Primary Submission Category: Sensitivity Analysis
Causal progress with imperfect placebo treatments and outcomes
Authors: Chad Hazlett, Adam Rohde,
Presenting Author: Chad Hazlett*
In the quest to make defensible causal claims from observational data, investigators may leverage information from “placebo treatments” and “placebo outcomes” (aka “negative control outcomes”). Existing approaches focus largely on point identification and require two difficult assumptions: (i) “perfect placebos” (placebo treatments have precisely zero effect on the outcome; treatment has precisely no effect on placebo outcomes); and (ii) “equiconfounding” ( the treatment-outcome relationship where one is a placebo suffers the same amount of confounding as does the real treatment-outcome relationship, on some scale). By contrast, using an omitted variable bias framework, we consider degrees of placebo imperfection (non-zero effects of placebo treatment on real outcomes or of real treatments on placebo outcomes), and non-equiconfounding (different strengths of confounding suffered by a placebo treatment/outcome compared to the true treatment-outcome relationship). Postulated values for these quantities identify or bound the linear estimates of treatment effects. While applicable in many settings, one broad use-case for this approach is to employ pre-treatment outcomes as (perfect) placebo outcomes. In this setting, our approach offers a credibility-enhancing relaxation of the parallel trends assumption of DID. We demonstrate the use of our framework with two applications, employing an R package that implements these approaches.