Within-center matching performed better when using propensity score matching to analyze multicenter survival data: empirical and Monte Carlo studies

Objective: Propensity score (PS) methods are applied frequently to multicenter data. To date, methods for handling cluster effect when analyzing PS-matched data have not been assessed for survival data. Accordingly, the objective of the present study was to determine the optimal PS-model to account for a potential cluster effect when analysing multicenter observational data.

Study design and setting: In the current study, five strategies were compared. One analyzed the original sample and four used global or within-cluster matching using a global or a cluster-specific PS. All were applied to simulated data sets and to two cohorts.

Results: Failing to account for clustering in the PS model led to a biased estimate of the treatment effect and to an inflated test size. Within-cluster matching using either a global or a cluster-specific PS led to the lowest mean squared error and to a test size close to its nominal value. However, the cluster-specific approach led to a drastic reduction of sample size compared with the global PS one. Analyses of the cohorts confirmed that the latter model led to the smallest sample size, but also necessitated the discard of a high number of clusters from the matched sample.

Conclusion: In the considered simulation scenarios, within-cluster matching using a global PS presented the best balance between sample size and bias reduction, and it should be used when applying PS methods to clustered observational survival data.