Results with U.S. Wage Data

with more education. 8 This is an interesting pattern in its own right, but it also suggests that it is reasonable to assume, as we did earlier, that the return to migration experience does not signifi cantly increase with skill.

B. Results with U.S. Wage Data

If we had access to panel data on wages or income in Mexico before and after migra- tion to the United States, we could use traditional fi xed- effects methods to control for permanent unobserved components of skill. While the MMP data does not contain repeated observations of earnings in Mexico, we do observe data on the hourly wage rates that migrants earn on their last trip to the United States. 9 We assume that wages refl ect worker skill plus shocks to productivity in the United States. One can therefore use the wage in the United States as a noisy measure of skill, and include this in Mexi- can income regressions as a control for unobserved skill. Although we could directly include the U.S. wage in the Mexican income regres- sions, we would like to try to reduce some of the noise in this measure of skill by elim- inating variation from observable shocks. To do this, we fi rst regress an individual’s last U.S. wage on a set of controls, and then add the U.S. wage residual as a regressor to our basic Mexican earnings specifi cation. For a subsample of return migrants, we observe complete data on the hourly wage rate and other control variables related to the last trip to the United States. We discuss the construction of this sample in the appendix. After dropping the top and bottom 1 percent of U.S. wage observations, we are left with 1,240 return migrants with U.S. wage data. In the fi rst column of Table 4, we report the results of regressing the real hourly wage rate in the United States on age, the square of age, education, accumulated migration experience, as well as dum- mies for year, U.S. state, interactions between year and U.S. state, and an individual’s community of origin in Mexico. All of these variables refl ect characteristics at the time of the last migration. While the coeffi cient estimates from this regression are not of interest by themselves, we extract the residual from this regression as a measure of unobserved skill. In the next column of Table 4, we reestimate our basic Mexican earnings regression using only the 1,240 return migrants with U.S. wage data. Since we are restricting ourselves to migrants in this subsample, we drop the US dummy variable. With this smaller subsample of migrants, we estimate a return to U.S. migra- tion experience of about 2.3 percent, which is very close to the estimate obtained in the baseline specifi cation and reported in Column 3 of Table 3. In Column 3 of Table 4, we directly add the log of the real wage from the last migratory trip as a control. The coeffi cient on the log wage is small and imprecisely estimated, consistent with the idea 8. Earlier, we found that the negative relationship between education and migration experience was strongest for educated migrants. Indeed, for those with less than three years of education, there appears to be a positive relationship between education and migration experience. Thus, an alternate interpretation of the results in Column 5 is that the interaction term captures not heterogeneity in returns, but rather changes in the correla- tion between USExp and the unobserved components of income across education groups. However, when we restrict the sample to those with three years of education or more, we get very similar point estimates and again fi nd a decline in the return to migration experience as education rises. 9. The MMP asks individuals about the wages earned on the fi rst and last trips to the United States. We use the wage data for the last trip in this exercise for two reasons. First, we have many fewer missing observations for the last trip. Secondly, one might be concerned about recall error when asking retrospective questions about a number as specifi c as a wage rate. Using more recent wage data might lessen such error. that the log wage by itself is a very noisy measure of underlying skill. In Column 4, we add the normalized U.S. wage residual residual divided by the standard deviation of residuals to the Mexican earnings specifi cation. The coeffi cient on the U.S. wage residual suggests that a one standard deviation increase in the U.S. wage residual is associated with a 3.8 percent increase in earnings back in Mexico, and this estimate is signifi cant at the 0.10 level. This relationship suggests that the U.S. wage does contain information about an individual’s level of skill that is not captured by our observables. Crucially, when the U.S. wage residual is added, the coeffi cient on U.S. migration experience is virtually identical to the estimate in the previous column. If endogeneity were driving our results, we would expect the coeffi cient on years of migration experi- ence to drop substantially with the inclusion of a the U.S. wage residual as a control. This exercise provides some evidence against the claim that selection and endogeneity bias are driving our main results.

C. Comparison with the Population and Dwelling Count