How to adjust confounders in studies on observational comparative effectiveness: (3) approaches on sensitivity analysis for confounder adjustment

Abstract

Confounders are difficult to avoid in studies on observational comparative effectiveness. It is often unclear whether the confounders have been completely eliminated after controlling the measured or unmeasured potential confounding effects or if sensitivity analysis is needed when using the specific statistical methods, under given circumstances. This manuscript summarizes and evaluates the confounding sensitivity analysis methods. Based on different studies, sensitivity analyses need to use different approaches. The traditional sensitivity analysis can be applied for the measured confounders. Currently, the relatively systematic sensitivity analyses for unmeasured confounders would include confounding function, bounding factor and propensity score calibration. Additionally, more investigations are associated with Monte Carlo and Bayesian sensitivity analysis. Reliability of the research conclusion thus may largely be improved when the sensitivity analysis results are consistent with the main analysis.