RESOURCE POOLING WITHIN FAMILY NETWORKS: INSURANCE AND INVESTMENT

RESOURCE POOLING WITHIN FAMILY NETWORKS: INSURANCE AND INVESTMENT Manuela Angelucci 1 Giacomo De Giorgi 2 Imran Rasul 3 1 University of Michigan 2 Stanford University 3 University College London June 20, 2012

Spillover effects, risk sharing, investment, and the extended family in rural MX. A few related papers. 1 Extended family networks in rural Mexico: a descriptive analysis, July 2006, with G. De Giorgi, M. Rangel, and I. Rasul, forthcoming in CESifo Conference Volume on Institutions and Development, edited by Timothy Besley and Raji Jayaraman, Massachusetts: MIT Press. 2 Indirect effects of an aid program: how do cash injections affect ineligibles consumption, with G. De Giorgi, American Economic Review, 99(1), 486-508, March 2009. 3 Village Economies and the Structure of Extended Family Networks, with G. De Giorgi, M. Rangel, and I. Rasul, B.E. Journal of Economic Analysis and Policy, (9)1, Article 44, 2009. 4 Extended family networks and schooling outcomes: evidence from a social experiment, with G. De Giorgi, M. Rangel, and I. Rasul, Journal of Public Economics, 94(3-4), 197-221, April 2010. 5 Program evaluation and spillover effects, with V. Di Maro, Impact-Evaluation Guidelines Technical Notes No. IDB-TN-136, May 2010. 6 Resource pooling within family networks: insurance and investment, with G. De Giorgi and I. Rasul.

Adverse effects of lack of insurance/credit market for the poor Poor households in LDCs: high risk + no formal insurance/credit = smoothing consumption is difficult/costly; profitable investment not undertaken; long-term costs of under-investment (e.g. e.g. Nurkse (1953); Lewis (1954), McKinnon (1973); Fafchamps and Pender (1997)).

Resource pooling within one s family network provides insurance and investment Informal institutions: May provide insurance against idiosyncratic risk [smoother consumption] and relax capital constraints to high-return but lumpy investment [higher consumption]. Effect on investment if aggregate resources are high enough. In sum, to assess value of informal institution, look at both insurance and investment (or smoothness and level of consumption). Test these hypotheses in the context of the village-based extended family in rural MX.

Why the village-based extended family We consider the extended family because: High, but not full insurance in villages (e.g. Townsend 1994 and many others). Theory predicts full insurance more likely within small groups (e.g. Ambrus, Mobius, and Szeidl, 2010). Easy to get information/monitor; altruism/reciprocity/sanctions (e.g. La Ferrara 2003). Census of 506 MX poor rural villages with surname data.

Insurance and investment within the extended family Insurance: degree of insurance depends on transaction costs. No insurance if cost sufficiently high. With zero costs, full insurance. Hypotheses: (1) insurance only within the extended family; (2) full insurance within the extended family.

Insurance and investment within the extended family Investment: extended family members pool resources to finance lumpy investment, if aggregate endowment high enough. Given endowments, two households may invest differentially in high-return, lumpy good and low-return, divisible good. Hypotheses: (3) effect of income shock on investment differs for households with and w/o extended family; (4) resource pooling in extended family has long-term effect on investment/consumpiton, given the right economic circumstances.

Panel A: Returns on Investments Figure 1: Investment Returns on Investments (I p, I S ) (1+r S )I S (1+r)I p S I min I Panel B: Aggregate Investment in Schooling Aggregate Investment in I S (e.g. Schooling) Sharing Households (K) Zero Investment for K and O Autarkic Households (O) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 t K t O y j Panel C: Aggregate Investment in Poultry Aggregate Investment in I p (e.g. Poultry) Same (Average) Investment for K and O Autarkic Households (O) Sharing Households (K) 0 0. 1 0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 1 t K t O y j

