Migration, remittances and poverty in Ecuador

Transcription

1 Migration, remittances and poverty in Ecuador Simone Bertoli a and Francesca Marchetta a a CERDI, University of Auvergne and CNRS July 6, 2012 Abstract We analyze the impact of the recent wave of international migration on poverty in Ecuador using propensity score matching. We draw the data from a survey providing detailed information on migrants which is crucial for the estimation. The results evidence a limited impact of migration on poverty among migrant households, while the estimates obtained when focusing on recipient households reveal a significantly larger and more robust impact on poverty. We explore the factors that can account for this divergence in the results. Our analysis further strengthens the choice of defining migration as the treatment of interest. Keywords: remittances; household-level data; poverty; propensity score matching; Ecuador; Latin America. The authors are grateful to Sascha O. Becker, Simon Cueva, Francesca Francavilla, Mihails Hazans, José Hidalgo Pallares, Mauricio León, Jeannette Sanchez, Alessandra Venturini at to the participants to the Conference on Transnationality of Migrants, Riga, to the First Conference on Migration and Development, Lille, organized by the World Bank and the Agence Française de Développement, and to a seminar presentation at the IAB for their comments; they also gratefully acknowledge the contribution of Marco Quinteros, the director of the INEC, who allowed them to get access to the data; Simone Bertoli thanks Vicente Albornoz and Osvaldo Hurtado for the two months that he spent as visiting fellow at CORDES in Quito; the usual disclaimers apply. Centre d Études et de Recherches sur le Développement International, University of Auvergne, Bd. François Mitterrand, 65, F-63000, Clermont-Ferrand; simone.bertoli@udamail.fr, francesca.marchetta@udamail.fr (corresponding author), phone: , fax:

2 1 Introduction Human mobility across borders is driven by some irresistible forces such as enduring economic and demographic imbalances which suggest that it will further increase in scale (Pritchett, 2006), notwithstanding the restrictive policies adopted by destination countries. This implies that it is crucial to understand whether the optimistic expectations about the impact of international migration and of the ensuing flows of remittances on the countries of origin are solid-grounded. 1 Adams and Page (2005) rely on aggregate data to provide evidence that a 10 percent increase in the share of international migrants reduces the number of individuals at origin living below the $1 a day poverty line by 2.1 percent. Similar evidence has been provided by Acosta, Calderón, Fajnzylber, and López (2006, 2008) about the impact of migrants remittances on poverty in Latin American countries, and by Yang and Martinez (2007) for the Philippines using household-level data. 2 This paper analyzes the impact of migration upon the incidence of income poverty in Ecuador, a country which experienced an unprecedented wave of international migration induced by a severe economic and financial crisis at the end of the 1990s (Bertoli, Fernández- Huertas Moraga, and Ortega, 2011). We focus on the impact of migration rather than of remittances on poverty, following the recommendation by McKenzie and Sasin (2007). The groups of migrant and of recipient households largely overlap, but do not coincide: 3 international migration represents a household-level strategy which is not risk-free for the stayers, as migrants can experience spells of unemployment at destination that prevent them from sending back a steady flow of remittances. The share of households with at least one migrant that do not receive remittances stands, for instance, at 41 percent for Nicaragua in Barham and Boucher (1998) and at 37 percent for Guyana in Agarwal and Horowitz (2002). 4 We resort to propensity score matching to identify the effect of migration on poverty, as 1 As the development community continues the search for additional resources to finance the Millennium Development Goals, remittances - pro-poor and cyclically stable, compared to other capital flows - appear to be a promising source. (Maimbo and Ratha, 2005, p. 2). 2 The literature focuses on stayers, and this implies that these estimates do not account for the influence of migration on the poverty status of the migrants themselves, as discussed by Clemens and Pritchett (2008) and Clemens (2011). 3 We use in this paper the expression migrant households to denote the households which sent at least one of their members abroad. 4 Alcaraz, Chiquiar, and Salcedo (2012) provide evidence of the dramatic impact of the recent economic crisis in the United States on the share of recipient Mexican households. 2

