Seasonal migration and improving living standards in Vietnam.

To understand whether asset holdings or other observable coefficients affect our estimate of [[beta].sub.1], we also control for vectors of variables measured in 1993 at the commune (C) and household level (K). The commune level variables include general economic conditions as well as the distance to Hanoi and HCMC, and the household level variables include measures of both human and physical capital. The most general equation we estimate is:

(4) [r.sub.hvp] = [[tau].sub.p] + [[beta].sub.1] [DELTA] [M.sub.hvp] + [[beta].sub.2][DELTA][X.sub.hvp] + [psi][C.sub.vp] + [eta][K.sub.hvp] + [DELTA][epsilon.sub.hvp].

Identifying Migration

Because our theoretical model suggests that unobservables that affect migration will also affect household expenditures, we attempt to identify migration using measures of informational factors that affect migration but would not directly affect household expenditures. The literature suggests that migration networks increase the amount of information that certain households have about opportunities away from the commune, lowering the implicit costs of migration (Stark 1991; Carrington, Detragiache, and Vishnawath 1996). We use two variables measuring different types of networks that rural Vietnamese households may use to find employment away from the village. First, we use a standard measure of migration networks from the commune survey, which is the percentage of the commune that was seasonally migrating in 1993. Migrants who left communes soon after restrictions of movement were loosened may have given those communes an inherent advantage in migration. Second, we use a measure of networks unique to Vietnam, the percentage of people in the commune who were born in either Hanoi or HCMC before 1975. Households in communes with more members born in either Hanoi or HCMC might have an advantage in finding jobs in the city.

Neither instrument is necessarily randomly distributed across communes, so we are concerned that the instruments may have a more direct correlation with growth. For example, villages with early migrant networks could have had better economic conditions at the beginning of the 1990s, and therefore they were inherently able to grow faster than other communes. Along similar reasoning, individuals who were born in Hanoi or HCMC may have had access to better education, and therefore these households would be in a better position to grow as the economy liberalized during the 1990s.

To ensure our instruments are largely uncorrelated with observables at the commune level, we regress our instruments on a number of observables at the commune level that could be correlated with inherently higher growth rates (table 3). We find no significant correlations between the percentage of the commune population who were seasonal migrants in 1993 and any of the observable commune characteristics (column 1). In fact, we cannot reject the hypothesis that all of the coefficients are jointly zero. The proportion of commune population that was born in Hanoi or HCMC is somewhat correlated with whether or not a commune has electricity and is negatively correlated with both the distance to Hanoi and to HCMC. Other estimated coefficients are statistically insignificant (column 2). As a result, we include a variable that measures the presence or lack of electricity, as well as the distance variables, as part of the commune characteristics vector C.

Second, as Vietnam's agricultural sector became more marketized during the 1990s, farmers who were educated in Hanoi or HCMC may have received a higher quality education and were therefore better positioned to grow. To test whether or not farm income was affected by higher quality education, we regressed the logarithm of farm income on the household head's years of schooling, an interaction between schooling and the percent of the commune born in Hanoi/HCMC, a vector of variables that should affect farm income (land, physical capital, and human capital), and a set of commune dummies. If farmers with an urban education were able to benefit more from marketization more than other farmers, the coefficient on the interaction term would be positive. We found no evidence of a positive coefficient, so we feel comfortable that the Hanoi/HCMC instrument is not correlated with growth of a major portion of income.

Estimation and Results

We first estimate equations (3) and (4) using OLS and an instrumental variables, Generalized Method of Moments (IV-GMM) estimator (table 4). (6) The IV-GMM estimator we use incorporates a weighting matrix that accounts for arbitrary heteroscedasticity and intracluster correlation, and is asymptotically efficient in the presence of heteroscedasticity (Wooldridge 2002; Baum, Schaffer, and Stillman 2003).

Estimating the effect of migration on expenditure growth using OLS, we find a small (0.004), statistically insignificant coefficient (model 0). However, unobservables should be correlated with both the migration decision and expenditure growth. For example, households with poorer business skills, or less ability to market crops, might have been more likely to send out migrants, because they could earn a wage away from the commune whereas they would have to farm or run a business within the commune. In that case, the OLS coefficient would both measure the effect of migration and poor business acumen on expenditure growth, and the coefficient estimate would be lower than the true effect of migration on expenditure growth.

When we estimate equations (3) and (4) using the IV-GMM estimator, we find that the both instruments have a strong correlation with the migration variable. In all specifications, the cluster corrected F statistic testing the hypothesis that the coefficients on both estimates are jointly zero is larger than 8 (Appendix table B). In our base IV regression (model 1), the estimates coefficient is 0.063 with a z-statistic of 2.39. Among households that were likely to respond to migrant networks, our estimate suggests that an additional migrant is associated with an expenditure growth rate that is 6.3 percentage points higher than it would have been without migration participation. (7) The overidentification tests indicate that the results do not differ when each instrument is used separately, as the Hansen J statistic indicates we cannot reject the overidentifying restrictions.

When we add commune characteristics (C) measured in 1993 to the model (model 2), the estimated coefficient drops to 0.053, but remains significantly different from zero at the 5% level. Adding commune characteristics that may have affected commune-specific growth rates to the model, we find that among households affected by the instruments, an additional migrant is associated with 5.3 additional percentage points of annual per capita expenditure growth. Controlling for human capital characteristics of the household head (model 3) and household land holdings and productive assets (measured in 1993, model 4), we estimate a slightly lower coefficient of 0.052. (8) Since the consumption data exhibit mean reverting measurement error, these estimates should all be considered lower bounds.

Robustness Checks

Our results imply that demographics matter when explaining household expenditure growth. Therefore, we might be concerned that either measurement error in consumption is correlated with household size and affects our results, or that we mismeasure the amount of consumption allocated to each individual in the household. Therefore, we re-estimate equation (3) using both nonfood expenditures and expenditures per adult equivalent as the dependent variable (table 5, models 1 through 4). When we use per capita nonfood expenditures as the dependent variable, we find that among households likely to respond to migrant networks, an additional migrant is associated with 7.9% higher nonfood expenditure growth per annum (models 1 and 2). The higher coefficient is consistent with our previous findings, as nonfood expenditures grew much faster than food expenditures over the study period. When expenditures per adult equivalent are used instead of per capita expenditures (models 3 and 4), the estimated coefficient is approximately 0.054, or slightly higher than the same specification using per capita expenditures (0.052). These findings strengthen the view that migration improves the monetary well-being of households; households increasing their participation in migration between the two surveys also experience a large increase in their well-being, measured in terms of consumption.

Since attrition did not occur randomly across communes, we also test whether correcting for attrition bias affects the results (table 5, models 2, 4, and 6). Using a procedure described in Wooldridge (2002), we first use a probit model to estimate the probability that each household surveyed in 1993 stays in the data set, using our full set of variables as explanatory variables. We then re-estimate each model, using inverse probability weighting to weight each observation by the estimated probability that it remains in the data set. The weighted estimates are virtually the same as the unweighted estimates, so we can conclude that attrition bias does not affect our results.