WPS7838 Policy Research Working Paper 7838 When the Cat’s Away The Effects of Spousal Migration on Investments on Children Lucia Rizzicas Development Economics Vice Presidency Operations and Strategy Team September 2016 Policy Research Working Paper 7838 Abstract Household expenditures for children-related goods may variables derived from the model. Results show that when change when one of the parent migrates and do so differ- children are left with fathers, the household budget is signif- ently depending on whether it is the mother or the father icantly diverted toward the purchase of adult private goods, that leaves. A sequential model that explains migration and but the share of budget devoted to children remains unaf- budget allocation choices is proposed and its predictions are fected because mothers compensate by giving up their own tested on data from Indonesia. Selection of households into private consumption and sending home more remittances. female migration is accounted for using a set of instrumental This paper is a product of the Operations and Strategy Team, Development Economics Vice Presidency. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The author may be contacted at lucia.rizzica@bancaditalia.it. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team When the Cat’s Away The Effects of Spousal Migration on Investments on Children Lucia Rizzica JEL codes: F22, J13, O15. Keywords: Children left behind, household decision making, Indonesia, migration. Lucia Rizzica is an economist at the Directorate General for Economics, Statistics and Research of the Bank of Italy, via Nazionale 91, 00184, Rome, Italy. Her email is lucia.rizzica@bancaditalia.it. The opinions expressed in this paper are those of the author alone and do not necessarily reflect those of the Bank of Italy. The author thanks the editor Andrew Foster, two anonymous referees, Francesca Carta, Riccardo De Bonis, Andrea Ichino, Italo Lopez Garcia, Steve Machin, Imran Rasul, and Marcos Vera-Hernandez as well as audiences at seminars at IFS, UCL, Royal Holloway, the International Conference on Migration and Development, the Cream-Northface Conference and the European Economic Association annual meeting for all their helpful comments and suggestions. In a context of global increase of international migration of workers, a more recent phenomenon is represented by female independent migration, sometimes referred to as feminization of migration. The most significant flows of female migrants are those of women from less developed countries who leave their families behind and migrate on their own to richer countries where they are usually employed as domestic workers and where they remain for a few years before going back to their country of origin to rejoin their families. Scholars have suggestively labeled this phenomenon as “the servants of globalization” (Parrenas 2001) or “the global nanny chain” (Lan 2006) or “the globalization of household production” (Kremer and Watt 2006). In the light of the rising importance of this phenomenon, it becomes imperative to understand its consequences for the household left behind and, in particular, for the children. The present work contributes to answering this question by analyzing the differences in the effects of parental migration on investments on children depending on whether it is the mother of the child that leaves or the father. From a policy perspective, understanding whether and how migration of the father or of the mother differently affects the children left behind can have important implications: for instance it can help governments as well as nongovernmental organizations decide about how to target financial and nonfinancial support to the families of migrants,1 or provide useful insights for the regulation of migration both in sending and receiving countries; indeed, while receiving countries increasingly adopt policies that allow the immigration of female domestic workers from developing countries to face the aging of their population and to encourage the labor force participation of local women,2 sending countries are starting to perceive the dangers entailed by the massive outflows of women and react by putting legal limits to emigration.3 This paper also provides an original contribution to the existing literature on migration: while numerous studies have analyzed the impact of migrant inflows on the destination countries’ labor markets (Card 1990; Altonji and Card 1991; Borjas 1994, 1999; Kremer and Watt 2006), considerably fewer have considered the effects on the sending country and on the households of origin (Cox-Edwards and Ureta 2003; Yang 2008). In this context, some authors have provided estimates of the impact of parental migration on the well-being of children left behind (Hanson and Woodruff 2003; McKenzie and Hildebrandt 2005; Mansuri 2006; Chen 2013), but these typically considered cases of paternal migration. Extending the results found for fathers to mothers is nontrivial for several reasons: first, because the migration experience of men and women is generally different, for example, women tend to migrate for shorter periods of time and to maintain stronger links with the household of origin (de la Briére et al. 2002); second, because mothers typically provide more care to the children, so that when the mother is away 1 UNICEF, for example, promotes policy research on migration and children left behind with a special focus on gender issues. 2 Similar policies are for instance in place in Hong Kong and Singapore, as analyzed in Kremer and Watt (2006). 3 In 2008, the government of Sri Lanka, passed a law that banned the international migration of mothers of children under the age of five. 2 children may suffer a more severe emotional loss (Cortes 2015); finally, because men and women have different preferences over investments on children (Thomas 1990; Duflo 2003; Qian 2008), so that the change in the structure of the household induced by migration of one of the spouses is likely to have different effects depending on who migrates and who stays.4 The present paper moves in the latter direction and complements the existing literature by investigating the allocation of resources within the household in order to understand whether spousal migration generates diversions of the household budget that affect children. Answering this question requires understanding what are the determinants of female migration. The few existing papers that investigated this aspect seem to suggest that female migration would be a means to provide the family left behind with a more stable and reliable source of income than what would be in case of male migration, this would happen because the jobs chosen by migrant women are typically less risky than those chosen by men (Lauby and Stark 1988) and because women are intrinsically more attached to the family left behind and thus tend to send more remittances (de la Briére et al. 2002). Building on this literature, I propose a sequential model in which first the two spouses cooperatively decide who should migrate so as to maximize the expected returns from migration while minimizing the associated risk, second, the migrant spouse chooses how much to send back home through remittances, and finally the spouse left behind chooses how to allocate the available household budget across different commodities. The resulting subgame perfect Nash equilibrium will be one in which diversions of the budget off of children-related expenditures are offset by a strategic change in the amount of remittances that the migrant sends back, so that the share of budget devoted to children- related expenditures is the same no matter whether it is the mother or the father that migrates, even if the two have different preferences over investments on children. The empirical analysis, based on Indonesian data, shows indeed that, when the mother migrates and the husband is left behind with the children, the share of income spent for exclusively adult goods increases but that devoted to children-related expenditures does not change significantly. The results are obtained through a two-stage least squares estimation procedure that exploits a set of instrumental variables derived from the theoretical setup, that is, the expected value and volatility of previous migrants’ earnings. Note that these results do not necessarily imply that children are no worse off when left behind with the father instead of the mother because other mechanisms, different from that of the allocation of the budget may play a crucial role; to shed light on this, I further estimate the effects of maternal versus paternal migration on children’s health and education indicators and find no difference in health measures but a significant decrease in the amount of time spent at school for children left with their fathers mirrored by an increase in the amount of time devoted to housework. This type of disruptive effect is likely to explain the result found by Cortes (2015) that children of migrant mothers are more likely to lag behind at school than children of migrant fathers. 4 A similar reasoning has been applied to the case in which one of the spouses passes away by Gertler et al. (2004). 3 The remainder of the paper is structured as follows: Section I describes the model of migration choice and of intra-household allocation of resources of the household; section II introduces the data; section III is dedicated to the identification and estimation strategy; section IV shows the estimation results; section V provides some robustness checks; and section VI concludes. I. THEORETICAL FRAMEWORK I propose a three-stage decision process in which first the spouses cooperatively5 decide whether and who should migrate so as to insure the household against income shocks; second, the spouse who migrates decides how much of her income to remit back to the household of origin; finally, the spouse left behind allocates all the available budget, made of her own income plus the remittances received, among different categories of goods. General Setup Consider a rational, utility optimizing, risk averse household faced with a risky source of income. Following Levhari and Stark (1982) and Lucas and Stark (1985), the household decides to reduce this risk by diversifying its income sources and specifically by placing one member of the household to a different location so that her income streams are less correlated with those of the household left behind. In a household composed of parents and children the choice becomes who to send away between the two spouses. We can call this a portfolio choice (Markowitz 1952) where “woman migrates” and “man migrates” are two risky assets that can be combined with other two risky assets that are “man stays” and “woman stays,” so that either both spouses stay or one of the two migrates and the other stays.6 I refer to the latter type as mixed migration portfolios. I assume that the household decides whether someone should migrate and who migrates but does not decide where the migrant will go, just men and women may migrate to different predetermined destinations. This assumption is supported by the sociological and economic literature that has widely documented that migrants tend to show very little variation in the choice of their destinations, following instead quite stable patterns of migration from one place to the other (Bartel 1989; Altonji and Card 1991). In this setting the expected returns associated with each of the three possible portfolios will be the sum of the expected wages of the two spouses, specifically: 5 Individual migration choices are similar to individual labor supply choices, for which collective models have been proposed (Chiappori 1992). 