(195) Optimal matching ratios for the analysis of time-to-event outcomes in matched cohort studies in EHR research: A simulation study

Background: Covariate-matched cohort designs are increasingly used in electronic health record (EHR)-based pharmacoepidemiology. Matching on key preselected factors can help to reduce confounding, and offers other practical advantages. However, the optimal design parameters for implementation remain uncertain.

Objectives: To determine the optimal matching ratio and analysis models for matched cohort designs with time-to-event outcomes in EHR research.

Methods: We simulated 2,000 longitudinal datasets of 100,000 individuals with a 1% prevalence rate and a binary time-to-event outcome that had no association to the exposure (true log rate ratio = 0). Individuals were censored at the earliest of the outcome date or an administrative censoring date, to mimic real world EHR applications.
Matching was conducted at the date of exposure and based on three variables that emulated age, sex and GP practice, at different unexposed to exposed matching ratios (1:1, 3:1, 5:1 and 10:1), without replacement.
For each scenario, we evaluated common time-to-event analysis models, including Poisson regression (unadjusted and adjusted for matching variables), conditional Poisson regression, Cox regression (unadjusted and adjusted for matching variables), and Cox regression stratified by matched sets. For each scenario, we evaluated bias of the estimate; the empirical standard error (SE); relative % error between the average model-based SE and the empirical SE; and coverage of the 95% confidence intervals (CIs).

Results: We observed no evidence of bias for Cox models stratified by matched set across all ratios, but statistical evidence of bias at higher matching ratios for all other models. To note, Conditional Poisson models exhibited bias across all matching ratios. Precision generally increased with matching ratio (empirical SE [stratified Cox model] (0.042 for 1:1, falling to 0.033 for 10:1). The relative % error was close to 0 in all scenarios with the exception of the Conditional Poisson models. 95% CI coverage was adequate in all scenarios except for Conditional Poisson models, where coverage was poor across all ratios.

Conclusions: In our simulation scenario, employing 5 controls per exposed individual appeared optimal. Increased biases and limited precision gains at higher ratios could be due to increased numbers of exposed patients with no available unexposed matches in the dataset. Analysis using Cox models stratified by matched set performed the best, while conditional Poisson models performed badly across all scenarios. Further work will explore how these findings generalise across a range of dataset specifications.