Evidence synthesis using observational and randomized studies: empirical results, statistical methods, and practical recommendations
Issa J Dahabreh, Brown University
Keywords: observational studies, randomized trials, evidence synthesis, cross-design synthesis
Large-scale observational data are increasingly used to address comparative effectiveness research questions. As a result, decision-makers examining the effects of medical interventions are often faced with bodies of evidence that include both observational and randomized studies. In this talk, we review empirical results regarding the use of observational studies in evidence syntheses that aim to inform clinical and policy decisions on treatment effects. We focus on recent investigations that use exposure modeling methods (e.g., propensity scores) to reproduce the findings of randomized trials in observational data, with an emphasis on identifying potential reasons for failure to obtain similar estimates. We propose a taxonomy of statistical methods for the synthesis of observational and randomized data, focusing on methods that leverage empirical results to attempt “bias adjustment.” Finally, based in part on the deliberations of an AHRQ-sponsored workgroup on “integrating bodies of evidence: randomized and nonrandomized studies,” we present a set of tentative recommendations for the conduct of cross-design evidence synthesis.