Association for Public Policy Analysis & Management

Earnings differences across by major and institution are in many cases nearly as large as the college-high school wage gap (Altonji et al. 2012; Altonji et al. 2016; Chetty et al. 2017). Though the magnitude of these differences suggests that college and major choice may play an important role in earnings determination, interpreting them is difficult because students may select into colleges and majors in ways that are correlated with both earnings levels and match-specific earnings effects.

We link nearly fifty years of admissions records for the population of selective universities in Chile to population tax data. Using these data, we estimate two sets of specifications. The first are OLS specifications that describe how earnings differ across college, majors, course content (such as STEM and Humanities credits), and observable a rich set of student characteristics. These conditional means provide a broad picture of earnings outcomes in the population of students, but yield biased estimates of causal effects if students choose college and majors based on determinants of earnings levels or gains that we do not observe. Our second set of specifications exploits an institutional feature of the admissions process that generates thousands of regression discontinuities across college by major combinations to estimate the effect of going to one option relative to some other option for students at that margin. Our regression discontinuity specifications crack the relationship between unobservable earnings determinants and degree assignment for the subset of students at admissions margins.

We combine these two sets of estimates into a rich picture of choice and earnings determination using a structural model of college and major choice and earnings determination. The earnings model allows for unobservable preferences and talents to play a role in degree choice. We estimate the model using indirect inference: we choose the parameters governing the choice equation, earnings equation and the distribution of unobservables so as to best replicate the OLS and RD estimates as well as observed college application behavior.

Results from OLS and RD estimation suggest that there is considerable heterogeneity in earnings outcomes across majors and colleges. Majors with high concentration of STEM course content have higher averages earnings even after controlling for test scores in a variety of subjects, GPA, gender and SES. We also find evidence of comparative advantage across observable match effects for students with high math at high STEM majors and high language at high humanities majors. Taken all together, at the time of application, the predicted change in expected earnings obtained from OLS is correlated positively with RD estimates. This indicates that after controlling for a wide range of observables, the earnings differences across colleges and majors cannot be attributed solely to selection on unobservable student characteristics. Our findings suggest that in the presence of rich control sets, OLS estimates of earnings outcomes provide a reasonable guide for policy that looks to regulate higher education markets.