Control Outcomes and Exposures for Improving Internal Validity of Nonrandomized Studies.

Abstract

Control outcomes and exposures can improve internal validity of nonrandomized studies by assessing residual bias in effect estimates. Control outcomes are those expected to have no treatment effect or the opposite effect of the primary outcome. Control exposures are treatments expected to have no effect on the primary outcome. We review examples of control outcomes and exposures from prior studies and provide recommendations for conducting and reporting these analyses. Review in Google Scholar and Medline of research studies employing control outcomes or exposures. We abstracted publication year, control outcome, control exposure, primary outcome, primary exposure, control outcome/exposure effect, proposed source of bias, and causal criteria. There is inconsistent terminology for these concepts, making study identification challenging. Six of 11 studies found null associations between treatments and negative control outcomes/exposures, providing greater confidence that the primary study findings were not biased. Five studies found unexpected associations, suggesting bias in the primary association. The rigor of nonrandomized studies can be improved with inclusion of control outcomes and exposures for bias detection. Given ongoing concern about clinical and policy inferences from nonrandomized studies, we recommend adoption of these measurement tools.