We estimate heterogeneous returns to STEM education by leveraging relative distances to technical versus general universities in Switzerland. While individuals who choose a STEM education gain on average, a declining marginal treatment effect curve indicates positive selection on gains, suggesting that low-resistance individuals benefit the most. Through policy simulations aimed at increasing STEM enrollment and estimating corresponding policy-relevant treatment effects, we demonstrate that these policies' effectiveness critically depends on both observable and unobservable characteristics of affected individuals. Furthermore, we highlight how policies should be designed to both increase STEM enrollment and generate positive returns for targeted groups, particularly women.