the resource-sharing ones. This implies future aggregate consumption will be higher for resource-sharing than for autarkic households. Fourth, investment is less sensitive to changes in own endowment for resource- 2 We choose to work with the simplest possible model, however Progresa is not purely an exogenous shock to income, rather it comprises to parts, an unconditional transfer and a subsidy to schooling. We have also consider such a model and indeed the main points made go through, but one important difference: given the large subsidy to part of the network it is perfectly plausible that the extended family is going to send subsidized kids to school while moving more unsubsidized kids to the labor market. We discuss this point further in the empirical section.v 14

16 sharing households. However, if there are non-convexities or fixed costs in investment it is possible that the marginal effect of own endowment be positive for the connected and zero for the isolated. Estimating these marginal effects is a way to empirically distinguish between the two investment models. [HP 2] : V ar(c F I) < V ar(c AUT ); (32) [HP 3] : E[I F I ] > E[I AUT ]; (33) [HP 4] : I F I h y t h < IAUT h yh t. (34) Our tests have the following structure. For hypotheses one and three, we regress household consumption and received transfers and loans on some measure of their relatives income. For hypothesis four we look at how household investment changes when their income changes. Since income is typically endogenous, we use some exogenous income component. For hypothesis one, the eligibility to the program cash transfers, which is allocated to villages on a random basis. For hypothesis three we also use the illness of the household head We will explain how to test these hypotheses after describing the data. 3 Progresa and its evaluation Progresa is an anti-poverty program that targets poor households in rural Mexico. The average transfer is 200 pesos, equivalent to 22% of eligible households monthly income (De Jainvry and Sadoulet, 2006) and to 25% of pre-program household s food consumption, which averages 160 pesos for adult (equivalent) in recipients households (Angelucci and De Giorgi, 2009). About 75% of households are classified as eligible based on their poverty status as computed in October 1997, although only a smaller share are treated initially. 3 The transfer is conditional on eligible members attending nutrition and health classes and having regular health checks. It has a fixed component of 100 pesos per month and a variable component conditional on children attending classes between 3rd and 9th grade. These schol- 3 The initial allocations of households between eligible and ineligible that placed 50% of the households in the eligible group was subsequently revised just before the roll out of the program; however, in the first year of the program most of the re-classified households, typically elderly, did not receive any transfer for administrative reasons. In our analysis we keep those observations as treated. 15

17 arships vary between 140 and 510 pesos per child, increase with school grade and are larger for girls than for boys for grades 7 to 9. The transfer is capped at 625 pesos per month (all these values are at November 1998 prices). While nominally conditional, a substantial component of the grant is in defacto unconditional. This is because pre-program enrollment rates up to 6th grade (corresponding to primary school) are higher than 90% and the health checks are infrequent for most eligible persons (e.g. they are annual for adults) and secondary school rate (grades 6-9) are about 65%. This part of the transfer has a pure income effect for its recipients. Conversely, the conditionality for the part of the grant linked to secondary school attendance is actually binding for households whose eligible secondary school-age children would have worked in the absence of the program. This is a non-negligible group, as pre-program attendance to 7th to 9th grade (corresponding to secondary school) is about 65%. These households may incur a net financial loss from sending their children to school despite receiving the program transfer, because the secondary school transfer amounts to only about two thirds of full-time child wage (Schultz, 2004). For these families, the program transfer is a partial subsidy. About one third of the eligible households have no child in the subsidized year/grades in November 1998, however they still receive the fixed grant component. Out of the remaining two thirds that are entitled to a larger transfer, most families have some primary school eligible children - about 87% compared to 57% of families with at least one secondary school child. On average, 30% of the total potential grant is associated with scholarships for grades 3 to 6. The partial grant conditionality and its relatively low monetary value, compared to the opportunity cost of secondary school attendance, suggests that the program may not provide some eligible families with enough resources to increase secondary school attendance, unless the unconditional component is large. This distinction is going to be important to assess how the program eligibility differentially affects secondary school attendance for connected and isolated households. Our data have three important characteristics. First, they provide a complete census of all households from 506 rural villages in 7 different states. 4 The data are longitudinal, with an initial wave collected in October 1997 and then approximately every 6 months until November 2000, and lastly in November We have income data in all cross sections, and consumption 4 Guerrero, Hidalgo, Michoacan, Puebla, Queretaro, San-Luis Potosi and Veracruz; mostly in central Mexico. 16

