Estimation of causal effects in observational studies: Applications to training programs and labor migration decisions

This thesis concerns the estimation of causal effects in observational studies. A causal effect is defined as a treatment which brings about a change in the variable of interest compared to some baseline called the control. The problem of identifying a causal effect is that variables of interest are observed under either the treatment or control, but never both. For example, the researcher observes an individual's earnings after a training program (a treatment) or the earnings if no training program is received, but not both. In many developing economies agricultural households send a member to work abroad for a higher wage, allowing the household to make economic decisions such as increasing the education of its children (an outcome). The children's education had there been no migrant is not observed.;The thesis draws on statistical methods of causal inference analysis that extend the theory of classical randomized experiments in observational studies to estimate the treatment effect when selection is on observable characteristics. The method uses the propensity score--the probability of receiving treatment given observed characteristics--to reduce the dimensionality problem for a non-parametric estimation of the treatment effect. The first and second papers use the National Supported Work (NSW) Demonstration Program, a randomly assigned training program, to contrast the proposed non-parametric estimates of the treatment effect with more conventional regression-based estimates. The results demonstrate that when applied to observational studies the method closely replicates the experimental training effect.;The third paper examines the effect of temporary labor migration in agricultural households on gender differences in primary and secondary education using household data from rural Egypt. Temporary labor migration provides a well defined setting in which to examine the causes and extent of gender differences in schooling, a persistent phenomenon in rural areas of many developing economies. With an absent worker, household labor participation increases and children's participation is also likely to increase adversely affecting school enrollment. Using the estimation approach described in the first paper, results indicate that while the migrant is abroad secondary school enrollment for girls compared to boys falls substantially but increases after a migrant returns. This suggests that resource constraints faced by poor rural households may be the most important factor causing gender differences in school enrollment.;The techniques discussed in the thesis complement standard econometric tools for estimating causal relations. They are powerful enough to sort out which observations from a pool of potential controls are relevant comparisons to the treated units under consideration. The argument is: before recourse to modeling through assumptions on functional forms and distribution, assumptions on unobservables which are difficult to test in the data, there is substantial reward in exploring first the information contained in the variables that are observed.