6 Note that the model would potentially allow for the definition of a fourth portfolio, that is, a full migration portfolio in which both spouses migrate. Yet this portfolio will be excluded from the analysis in that it would not be feasible for a household with children because migrants cannot typically bring their children with them, given the types of jobs and accommodation they get upon migration, and cannot leave their children alone either. 4 E ( R0 )  h E ( wm )  E ( whf ) E ( RM )  E ( wm )  E ( w f ) d h (1) E ( RF )  f )  E ( wm ) E ( wd h where E ( R0 ), E ( RM ) and E ( RF ) are, respectively, the expected returns to the household when none of the spouses migrates, when the man only migrates and when the woman wmd wd only migrates; and f are the wages of men and women upon migration (at h wh destination d) and wm and f the wages of men and women if they do not migrate (in the home village h). The risk associated with each of the three portfolios is instead: 0 2  f )  Var ( wm )  2 Cov ( wm , w f ) Var ( wh h h h M  2 f )  Var ( wm )  2 Cov ( wm , w f ) Var ( wh d d h (2) F 2  h Var ( wm f )  2 Cov ( w f , wm ) )  Var ( wd d h If the household’s utility is increasing in the expected returns of the portfolio chosen and decreasing in the associated risk, the migration choice will also be affected by the household’s degree of risk aversion (RAh) so that more risk averse households will choose the migration portfolio with lower risk, even if the expected returns are lower. The Allocation of Resources within the Household Each of the two risk averse spouses, i, obtains utility from the consumption of some private good Xi and from that of a common good Z that yields utility to both of them. The vector of common goods Z will contain all children-related expenditure, that is, both parents care about investment on children. Assuming Cobb-Douglas preferences, I express the preferences of men and women in the following way: U m   log X m  (1   ) log Z U f   log X f  (1   ) log Z where (1   ) and (1   ) , respectively, indicate the man’s and the woman’s willingness to contribute to the common good, and thus can be interpreted as measures of their generosity toward children.7 No migration portfolio. In the case in which none of the spouses migrates, they cooperatively choose how to allocate the available resources under income pooling so as 7 A large and well established literature, among which Thomas (1990), Duflo (2003), Qian (2008), documented that women have stronger preferences for investing on children than men, this paper will provide an indirect test of this hypothesis. 5 to maximize the sum of their private utilities with equal weights assigned to each of the two: max X f , X m , Z  log X m  (1   ) log Z   log X f  (1   ) log Z s.t.: X m  X f  Z  E ( R0 ) The household utility maximization problem yields the following optimal allocations:8  Xm 0  E ( R0 ) 2  f  X0 E ( R0 ) (3) 2     Z0   1  E ( R0 )  2  Mixed migration portfolios. Suppose now that the household prefers a mixed migration portfolio, for illustration, that in which the wife migrates and the husband stays with the children. In this case, the wife chooses how much remittances to send back home and the husband decides how to spend the available resources. The absence of the spouse induces the spouse left behind to deviate from the Pareto efficient equilibrium that would arise if they were together and divert resources toward his own private good. This situation can thus be seen as a noncooperative sequential game in which the subgame perfect Nash equilibrium (SPNE) can be obtained by backward induction: the husband decides how to allocate the budget available to him according to his own preferences; the wife anticipates this allocation and incorporates the husband’s choice in her decision problem of how much of her income to send back in remittances (R). Thus at the final stage of this game the husband left behind solves: max  log X m  (1   ) log Z X m ,Z s.t.: X m  Z  E ( wm h )R which yields: Z*  (1   )( E ( wm h )  R) Xm *   ( E ( wm h )  R) In the previous stage of the game, instead, the migrant wife anticipates the husband’s choice and decides how much to send back through remittances by solving: 8 These are Pareto efficient (Chiappori and Donni 2009). 6 max  log X f  (1   ) log Z R s.t.: f )  R0 X f  E ( wd Z  Z  (1   )( E ( wm * h )  R) The wife’s problem allows for only one possible internal solution.9 In this case, the migrant will choose to send remittances: f )   E ( RF ) R*  E ( wd and the resulting allocations will be: Xm *   (1   ) E ( RF ) Xf  *  E ( RF ) (4) Z  * (1   )(1   ) E ( RF ) Symmetrically, when it is the husband that migrates and sends back remittances to the wife, the resulting equilibrium allocations will be: Xm **   E ( RM ) Xf  **  (1   ) E ( RM ) (5) Z  ** (1   )(1   ) E ( RM ) with remittances sent by the husband: R**  E ( wm d )   E ( RM ) Clearly, these allocations are second best compared to the allocations that would be chosen if the spouses decided together how to spend the (higher) available budget. The first best allocation would be analogous to those in 3. The Migration Choice The subgame perfect Nash equilibrium (SPNE) of the three-stage migration game is found by comparing the level of utility that the household obtains from the three portfolios. Foreseeing that the final allocations will be 3, 4, and 5, the two spouses cooperatively decide, at the first stage, which migration portfolio maximizes the joint expected utility. 9 A corner solution arises instead when the migrant’s expected income is too low to send remittances; this case is discussed in the appendix. 7 max {  EU i ( X i0 , Z 0 ),  EU ( X i * i , Z * ),  EU ( X i ** i , Z ** ) } i  f ,m i  f ,m i  f ,m Pairwise comparison of the expected values of indirect utility of the three portfolios delivers the following: Proposition 1 When the spouses cooperatively choose the migration portfolio that ex-ante maximizes their joint expected utility, 1. Female migration is preferred to male migration if: 2  E ( RF )  (1   )  (6)   E ( RM )   (1   ) 2. Female migration is preferred to no migration if: 2  E ( RF )  (2     ) 2     (7)   E ( R0 )   4(1   ) 2   (1   ) 2    3. Male migration is preferred to no migration if: 2  E ( RM )  (2     ) 2     (8)   E ( R0 )   4(1   ) 2  (1   ) 2    These conditions highlight that the main driver of the choice is the expected financial gain from migration. Indeed, if the expected returns from migration do not exceed those of no migration, the spouses will prefer to stay in their village of origin. Similarly, if the expected gains from female migration significantly exceed those from male migration, households will prefer the former. The comparisons above neglected the risk dimension of the three portfolios, which is implicit in the structure of individual and household preferences. Proposition 1 is indeed based on the comparison between levels of utility of the expected returns of each portfolio. But because individuals are risk averse the difference between utility of the expected returns may differ from that between levels of expected utility. For example, a riskier portfolio, that is, one with more volatile returns, may give lower expected utility than a safer one even if the associated expected returns are higher. Therefore, the final migration portfolio choice will crucially depend on relative expected risk too. Finally, the spouses’ individual preferences may still play a key role in determining the optimal migration portfolio. Figure 1 shows how the optimal choice varies depending on α (horizontal axis) and β (vertical axis) when the expected returns from male and female 8 migration are the same and are 50% higher than in the case in which no one migrates. It turns out that when both spouses are very selfish (high values of α and β), it will be optimal not to migrate. This happens because the spouse left behind would allocate very little resources to the common goods and the migrant spouse would not be willing to compensate for this by sending more remittances. This relation is not linear though: when the spouse left behind is very selfish, no migration will be preferred also in the case in which the migrant is very generous, because in this situation, the utility of the latter will be too low compared to the case of no migration. Finally, condition (6) also implies that when the expected returns from male and female migration are the same, the spouses will prefer the more generous spouse to leave as she would extract less private gains from her first mover advantage. Figure 1. Optimal Migration Portfolios, by parental preferences. ER  ER  1.5 ER f m 0 Notes: . Gray area is where male migration is preferred to no migration. Horizontal dashed area where female migration is preferred to no migration. Dotted area is where female migration is preferred to male migration. Comparative Statics Having derived the optimal allocations in each scenario and the conditions under which the corresponding portfolios are SPNE, I can examine some comparative statics to understand how much of the (maximized) household income is spent on which types of goods and who benefits the most from the increase in the available budget that results from migration. Indeed, let small letters denote the share of household income devoted to each type of consumption goods. The model delivers the following prediction: 9 Proposition 2 The share of total household income spent on expenditure on children will be the same no matter which of the parents migrates: z*  z** (9) Z* Z ** z  * z  ** where E ( RF ) and E ( RM ) . Proposition 2 results from the ability of the migrant to anticipate the allocation that will be chosen by the spouse left behind and offset the implied shift of resources away from the children by sending back more remittances. Note that, because these are shares of different amounts of income, proposition 2 does not imply that the amount spent for children is the same in the two scenarios but only that their share, relative to the available income, is.10 On the other hand, when comparing the shares of income spent for the private consumption of the adults I obtain the following: Proposition 3 The share of total household income spent for adult private goods is higher when the wife migrates and the husband stays than when the husband migrates and the wife stays if and only if women have stronger preferences for expenditure for children than men (   ) :  0 if    xm f   (1   )   (1   )        x** * (10)  0 otherwise This result derives from the fact that the share of income available for private consumption of the spouse left behind depends on the “generosity” of the migrant so that if women are more generous than men, the latter can enjoy more private consumption out of the wife’s remittances. Symmetrically, the implied loss of private good consumption of the migrant will be larger if she is more generous than the spouse left behind. Finally, I can compare the allocations resulting from the mixed migration portfolios with those chosen when no migration takes place. It turns out that: Proposition 4 The migrant spouse consumes a larger share of the available income with respect to the situation in which no one migrates: f  xf  0 x* 0 (11) 10 This result depends on the assumption that preferences have either unitary elasticity of substitution, that is, they are Cobb-Douglas, or null elasticity of substitution, that is, they are Leontief. 