Estimation and Inference in Social Experiments

This paper develops a framework for analyzing the outcome of experiments carried out on forward-looking subjects. Natural experiments, unexpected policy changes, and true experiments are all included in the framework as special cases. These concepts are defined in conjunction with explicit notions of controlled and randomized experiments. The persistent issues of sample-selection bias and heterogeneous impacts that surround interpretations of experiments are endogenous to the model. Special attention is given to interpreting empirical impact of the treatment within the model. The environments in which estimated mean impacts correspond to mean subjective impacts are specified, and they are found to be a small, uninteresting subset of environments contained within the framework.