Combining evidence from multiple electronic health care databases: performances of one-stage and two-stage meta-analysis in matched case-control studies

Purpose: Clustering of patients in databases is usually ignored in one-stage meta-analysis of multi-database studies using matched case-control data. The aim of this study was to compare bias and efficiency of such a one-stage meta-analysis with a two-stage meta-analysis. Methods: First, we compared the approaches by generating matched case-control data under 5 simulated scenarios, built by varying: (1) the exposure-outcome association; (2) its variability among databases; (3) the confounding strength of one covariate on this association; (4) its variability; and (5) the (heterogeneous) confounding strength of two covariates. Second, we made the same comparison using empirical data from the ARITMO project, a multiple database study investigating the risk of ventricular arrhythmia following the use of medications with arrhythmogenic potential. In our study, we specifically investigated the effect of current use of promethazine. Results: Bias increased for one-stage meta-analysis with increasing (1) between-database variance of exposure effect and (2) heterogeneous confounding generated by two covariates. The efficiency of one-stage meta-analysis was slightly lower than that of two-stage meta-analysis for the majority of investigated scenarios. Based on ARITMO data, there were no evident differences between one-stage (OR = 1.50, CI = [1.08; 2.08]) and two-stage (OR = 1.55, CI = [1.12; 2.16]) approaches. Conclusions: When the effect of interest is heterogeneous, a one-stage meta-analysis ignoring clustering gives biased estimates. Two-stage meta-analysis generates estimates at least as accurate and precise as one-stage meta-analysis. However, in a study using small databases and rare exposures and/or outcomes, a correct one-stage meta-analysis becomes essential.