10 The spouse left behind gains from migration of the spouse, in terms of shares of income spent on his own private consumption, only when the migrant spouse is “generous enough” 1 xm *  xm 0  0 if   (12) 2 Children always receive a larger share of income when none of their parents migrates, because either of them has incentives to shift resources onto own private consumption when alone. z*  z**  z 0 (13) These results should not be interpreted as suggesting that, for example, children are always worse off when either of their parents migrates, or that the spouse left behind would prefer no migration to take place, because the allocations compared are in terms of shares of the total household budget available, which is generally larger than the budget available in the case of no migration, as suggested in proposition 1. Yet they highlight how, in the absence of a moral hazard problem, children of migrant parents would instead receive a larger share of the household income. II. DATA The empirical analysis is based on data from the Indonesia Family Life Survey (IFLS), an ongoing longitudinal survey of Indonesian households, which started in 1993, was repeated in 1997, 2000, and 2007, and contains a sample that is representative of about 83% of the total Indonesian population, including over 30,000 individuals living in 13 of the 27 provinces of the country.11 By tracking individuals over time, these data allow detecting migration. Indeed, for all individuals who appeared in the first wave of the survey the IFLS roster provides information on where they currently are (if they are not in the household anymore), why and when they left and how much they earned in the past twelve months. I define migrants as those adult people who have left the household and are reported to having done so for work reasons or explicitly to help the family. Table 1 shows that from the time of the first interview in 1993, fourteen years later more than one household out of four had at least one member that had migrated and not come back while, among individuals, migrants represented about 9%. Table 1. Migration in the IFLS. Households Migrant % Wave Individuals Migrants % Households 11 For a more detailed description of the dataset see Thomas et al. (2012). 11 1993 33,081 - - 7,224 - - 1997 39,714 1,701 4.28 7,699 1,304 16.94 2000 54,991 2,835 5.16 10,435 2,022 19.38 2007 73,016 6,352 8.70 13,536 3,787 27.98 Table 2 shows the gender composition of these migration flows: almost two thirds of those recorded as migrants in 2007 are men, but women are more than twice as likely as men to migrate internationally, and this is particularly true for mothers who leave their family behind. On the other hand women, especially those who leave their children behind, tend to stay away for a period of time that is significantly shorter than that of men.12 Table 2. Migrants by Gender. IFLS 2007. Number % of total % international Migration spell migrants (months) Migrants 6,352 - 11.04 62.89 Men 4,046 63.70 7.74 64.83 without children 3,255 51.24 7.74 64.10 with children along 493 7.76 2.44 78.23 with children left behind 298 4.69 16.55 51.23 Women 2,306 36.30 17.00 59.48 without children 1,892 29.79 13.27 59.76 with children along 222 3.49 10.36 76,43 with children left behind 192 3.02 60.32 37.48 To identify households who have chosen the different migration portfolios described in section I, I check, for every child under the age of 18, whether his father or mother migrated and the other parent is in the household. Table 3 shows the partition of households with children in the 2007 IFLS sample: the vast majority of them, over 95% chose a no migration portfolio, while 1.6% chose the female migration portfolio and 2.7% the male migration portfolio. A negligible share, about 0.4% chose a portfolio in which both parents migrated leaving their children behind with other relatives. Table 3. Migration Portfolio Choices of Households with Children. IFLS 2007. Mother migrates Mother stays Father migrates 40 258 Father stays 152 9186 Table 4 provides a comparison between households that have chosen different migration portfolios: column (1) reports the characteristics of the households in which none of the parents migrated, column (2) refers to households in which either the mother or the father 12 Table S.1 in the supplemental appendix reports descriptive statistics of the characteristics of migrants by gender, comparing those who have no children with those who migrate with their families and those who leave their family behind. While there are sizable differences between migrants who have children and those who don’t, parents who bring their children along when they migrate do not appear to be significantly different from those who leave them behind, except for the fact that those who bring their children along are generally more educated. 12 migrated, while columns (3) and (4) split this sample between male and female migration. It appears that the differences between nonmigrant households (column 5) and migrant households are very stark: households with no migrants are more likely than migrant households to reside in urban areas, they generally have less children and are richer and more educated. The comparison between households in which the mother migrated and the father stayed and households in which the father migrated and the mother stayed (column 6) further reveals that these differences are amplified for households with migrant mothers; these are indeed more likely to live in rural areas, to be less educated and poorer, and to have more children, though these are on average older than those of migrant fathers. The last row of table 4 finally shows differences in terms of degree of risk aversion of the head of household.13 This measure is a 1–4 score derived from the answers to a sequence of questions in which respondents are faced with hypothetical lotteries with high stakes.14 It appears that migrant households are on average more risk averse than nonmigrant ones, while there is no significant difference between households who choose female migration and those who choose male migration. Table 4. Descriptive Statistics, Households with Children. IFLS 2007. (1) (2) (3) (4) (5) (6) Nonmigrant Migrant Migrant Migrant Difference Difference households households father mother (1)-(2) (3)-(4) Size of household 4.623 4.590 4.570 4.558 0.033 0.012 (1.797) (1.933) (2.014) (1.790) Rural 0.545 0.643 0.587 0.711 -0.098*** -0.124* (0.498) (0.480) (0.493) (0.455) Number of children 1.728 2.066 2.050 2.079 -0.338*** -0.029*** (1.009) (1.114) (1.135) (1.077) Age of children 8.456 10.05 9.680 10.39 -1.594*** -0.71* (4.701) (3.502) (3.706) (3.299) Share of female 0.494 0.474 0.462 0.499 0.02 -0.037 children (0.403) (0.363) (0.363) (0.360) Mother’s years 8.301 7.250 7.325 6.994 1.051*** 0.331* of education (4.333) (3.364) (3.593) (2.789) log total expenditure 15.22 15.08 15.15 14.93 0.14*** 0.22** (0.937) (0.849) (0.905) (0.659) Risk aversion of head 3.693 3.757 3.738 3.784 -0.064* -0.046 13 Or the spouse when no measure for the head is available. Indeed the data show that the head’s and her spouse’s degree of risk aversion are significantly positively correlated,   0.34 . 14 For example the first question asks: “Suppose you are offered two ways to earn income. With option 1, you are guaranteed an income of Rp 4 million per month. With option 2, you have an equal chance of earning either the same income, Rp 4 million per month, or, if you are unlucky, Rp 2 million per month, which is less. Which option will you choose?” 13 (0.794) (0.706) (0.732) (0.666) N 9186 450 258 152 Notes: Standard deviations reported in parenthesis. Differences computed with standard errors clustered at village level. *** p<0.01, ** p<0.05, * p<0.1 III. EMPIRICAL STRATEGY The equation of interest is one in which I estimate the shifts in the shares of total household expenditure from one category of consumption goods to the other when the parent that migrates is the mother rather than the father. This is: wih   ih   ln nh   Fh   X h  uih (14) where on the left-hand side is the share of total household income allocated to expenditure for commodity i by household h, and on the right-hand side is the (log) number of members in the household nh together with several household’s observable characteristics Xh and a term Fh equal to one if the household is one in which the mother of the children has migrated, while the father did not, and zero if instead it was the father who is away and the mother remained with the children. The coefficient of interest is γ, associated with the term Fh; this provides us with a measure of the difference between the budget allocation of households with migrant mothers and households with migrant fathers. Equation (14) provides an empirical test of propositions 2 and 3. We will consider that there are some types of goods that are exclusively consumed by adults (Xm and Xf), while all other goods are, to different extents, consumed by both adults and children (Z). Specifically, expenditure for food, health, education, and durables will be proxies of Z, where education is a purely child-related expenditure, while food, health, and durables also partly pertain to adults; on the other hand, alcohol, tea, coffee, and tobacco will proxy for adults’ exclusive private consumption, following Deaton (1989).15 Estimating the effects of migration and how they differ depending on the gender of the migrant spouse entails problems of endogenous selection into treatment of two types, as suggested in table 4: first there is a problem of selection into migration as households that decide to send some member out for migration will likely differ from the others on both observable and unobservable characteristics; second, there is a problem of selection into female migration because households from which it is the mother that migrates are likely to differ from those from which the father migrates in a number of unobservable factors that might as well influence the outcomes of interest. 