18 data in the 5 cross sections between November 1998 and 2003 (6 if one is willing to use preprogram expenditure data, collected in March 1998). We have complete data for about 22,500 households per wave up to November 1998, about 20,000 up to November 2000, and 19,000 in We provide further details of the the key variables and their availability in the various data waves in the Appendix. Second, between May 1998 and November 1999 the program is offered only in a random group of 320 villages. The remaining 186 villages receive their first transfers only at the end of Therefore we have three data waves, November 1998, May 1999, and November 1999, in which only a random subset of our sample of villages was treated. 5 We have information on current and potential eligibility, which depends on a time-invariant wealth measure, for all households in our 506 villages. Third, we observe people s last names, which we use to match related households, as we explain in the next section. These three unique features enable us to group households according to their eligibility status, village of residence, and connectedness, as shown in Figure 1. This figure shows our data provide a partial-population experiment between November 1998 and 1999 (Moffitt, 2001), as the treatment is offered only to a subset of the villagers. Thus, for that time period we can measure the effect of Progresa on consumption and transfers to connected and isolated ineligible households. We also have self-reported data on illness of the household head during the previous month. We use this variable as our measure of negative health shock. While the head s health status may not be random, it is uncorrelated with the household wealth index, once we control for predetermined or exogenous variables such as regional dummies, age, gender, ethnicity, and village weather shocks. 6 Moreover, the incidence of household head illness is not statistically different for connected and isolated households, as we show below. 5 In fact in November 1999 some households in control control villages started to receive transfers. Thus for November 1999 we can only estimate lower bounds (in absolute value) to the true treatment effects. 6 Results available upon request. 17

19 4 Family networks in rural Mexico 4.1 Network creation To construct extended family links between households in the same village we exploit information on surnames provided in the third wave of data and the fact that Mexicans use two surnames the first is inherited from the father s paternal lineage and the second from the mother s paternal lineage. For example, former Mexican president Vicente Fox Quesada would be identified by his given name (Vicente), his father s paternal name (Fox) and his mother s paternal name (Quesada). Hence a household that is headed by a husband and wife, has four associated surnames the paternal and maternal surname of the head, and the paternal and maternal surname of his wife. We show the assumed links between different households in Figure B.4. We use these data to create within-village family networks, starting from distance one (each household head and spouse s parents, offspring, and siblings) and proceeding to more distant relatives (grand-parents, aunts and uncles, etc.). All family links are defined across households on the basis of two surname matches. We believe our matching algorithm to be reliable based on the following evidence. First, consider the possibility that names were measured with error. If this occurred, we might find spouses with identical surnames in all four positions, however this happens only in 1.6% of the household. Further, 97% of the households are linked to no more than two parental households. More generally, comparing our data with the network data from the Mexican Family Life Survey (MxFLS), collected in 2001 shows that the links we identify in our data are indeed no larger than those measured in the MxFLS, which uses relatives from any location and not only from within the village. This applies also when we compare different types of link (e.g. the number of sibling links we identify is not larger than the reported number of siblings in the MxFLS). Two further characteristics of our surnames support the view our matching algorithm is consistent. First, there are fewer paternal surnames reported by male household heads than for the other types of surname, including those reported as the wives paternal surname. This is exactly what one would expect, because the patronymic naming convention implies wives paternal surnames have lower survival rates across generations than those of male heads of household. Second, 85% of spouses that have their parents present in the village report remain- 18

20 ing in the same village at the time of marriage. The corresponding figure for spouses that have no parental links in the village is only 59%. This can be explained by female marriage migration - i.e. moving to the husband s village upon marriage, which corresponds with anecdotal evidence (as well as empirical evidence from other poor countries, e.g. (Rosenzweig and Stark, 1989). To conclude this discussion, we consider the implications of having wrong matches (either erroneously matching two unrelated households or failing to match two relate ones). First, as we explained above, the available evidence suggests this is not a serious issue in our data, once we omit single-headed households. Indeed, this group of families is the one for which mismatches are most likely. This happens because we do not observe the missing spouse s last names, so we cannot match the single-headed household to all its links. Therefore, some single-headed households may be erroneously classified as having no relatives in the village. Call households with relatives in the village and without connected and isolated. Suppose only related households share resources and we want to test this. That is, as household i s endowment changes, this will affect consumption and investment of its relatives only. Including the single-headed adds some connected households to the isolated pool. So, if, for example, we want to estimate the effect of a government grant to household i on the consumption and investment of its related and unrelated households, including the single-headed would result in over-estimating the effect of the grant for the unrelated. That is, the difference between the grant effect on the connected and the isolated is a lower bound of the true difference. In a way, then, this type of bias actually strengthens our results (if we find a significant difference anyways). In general, if our networks data were contaminated by large measurement error we should not find any evidence of the proposed mechanism and in particular no significant differences in consumption smoothing and investment behavior between isolated and connected households. We deal with single-headed households in the following way. When our tests compare connected and isolated households, we drop the single-headed. However, when we compare outcomes within connected households we include both the couple-headed and the single-headed. We discuss the rationale for this choice in the Appendix. Angelucci, De Giorgi, Rangel, and Rasul (2006, 2007) discuss the creation of these family networks, the characteristics of the matching algorithm, and the possibility and extent of 19

21 measurement error in greater details. There are 7.8 households within the same family network. The average distance between any two households in the same network is The degree of the average network, defined as the number of households each is directly connected to, is The diameter of the network, defined as the longest distance between any two households in the network, is around 2.5 on average. Hence family networks are unlikely to span across more than three generations. About 50% of all networks and two thirds of networks of more than two households are a mix of eligible and ineligible households. 4.2 Correlates of Extended Family Table 1 shows the means of key demographic and socioeconomic variables from the 1997 data. Once we drop single-headed households and compare both demographic and economic outcomes, which we do in the first 5 columns, connected and isolated households appear remarkably similar. 7 The main differences between the two groups is that isolated households heads are 2.5 years older, have fewer young children, and a 9% higher share of illiteracy than connected heads. Most of the other variables do not differ between the connected and isolated. In particular, income, food and non-food expenditure, and wealth index are almost identical, as well as the total amount of land held, adult and child labor, and most assets and livestock. The few exceptions are that the isolated own 12.5% more poultry and are 6% and 9% less likely to own stoves and TV sets. The correlation between the age of the head and spouse of a household and the likelihood that their parents, adult children and siblings live in the same village in a separate household is somewhat mechanical. Angelucci et al. (2007) show that the rate at which a household head loses links as he ages (i.e. parents passing away) is faster than the rate at which he acquires new links (through children being born). This makes households with older heads more likely to be isolated. Thus, these difference seem overall attributable to life cycle and cohort effects: isolated households are older, therefore have fewer infants and a lower education level. Because they 7 Single-headed households are poorer, less educated, and larger than couple-headed households. We discuss their inclusion and omission in the Appendix. 20

22 are more traditional, they might own fewer TV sets and stoves. We also performed a Pearson s chi squared test of the difference in the distribution of unemployment and occupation for all members of connected and isolated households at least 8 years old. Different employment patterns may reflect different risk preferences or need for insurance. 8 We find no significant difference in the occupational distributions for isolated and connected households. Since a number of the tests we run exploit the program randomization, the last three columns of Table 1 report the p-values of the differences in observables between households in treatment and control villages. We do that separately for the connected and the isolated. Most of the variables do not have statistically different means in control and treatment villages, not surprising, given the randomization. A few notable exceptions for the isolated are the differences in the number of oxen and goats owned and the share of year old children enrolled - all of which are lower in control villages. For the connected, the number of oxen owned is slightly lower in control villages. The share of households potentially eligible for Progresa is also statistically different in control villages, unless we include the single-headed, in which case this difference is no longer significant at conventional significance levels. We address this issue by controlling for these initial differences when estimating treatment effects on these outcomes. Moreover, we control for the set of variables that are unbalanced between treatment and control villages at baseline. Comparing means for connected and isolated households is somewhat misleading if there is assortative matching. For example, all the wealthiest and poorest households may belong to distinct family network, but by aggregating all families we would not detect it. We test for assortative matching in wealth by computing the standard deviation of the wealth index used to determine program eligibility for each family network and each village. If there is positive (negative) assortative matching then the ratio of network over village standard deviation, sdn sd v, would be less (more) than one. The computed ratios, available upon request, are centered around one, rejecting the hypothesis of assortative matching. This confirms the previous conclusion, namely that one cannot easily predict whether a household is connected or isolated by looking at its observable characteristics. 8 The main labor force categories are unemployed (68%), daily worker (15%), non-agricultural manual worker (4.5%), self-employed without (4.3%) and with personnel (0.12%), working for a family business without pay (4.4%), and ejidatario (2.5%). 21

