Causal Ablation Methods for Comparative Effectiveness Research

About This Grant

Project Summary Comparative effectiveness research is a critical component of medical research that relies on randomized experiments and observational study evidence to identify interventions that improve healthcare outcomes. In many cases, the analysis of the data obtained from such studies is complicated by truncation by death, time-varying confounders, and zero-inflated outcomes. A common approach to address these difficulties is to use models that rely on unrealistic, untestable assumptions. In many cases, the quality of the evidence is low, since it is hard to assess which assumptions are necessary to make inferences. To that end, we develop ablation methods for causal inference. The result is an ablation framework that allows analysts to understand which assumption or combination of assumptions lead to informative conclusions. For instance, analysts will be able to identify the minimum set of assumptions that are needed for either point identifi- cation or the sign of the effect. We develop three different causal ablation frameworks for three specific applications that are common in comparative effectiveness research: truncation by death, time-varying confounders, and zero-inflated outcomes. Under each aim, for each application, we develop (1) appropriate assumption sets that meet a compatibility criterion, (2) a grid search tai- lored to the assumption sets, (3) a corresponding set of point and partial identification estimators, and (4) a summary step that reports results by order of assumption credibility.