15 Note that these adult goods are typically consumed more by men than women. In female headed households, the share of total income spent on these items is about 2% for each adult member, while in male headed households over 4%. This difference is larger if one compares households with no adult men with households with no women, in the former the share of total income spent on alcohol, tea, coffee, and tobacco is about 2.6% per adult, while in all men households it exceeds 9%. These figures, therefore, confirm that men have stronger preferences than women for this type of goods but also show that women, too, consume positive amounts of them. 14 The focus of the paper will be addressing the second type of selection so as to be able to compare households with migrant mothers to households with migrant fathers. On the other hand, the results of proposition 1 will provide a guide for interpreting the direction of the bias that may arise from conditioning on selection of households into migration. I thus build on the intuition of the theoretical model introduced in section I to derive a set of instrumental variables that may influence the decision of migrant households about which of the spouses to send out for migration but will not have any direct effect on the outcome variables of equation (14). The model of section I is translated to the data by first assigning “counterfactual” gender specific destinations to each household. To this purpose I identify, for each household, the year in which the migration decision has been taken as that in which the migrant has departed; I then consider the destinations chosen by previous migrants from the same village and take the destination that was most common among female migrants as counterfactual destination for women and the one that was most common among male migrants as counterfactual destination for men. Counterfactual destinations will vary by gender, time of departure, and village. The choice of assigning the most common destination of migrants from the same village to perspective migrants is supported by the evidence reported in table 5, which shows that over 60% of migrants from the same village had chosen the same destinations. This is in line with the literature that has documented the formation of networks of migrants (Bartel 1989; Altonji and Card 1991; Lafortune and Tessada 2012; Patel and Vella 2013) but can also, in this particular context, be explained by the widespread use in South Asia of recruiting agencies that are connected to other agencies in a foreign country and, therefore, typically send all of the people of the village they visit to the same destination (Parrenas 2001; Suradji 2004). Table 6 shows the gender specific destinations assigned to each household for the whole IFLS sample and for the restricted sample of households with a migrant parent. The only significant difference between the two samples is that migrant mothers are much more likely to be assigned Saudi Arabia as counterfactual destinations than the full sample of women, this is in line with the evidence in table 2 that women who leave children behind are much more likely to migrate internationally than any other type of migrant. Table 5. Migrants per Village. IFLS 2007. Men Women Adults per village 55.31 58.70 (40.99) (42.29) Migrants per village 15.64 9.39 (14.002) (9.185) % Migrants at same destination .617 .622 (0.231) (0.234) Notes: Standard deviations reported in parenthesis. 15 Table 6. Households’ Counterfactual Gender Specific Destinations, 2007. Counterfactual Men Women destinations Full Restricted Full Restricted sample sample sample sample West Java 24.19 34.99 19.49 21.53 East Java 17.2 8.68 19.74 9.53 Central Java 13.43 14.66 14.41 14.48 Jakarta 12.02 15.69 12.49 15.08 North Sumatra 6.16 2.87 5.3 6.7 Lampung 4.44 5.76 2.93 2.73 South Sumatra 3.93 2.99 3.06 2.58 South Sulawesi 3 2.09 2.94 2.75 West Nusa Tenggara 2.29 3.32 1.96 3.10 Bali 2.16 1.07 1.86 1.08 South Kalimantan 2.08 0.52 2.06 0.53 West Sumatra 1.97 0.95 2.28 3.7 Yogyakarta 1.86 0.45 1.89 0.43 Riau 0.74 0.30 0.39 0.33 Banten 0.35 0.1 Southeast Sulawesi 0.24 0.50 East Kalimantan 0.18 0.37 0.49 0.50 Central Kalimantan 0.09 N Aceh Darussalem 0.03 0.07 Irian Jaya 0.02 Sumatra 0.01 Bengkulu 0.01 West Sulawesi 0.01 Central Sulawesi 0.28 North Sulawesi 0.11 Malaysia 3.43 4.74 1.98 3.83 Singapore 0.03 Taiwan 0.02 Saudi Arabia 0.01 5.65 18.60 Timor Leste 0.03 United Arab Emirates 0.56 0.43 N 13273 335 13134 336 Once I have assigned gender specific destinations to each household, I exploit again the information about previous migrants and generate, for every destination and year of migration decision, a measure of expected returns and risk by taking the mean and standard deviation of the incomes of all male and female migrants that previously migrated to that destination, and combining them as in equations (1) and (2). Note that these instrumental variables will take different values depending on the village of origin and time of migration, so that they will eventually be household specific, in that, once households from the same village have been ordered on the basis of the time of departure, the values of wages of previous migrants will be different from one another. For what concerns the covariance between income at home and income at destination, instead, I exploit the longitudinal dimension of the data, and compute the covariance for every village-destination pair across waves. Table 7 shows that, on average, the expected wages 16 of men upon migration are higher than those of women, but also more volatile. The difference between men’s and women’s expected wages is even larger if they remain in the home village, while the difference in variance remains similar. In terms of covariance between migrant’s and spouse’s wages, men’s migration is associated with higher risk. Finally, when combining the expected wages and the associated risk as in expressions (1) and (2), the difference between the two portfolios becomes much less clear with male migration being slightly less remunerative and riskier. Table 7. Expected Returns and Risk from Migration Portfolios (Log). Men Women (1) (2) (3) (4) (5) (6) Full Restricted Full Restricted Difference Difference sample sample sample sample (1)-(3) (2)-(4) E ( wid ) 16.218 16.205 15.945 16.067 0.273*** 0.138*** (0.516) (0.403) (0.361) (0.353) E (wih ) 16.056 15.964 15.641 15.647 0.415*** 0.317*** (0.494) (0.490) (0.614) (0.604) Var ( wid ) 16.390 16.345 15.912 15.934 0.478*** 0.411*** (0.881) (0.766) (0.556) (0.529) Var ( wid ) 16.133 16.007 15.723 15.779 0.41*** 0.228*** (0.731) (0.734) (0.766) (0.834) h Cov( wid , w i) 32.051 32.094 32.030 31.928 0.021*** 0.166*** (1.235) (1.110) (0.903) (0.947) E ( Ri ) 16.722 16.719 16.739 16.755 -0.017*** -0.036* (0.408) (0.365) (0.347) (0.336) σi 16.870 16.905 16.799 16.744 0.071*** 0.161*** (0.788) (0.728) (0.567) (0.562) N 13507 337 13457 337 Notes: Standard deviations reported in parenthesis. *** p<0.01, ** p<0.05, * p<0.1 The first stage specification for selection into female migration will include relative expected returns of the two migration portfolios as suggested in proposition 1, but also relative risk of the two portfolios and risk aversion of the household. The latter will be included as a control variable on its own and then as an instrumental variable when interacted with the level of risk: 17  ER f  f  f  Fh  ah  b1    b2    b3 RAh     c1 RAh  c2 ln nh  c3 X h  vh (15)  ERm  h  m  h  m  h The coefficients of the instrumental variables are b1, b2 and b3, while the other variables will not be excluded from regression (14). In terms of validity of these instruments, I imagine that households will make their migration decision based on their information about the possibilities they might have at destination; thus, it is reasonable to believe that the experience of previous migrants best represents the information set available to potential migrants. The implicit assumption is that migrants’ wages are not determined by the individual characteristics of the migrant but are somehow exogenously set so that nonmigrants would be faced with the same wages if they migrated. This assumption is supported by the data reported by the Indonesian Statistical Office (Badan Pusat Statistik 2007) about jobs of international migrants. These show that there is strikingly little variation in the types of jobs Indonesian migrants get at destination: 53% of migrants who were abroad in 2007 worked as domestic helpers, while 42% as either construction, factory, or plantation workers. Although this information is not available by gender, one can see a clear difference between typical female jobs and typical male jobs: domestic workers are very likely to be the female migrants (whose total percentage is in fact around 55% of total international migrants in the IFLS), while the construction, factory, and plantation workers are likely to be the men. As these are all very low skilled jobs, it becomes reasonable to assume that wages are fixed and exogenous. Moreover, at least for women, there is vast anecdotal evidence that they are hired to go work abroad as domestic workers under standard contracts that specify the same wage and duration of employment for all (Lan 2006; Parrenas 2001; Suradji 2004). IV. RESULTS The outcomes of interest are the shares of total household income spent for a given type of commodity. To compute the denominator, I follow Dai et al. (2011) who have estimated the distribution of household income using the same data from the fourth wave of the IFLS. Income is thus obtained summing up five components: labor income; income from agricultural business; income from nonagricultural business; household nonlabor income (scholarships, pensions, other transfers); household assets income. Moreover, to correct for attrition, a two step Heckman procedure is applied, in which the probability of response is predicted by dummy variable for whether the respondent is the head of the household. Levels of (log) income and predicted income are reported in table 8 suggesting, as in table 4 that migrant households are on average poorer than nonmigrant households and households with migrant mothers poorer than households with migrant fathers. Table 8 further shows descriptive statistics for the shares of income allocated to the various types of commodities, distinguishing, in particular, between common goods, that is, food, education, health, and durables and exclusively adult goods, which include 18 alcohol, tobacco, coffee, and tea.16 It appears that generally migrant and nonmigrant households devote similar shares of household income to the various commodities, with the exception of adult goods, whose share is significantly larger in nonmigrant households. This is somehow suggestive of the fact that both parents consume alcohol, tobacco, coffee, and tea, so that when there is only one of them in the household, the share of income spent on these items falls. Yet this drop is larger in the case in which the husband migrates, because men generally have stronger preferences for these goods. Indeed, the comparison between households with migrant mothers and households with migrant fathers reveals that the latter devote a significantly larger share of income to food expenditure and less to adult goods, reflecting differences in tastes between men and women. Table 8. Household Income Levels and Shares. Nonmigrant Migrant Migrant Migrant Difference Difference households households father mother (1)-(2) (3)-(4) Household income 16.56 16.26 16.25 16.26 0.3*** -0.01 (log) (0.711) (0.704) (0.717) (0.621) Predicted household 16.47 16.36 16.42 16.25 0.11*** 0.17** income (log) (0.711) (0.644) (0.687) (0.500) Shares of income spent on: Food 0.634 0.711 0.764 0.582 -0.077 0.182** (4.717) (0.800) (0.858) (0.668) Education 0.153 0.161 0.171 0.137 -0.008 0.034 (1.460) (0.255) (0.304) (0.158) Health 0.0513 0.0481 0.058 0.027 0.003 0.031 (0.586) (0.228) (0.292) (0.094) Durables 0.0139 0.130 0.0202 0.001 -0.116 0.019 (0.528) (2.188) (0.377) (0.007) Adult goods 0.117 0.080 0.060 0.108 0.037*** -0.048*** (0.826) (0.135) (0.133) (0.133) Notes: Standard deviations reported in parenthesis. Differences computed with standard errors clustered at village level. *** p<0.01, ** p<0.05, * p<0.1 Equation (14) is first estimated through OLS; the results are reported in table 9, controlling only for household size in the first panel, and adding an indicator for rural households, the number of years of education of the mother and the (log of) total household expenditure in the second panel. To correct for correlation of errors within villages, standard errors are heteroskedastic robust and clustered at village level in all specifications. The OLS results confirm that households in which mothers have migrated spend significantly more on adult goods, the shift being about 6 percentage points, and less on common goods, 16 Note that these shares do not add up to one for two reasons: the first is that the denominator is a measure of predicted income and not total household expenditure; this is done to better replicate the theoretical predictions. Second, there is a residual, unobserved category of expenditure, which is that for private consumption of the migrant. 19 although these coefficients are not statistically significant once I control for those characteristics that most differ between the two groups of households. Indeed, the inclusion of control variables seems to amplify the difference in adult goods consumption revealed in table 8 and to reduce the difference in terms of food consumption. Table 9. OLS Estimation. Household Level. Shares of household income spent on: (1) (2) (3) (4) (5) Food Education Health Durables Adult A. No controls Migrant mother -0.206** -0.036 -0.034 -0.024 0.057*** (0.096) (0.029) (0.023) (0.018) (0.017) Size of household -0.402*** -0.078*** -0.105* -0.004 -0.019 (0.104) (0.023) (0.062) (0.005) (0.020) Observations 337 337 337 337 337 R2 0.053 0.019 0.033 0.001 0.045 B. Controls included Migrant mother -0.104 -0.012 -0.019 -0.024 0.062*** (0.079) (0.022) (0.019) (0.019) (0.017) Size of household -0.032 -0.036 -0.080 0.000 0.000 (0.099) (0.031) (0.052) (0.006) (0.019) Observations 337 337 337 337 337 R2 0.376 0.097 0.064 0.003 0.078 Notes: Controls are rural household, years of education of mother, log total household expenditure. Standard errors robust to village level clustering in parenthesis. *** p <0.01, ** p <0.05, * p <0.1 If households from which mothers migrate are on average poorer, more rural and less educated than households from which fathers migrate (table 4), the OLS coefficient associated with common goods will likely be downward biased. For this reason, in order to control for the possibility that households from which mothers migrate differ from those from which it is the father who leaves, I proceed to estimate equation (14) by two-stage least squares (TSLS). Table 10 shows the results of the first stage regression: columns (1)–(3) use the measures of expected returns and risk of migration portfolios constructed according to equations (1) and (2). The signs confirm the existence of a trade off between expected returns and expected risk and acknowledge the amplifying effect of risk aversion. Moreover, more risk averse households tend to prefer female migration thus confirming the idea, suggested in Lauby and Stark (1988) and de la Briére et al. (2002), that female migration is perceived as a “safer” investment than male migration. Columns (4)–(6) then use only the destination side of the migration portfolios; this is done to reduce the number of instruments and their potential collinearity and hence increase the power of the first stage 20 predictions. These regressions confirm that female migration is more likely when it is associated with higher expected returns and lower variance relative to male migration. For all specifications, the last three rows of the table report the values of the F statistic of excluded instruments. A first look at these values convinces us that the most efficient specification is that of column (7), which displays the highest level of the F statistic and includes the control variables. Table 10. First Stage Regression. Migrant mother (1) (2) (3) (4) (5) (6) (7) (8) E ( R f ) / E ( Rm ) 0.059 0.241 0.362** 0.379** (0.069) (0.185) (0.180) (0.174)  f / m -0.089 -0.113 0.000 (0.081) (0.078) (0.127) RA   f /  m -0.029 (0.031) E ( wd d f ) / E ( wm ) 0.149*** 0.297*** 0.316*** 0.308*** (0.057) (0.101) (0.100) (0.100) Var ( wd d f ) / Var ( wm ) -0.092* -0.090* 0.039 (0.048) (0.047) (0.106) RA  Var (wd d f ) / Var ( wm ) -0.032 (0.026) RA 0.032 0.013 (0.040) (0.043) Controls n n y y n n y y Observations 377 369 334 318 379 372 337 321 R2 0.002 0.010 0.053 0.067 0.026 0.046 0.081 0.090 F statistic 0.746 1.919 3.941 3.232 10.03 8.840 9.374 6.297 E ( Ri ) are the expected returns from migration of individual I, σ is the associated risk, RA is the degree of risk aversion, Notes: f d E ( w ) are the expected wages of migrant i at destination, Var ( wd i f ) is their variance. Controls are: size of household, rural household, years of education of mother, log total household expenditure. Standard errors robust to village level clustering in parenthesis. *** p <0.01, ** p <0.05, * p <0.1 Table 11 shows the results of the TSLS estimation of equation (14). Comparing these with those of the OLS, we observe that the increase in adult goods expenditure is larger than it was in the OLS, while the effect on all types of common goods is not statistically different from zero. Table 11. Two-Stage Least-Squares Estimation. Household Level. Shares of household income spent on: (1) (2) (3) (4) (5) 21 Shares of household income spent on: Food Education Health Durables Adult A. No controls Migrant mother 0.026 -0.082 0.089 -0.007 0.165** (0.335) (0.112) (0.150) (0.019) (0.083) Size of household -0.419*** -0.074*** -0.114 -0.005 -0.027 (0.102) (0.023) (0.070) (0.005) (0.022) Observations 337 337 337 337 337 Uncentered R2 0.443 0.278 0.0163 0.00285 0.162 F statistic 8.604 8.604 8.604 8.604 8.604 Hansen J statistic 0.847 1.025 1.018 2.127 0.812 p-value 0.357 0.311 0.313 0.145 0.367 B. Controls included Migrant mother -0.293 -0.082 0.082 -0.017 0.149** (0.337) (0.102) (0.135) (0.020) (0.074) Size of household -0.027 -0.034 -0.083 0.000 -0.002 (0.102) (0.030) (0.054) (0.005) (0.021) Observations 337 337 337 337 337 Uncentered R2 0.633 0.329 0.0652 0.00462 0.230 F statistic 9.374 9.374 9.374 9.374 9.374 Hansen J statistic 0.723 0.775 0.771 1.938 0.637 p-value 0.395 0.379 0.380 0.164 0.425 Notes: Controls are rural household, years of education of mother, log total household expenditure. Standard errors robust to village level clustering in parenthesis. *** p <0.01, ** p <0.05, * p <0.1 These findings confirm the predictions of the model in section I for which the difference in the share of household income devoted to investment on children, here expenditure on food, health, education, and durables, is the same in households from which the father migrated and households from which it was the mother that migrated, z  z . This would * ** be due to the fact that the migrant spouse can use remittances as a way to offset the shifts in the allocation of the household budget that would be made by the spouse left behind. On the other hand, proposition 3 was predicting that the difference in the share of household income devoted to adult’s private consumption between the case in which the x*  x** mother migrated and that in which the father migrated, m f , was equal to    , positive if women have a stronger preference for investment on children with respect to men. The estimated coefficient thus suggests that men’s preferences are such that they would like to spend around 15% more than women on private consumption rather than on expenditure on common goods. These results also implicitly indicate that the migrant 22 mother suffers a loss of private consumption that is larger than that suffered from a migrant father.17 The estimates may be biased by selection into migration if households that choose to send one of the spouses away for migration systematically differ from nonmigrant households in terms of spouses’ preferences.18 Indeed, figure 1 suggested that, if the returns to migration are not high enough, a mixed migration portfolio may not be optimal if either of the spouses was excessively selfish, that is, if either α or β were particularly high. In this sense, the estimated difference would be a lower bound of the actual difference between men’s and women’s generosity toward children.19 That women have stronger preferences for consumption on common goods than men has been highlighted by many studies: Thomas (1990), Lundberg et al. (1997), Duflo (2003), Qian (2008), and Ashraf (2009) all show that income accruing to women generates larger benefits for children than that accruing to men. Unfortunately, it is generally difficult