23 5 Do family networks share resources? 5.1 Tests and identification This set of tests exploits the random variation in income resulting from the experimental design of Progresa. For the first 18 months of its existence Progresa was implemented only in a subset of the sampled villages. Therefore, we think of the program transfers as an exogenous income shock to some households (household 2 in our model). If we abstract from its effect on secondary school enrollment, which increased by about 10 percentage points among eligible households without affecting the enrollment rates in ineligible households (Angelucci et al., 2007), the experimental variation we observe in the data fits our model. In particular, as we mentioned the transfer conditionality does not bind for a large group of households. As such, we can observe how an exogenous income increase for eligible households changes consumption and transfer or informal loans for different groups of ineligible households depending on their relationship to the eligibles. Although it has some unique features that distinguish it from the previous poverty alleviation programs, Progresa is just the last of many aid programs targeting the poor in rural Mexico. The recipients are used to receiving government assistance of different forms; the possibility of such policies should then be considered part of the information set of those agents. 9 Thus, we think of Progresa as one of the states of the world on which households write implicit contracts, rather than an entirely new occurrence. 10 Since some of our tests estimate treatment effects on ineligible and eligible households, we briefly define these parameters and discuss their identification. Define Y 1i and Y 0i as the potential outcomes for household i in treatment villages (P i = 1) in the presence and in the absence of the treatment, where the treatment is the existence of Progresa transfers to eligible households (N i = 0) in treatment villages (P i = 1). We call the average effects of the program on 1) eligible households and 2) ineligible households (N i = 1) living in treatment villages, 9 For example, at the time Progresa is implemented, qualifying households receive basic consumer goods at subsidized prices (DICONSA), free tortillas (TORTIBONO), free breakfast for children (DIF), food packages (PASAF), free school supplies (CONAFE), lodging and education grants for indigenous students (INI), other school grants for all poor children (Ninos de Solidaridad), financing of productive projects (FONAES), temporary employment (PET), training scholarships for the unemployed (PROBECAT), and cash transfers to farmers producing specific crops (PROCAMPO) (Skoufias, 2005). 10 Angelucci and De Giorgi (2009) find no difference in the mean and distribution of the longitudinal variation of consumption in treatment and control villages, which would be the case if Progresa changed the amount of risk-sharing. 22

24 Average Treatment Effect on the Eligibles (ATE) and Indirect Treatment Effect (ITE), and we define them as follows: AT E(Y ) f = E(Y 1i P i = 1, N i = 0, f = 1) E(Y 0i P i = 1, N i = 0, f = 1) IT E(Y ) f = E(Y 1i P i = 1, N i = 1, f = 1) E(Y 0i P i = 1, N i = 1, f = 1) f = {K, I} Under the assumption of random assignment and in the absence of program spillover effects to control villages, the expected value of the potential outcome in the absence of the treatment, Y 0, is the same in both treatment and control villages, i.e. E(Y 0i P i = 1, N i = j, f = 1) = E(Y 0i P i = 0, N i = j, f = 1), for j = {0, 1} and f = {K, I}. Therefore, the differences AT E(Y ) f = E(Y i P i = 1, N i = 0, f = 1) E(Y i P i = 0, N i = 0, f = 1) (35) IT E(Y ) f = E(Y i P i = 1, N i = 1, f = 1) E(Y i P i = 0, N i = 1, f = 1) (36) identify the ATE and the ITE. The eligibles and the ineligibles in treatment villages are the agent 2 and agent 1 households from the model, and the y 2 is the Progresa transfer. Therefore, we can formulate our first set of testable hypotheses in the following way. 1.1 If the connected households (K) share risk only between family members and not with the isolated (I), only connected ineligible households will increase consumption because of Progresa, i.e. IT E(c) K > 0, but IT E(c) I = Among connected ineligible households, the IT E on consumption is positive only if these households are connected to some eligible household. 1.3 For a given Progresa transfer and similar Pareto weights, the treatment effect on consumption for connected eligible households is lower if these households are related to the ineligible with whom they share part of their transfer, i.e. AT E(c) K < AT E(c) I. 1.4 The increase in consumption for the ineligibles is a function of available resources at the network level. We consider monthly food consumption per adult equivalent. We compute this variable from seven-day recall data on 37 different food items. For each item, we know the quantity 23