Unknown confounders did not bias the treatment effect when improving balance of known confounders in randomized trials

Objective: The objective of the study was to measure if improving balance in known and observed confounders by propensity score (PS) matching yields different treatment effect estimates in randomized controlled trials (RCTs), thus indirectly measuring the influence of unknown confounders. Study Design and Setting: This is an analysis of individual patient data of 26 large RCTs and comparison of agreement between PS-matched samples and the RCT results on one hand with the agreement between subsamples of RCTs (with sample sizes equal to the sample sizes of the PS-matched samples) and RCTs by Bland-Altman plots and corresponding intraclass correlation coefficients on the other. Results: We included data on 213 outcomes from 37 treatment comparisons with 193,620 patients from 26 trials. Bland-Altman plots and intraclass correlation coefficients showed better agreement between PS-matched analysis and RCTs than between reduced RCTs and RCTs. Conclusion: We found no indication for a detrimental influence of unknown confounders in PS-matched samples of RCTs. (C) 2020 Elsevier Inc. All rights reserved.