to compare the magnitude of these estimates with those found in this paper because in most cases both the outcome and the explanatory variables are defined differently. Yet there are at least three papers that contain comparable estimates as they use as outcome variables the shifts in the shares of household expenditure. Aggregating their shares in a way that is comparable to the one used in this paper, Hoddinott and Haddad (1995) show that the shares of the household budget allocated to adult goods are between 3.2 and 6.6 percentage points lower in the case in which the woman earns the whole household budget with respect to the case in which it is the man. Similarly, Attanasio and Lechene (2002) find 17 An alternative mechanism underlying these results could be that women and men, when migrating, change their preferences over investment on children. Specifically, if women experienced very low returns to education when migrating, relative to what they observed before migrating, it may be that they decide to invest less on children’s education. This would explain the result that investments on children’s education are not affected by female versus male migration. To rule out this mechanism, I computed returns to education at each counterfactual destination and compared them to returns to education in the migrant’s home village. Results are reported in table S.4 in the supplemental appendix and show that there is no significant difference between returns to education observed in the home village and those observed by the migrant at destination. It is thus unlikely that the pattern of results observed can be explained by changes in preferences over investment on children’s human capital generated by the different migration experience of men and women. 18 . Another source of selection bias may be due to differences between migrants who bring their children along and those who leave them behind. Table S.1 in the supplemental appendix shows that these are very similar, the only significant difference being the level of education, which is controlled for in all specifications. 19 Tables S.2 and S.3 in the supplemental appendix report the results of two different empirical specifications which are meant to account for selection into migration. The first is panel fixed effect estimation in which I distinguish the effect of having one spouse away from that of having the mother instead of the father. The estimates show that households with migrant mothers still spend significantly more on adult goods. The difference (+6%) is smaller but not statistically different from that of table 11. Table S.3 in the supplemental appendix, instead, reports the estimates of a three-stage estimation (Wooldridge 2010) in which risk aversion is used in a first stage as instrument for selection into migration, and then, in a second stage, the usual instruments are used to correct for selection into female migration and the Inverse Mills Ratio from the first stage is included. Unfortunately, the inclusion of the IMR, weakens the instruments so that the final coefficients are very imprecisely estimated. The coefficient for adult goods yet is now considerably larger in magnitude. 23 that a 100% increase of the woman’s household income share generates a decrease in the share of expenditure allocated to alcohol and tobacco between 19 and 40 percentage points. The two papers just cited do not provide an exact test of the model introduced in this paper because the presence of the spouse, even when she does not contribute to the household’s income at all, is likely to affect the choice of how to allocate it and because in the case of migration the migrant spouse does no typically earn the full household income although generally becomes the main income earner. Finally, the paper by Ashraf (2009) offers an interesting comparison: she uses an experimental setting in the Philippines to test whether husbands and wives have different preferences over the allocation of the household budget and how information and communication affect their choices. Interestingly, she shows that in situations in which one of the spouses receives a temporary shock to income and the other spouse is not able to control how she spends this extra budget (in the setting of her experiment this is the “Private” treatment), 60.4% of men versus 52.1% of women choose to deposit that money on their own private account rather than converting it into food vouchers. This is a rough test of the difference in the “generosity” parameters included in the parents’ utility functions as described in section I and suggests that     8.3% , a number lower but still close enough to the estimates of table 11. V. ROBUSTNESS CHECKS Table 12 reports the results obtained under several alternative empirical specifications together with the main OLS and TSLS estimates. The first alternative specification is a method proposed by Wooldridge (2010) to improve efficiency of TSLS estimates. This consists in estimating first a binary response model P( Fh  1| x, z )  G ( x, z ,  ) by maximum likelihood (typically probit), obtaining the fitted probabilities and then estimating the outcome equation by IV using as instrument . The estimates obtained through this method are very similar in magnitude to the TSLS ones but more precise. The main concern, though, relates to the possibility that the instruments employed may be weak and thus the TSLS estimates would be biased toward the corresponding OLS estimates. Indeed, the F statistic of the TSLS only exceeds the critical value corresponding to 20% size of test as of Stock and Yogo (2002). To test the robustness of the TSLS results to instruments’ weakness I first estimate equation (14) using only the most powerful E(wd d f ) / E(wm ) instrument I have, namely . As suggested by the higher value of the corresponding F statistic, these estimates are less biased toward the OLS. The coefficient for adult goods is larger in magnitude than the TSLS one but not statistically different from it. Second, in the case of over identified models, the limited information maximum likelihood estimator (LIML) and the Fuller-k estimator (with α = 1) are less subject to weak instruments bias than TSLS, as showed by Stock et al. (2002). Indeed, the LIML estimates reported in table 12 turn out to be associated with the lowest critical values of the F statistic and are thus the most unbiased as the corresponding value of the F statistic exceeds the 10% critical value. As predicted by Blomquist and Dahlberg (1999), the absolute magnitude of the coefficients estimated through LIML is slightly larger than the 24 TSLS estimates, as are the standard errors, but the fact that the difference between the coefficients estimated with the two methods is negligible reinforces the hypothesis that the instruments have enough predictive power. Indeed, if they were weak, the TSLS estimates would have been much closer to the OLS ones than to the LIML ones. Also, the Fuller specifications20 with α = 1 and α = 4 produce lower threshold values for the F statistic and thus are less biased than the TSLS coefficients, and their magnitude is almost identical to the TSLS. Table 12. Robustness Checks: Weak Instruments. Shares of household income spent on: Stock-Yogo weak ID (1) (2) (3) (4) (5) F test critical values Food Education Health Durables Adult statistic 10% 20% 30% OLS -0.104 -0.012 -0.019 -0.024 0.062*** (0.079) (0.022) (0.019) (0.019) (0.017) TSLS -0.293 -0.082 0.082 -0.017 0.149** 9.374 19.93 8.75 7.25 (0.337) (0.102) (0.135) (0.020) (0.074) Wooldridge -0.096 -0.079 0.074 -0.028 0.143** 18.61 16.38 6.66 5.53 (0.251) (0.094) (0.131) (0.020) (0.071) IV -0.153 -0.017 0.150 -0.053 0.186** 14.03 16.38 6.66 5.53 (0.342) (0.126) (0.158) (0.041) (0.087) LIML -0.299 -0.086 0.088 -0.017 0.154** 9.374 8.68 4.42 3.92 (0.348) (0.107) (0.143) (0.020) (0.078) Fuller (α = 1) -0.288 -0.081 0.082 -0.018 0.149** 9.374 10.89 9 7.49 (0.329) (0.101) (0.135) (0.019) (0.073) Fuller (α = 4) -0.262 -0.071 0.067 -0.019 0.136** 9.374 10.89 9 7.49 (0.282) (0.087) (0.116) (0.016) (0.062) Notes: Controls are size of household, rural household, education of mother, log total household expenditure. Standard errors robust to village level clustering in parenthesis. *** p <0.01, ** p <0.05, * p <0.1 In terms of exogeneity of the instruments, as those I am employing are time (of migration) and village specific, one may be concerned that some villages have unobserved characteristics that simultaneously affect the share of household income devoted to a certain commodity and the destination, or the wages at destination, of the migrants. For example, suppose a village has a very good nursing school, women there will mainly be nurses and thus obtain, when they migrate, a certain specific wage. Yet the fact that the 20 . When errors are normally distributed and instruments are fixed, the Fuller-k with α = 1 is best unbiased to second order (Rothenberg 1984). While the critical values for the F statistic are not significantly lower than those for TSLS, this estimator has been proved to yield more precise estimates than both TSLS and LIML when instruments are weak. 25 village hosts a good nursing school will presumably be correlated to a higher taste for expenditure on health. To rule out similar possibilities, I regress a number of village specific characteristics on the instruments to check that there is no correlation between the two. The results are reported in table 13 and show no clear pattern of correlations that give enough confidence about the exogeneity of the instruments. Table 13. Robustness Checks: Exogenous Instruments. (1) (2) (3) (4) (5) (6) (7) % of junior number distance to distance households elementary health high of coach to with schools posts schools midwives station post office electricity E ( wd d f ) / E ( wm ) -5.901 -0.599 0.316 -0.115 0.950 23.984 -0.950 (4.430) (0.499) (0.283) (0.106) (1.233) (21.131) (1.940) d d Var (w ) / Var ( w ) f m 3.782** 0.293 -0.336** 0.015 0.264 -11.480 -1.190 (1.815) (0.226) (0.163) (0.048) (0.540) (8.781) (0.872) Observations 12,217 12,766 12,686 8,796 12,190 9,229 10,183 R2 0.0188 0.0046 0.0217 0.0056 0.0155 0.0105 0.0299 Notes: Unit of observation is household. Standard errors robust to village level clustering in parenthesis. *** p <0.01, ** p <0.05, * p <0.1 As a last check I performed some Montecarlo simulations to assess the robustness of the estimates. I have drawn 1,000 random samples, adequately calibrated to reproduce the correlation between endogenous and outcome variable of the real sample, and I estimated the TSLS and the LIML coefficients for each drawn. Table 13 reports the averages and standard deviations of the coefficients estimated, together with the corresponding F statistic, the p-value of the test of over identifying restrictions and the confidence interval corresponding to the test of equality between the coefficient estimated from the simulated sample and the ones estimated from the real sample and reported in table 12. The coefficients estimated are very similar to those of table 11, I do not reject that they are equal in more than 95% of the cases (coverage). Moreover, the F statistic is now systematically larger than 10, which is above the 20% level of the Stock and Yogo (2002) critical values for the TSLS, and above the 10% one for the LIML estimator. Finally, the Hansen J test of overidentifying restrictions leaves us little doubt about the possibility that the instruments are not exogenous. 