Online Program

Multiple-Bias Modeling for Credible Causal Inference in Health Policy Studies (307876)

Keywords: bias analysis, unmeasured confounders, measurement error, selection bias, multiple-bias modeling, sensitivity analysis, clinical outcomes, sleep apnea

Causal inference requires strong assumptions about the absence of uncontrolled confounding, selection bias, and measurement error. Some combination of these sources of bias can lead to substantial bias during causal analysis. This study gives intuitive and graphical illustrations of the separate and combined consequences of unmeasured confounders, selection bias, and measurement error including misclassification on treatment effect identifiability and estimation. We introduce modern multiple-bias modeling using appropriate bias formulas for each bias source that account for the order of bias occurrence. We then develop and demonstrate simulation tools for joint probabilistic modeling of multiple sources of bias. We discuss how to obtain, specify, and use bias parameters in multiple-bias modeling. Finally, we illustrate the simulation-based multiple-bias modeling method including its interpretation and reporting using numerical examples from a large cohort study of the effects of incident and treated and untreated obstructive sleep apnea on clinical outcomes among 3,079,514 US veterans. We conclude that multiple-bias modeling makes large causal inference studies more credible.