3 we lack a credible instrument for migration, 5 and as this estimation technique does not require to introduce assumptions on the functional form of the relationship between household characteristics, migration and poverty. We test the sensitivity of our estimates to possible departures from the identifying assumption following Becker and Caliendo (2007). We draw the data from the the Encuesta Nacional de Empleo, Desempleo y Subempleo, ENEMDU, conducted by the Ecuadorian National Statistical Institute in December This labor market survey allows us to derive income-based definitions of poverty and it provides detailed information on migrant members, including the year of migration, age, gender, education and the amount of remittances sent to their household over the previous 12 months. The availability of data on migrants represents a key value added of this survey, as our analysis exploits the information on the timing of the various migration episodes, the information on the individual characteristics of the migrants, and the length of the recall period over which we have data on remittances for migrant households. Specifically, the ability to identify migrant and not only recipient households allows us to follow the analytical recommendation by McKenzie and Sasin (2007), and the information on the timing of the migration episodes allows us to define our treatment variable as having sent at least one member abroad after 1998, the year which marks the start of the Ecuadorian crisis (Beckerman and Cortés-Douglas, 2002; Jácome, 2004). This allows us to greatly reduce the heterogeneity of the impact on poverty which would arise if the group of treated households was based on migration episodes which are spread in time, possibly over several decades, and it also reduces the strength of the effects of migration on the supply (Mishra, 2007) or on the demand side (Woodruff and Zenteno, 2007; Piracha and Vadean, 2010; Marchetta, 2012) of the labor market at origin which can indirectly influence non-migrants. Our empirical analysis reveals that the recent Ecuadorian migration produced a limited impact on poverty among migrant households, which experienced a decline in the incidence of poverty by at most 20.8 percent, which appears to be sensitive to violations of the hypothesis of selection on observables. The ENEMDU 2005 survey also allows us to identify recipient households independently from the questions related to migrants, as the income section of the questionnaire provides 5 Calero, Bedi, and Sparrow (2009) use the number of Western Union branches in each Ecuadorian province as an instrument in their analysis of the impact of remittances on school enrollment and child work, and Acosta, Calderón, Fajnzylber, and López (2007, 2008) rely on the interaction between the share of recipient households at the county level and the number of adult household members. 3

4 information related to the receipt of remittances from abroad over the month before the survey. When, as the literature does, 6 we focus on recipient households, we find sharply different results, as the receipt of remittances is estimated to induce a decline in poverty among recipient households by nearly 60 percent, with this result also appearing to be notably robust to possible violations of the identifying assumption. Three key factors contribute to explain this significant divergence in the results; specifically, the definition of migration rather than of the receipt of remittances as the treatment of interest allows us to (i) account for the non-negligible share of non recipient migrant households, (ii) measure remittances over a longer recall period, and (iii) rely on the reported characteristics of the migrants rather than introduce assumptions on them. We will argue that these factors strengthen the case for focusing on migrant rather than on recipient households, as the latter analytical choice would actually lead to unduly optimistic conclusions on the impact of international migration on poverty in Ecuador. This paper is related to four different strands of literature. First, it is related to the papers that analyze the impact of migration and remittances on poverty at origin with either macro (Adams and Page, 2005; Acosta, Fajnzylber, and López, 2007) or micro data (Yang and Martinez, 2007; Acosta, Calderón, Fajnzylber, and López, 2006, 2007, 2008; de la Fuente, 2010). 7 Second, it is also related to the papers that analyze the determinants of migrants selection (Chiquiar and Hanson, 2005; McKenzie and Rapoport, 2010; Bertoli, 2010) as the factors influencing international migration decisions also are likely to shape the outcomes of international migration and remittances (Taylor, 1999, p. 64). Third, it is connected to the vast literature on propensity score matching originating from the seminal contribution by Rosenbaum and Rubin (1983), and in particular to the papers that deal with the use of sampling weights in the estimation (Frölich, 2007; Zanutto, 2006) and with departures from the identifying assumptions (Becker and Caliendo, 2007; Ichino, Mealli, and Nannicini, 2008; Nannicini, 2007). Fourth, this paper contributes to the strand of literature analyzing the determinants and the effects of the recent wave of Ecuadorian migration (Beckerman 6 This analytical choice is often constrained by the available data, as household surveys conducted in the countries of origin seldom provide information on migrant members even when they allow to identify recipient households; for instance, only two of the ten surveys used by Acosta, Calderón, Fajnzylber, and López (2008) gathered data on migrants. 7 See Adams (2011) for a survey of the analytical challenges posed by the use of household-level data to analyze the impact of international migration on migrant-sending countries. 4

5 and Cortés-Douglas, 2002; Jácome, 2004; Gratton, 2007; Calero, Bedi, and Sparrow, 2009; Bertoli, 2010; Bertoli, Fernández-Huertas Moraga, and Ortega, 2010, 2011). The rest of the paper is structured as follows: Section 2 briefly describe the propensity score matching technique, and Section 3 discusses its implementation to the analysis of the impact of international migration upon income poverty in Ecuador. The source of our data and the relevant descriptive statistics are presented in Section 4. Section 5 provides the results from the empirical analysis, and it explores the difference between the results obtained when focusing on migrant or on recipient households, and Section 6 draws the main conclusions. 2 Propensity score matching Assume that units in a sample can be either subject to a treatment, z i = 1, or not, z i = 0; let T denote the subsample of treated units, and U the subsample of untreated units. An outcome variable y can be influenced by the assignment to treatment; specifically, y i1 is the value of the outcome variable y when z i = 1, and y i0 represents its value when z i = 0. The observed value of the outcome variable y i is related to its potential outcomes by the observation rule: y i = z i y i1 + (1 z i )y i0 The treatment effect on the unit i is defined as τ i = y i1 y i0. The average impact of the treatment on the subsample T of treated units, ATET, can be defined as: E T (τ i z i = 1) = E T (y i1 z i = 1) E T (y i0 z i = 1) (1) where E T denotes the average on treated units only. The observational rule for y i precludes the estimation of the ATET, as y i0 is not observed when z i = 1. In a experimental setting, where randomization ensures that there is no systematic difference between treated and untreated units, we have that E U (y i0 z i = 0) = E T (y i0 z i = 1), so that observed outcomes for the untreated units can substitute for the unobserved outcomes y 0 for treated units. With non-experimental data, this does not hold true because the assignment to the treatment can be influenced by a vector x of covariates which also have an impact on the outcome variable y. 5

