Answers are grounded in this paper and cite their sources.
Covariate matching is widely used in cohort studies, particularly in EHR-based research, to balance comparison groups and estimate exposure effects in time-to-event analyses. Yet poorly chosen matching or analytical strategies can introduce inefficiency, bias, or incorrect standard errors, and guidance on optimal approaches remains limited. We conducted a simulation study to evaluate effective matching and analysis strategies for matched cohort designs. We compared analytic performance across matching ratios and whether controls were reused. We assessed several approaches to account for prespecified variables: unadjusted and adjusted Poisson models, conditional Poisson models, unadjusted and adjusted Cox models, and Cox models stratified by matched sets. We also examined whether results differed by timescale (time-in-study vs. age). The best-performing scenarios were re-tested under varying exposure prevalence (to reflect control availability) and different levels of confounding. Across scenarios, matching without replacement was preferred, giving accurate SEs and aligning with clinical reasoning. Using multiple controls per case was beneficial when the proportion of unmatched exposed cases was minimal. Adjusted models and stratified Cox models produced unbiased estimates with small standard errors: adjustment improved precision, while stratification handled complex or multidimensional variables well. Conditional Poisson models performed poorly and are not recommended for matched cohort time-to-event analyses.
No takes yet. Share an insight, caveat, or question.
Andresen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: