4G-01 - The Value of the Pattern of Change in ATT and ATU After Trimming When Interpreting Trimmed Estimates: An Example Comparing New Users of SSRIs and SNRIs

Background: Propensity score (PS) trimming is a useful tool to improve the precision of estimates and eliminate positivity violations. While it can be challenging to determine whether estimates change after trimming due to changes in unmeasured confounding or in the prevalence of measured effect measure modifiers (EMMs) in the sample, evaluating changes in estimates of the effect in treated and untreated populations may be helpful. The limited available heuristics have only been studied in simulations.

Objectives: Evaluate changes in the average treatment effect on 1-year mortality in the treated (ATT) and average treatment effect in the untreated (ATU) after 5% asymmetric PS trimming in real-world patients initiating selective serotonin reuptake inhibitors (SSRIs) versus selective norepinephrine reuptake inhibitors (SNRIs) in the Clinical Practice Research Datalink (CPRD) and identify if those could be due to PS-positive (i.e., those who benefit are more likely to be treated) or - negative (i.e., those who benefit are less likely to be treated) heterogeneity.

Methods: We identified patients initiating SSRIs or SNRIs for the first time in the CPRD with a minimum of 1 year of look-back from 2019 to 2022 and estimated the PS for initiating an SSRI conditional on an array of baseline characteristics. We used weighting to estimate the average treatment effect as well as the effect in SSRI initiators (ATT) and SNRI initiators (ATU). We calculated hazard ratios (HRs) for all-cause mortality comparing SSRI and SNRI initiators using each weighting strategy before and after 5% asymmetric PS trimming with 95% confidence limits from the sandwich estimator.

Results: There were 335,469 SSRI initiators and 24,318 SNRI initiators in the study sample. 259,178 SSRI and 20,294 SNRI initiators remained after trimming. The estimated ATE HRs were 0.76 (95% CI: 0.76, 0.82) before and 0.84 (95% CI: 0.77, 0.90) after trimming, ATT HRs were 0.77 (95% CI: 0.70, 0.86) before and 0.81 (95% CI: 0.75, 0.88) after trimming, and ATU HRs were 1.24 (95% CI: 1.17, 1.31) before and 1.17 (95% CI: 1.09, 1.24) after trimming. The ATT becoming less favorable and the ATU becoming more favorable for SSRIs is compatible with PS-positive heterogeneity but not PS-negative heterogeneity. Further investigation showed that older age (a major confounder and EMM) was inversely associated with initiating SSRIs and also associated with a more harmful side-effect profile, creating PS-positive heterogeneity.

Conclusions: The changes in ATT and ATU estimates observed after 5% asymmetric trimming are compatible with PS-positive heterogeneity in the data. Changes in the ATT and ATU before and after trimming can help rule out whether changes in estimate from trimming could result from a specific type of heterogeneity across the PS.