6 Assume that the vector x includes all covariates which have a simultaneous influence on the treatment and on the outcome, so that the potential outcome y 0 is independent from the assignment to treatment conditional upon x. Formally: y 0 z x (2) where the symbol denotes statistical independence. Condition (2) implies that assignment to treatment is random conditional upon the vector of covariates. Let f(x) represent the probability of assignment to treatment, i.e. f(x) = Pr(z = 1 x), which is also called the propensity score. The seminal contribution by Rosenbaum and Rubin (1983) demonstrates that if the probability of assignment is bounded away from zero, f(x) (0, 1], so that all units in the sample can be exposed to treatment with a positive probability, and (2) holds, then we also have that: y 0 z f(x) (3) The outcome y 0 is independent from the assignment to treatment z conditional upon f(x), as f(x) represents a balancing score which ensures that x z f(x) (Rosenbaum and Rubin, 1983, p. 44). This, in turn, implies that the expected value of the unobserved outcome y 0 for treated units conditional upon f(x) coincides with the expected value of the observed outcome y 0 for untreated units, that is: E T [y i0 z i = 1, f(x) = p] = E U [y i0 z i = 0, f(x) = p] Hence, the ATET can be estimated through an iterative averaging as follows: Ê T (τ i z i = 1) = 1 0 [ ET (y i1 z i = 1, f(x) = p) E U (y i0 z i = 0, f(x) = p) ] g(p)dp (4) where g(p) denotes the distribution of the propensity score f(x) in the subsample T of treated units. The expression in (4) implies that (i) the average outcome for treated and untreated units is estimated at each value of the propensity score, and then (ii) the difference between the two average outcomes is again averaged over the distribution g(p) of the propensity score. 6

7 3 Implementation We discuss here the key steps of the implementation of propensity score matching to the analysis of the impact of migration on Ecuadorian households. 8 The treatment z i is represented by having an household member who moved out of Ecuador, while the outcome variable y i is given by the income poverty status of the household. 3.1 Selection of the covariates The first step is related to the identification of the relevant variables to be included in the vector x of covariates which is used to estimate the propensity score. This vector has to include only variables that have a simultaneous influence on the probability to have a migrant member and upon the poverty status of the household. Variables which only have an impact on the probability to migrate should not be included, as the objective of the estimation of the propensity score is not to maximize the fit of the model (Caliendo and Kopeinig, 2008). The inclusion in the vector of covariates of variables which only have an influence on the treatment would actually reduce the ability of the estimated propensity score to serve as a balancing score, and it would also worsen the overlap of the distributions of the propensity score for migrant and non-migrant households. The effect of migration on poverty is identified under the assumption that, conditional upon x, the decision to migrate is not systematically related to the poverty status of Ecuadorian households. As this is a strong identifying assumption, we discuss below in Section 3.6 how to assess the sensitivity of our estimates to violations from the assumption of selection on observables following Becker and Caliendo (2007) Post-treatment measurement of the covariates The ENEMDU survey does not provide retrospective information which could have been used to measure relevant household characteristics in 1998, before some households got exposed to the treatment. Still, a distinctive feature of the December 2005 wave of the survey is that it provides detailed information on migrant members, including their age, gender and education. This implies that all the household characteristics that we included in the vector 8 Caliendo and Kopeinig (2008) provide an excellent survey of the analytical choices involved in the implementation of this estimation method. 7

8 of covariates are measured post-treatment, but in a way that is not dependent upon the migration status of the household, and this is crucial to ensure the unbiasedness of our estimates. While Rosenbaum and Rubin (1983) write that the analysis should be based only on variables which are measured before the treatment so to avoid any endogeneity with respect to the exposure to the treatment itself, 9 Lechner (2008) actually demonstrates that this requirement can actually be relaxed, specifying the conditions under which the reliance on post-treatment covariate does not bias the estimate of the ATET. One of the two conditions derived by Lechner (2008) requires that the influence of the treatment on the covariates should be non-systematic, so that measuring them after the treatment only induces a measurement error in x. If the distribution of the elements of x depends on the exposure to the treatment, than the estimate of the ATET will be biased. This is why the availability of data on migrant household members in the survey is crucial for the analysis: if migrants are not randomly selected within the household with respect to their gender, age and education, 10 then any measure of the demographic structure of the household or of its average level of education will be endogenous to migration, precluding its inclusion in the analysis. 11 Any demographic event other than migration, such as births, deaths and marriages, can also drive a wedge between the household characteristics measured before or after our reference period, but there is no reason to expect that these events are systematically correlated with migration, so that we can draw on Lechner (2008) to dismiss the concerns related to the post-treatment measure of household characteristics Rosenbaum (1984) represents the first paper which analyzed the implications of the inclusion an endogenous variable in the vector of controls. 10 Bertoli (2010) provides evidence on the determinants of the selection in observables of Ecuadorian migrants, which demonstrates the relevance of gender, age and schooling in influencing the probability to migrate. 11 An alternative approach to reduce the concerns about endogeneity is to introduce explicit assumptions about the migrant members, when these are unobserved; for instance, Acosta, Calderón, Fajnzylber, and López (2008) assume that remittances are sent by an adult male family member, who has the average years of education of other adults in the household (p. 99) in their analysis of the impact of remittances on poverty in Latin American countries. 12 The sudden character of the Ecuadorian crisis which triggered the recent wave of migration plays a key role in our analysis, as it allows us to credibly regard education, for both migrants and stayers, as exogenous with respect to the prospect to migrate. 8

