Online Program

Evaluating Observational Data Analyses: Confounding Control and Treatment Effect Heterogeneity

*Issa J Dahabreh, Brown University 
David M Kent, Tufts Medical Center 

Keywords: causal inference, propensity scores, effect measure modification, treatment effect heterogeneity

Large-scale observational data are increasingly available for comparative effectiveness research. Exposure modeling techniques (e.g., propensity score methods) are popular for addressing confounding bias in observational studies. Theoretical work and simulation studies have explored the properties of propensity score methods for estimating average treatment effects of point (or fixed) exposures. However, recent attempts to "replicate" the findings of randomized clinical trials (RCTs) using propensity scores in observational data sets have produced mixed results. In addition, empirical assessments of propensity score methods when research questions pertain to heterogeneous (across subpopulations) treatment effects are limited. We present the rationale, methods, and preliminary results of a large-scale attempt to "replicate" the findings of RCTs using observational data sets. We discuss how this work overcomes some of the limitations of previous attempts to compare the results of RCTs and observational studies. Finally, we present the results of simulation studies of the performance of propensity-score based estimators under treatment effect heterogeneity.