I use Swedish registry data to estimate lifetime earnings returns to college and how they vary with observed and unobserved characteristics. The richness of the data also allows me to examine heterogeneity with respect to cognitive and noncognitive ability and parental earnings. Local instrumental variable analysis is used to recover marginal and average treatment effects under selection on gains. The findings support the notion of self-selection, but mainly on observed characteristics. Returns vary little with parental earnings but substantially with respect to both cognitive and noncognitive ability, thus suggesting important complementarities between formal schooling and informal skills.