9 While the literature often includes variables which relate to the household head, 13 we chose not to do so as household headship can be endogenous to migration, as observed by Cox-Edwards and Rodríguez-Oreggia (2009), and the ENEMDU 2005 does not provide information which could be used to identify the household head in the counterfactual no migration scenario. 14 The Encuesta de Condiciones de Vida conducted by the INEC in reveals that 20.8 percent of the Ecuadorian migrants were household heads before migrating. 15 For similar reasons, we also opted for omitting measures of asset holdings from the vector of covariates, as Bertoli (2010) provides evidence of their endogeneity with respect to the time elapsed since migration Estimation of the propensity score The propensity score f(x) is not known, and it is estimated through a logit model, so that the probability of assignment to the treatment is estimated as: f(x) = 1 + e x β The coefficients of the estimated propensity score do not have a behavioral interpretation (Dehejia and Wahba, 2002), so that they should not be interpreted as reflecting the effect of the elements in the vector x upon the probability to have a migrant member. The functional specification of the estimated propensity score f(x) is only meant to ensure that f(x) acts as a balancing score of the covariates, and this can call for the inclusion of higher-order and interaction terms between the elements of x (Caliendo and Kopeinig, 2008). For the same reason, sampling weights are not used in the estimation of f(x), as this is not meant to support inferences about the whole population from which the sample has been drawn (Zanutto, 2006) See, for instance, Acosta, Calderón, Fajnzylber, and López (2006, 2008) and Calero, Bedi, and Sparrow (2009) for with respect to Ecuador. 14 Binzel and Assaad (2011) propose an approach to side-step the endogeneity of headship which applies to cases where migration episodes are almost exclusively related to only one of the two genders. 15 This survey provides information only on the migrants who left Ecuador after January Specifically, the asset index, which is built only on the basis of the assets which are more likely to reflect past savings (Acosta, 2011, p. 919), is an increasing function of the years passed since the household has sent one of its members out of Ecuador; this might reflect either the positive effect of remittances on household assets or the depletion of assets to cover the monetary costs of migration, or both. 17 Frölich (2007) also argues that sampling weights have not to be used in the estimation of the propensity ex β 9

10 3.3 Check of the balancing property The literature offers different approaches to the necessary evaluation of the ability of f(x) to serve as a balancing score. One can perform a t-test on the null hypothesis of the equality of the mean, conditional on the value of f(x), of each of the elements in x in the groups of treated and untreated units. 18 Still, this approach is exposed to two critiques. First, the balancing property should be verified not on the whole sample of observations, but on the subsample of treated and control units which is used to estimate the ATET, so that the ability of f(x) to serve as a balancing score is closely intertwined with the choice of the matching method (Lee, 2012). Second, Imai, King, and Stuart (2008) regard the reliance on hypothesis testing as a balance test fallacy, as balance is a characteristic of the sample, not some hypothetical population, and so, strictly speaking, hypothesis tests are irrelevant in this context (Imai, King, and Stuart, 2008, p. 497). The logic that underlies hypothesis testing is that there is threshold level below which the imbalance of the covariates can be accepted, while imbalance with respect to observed pre-treatment covariates [...] should be minimized without limit where possible (Imai, King, and Stuart, 2008, p. 497), and parametric methods could be used to adjust for any residual imbalance (Ho, Imai, King, and Stuart, 2007). Hence, in order to evaluate the extent to which matching satisfies the necessary balancing property, we follow Sianesi (2004) and re-estimate the propensity score on the matched sample alone: the difference between the pseudo-r 2 on the unmatched and matched sample gives us a measure of the extent to which the estimated propensity score f(x) effectively balances the covariates. If f(x) balances the covariates in the subsample of treated and control observations, then the logit model should be poorly able to predict assignment to treatment in the subsample. Following the arguments by Imai, King, and Stuart (2008), we also compute the ATET with the adjustment for the residual imbalance of the covariates x proposed by Abadie, Drukker, Herr, and Imbens (2004) as discussed in Section 3.5 below. score when data for treated and untreated units are drawn from the same survey, and this represents a more general necessary condition for a sound application of matching estimators (Heckman, Ichimura, and Todd, 1997). 18 See, for instance, the Stata command pscore by Becker and Ichino (2002) which performs the t-tests within each stratum of the f(x), and it warns the user to modify the specification of the propensity score if the null hypothesis is rejected. 10

