We estimate the returns to college using administrative data on both college enrollment and earnings. Exploiting that colleges dismiss low-performing students on the basis of exact GPA cutoffs, we use a regression discontinuity design to estimate the earnings impacts of college. Dismissal leads to a short-run increase in earnings and tuition savings, but the future fall in earnings is sufficiently large that 8 years after dismissal, persisting students have already recouped their up-front investment with an internal rate of return of 4.1%. We provide a variety of evidence that manipulation of the running variable does not drive our results.

We highlight that analyses using interaction terms to study treatment effect heterogeneity are susceptible to a form of omitted variable bias that is often overlooked in economics. Unlike most instances of omitted variable bias, the omitted variables in this case are available to the researcher but were not included in the model. We demonstrate that this exclusion matters based on a replication of 205 estimates across seventeen papers published in the American Economic Review over a five-year period. For approximately 60% of these papers, failing to account for the omitted variables changes the majority of estimates by more than 100%.