26 Table 14. Robustness Checks: Montecarlo Simulations. Shares of household income spent on: (1) (2) (3) (4) (5) Food Education Health Durables Adult OLS -0.104 -0.011 -0.020 -0.032 0.062 (0.079) (0.033) (0.028) (0.066) (0.016) TSLS -0.113 -0.084 0.068 -0.042 0.14 (0.348) (0.141) (0.124) (0.283) (0.075) Coverage [0.969] [0.964] [0.964] [0.967] [0.951] F statistic 10.669 10.723 10.719 10.654 10.624 Hansen J Statistic - p value 0.502 0.503 0.502 0.993 0.504 LIML -0.113 -0.092 0.074 -0.043 0.146 (0.393) (0.164) (0.139) (0.319) (0.084) Coverage [0.968] [0.965] [0.968] [0.966] [0.956] F statistic 10.669 10.723 10.719 10.654 10.624 Hansen J Statistic - p value 0.504 0.506 0.505 0.957 0.508 Notes: Standard deviations in parenthesis. Number of iterations = 1,000. Coverage is the frequency with which the hypothesis of equality between coefficient estimated from actual sample and coefficient estimated from simulated sample has not been rejected. Controls included. Standard errors robust to village level clustering. VI. CONCLUDING REMARKS This paper explores the effects of parental migration on investments on children left behind. The main concern is that migration of one of the spouses may be associated with a shift of resources away from the children due to a moral hazard problem: as the migrant loses the ability to observe the behavior of the spouse left behind, this creates for the latter incentives to shift resources away from the common goods, which include investments on children, onto the private ones. I showed that in a model in which first the two spouses cooperatively choose whether and who should migrate, then the migrant decides how much to send back home in the form of remittances and, finally, the spouse left behind chooses how to allocate the available budget within the household, the subgame perfect Nash equilibrium is one in which the share of total income devoted to children is the same no matter which of the parents migrates. This is because remittances act as a device in the hands of the migrant to offset the decisions of the spouse left behind. I tested the predictions of this model on data from Indonesia, where female migration is particularly widespread. To account for endogenous selection into migration, I exploited information about previous migrants’ earnings and built measures of expected returns and risks of, respectively, male and female migration to use as instrumental variables. I thus showed that households prefer to send away the member who is expected to earn more upon migration and whose earnings will be less volatile. The estimation results revealed 27 that the share of total income devoted to children-related expenditure does not change significantly between the case in which the father migrates and that in which it is the mother that leaves. On the other hand, the difference in the share of total household income devoted to private adult consumption between the case in which the mother migrates and that in which it is the father is positive and reflects the difference in tastes for investment on children. This would be about 15 percentage points. The findings of this paper indicate that, at least through the channel of household expenditure, mother’s migration has no detrimental effects on children compared to father’s migration as long as the migrant mother has the possibility of compensating the shift of resources away from children by sending more remittances back home. As this comes at the expenses of the mother’s own private consumption, it may be desirable to improve the control of migrants over the money they send back home, for example by allowing them to directly pay for children’s schools and cares. A similar policy would reduce the moral hazard problem that arises when one of the spouses migrates and generate an allocation of the household budget closer to the first best one. Moreover, if migrants were better able to control the allocation of the remittances they send back home, more people, especially women, would be willing to migrate to increase the household budget and, through this, investment on children. Clearly, parental migration may well affect children’s wellbeing through channels that go beyond the financial resources spent on their health or education. While quantifying these “non-monetary” effects of parents’ migration on children is beyond the aim of this paper, in table 15 I estimated the direct effects of mother’s versus father’s migration on several indicators of children welfare, which I split into health, education, and work measures. These estimates reveal no differences in the effects on conventional anthropometric measures, nor on cognitive test scores. On the other hand, there appears to be a significant drop in the amount of time children spend at school, mirrored by an increase in the time they spend on house work when the children’s mother migrates. This disruption effect may have long lasting consequences on human capital accumulation that are not observable in the short term (as, for example an increase in grade repetition as documented by Cortes [2015]). It may therefore be desirable to provide more assistance to households of migrant mothers in terms of housework aid and school support for children. Further research should investigate the long term effects on children’s human capital so as to provide a broader assessment of the effects of the feminization of migration. This would also require some analysis of the behaviors of female migrants upon return to their country of origin and to their household. The experience acquired by these women during their migration spell may, indeed, induce significant changes in the household’s decision making process and thus on children’s welfare. 28 Table 15. Effects on Children. IFLS 2007. (1) (2) (3) (4) (5) (7) (6) (8) Hours Hours Hours Weight Height Weight Hours at work Cognitive work for work for height for age for age school family test score wage househol z-score z-score z-score weekly business weekly d weekly weekly OLS Migrant mother -0.008 0.057 0.017 -3.699** -0.014 0.063 0.382 0.144 (0.034) (0.113) (0.108) (1.468) (0.023) (0.437) (0.655) (0.730) Age 0.083*** -0.018* -0.020** 1.260*** 0.022*** 0.142 0.222*** 0.462*** (0.003) (0.010) (0.009) (0.252) (0.003) (0.094) (0.069) (0.108) Female 0.035 -0.307*** -0.038 0.939 0.032* 0.222 0.640 1.616*** (0.032) (0.098) (0.087) (1.225) (0.019) (0.433) (0.662) (0.543) Controls y y y y y y y y Observations 610 611 610 348 439 404 404 404 R2 0.546 0.042 0.033 0.110 0.173 0.013 0.049 0.082 TSLS Migrant mother -0.139 0.511 -0.070 -15.505** 0.010 0.605 0.203 6.406** (0.106) (0.373) (0.370) (6.409) (0.078) (1.199) (1.932) (3.177) Age 0.084*** -0.019* -0.014 1.174*** 0.022*** 0.151 0.226*** 0.474*** (0.003) (0.012) (0.010) (0.279) (0.003) (0.100) (0.072) (0.121) Female 0.027 -0.327*** -0.079 1.789 0.036* 0.183 0.713 1.367** (0.034) (0.104) (0.087) (1.370) (0.020) (0.517) (0.649) (0.589) Controls y y y y y y y y Observations 566 567 566 324 404 378 378 378 R-squared 0.537 0.009 0.027 -0.081 0.166 0.009 0.049 -0.202 Uncentered R2 0.769 0.0334 0.0595 0.710 0.923 0.0200 0.0741 0.00585 F statistic 22.57 22.59 22.57 10.62 21.23 11.49 11.49 11.49 Notes: Controls are size of household, rural household, mother’s years of education, log total household expenditure. Standard errors robust to household level clustering in parenthesis. *** p <0.01, ** p <0.05, * p <0.1. 29 References Altonji, J. G., and D. Card. 1991. “The Effects of Immigration on the Labor Market Outcomes of Less-skilled Natives,” in “Immigration, Trade and the Labor Market” NBER Chapters, National Bureau of Economic Research, Inc., pp. 201–34. Ashraf, N. 2009. “Spousal Control and Intra-household Decision Making: An Experimental Study in the Philippines,” American Economic Review 99 (4): 1245–77. Attanasio, O., and V. Lechene. 2002. “Tests of Income Pooling in Household Decisions,” Review of Economic Dynamics 5 (4): 720–48. Badan Pusat Statistik. 2007. “Welfare Statistics 2007,” National Socio Economic Survey Katalog BPS: 4101002, Jakarta. Bartel, A. P. 1989. “Where Do the New U.S. Immigrants Live?,” Journal of Labor Economics, 1989, 7 (4): 371–91. Blomquist, S., and M. Dahlberg. 1999. “Small Sample Properties of LIML and Jackknife IV Estimators: Experiments with Weak Instruments,” Journal of Applied Econometrics 14 (1): 69–88. Borjas, G. J. 1994. “The Economics of Immigration,” Journal of Economic Literature 32 (4): 1667–717. ———. 1999. “The Economic Analysis of Immigration,” in O. Ashenfelter and D. Card, eds., Handbook of Labor Economics, Vol. 3 of Handbook of Labor Economics, Elsevier, chapter 28, pp. 1697–760. Card, D. 1990. “The Impact of the Mariel Boatlift on the Miami Labor Market,” Industrial and Labor Relations Review 43 (2): 245–57. Chen, J. J. 2013. “Identifying Non-cooperative Behavior among Spouses: Child Outcomes in Migrant-Sending Households,” Journal of Development Economics 100 (1): 1–18. Chiappori, P. A. 1992. “Collective Labor Supply and Welfare,” Journal of Political Economy 100 (3): 437–467. Chiappori, P. A., and O. Donni. 2009. “Non-unitary Models of Household Behavior: A Survey of the Literature,” IZA Discussion Papers 4603, Institute for the Study of Labor (IZA) 2009. Cortes, P. 2015. “The Feminization of International Migration and its Effects on the Children Left Behind: Evidence from the Philippines,” World Development 65 (C): 62– 78. 30 Cox-Edwards, A., and M. Ureta. 2003. “International migration, remittances, and schooling: evidence from El Salvador,” Journal of Development Economics 72, 429–61. Dai, J., S. Sperlich, and W. Zucchini. 2011. “Estimating and Predicting Household Expenditures and Income Distributions,” MAGKS Papers on Economics 201147. de la Briére, B., E. Sadoulet, A. de Janvry, and S. Lambert. 2002. “The Roles of Destination, Gender, and Household Composition in Explaining Remittances: an Analysis for the Dominican Sierra,” Journal of Development Economics 68 (2): 309–28. Deaton, A. 1989. “Looking for Boy-Girl Discrimination in Household Expenditure Data,” The World Bank Economic Review 3 (1): 1–15. Duflo, E., 2003. “Grandmothers and Granddaughters: Old-Age Pensions and Intrahousehold Allocation in South Africa,” World Bank Economic Review 17 (1): 1–25. Gertler, P., D. I. Levine, and M. Ames. 2004. “Schooling and Parental Death,” The Review of Economics and Statistics 86 (1): 211–25. Hanson, G. H., and C. Woodruff. 