11 3.4 Matching methods The propensity score greatly reduces the curse of dimensionality which characterizes matching methods (Caliendo and Kopeinig, 2008), but the exact matching on f(x) which would be required by the estimation of the ATET according to (4) is nevertheless unfeasible, and we need to resort to approximate matching techniques. Specifically, we rely on n-nearest neighbor matching, adjusting the matching technique to account for the sampling weights w i associated to each household in the ENEMDU 2005 following Abadie, Drukker, Herr, and Imbens (2004). Specifically, let w T represent the average sampling weight of migrant households in the sample. With n-nearest neighbor matching, with n 1, each migrant household i is matched with a set C n (i) non-migrant households whose estimated propensity score is nearest to f(x i ), and whose sum of sampling weights is equal to nw T. 19 Matching is performed only on the subsample of treated and untreated units which belong to the common support, defined as the closed subset of the interval [0, 1] where the density of the estimated propensity score f(x) is positive both for migrant and non-migrant households. Restriction to the common support ensures that the set of control households C n (i) does not include non-migrant households which represent a poor match for household i Estimation of the ATET Sampling weights from the ENEMDU 2005 are not used in the estimation of the propensity score, as discussed in Section 3.2 above, while the weights w are used in the estimation of the ATET. This, as described in (4), is the result of an iterative averaging procedure: following Zanutto (2006) and Frölich (2007), sampling weights w are used when we compute the counteractual poverty status y 0 for each migrant household. Specifically, we compute it as: 19 More formally, and assuming no ties in the estimated propensity score, a non-migrant household j belongs to the set C n (i) of controls for the migrant household i if the sum of the sampling weights of the non-migrant households with an estimated propensity score which differs from f(x i ) by less than f(x j ) f(x i ) is smaller than nw T. 20 We do not impose a caliper on the distance between f(x i ) and the propensity score f(x j ) for j C n (i), and rely on the bias-adjustment procedure by Abadie, Drukker, Herr, and Imbens (2004) to correct for the residual imbalance in the covariates that this might induce. 11

12 ŷ i0 = j C n(i) w j y j0 j C n(i) w j Let T represent the subset of migrant household for which the set of matched control units is non-empty, i.e. i T if and only if i T and C n (i). Then, the ATET is estimated as: Ê T (τ i z i = 1) = i T w i(y i1 ŷ i0 ) i T w i The estimation of the impact of migration on poverty following (5) is the outcome of a two-step procedure, and the estimation of the standard error of ÊT (τ i z i = 1) should also reflect the uncertainty which is due to the estimation of the propensity score f(x). Abadie and Imbens (2008) argue that the reliance on bootstrapping to derive the standard error associated to (5) lacks theoretical justifications, and it can fail to produce an unbiased estimate of the true standard error. Hence, following their suggestion, we rely on the analytical standard errors proposed by Abadie and Imbens (2008) and implemented in Stata by Abadie, Drukker, Herr, and Imbens (2004) to derive correct confidence intervals around our point estimates of the impact of international migration on poverty in Ecuador. We follow Abadie, Drukker, Herr, and Imbens (2004) also with respect to the estimation of the ATET through an OLS regression of the observed outcome variable y i on the treatment z i and on the vector x i on the subsample of matched households in order to correct for the residual imbalance in the covariates. 21 (5) 3.6 Sensitivity to departures from selection on observables The identification of the effect of the treatment z through propensity score matching is based on the assumption of selection on observables or of conditional independence, reflected in (3): this requires that any variable that simultaneously influences the probability of selection into treatment and the outcome has to be included in the vector x. The plausibility of the assumption that selection is driven only by observable covariates included in x can be defended on the basis of the relevant theoretical and empirical literature, but it cannot be 21 This choice is in line with McKenzie, Stillman, and Gibson (2010), who observe that among the other non-experimental methods, difference in-differences and propensity score matching with bias-adjustment work best (p. 942) when estimating the income gains from migration. 12

13 tested, as observed data are uninformative about the relationship between the treatment z and the potential outcome y 0. Nevertheless, it is possible to assess the robustness of the estimated ATET with respect to possible violations of the conditional independence assumption, following the approach proposed Becker and Caliendo (2007). 22 Specifically, Becker and Caliendo (2007) assume that the distribution of a binary outcome y 0 conditional on the propensity score f(x) is not independent from the assignment to treatment z, while independence would hold conditional on the propensity score estimated on x plus an unobserved dichotomous variable u: y 0 z f(x, u) (6) This implies that the ATET estimated on the basis of matching on f(x) does not represent the true causal effect of the treatment z upon the outcome y for the treated units, as it also captures the influence on the outcome y on non-random selection in the unobservable u of the treated households. If: ex β+γu f(x, u) = 1 + e x β+γu then, for two households with identical values of the covariates x, we have that the ratio of their actual odds of exposure to the treatment z belongs to the interval [e γ, e γ ]: only if the unobservable u has no impact, i.e. γ = 0, on the probability of exposure to the treatment, then the two observationally identical households have the same probability of exposure to z. Becker and Caliendo (2007) rely on the test statistic proposed by Mantel and Haenszel (1959) to evaluate the effect of the unobserved variable u on the significance of the estimated ATET for different values of e γ, which reflect different assumptions about the possible impact of u upon the probability of exposure to treatment. 22 A similar approach is proposed by Ichino, Mealli, and Nannicini (2008); a different strategy is proposed by Imbens (2004) who suggests to estimate the effect, within the group of controls, of a treatment z that should have no effect on the outcome y under the assumption of selection on observables: for instance, z could be represented by a reported intention to migrate: if, conditional upon the observables x, this treatment z has a significant influence on the incidence of poverty among non-migrant households, then the plausibility of the conditional independence assumption is significantly weakened, and non-random selection in unobservables could represent a serious threat to identification. Regrettably, the ENEMDU 2005 provides information on the intention to migrate only for households who already have a migrant, so that this approach cannot be applied. 13