Identifying the Extended Family Using Surnames Figure 1: Family Tree Parents (F1, f1) (F2, f2) Son (F1, F2) Spouse (F3, f3) Son (F1, F2) Spouse (F4, f4) Daughter (F1, F2) Husband (F5, f5) Daughter (F1, F2) Husband (F6, f6) Son (F1, F3) Spouse (F7, f7) Daughter (F1, F3) Husband (F8, f8) Son (F1, F4) Spouse (F9, f9) Son (F5, F1) Spouse (F10, f10) Daughter (F5, F1) Husband (F11, f11) Daughter (F6, F1) Husband (F12, f12) Notes: We use the convention that the head's surnames are written in standard (black) font, and those of his wife are written in (red) italics. Paternal surnames are indicated in upper case (F1, F2 ) and maternal surnames are indicated in lower case (f1, f2 ). First names are not shown as they are not relevant for the construction of extended family ties. Each household in the family tree is assumed to be couple headed purely to ease the exposition.

Why we think our data are reliable Never exploit single-name matches across households; 97% of households have at most two parental households in the village; external validity: consistent with evidence from MxFLS; females have fewer links than males [Marriage migration, Rosenzweig and Stark 1989]; Bjorkman, La Ferrara, Schuster (2012) validate our algorithm; Details on creation, measurement error, checks, etc. in paper (1).

Connected (K) and isolated (O) households: definition K = connected households; extended family members; households with at least one first-degree relative in the village (parent/offspring, siblings). O = isolated households; households without first-degree relatives in the village. K and O equally likely to be eligible for PROGRESA. Most networks are a mix of eligibles and ineligibles for PROGRESA.

K and O have similar baseline pscore [P(O = 1 X, 1997)] -.15 -.1 -.05 0.05.1.15 0.2.4.6 all households - couple-headed

Baseline differences b/w K and O consistent with our theory K and O have similar average characteristics in 1997. Same income, non-durable consumption, wealth score, land, employment. However, the isolated have: 1 more low-return, low-cost investment (poultry: 8 vs. 7); 2 lower high-return, high-cost investment (temporary migration: 3.8% vs. 5.8%); 3 fewer durables (TVs: 46% vs. 49%; stoves: 29% vs. 32%). More details in paper (3).

Data: poor households in rural MX; exogenous income shock (CCT) PROGRESA: cash transfers to the poor. Transfer is big: 20 USD per household per month (22% of income); pre-program monthly consumption is 16 USD per adult (20 USD for ineligibles; 10 pesos 1USD). 75% of households are eligible. Program only in randomized-in villages between April 1998-Nov. 1999. 1997-2003 data for other tests. 20,000 households; observe consumption, investment. Up to 8 waves. Detailed food consumption data.

Transfer conditionality by school grade Cash transfer is de facto unconditional for primary school children (counterfactual enrollment > 90%). Cash transfer is more likely to be conditional for secondary school children (counterfactual enrollment 66%). Transfer covers only 2/3 of full cost of secondary education.

Data structure and H 1 : resource-pooling only for K. 506 villages Treatment (320) Control (186) Eligible Ineligible Eligible Ineligible K O K O K O K O

H 1 : resource pooling only for K. Show that, when eligible K receive PROGRESA, their ineligible relatives (K) consume more, but not their non-relatives (O); i.e. ITE(c) K > 0 and ITE(c) O = 0; ITE identified if randomization worked and absent local GE effects and spillover effects to control villages. Identify causal effect of connectedness on consumption under assumption that all sources of potential endogeneity are additive.

Previous findings from paper (2) 1 PROGRESA: increases average consumption and net transfers for ineligible households; does not change income or prices; is one of many existing government programs so it is unlikely to change incentives to share risk (at least in short term). 2 H 0 of full insurance rejected at the village level.

