|
Abstract:
|
While assignment to treatment and control groups are randomized in experiments (e.g., clinical trials), selection into the trial itself is most often not. This makes it more difficult to generalize experimental results to the wider population. Current techniques to model trial participation generally require population-level auxiliary data to supplement the sample information. We take a different approach. Our strategy is to use an instrumental variable, such as distance from the trial site, to model these trial participation probabilities without confounding. Such a variable avoids the omission bias by being correlated with trial participation but not directly with trial outcomes or other covariates. From this foundation, we develop a statistical test to check for generalizability.
|
