Abstract:
|
In randomized trial analyses, the average treatment effect is often used to compare treatment with control. For patients, another parameter of interest is the fraction who benefit from treatment, which is defined as the proportion of patients whose potential outcome under treatment is better than that under control. The fraction generally is only partially identified, which makes inference a challenging problem. We propose a method for constructing a confidence interval for the fraction, using randomized trial data. We consider the case of an ordinal outcome, with binary outcomes as a special case. Our method does not require any assumptions about the joint distribution of the potential outcomes. However, it can incorporate user-defined restrictions on the support of the joint distribution, such as the no harm assumption. We prove pointwise consistency of the proposed confidence interval. Through simulation, we compare it to the m-out-of-n bootstrap, with respect to coverage probability, width, and computational efficiency.
|