Households share PROGRESA transfer with their close relatives only: ITE(c) K > 0 and ITE(c) O = 0 Table 2: Do Family Networks Share Resources? Dependent Variable: Food Consumption in Pesos, per Adult Equivalent Difference-in-difference OLS estimates, standard errors clustered by village (2) Ineligibles With and Without (1) Ineligibles Eligible Extended Family Members ITE K [connected] 26.4** (11.1) ITE O [isolated] -15.5 (20.2) ITE [connected with eligible relatives] 26.9** (11.7) ITE [isolated and connected w/o eligible relatives] -4.4 (16.2) ITE 41.9** 31.4* (20.9) (17.3) Connected households (11/99) 2843 Isolated households (11/99) 712 Connected households with eligible relatives (11/99) 2482 Isolated and connected w/o eligible family (11/99) 1073 Observations 11015 11015 Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The dependent variable is household food consumption per adult equivalent, measured in November 1998 pesos. Standard errors are clustered by village. The sample covers ineligible couple-headed households, the data waves Angelucci, used are March De1998, Giorgi, May 1999, Rasul and November FAMILY, 1999. The INSURANCE, number of households AND reported INVESTMENT in each column refers to November 1999. The following controls are included in each specification: household head's age, gender, ethnicity,

H 2 : K households share risk efficiently ln c K ht = β 1 lny K t + β 2 lny K ht + uk ht

(Almost) full insurance for K Table 3: Full Insurance Within Extended Family Networks Dependent Variable: Growth in Food Consumption per Adult Equivalent, Nov. 1998 pesos Standard errors are clustered by network (1) OLS (2) IV (health shock) (3) IV (lagged income) PANEL A: Aggregate resources as network consumption Log aggregate resources (Y).973***.970***.967*** (.005) (.008) (.009) Log household resources (y).021***.007.010 (.001) (.021) (. 009) PANEL B: Aggregate resources as network by wave dummies Log household resources (y).030***.012.016 (.002) (.034) (.085) Network by wave dummies yes yes yes Number of observations 76517 60012 56646 Notes: *** denotes significance at 1%, ** at 5%, and * at 10%. The dependent variable is household food consumption per adult equivalent, measured in November 1998 pesos. Within group estimator. Standard errors are clustered by village. Clustering at the network level doesn't change the results. The sample covers eligible and ineligible couple-headed connected households. The data waves used are November 1998, May 1999, November 1999, November 2000, and November 2003, the health shock is not available in the last wave of our data. The adult equivalence scale used for consumption is one for members 18 or older, and 0.73 otherwise. In Column 2 the IV for household income is a dummy for the head of household being sick. In Column 3 the IV for household income is the first lag of household income. Extreme values are trimmed in the regressions.

Data structure and H 3 : differential effect of shock on investment for K and O. 506 villages Treatment (320) Control (186) Eligible Ineligible Eligible Ineligible K O K O K O K O

H 3 : investment differentially affected by shocks for K and O Easier to use PROGRESA transfer to invest in lumpy, high-return goods for K than O. O may have enough capital to invest in divisible but low-return goods only Estimate effect of PROGRESA eligibility on 1 eligible O household that receives the shock, ATE(I) O ; 2 and compare with network of eligible and ineligible K households, NATE(I) K. H 3 : 1 NATE(I) K > ATE(I) O for high-return, lumpy good; 2 NATE(I) K < ATE(I) O for low-return, divisible good.