14 For instance, if we estimate a negative impact of international migration upon the incidence of poverty among Ecuadorian households, it is interesting to test whether this result might reflect a positive selection of migrant households in an unobservable characteristic, such as ability or talent, which is positively correlated with their income generating capacity. The Mantel and Haenszel (1959) test statistic tells us how strong can be the influence of this unobservable u before we are induced not to reject the null hypothesis that the causal effect of international migration upon poverty is actually zero. This test does not tell us whether such a bias due to an unobservable factor does exist, but only how strong such a possible bias would need to be in order to make the estimated results sensitive to departure from the underlying identifying assumption of selection on observables. 4 Data and descriptive statistics This ENEMDU 2005 survey, which was conducted on a sample of 18,357 households that is representative at the provincial level, 23 contains a module which provides information on household members who had moved abroad and were absent at the time of the survey. The data on migrants include age, gender, level of education, year of migration and country of destination; this allows us to identify all Ecuadorians who left after the late 1990s economic crisis, provided that at least one household member was still in Ecuador at the time of the survey and he or she was willing to disclose the information about the migrant. 24 Whole household migration leads to an undercount of recent Ecuadorian migrants, which does not represent a reason for concern given that, as the literature does, we are interested in identifying the impact of migration on the incidence of poverty among stayers. 25 Furthermore, interviewees might have been reluctant to disclose information on migrant members; reassuringly, Bertoli (2010) demonstrates that the observable characteristics of the Ecuado- 23 The survey was conducted in the 21 Ecuadorian in-land provinces; the Galapagos Islands, which host 0.15 percent of the Ecuadorian population, are excluded from the survey. 24 The ENEMDU 2005 suggests that around 250,000 Ecuadorians left the country between 1998 and 2005, while survey data collected in Spain and in the United States reveal that approximately 450,000 Ecuadorians moves to these two destinations (Bertoli, Fernández-Huertas Moraga, and Ortega, 2011), which absorbed approximately 85 percent of total migration flows. 25 The undercount induces us to focus on the impact of migration on poverty among migrant households rather then on its impact on the country-wide incidence of poverty, as the latter is sensitive to the undercounting. 14

15 rian migrants obtained from the ENEMDU 2005 do not differ from those that can be obtained from US or Spanish sources, so that the undercount of the migrants does not pose a threat to identification. We restrict the sample to households that (i) do not have any returnee or foreign-born among their members, and that (ii) do not have a migrant who left before our period of analysis ( ). This ensures that our sample includes only Ecuadorian households with no migration experience before the late 1990s economic crisis. We further restrict the sample to (iii) non migrant households who report not to have received remittances in the month before the survey, 26 outlying data on income. and we exclude from the sample households with missing or The ENEMDU 2005 provides information on labor earnings, both in cash and in kind, and on non-labor incomes, including public and private transfers. The questionnaire contains two distinct questions with respect to remittances: first, all households are asked whether they received remittances from abroad, and the reported amount refers to the same recall period as for all other income sources, namely the month before the survey. Second, the households that report to have at least one migrant member are asked about the amount of remittances sent to them by each migrant over the previous 12 months, and about the number of transfers over which this amount was distributed. Total income for migrant households is defined as the sum of all incomes, excluding remittances, reported for November 2005 plus the average monthly amount of remittances received over the year before the survey. The survey reveals that the migrants who send remittances to their households in Ecuador do on average 7.0 transfers per year, and this implies that the remittances received in the month before the survey can be an imprecise measure of what households receive on average. More specifically, the high and regressive transaction costs that are usually charged on money transfer operations can induce migrants to concentrate their remittances in a limited number of transfers of larger amount, 27 so that the value of a single transfer exceeds the monthly average of the transfers The exclusion of recipient households with no migrants is motivated by the idea that a large share of these households are likely to be mis-reporting with respect to the existence of a migrant; the results of our analysis are unaffected by this sample selection criterion. 27 See Freund and Spatafora (2008) for an analysis of the relationship between transaction costs and remittances. 28 Section explores the implications for the empirical analysis of measuring remittances on the shorter recall period of one month. 15

