Emulating Stratified Randomized Trials by Propensity Score Methods

Background: Stratified randomized controlled trials (RCTs) ensure balance on key risk factors across allocated treatments, both overall and within important subgroups. Emulating stratified RCTs with propensity score (PS) methods could allow us to address confounding separately within each stratum of strong risk factors of the outcome, leading to improved confounding control.

Objectives: To compare the stratified vs. non-stratified PS approaches in estimating the cardiovascular effectiveness of empagliflozin (EMPA) vs. a dipeptidyl peptidase-4 inhibitor (DPP-4i) in patients with type 2 diabetes.

Methods: Using Medicare data (2014-18), we identified older adults >65 years with type 2 diabetes initiating EMPA or DPP-4i, to assess a composite outcome of myocardial infarction or stroke hospitalizations. We estimated hazard ratios (HR) and rate differences (RD) per 1,000 person-years, accounting for death as a competing risk, overall and in subgroups with and without baseline cardiovascular diseases (CVD), including atherosclerotic CVD and heart failure. We implemented 4 PS strategies based on 143 baseline covariates: PS estimation and matching in the total cohort (‘total cohort PS') — with or without interactions between a subgroup and all other covariates, using the total cohort PS to match within each subgroup and pooling across subgroups, and PS estimation and matching within each subgroup and pooling data across subgroups (‘stratified PS‘ — that emulates the stratified RCT).

Results: Covariates were balanced between treatments across all PS strategies. CVD conditions included in a subgroup were slightly more balanced after stratified PS matching. For the overall EMPA vs. DPP-4i cohort, total cohort PS matching (without interactions) gave HR of 0.75 (0.64, 0.88) and RD of -6.12 (-9.43, -2.83). Including interactions in the total cohort PS and using the total cohort PS to match within subgroups gave similar estimates. Stratified PS-matched HR was 0.85 (0.72, 1.00) and RD was - 3.29 (-6.47, -0.11), consistent with trial findings. As for subgroup estimates, total cohort PS matching (no interactions) gave HR: 0.72 (0.59, 0.87) and RD: -9.89 (-15.45, -4.36) in the CVD (+) subgroup and HR: 0.84 (0.63, 1.13) and RD: -2.08 (-5.69, 1.51) in the CVD (-) subgroup. After stratified matching, HR was 0.80 (0.66, 0.97) and RD was -6.39 (-11.82, -1.00) in the CVD (+) subgroup, while HR was 0.98 (0.72, 1.32) and RD was -0.33 (-3.75, 3.10) in the CVD (-) subgroup, both consistent with trial findings.

Conclusions: The stratified PS improved confounding control, compared to the total cohort PS, by prioritizing balance in important risk factors of the outcome, and produced effect estimates close to trial findings, both overall and within subgroups.