Effect of PROGRESA: K invest more in education, less in poultry, than O. More in paper (4). Table 4: Investment and Resource Shocks OLS regression estimates, standard errors are clustered by village Effect of Progresa Eeigibility on investment for (1) eligible isolated households (ATEo), (2) family networks with at least one eligible household (NATEk), and their subgroups of (3) eligible (ATEk) and (4) ineligible (ITEk) connected households Human Capital Investment (1) School Enrollment Rate (aged 11-16) Agriculture (5) Poultry (6) Land used (7) Fertilizer/seeds ATEo [isolated] -0.020 0.275 0.085 0.060 [0.023] [0.094]*** [0.041]** [0.755] NATEk [networks w/ at least 0.050*** 0.072-0.263-0.190 one eligible household] [0.013] [0.065] [0.740] [0.688] NATEk-ATEo 0.070-0.203-0.250-0.057 [0.026]*** [0.100]** [0.898] [0.041] 28911 50380 22705 50236 Number of observations (NATEk and ATEo) Subgroups of NATEk: ATEk [connected eligibles] ITEk [ineligibles with eligible relatives] Number of observations (ATEk) Number of observations ( ITEk) 0.077 0.160 0.016 0.086 [0.014]*** [0.057]*** [0.017] [0.537] -0.052-0.253-1.133 0.038 [0.025]** [0.150]* [1.646] [0.047] 19541 32454 14614 32412 4841 10570 4716 10484 Average monthly transfer per adult equivalent: Isolated eligibles (household level) 48 Networks with at least one eligible household (network level) 34 Connected eligibles (household level) 48 Notes: *** denotes significance at 1%, ** at 5%, and * at 10%.Double difference estimates. We use September 1997 to November 1999 data for the effect of PROGRESA (with the exceptions

H 4 : long-term effect of income shock on investment, consumption higher for K If K respond to same income increase by investing more in the high-return good, long-term investment and consumption will be higher than for O. Compare change in y between 2003 (T = 1) and 1997 (T = 0). y=investment goods, consumption. y ht = κ 0 + κ 1 K h + κ 2 T t + κ 3 K h T t + κ 4 x h + u ht. Estimate O = κ 2 and K = κ 2 + κ 3. Test does not exploit randomization; stronger identification assumptions.

Bigger change in stock of education, poultry, consumption for K than O (1997-2003) Table 5: Extended Family Networks and Long Run Changes in Investment and Consumption OLS regression estimates. Standard errors clustered by village. (1) Share of Household Members With at Least 9th Grade (5) Poultry (6) Food Consumption All households Connected Household 0.049-0.678-8.937 [0.003]*** [0.038]*** [2.278]*** Isolated Household 0.038-0.833-14.061 [0.004]*** [0.057]*** [3.019]*** Difference 0.011 0.155 5.125 (Connected-Isolated) [0.004]*** [0.057]*** [3.193]* Number of observations 32306 32081 32088 Ineligible households Connected Household w/o 0.009-1.349-27.039 eligible relatives [0.020] [0.343]*** [9.558]*** Isolated Household 0.004-1.239-18.660 [0.010] [0.138]*** [9.261]** Difference 0.004-0.110-8.379 (Connected-Isolated) [0.020] [0.344] [12.158] Number of observations 1544 1531 1532 Notes: Angelucci, : *** denotes significance De Giorgi, at 1%, ** at Rasul 5%, and * at 10%. Estimates FAMILY, are double INSURANCE, differences in schooling, poultry, ANDand INVESTMENT consumption

To summarize 1. K share resources with each other but not with O (see Ambrus, Mobius, Szeidl 2010) 2. Effect of resource-sharing on insurance and investment. K: - are (almost) fully insured against idiosyncratic risk; - have higher high-return investment/lower low-return investment in response to same initial positive shock; - have higher long-term HK investment and consumption than O.

Takeaways 1 Resource sharing affects both smoothness and level of consumption (a twofer). 2 To measure value of informal institutions, look at both consumption smoothing and level. 3 The extended family uses aid more effectively in our data. 4 Need for the right economic conditions to find effect of family on investment.

Policy implications 1 Conditional cash transfer programs (CCTs) may increase consumption and investment. 2 Long-term effects. 3 Local institutions determine effectiveness of aid: need to understand/measure institutions; different interventions for different recipients (targeting). 4 Effects spill over to non-recipients, e.g. through informal institutions: underestimate effect of aid by looking at recipients only (e.g. 12% consumption underestimate, paper (2)); evaluations (randomization) should account for spillover effects, where relevant (partial population experiments [Moffitt 2001]; see paper (5)).