16 Our sample includes 16,089 households, with 832 households with at least one migration episode between 1998 and Table 1 presents the relevant descriptive statistics. Migrant households have on average 1.35 migrants abroad, with 63.4 and 23.1 percent of them having a migrant in Spain and in the United States respectively. The data reveal that 29.2 percent of the migrant households report not to have received remittances in the 12 months before the survey. We defined the subsample of migrant households as including all households that sent at least one migrant abroad between 1998 and 2005, and this implies that some migration episodes actually occurred close to the time of the survey. If Ecuadorians experience a spell of unemployment upon arrival at destination, then there might be a substantial lag between migration and the time when migrants begin transferring a steady flow of remittances back to their households of origin. 29 Indeed, 54.4 percent of the 92 households who sent their first member abroad in the two years before the survey do not report to receive any remittances, while the corresponding figure for the other migrant households stands at 26.0 percent. This suggests that the share of non recipients among migrant households is non-negligible even when we focus on Ecuadorian households who sent their members abroad at least two years before the survey. The incidence of income poverty, defined on the basis of the poverty line set by the INEC, 30 stands at 20.9 percent for migrant households; this figure is substantially below the 36.1 percent that we obtain when defining poverty on the basis of non-remittance income only. 31 Still, remittances do not represent a revenue which adds up to other exogenous income sources, and they at least partly compensate for the foregone domestic earnings of the migrants, 32 and the labor supply decisions of stayers can be endogenous with respect to migration (Chami, Fullenkamp, and Jahjah, 2005; Amuedo-Dorantes and Pozo, 2006; Hanson, 2007; Cox-Edwards and Rodríguez-Oreggia, 2009; Binzel and Assaad, 2011). This suggests to consider the incidence of poverty among non migrant households to get a first 29 The Encuesta Naciónal de Inmigrantes conducted in Spain in 2007 on a sample of 15,000 foreign-born reveals that 43 percent of Ecuadorian migrants remained unemployed for at least three months after their arrival in Spain. 30 The poverty line was set at $56.60 per month, which corresponds to $3.77 per day at purchasing power parity terms (World Bank, 2012). 31 Early contributions to the literature (Gustafsson and Makonnen, 1993; Leliveld, 1997) relied on the comparison between these two figures to assess the impact of migration and remittances on poverty. 32 If the migrant was the main breadwinner within the household, than the income per capita of the stayers falls after migration. 16

17 sense of the impact of migration on poverty. This, as reported in Table 1, stands at 32.2 percent; this implies that the share of poor households is 35.1 percent lower among migrant than among non migrant households. The descriptive statistics reveal that the two groups of households differ with respect to relevant observable characteristics which are likely to have an impact both on their poverty status and on the probability to have a migrant. Specifically, a smaller share of the households with at least one member who migrated between 1998 and 2005 reside in rural areas, they have a larger household size and a smaller dependency ratio, and their members are better educated. Migrant households have working age members with 9.6 years of schooling, while the corresponding figure for non migrant households stands at 8.4. Unsurprisingly, the households who recently sent one of their members out of Ecuador have also a better connection with migration networks, proxied by the share of households in each county with a migrant to the United States before The variables related to the demographic and schooling characteristics for migrant households can be defined either on the basis of all household members, as we did in Table 1, or on the basis of resident members alone. Table 2 reveals that the exclusion of the data on migrant members blurs most of the differences in observables between the two groups of households; beyond the mechanical impact on household size and on the number of working age members, Table 2 shows that migrants are positively selected with respect to education within the household, as the average number of years of schooling falls from 9.6 to 8.7 when we exclude migrant members, and the share of households with at least one college graduate falls from to percent, with the latter figure coinciding with the one for nonmigrant households reported in Table 1. This reinforces the empirical relevance of the need to have information on the individual characteristics of migrants, and it gives us a picture of Ecuadorian migrants which differs from the one assumed by Acosta, Calderón, Fajnzylber, and López (2008). 34 This, in turn, has relevant implications for the econometric analysis, as the number and level of education of the migrants clearly exerts an influence upon the poverty status of migrant households in the counterfactual scenario with no migration Networks in other destination countries, including Spain, were of limited size before the recent wave of migration. 34 As recalled in Section 3.1.1, Acosta, Calderón, Fajnzylber, and López (2008) assume that each household has one male adult migrant, with the same level of education as stayers. 35 Bertoli, Fernández-Huertas Moraga, and Ortega (2010, 2011) provide evidence on the substantial returns to schooling and to college and to the significant gender gap in wages in Ecuador. 17

