Interpreting IV Estimators in Information Provision Experiments

A growing literature measures "belief effects" -- that is, the causal effect of a change in beliefs on actions -- using information provision experiments, where the provision of information is used as an instrument for beliefs. In experimental designs with a passive control group, and under heterogeneous belief effects, we show that the use of information provision as an instrument may not produce a positive weighted average of belief effects. We develop an "information provision instrumental variables" (IPIV) framework that infers the direction of belief updating using information about prior beliefs. In this framework, we propose a class of IPIV estimators that recover positive weighted averages of belief effects. Relative to our preferred IPIV, commonly used specifications in the literature require additional assumptions to generate positive weights. And in the cases where these additional assumptions are satisfied, the identified parameters often up-weight individuals with priors that are further from the provided information, which may not be desirable.