Identifying Causal Effects in Information Provision Experiments

Information treatments often shift beliefs more for people with weaker belief effects. Since standard TSLS and panel specifications in information provision experiments have weights proportional to belief updating in the first-stage, this dependence attenuates existing estimates. This is natural if people whose decisions depend on their beliefs gather information before the experiment. I propose a local least squares estimator that identifies unweighted average effects in several classes of experiments under progressively stronger versions of Bayesian updating. In five of six recent studies, average effects are larger than-in several cases more than double-estimates in standard specifications.