2003. “Emigration and Educational Attainment in Mexico,” Technical Report, University of California San Diego. Hoddinott, J., and L. Haddad. 1995. “Does Female Income Share Influence Household Expenditures? Evidence from Cote d’Ivoire,” Oxford Bulletin of Economics and Statistics 57 (1): 77–96. Kremer, M., and S. Watt. 2006. “The Globalization of Household Production,” Working Paper 2008-0086, Weatherhead Center for International Affairs, Harvard University. Lafortune, J., and J. Tessada. 2012. “Smooth(er) Landing? The Dynamic Role of Networks in the Location and Occupational Choice of Immigrants,” Working Papers ClioLab 14, EH Clio Lab. Instituto de Economía. Pontificia Universidad Católica de Chile. Lan, P. C. 2006. Global Cinderellas: Migrant Domestics and Newly Rich Employers in Taiwan, Duke University Press. Lauby, J., and O. Stark. 1988. “Individual Migration as a Family Strategy: Young Women in the Philippines,” Population Studies 42, 473–486(14). Levhari, D., and O. Stark. 1982. “On Migration and Risk in LDCs,” Economic Development and Cultural Change 31 (1): 191–96. Lucas, R. E. B., and O. Stark. 1985. “Motivations to Remit: Evidence from Botswana,” Journal of Political Economy 93 (5): 901–18. 31 Lundberg, S. J., R. A. Pollak, and T. J. Wales. 1997. “Do Husbands and Wives Pool Their Resources? Evidence from the United Kingdom Child Benefit,” Journal of Human Resources 32 (3): 463–80. Mansuri, G. 2006. “Migration, School Attainment, and Child Labor: Evidence from Rural Pakistan,” Policy Research Working Paper Series 3945, The World Bank. Markowitz, H. M. 1952. “Portfolio Selection,” The Journal of Finance 7 (1): 77–91. McKenzie, D. J., and N. Hildebrandt. 2005. “The Effects of Migration on Child Health in Mexico,” Economia, Journal of the Latin American and Caribbean Economic Association 0, 257–89. Parrenas, F. S., Servants of Globalization: Women, Migration, and Domestic Work, Stanford University Press, 2001. Patel, K., and F. Vella. 2013. “Immigrant Networks and Their Implications for Occupational Choice and Wages,” The Review of Economics and Statistics 95 (4): 1249– 77. Qian, N. 2008. “Missing Women and the Price of Tea in China: The Effect of Sex- Specific Earnings on Sex Imbalance,” The Quarterly Journal of Economics 123 (3): 1251– 85. Rothenberg, T. J., “Approximating the distributions of econometric estimators and test statistics,” in Z. Griliches and M. D. Intriligator, eds., Handbook of Econometrics, Vol. 2 of Handbook of Econometrics, Elsevier, 1984, chapter 15, pp. 881–935. Stock, J. H., and M. Yogo. 2002. “Testing for Weak Instruments in Linear IV Regression,” NBER Technical Working Papers 0284, National Bureau of Economic Research, Inc. Stock, J. H., J. H. Wright, and M. Yogo. 2002. “A Survey of Weak Instruments and Weak Identification in Generalized Method of Moments,” Journal of Business & Economic Statistics 20 (4): 518–29. Suradji, 2004. “The Policy of Placement of Migrant Workers and Its Problems: Case Study of Indonesia,” Technical Report, The Research and Development Center for International Administration. Thomas, D. 1990. “Intra-Household Resource Allocation: An Inferential Approach,” Journal of Human Resources 25 (4): 635–64. Thomas, D., F. Witoelar, E. Frankenberg, B. Sikoki, J. Strauss, C. Sumantri, and W. Suriastini. 2012. “Cutting the costs of attrition: Results from the Indonesia Family Life Survey,” Journal of Development Economics 98 (1): 108–23. 32 Wooldridge, J. M. 2010. Econometric Analysis of Cross Section and Panel Data, The MIT Press. Yang, D. 2008. “International Migration, Remittances and Household Investment: Evidence from Philippine Migrants’ Exchange Rate Shocks,” Economic Journal 118 (528): 591–630. 33 Supplemental Appendix This Appendix provides supplemental material for Rizzica ( ) "When the cat’s away... The effects of spousal migration on investments on children", World Bank Eco- nomic Review. S. . Corner solution of SPNE When: d E (wf ) β m ≤ () E (wh ) 1−β the migrant will not be able to send remittances because her income at destination is not high enough to compensate for her selfishness (β ). In this case a corner solution arises, in which: R∗ = 0 ∗ m Xm = α E (wh ) ( ) ∗ f Xf =E (wd ) Z ∗ = (1 − α) E (wh m ) Comparing ex-ante the levels of utility achievable in each scenario, the following con- ditions are derived: Proposition S. When the spouses cooperatively choose the migration portfolio that ex ante maximizes their joint expected utility, and the expected wage of the migrant is “relatively low”: . Female migration is preferred to male migration if: α−β d E (wf ) αα (1 − α)2−α−β ≤ ( ) d ) E (wm β β (1 − β )2−α−β . Female migration is preferred to no migration if: β d E (wf ) β β (2 − α − β )2−α−β ≥ ( ) E (wh ) (1 − α)2−α−β . Male migration is preferred to no migration if: α d E (wm ) αα (2 − α − β )2−α−β ≥ ( ) E (wh ) (1 − β )2−α−β As in the case of the internal solution, the main driver of the choice is the expected gain from migration. The conditions above are derived for the case in which the expected wage in the home village is the same for men and women. In this case migration arises when the expected wage of the migrant is sufficiently larger than that of the spouse left behind. If it is equal or smaller, the spouses prefer no one to migrate. If the expected wage of the woman and the man upon migration is the same, the spouses prefer the more selfish of the two to migrate. This happens because in this case the allocation of resources to the public good is entirely in the hands of the spouse left behind. Figure S. shows the optimal choices for varying values of α (horizontal axis) and β (vertical axis). Migration with no remittances is preferred to no migration for very high values of β (female migration) and α (male migration). Indeed, the more the the migrant is selfish, the less she would contribute to the common good if she was in the household. Yet, if both are very selfish migration is less likely to be optimal because the spouse left behind would allocate too little resources to the common good. In terms of allocations, the difference between the share of expenditure for private adult good between the case of female and male migration is α − β (β − α for common goods). On the other hand, the difference in the share of total household income devoted d d h h to private adult goods will be equal to α − β only if E (wf ) = E (wm ) and E (wm ) = E (wf ). Figure S. : Optimal migration portfolios, by parental preferences. 1.0 Female migration 0.8 No migration 0.6 Male 0.4 migration 0.2 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Notes: E (wf d ) = E (w d ) = 1.1 E (w h ) = 1.1 E (w h ). Horizontal dashed m m f lines area is where Female migration is preferred to no migration. Vertical dashed lines area where Male migration is preferred to no migration. Dot- ted area is where Female migration is preferred to Male migration. Large horizontal dashed area is where condition ( ) holds, large vertical area is the equivalent for the case of male migration. S. . Supplementary tables Table S. : Descriptive statistics, migrants. IFLS . () ( ) ( ) ( ) ( ) Migrants without with children with children Difference Difference children along left behind ( )-( ) ( )-( ) Men Age . . . - . *** -. ** ( . ) ( . ) ( . ) Years of education . . . . *** . ** ( . ) ( . ) ( . ) Married . . . - . *** . ( . ) ( . ) ( . ) Rural . . . - . - . ( . ) ( . ) ( . ) Size of household . . . - . * - . *** ( . ) ( . ) (. ) Risk aversion . . . . . ( . ) ( . ) (. ) N Women Age . . . - . *** . ** ( . ) ( . ) ( . ) Years of education . . . . *** . ** ( . ) ( . ) ( . ) Married . . . - . *** . ( . ) ( . ) ( . ) Rural . . . - . *** - . ( . ) ( . ) ( . ) Size of household . . . - . - . *** ( . ) ( . ) (. ) Risk aversion . . . . . ( . ) ( . ) ( . ) N Notes: Standard deviations reported in parenthesis. Differences computed with standard errors clustered at village level. *** p< . , ** p< . , * p< . Table S. : Differences in returns to human capital investments () ( ) ( ) ( ) ( ) ( ) Difference Difference Difference βh β md βf d ( )-( ) ( )-( ) ( )-( ) . . . - . . . *** ( . ) ( . ) ( . ) [336] [333] [335] Notes: Standard deviations in parenthesis. Returns to years of schooling are computed from Mincerian equation with age and age2 as controls. β h are the returns to education observed in the home village; β md are the returns to ed- ucation observed in the counterfactual mens destinations; β f d are the returns to education observed in the counterfactual womens destinations. Sample re- stricted to households included in main regressions and sample size indicated in brackets. *** p< . , ** p< . , * p< . Table S. : Panel household fixed effects regression: all households with children, IFLS -IFLS . () ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Food Food Education Education Health Health Durables Durables Adult Adult Migrant Parent - . * - . . . . . ** - . *** - . *** - . *** - . *** ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) Migrant Mother - . * - . - . . ** . *** ( . ) ( . ) ( . ) ( . ) ( . ) Size of Household . ** . ** - . - . - . * - . * - . - . . . ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) Controls y y y y y y y y y y Year/village fixed effects y y y y y y y y y y Observations , , , , , , , , , , R-squared . . . . . . . . . . Number of id , , , , , , , , , , Notes: Controls are rural household, average years of education, log total household expenditure. Standard errors robust to village level clustering in parenthesis. *** p< . , ** p< . , * p< . Table S. : Three stage model: all households with children. IFLS . Shares of Total Household Income spent on: () ( ) ( ) ( ) ( ) ( ) ( ) Migrant parent Migrant mother Food Education Health Durables Adult Migrant Mother . . . - . . ( . ) (. ) (. ) ( . ) ( . ) Size of Household - . *** . - . - . - . . - . ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) ( . ) RA . * ( . ) d d E (wf )/E (wm ) . *** - . ( . ) ( . ) d d V ar(wf )/V ar(wm ) - . *** . ( . ) ( . ) IMR - . . . . - . . ( . ) (. ) ( . ) (. ) ( . ) ( . ) Controls y y y y y y y Observations , R-squared . - . -. - . - . - . Pseudo R . Uncentered R . - . - . - . -. F Statistic . . . . . Hansen J Statistic . . . . . p-value . . . . . Notes: Column ( ) is probit regression for selction into migration. Column ( ) is the first stage of columns ( )-( ), IMR is the Inverse Mills ratio calculated from column ( ) regression. The procedure follows ?. Controls are rural household, average years of education, log total household expenditure. Standard errors robust to village level clustering in parenthesis. *** p< . , ** p< . , * p< .