Inference about the causal effects of a policy intervention requires knowledge of what would have happened to the outcome of the units affected had the policy not taken place. Since this counterfactual quantity is never observed, the empirical investigation of causal effects must deal with a missing data problem. Random variation in the assignment to the policy offers a solution, under some assumptions. We discuss identification of policy effects when participation to the policy is determined by a lottery (randomized designs), when participation is only partially influenced by a lottery (instrumental variation), and when participation depends on eligibility criteria making a subset of participant and non-participant units as good as randomly assigned to the policy (regression discontinuity designs). We offer guidelines for empirical analysis in each of these settings and provide some applications of the methods proposed to the evaluation of education policies.