18 5 Estimates We first estimate the impact of migration on poverty among migrant households, and then we follow the standard approach in the literature, focusing on recipient households. The last part of this section explores the factors that contribute to explain the significant differences between the two sets of results. 5.1 Migration and poverty The treatment z is represented by having at least one household member who left Ecuador between 1998 and 2005, and the outcome y is represented by the poverty status of the households, defined on the basis of the national poverty line. 36 We retain the following household characteristics in the vector x of covariates which is used to estimate the propensity score f(x): the number of working age members, the dependency ratio, the share of female working age members, the average years of schooling, a dummy indicating if a household member completed tertiary education, a dummy signaling indigenous self-identification, the countylevel size of migration networks, a dummy for residence in rural areas. We also include dummies for the 21 Ecuadorian in-land provinces, to control for all unobservable factors which can simultaneously influence poverty and migration. The estimation of the logit model is done, as discussed in Section 3.2, without using the sampling weights, and the functional specification does not have a behavioral interpretation. The model has an overall goodness of fit, measured by the pseudo-r 2, equal to 0.124, and specification (1) in Table 3 reports the coefficients which are used to estimate the propensity score. The estimated propensity score f(x) is then used to define the subsample of non-migrant households that form the control group, and to estimate the average treatment effect on the treated. Table 4 reports the results obtained with nearest neighbor matching, with a number of matches n ranging from one to ten. The estimation of the propensity score on the subsample of matched households only is characterized by a pseudo-r 2 which is 89.8 to 98.3 percent lower than on the whole sample. The reduction in the pseudo-r 2 is increasing 36 Our results are robust to the adoption of the $2 a day poverty line, in purchasing power parity terms, used by the World Bank for cross-country comparisons in the incidence of poverty; the estimates are available from the authors upon request. 18

19 in the number of matches n, and it already stands at 96.3 percent when n = 3, and this is reassuring with respect to the ability of f(x) to act as a balancing score. Table 4 reports the estimated ATET: for n 3, migration induces a decline in the incidence of income poverty between 2.8 and 3.7 percentage points, a figure that stands substantially below the 11.1 percentage points of the difference between the incidence of poverty among migrant and non-migrant households. This suggests that differences in observable characteristics, which are controlled for through the matching procedure, can account for a relevant part of the lower incidence of poverty among the Ecuadorian households who sent at least one of their member abroad during the recent wave of migration. Furthermore, the null hypothesis that the true impact of migration on poverty is actually equal to zero can be just marginally rejected only for n 6. The estimated impact of migration on poverty increases by percentage points for n 3 when we adopt the regression-based approach to remove the bias due to the residual imbalance of the covariates following Abadie, Drukker, Herr, and Imbens (2004); the estimates for the bias-adjusted ATET range between -4.4 and -5.5 percentage points, and the null hypothesis that the true effect is zero can always be rejected at least at the 5 percent confidence level. The upper bound of this range implies that migration induced a reduction in the incidence of poverty by 20.8 percent among migrant households. 37 These estimates suggest that differences in observable characteristics can explain no less than half of the observed difference in the incidence of poverty between migrant and non-migrant households. 38 What about possible differences in unobservable characteristics? Table 4 reports the results from the Mantel and Haenszel (1959) test statistic; specifically, it reports the highest values of e γ that still allow to reject at the 5 or 10 percent confidence level the null hypothesis that the causal effect of migration is actually zero, and the estimated impact on poverty actually reflects only a positive selection in unobservables of migrant households. For n 3, the estimated ATET is still negative and significant at the 5 percent confidence level for values of e γ ranging between 1.15 and This, in turn, implies that an unobserved 37 The incidence of poverty among migrant households in the counterfactual scenario stands at 26.4 percent, and it is given by the difference between the observed poverty rate (20.9 percent) and the ATET (-5.5 percent). 38 The estimates are robust to conducting the analysis separately for urban and rural households; the results are available from the authors upon request. 19

20 variable that drives a wedge of 30 percent in the probability to select into migration for otherwise observationally identical households would suffice to account for all the estimated effect of migration on poverty. While the test proposed by Becker and Caliendo (2007) is silent about the existence of such a variable, as the conditional independence assumption in (3) is untestable, it suggests that the estimated impact of migration on poverty in Ecuador might reflect even a mild extent of non-random selection in unobservables. The exclusion from the vector x of covariates of any measure of household assets, which is due to the endogeneity of the asset holdings observed at the time of the survey (Bertoli, 2010) and to the absence of retrospective information in the ENEMDU 2005, implies that concerns about a possible non-random selection in unobservables cannot be readily dismissed. This, in turn, casts further doubts on the ability of the recent wave of migration to substantially reduce the incidence of income poverty in Ecuador. 5.2 Remittances and poverty The ENEMDU 2005 also allows us to define, as the literature does, the receipt of remittances as the treatment variable of interest. As discussed in Section 4, all households in the sample are asked whether they received remittances from abroad in the month before the survey, irrespective of whether they report to have a migrant member. We can then define the income of recipient households as the sum of incomes from all sources, including remittances, reported for November No information is provided on the characteristics of the sender of remittances; we introduce the same hypotheses as Acosta, Calderón, Fajnzylber, and López (2008), namely that remittances are sent by one male adult, with the same level of education as the adult stayers in his household. 39 The incidence of poverty among recipient households stands at 12.8 percent, significantly below the 20.9 percent which characterizes migrant households and 19.2 percentage points below the corresponding figure for non recipient households. Specification (2) in Table 3 reports the estimated propensity score f(x), and Table 5 reports the estimated average treatment effects on the treated. With respect to f(x), we 39 These hypotheses are also applied to the recipient households who report to have at least a migrant member; this implies that this part of the empirical analysis is not based on any actual information on migrants. 20