Primary Submission Category: Generalizability/Transportability
Sensitivity Analysis for Extending Inferences of a Binary Time-Fixed Treatment Effect Derived from a Randomized Controlled Trial to a Target Population When a Subset of Treatment Effect Modifiers Are Measured Only On Trial Subjects
Authors: Jay Xu, Marissa Seamans,
Presenting Author: Jay Xu*
The external validity of results from RCTs may be compromised when the distribution of effect modifiers (EMs) differs between the study population and the target population of scientific and/or policy interest. To better estimate real world and/or policy relevant causal effects of treatments or interventions, integrative methods that synthesize data from RCTs and observational data sources (e.g., EHR data) to infer causal effects for target populations represented by these observational data sources have been developed, which explicitly adjust for observed differences in covariate distributions between RCT and observational samples. Their utility of such methods, however, depends on the extent to which EMs that differ in distribution between the study and target populations are measured in both RCT and observational samples. We consider the theoretical scenario where all EMs are measured in the RCT sample, but a subset of them are unmeasured in the observational sample. We propose sensitivity analysis procedures to perform Frequentist or Bayesian inference for the target population average treatment effect of a binary time-fixed treatment under various user-postulated differences between the study and target population conditional distributions of the EMs measured only in the RCT given the dually measured covariates. We demonstrate the merits of the proposed sensitivity analysis procedures using a simulation study and illustrate their use on substance use clinical trial data.