Children who grow up in two-parent households score 0.2-0.4 standard deviations higher on covariate-adjusted cognitive skill measures than children in single-parent households. This paper analyzes the extent to which these differences reflect causal effects of marriage versus selection. We develop a dynamic Roy model that allows for selection into and out of marriage, with cognitive skill evolving according to a dynamic latent factor model. A long panel of marriage decisions permits identification of unobserved heterogeneity jointly affecting both marriage and skill development. Model estimates indicate that marriage improves skill development. A decomposition analysis suggests that differences in the technology of skill formation between the married and single states explains 30% of the skill difference between children growing up in two-parent and one-parent households while income differences explain 16%.