Policies impose lotteries of outcomes on individuals, since we never know exactly what the effects of the policy will be. In order to evaluate alternative policies, we therefore need to make some assumptions about individual preferences, even before social welfare functions are applied. Instead of making a priori assumptions about those preferences that are likely to be wrong, there are two broad ways in which experimental methods are used to evaluate policy. One is to use experiments to estimate individual preferences, valuations and beliefs, and use those estimates as priors in the evaluation of policy. The other approach is to undertake deliberate randomization, or exploit accidental or natural randomization, to infer the effects of policy. The strengths and weaknesses of these approaches are reviewed, and their complementarities identified.