Answers are grounded in this paper and cite their sources.
Retrospective case-control studies are susceptible to selection bias from differential participation. Knowledge of non-participation proportions by outcome and exposure status is difficult to estimate. This knowledge gap can be treated as a missing data problem through the implementation of imputation strategies to reduce potential bias. We used information about participants' exposure prevalence to inform non-participants' exposure prevalence and applied these estimates in probabilistic bias analysis to address the missing data problem. We conducted a simulation study using four probability distributions (uniform, triangular, trapezoidal, and beta) assigned to the exposure prevalence in non-participants for use in bias-adjustment, identifying the trapezoidal distribution as the most principled approach. We applied the trapezoidal distribution to a meta-analysis on passive smoking and breast cancer, with a recomputed summary OR=1.23 (95% CI, 1.00, 1.51) in retrospective case-control studies and a recomputed summary RR=1.07 (95% CI, 0.92, 1.24) in prospective cohort studies. Applying our bias adjustment to the retrospective ORs, we obtained a summary OR=1.16 (95% CI, 0.99, 1.35) from a meta-analysis of the retrospective studies; closer to the prospective studies' summary estimate. In the simulation and applied example, making defensible assumptions about the exposure prevalence of non-participants yielded more valid estimates compared with ignoring selection forces.
No takes yet. Share an insight, caveat, or question.
Bashir et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: