Testing attributable effects hypotheses with an application to the Oregon Health Insurance Experiment*

Following a randomized trial, the sum of the differences in the outcomes for the treated units compared to the outcome that would have been observed if the same units had been assigned to the control condition is known as the attributable effect. Most previous methods on testing hypotheses about the attributable effect require the outcome to be binary or ordinal. In this paper, we use a simple approximation to the distribution of a carefully selected test statistic under the hypothesis that the attributable effect is zero to expand attributable effects inference for count and continuous data. The method is efficient and performs well in a variety of simulations. We demonstrate the method using a large medical insurance field experiment.