Abstract Rationale Non-adherence to assigned treatment can compromise randomization and bias effect estimates in randomized trials. When participants deviate from their assigned therapy, the treated and control groups may no longer be exchangeable, and intention-to-treat (ITT) estimates tend to dilute the true causal effect. In previous work, we have used causal ML models to show that significant heterogeneity exists and can be explained through baseline variables using the SAVE trial, producing an individualized treatment effect (ITE); with higher the ITE reflecting higher treatment benefit. However, for ITE modeling, low adherence can mask true heterogeneity because differences in outcomes may reflect adherence behavior rather than treatment response. In this work, we address the impact of adherence on HTE models. Methods We analyzed SAVE participants with complete baseline data (N = 2,503). Among those randomized to CPAP, 510 had good adherence (≥4 h/night) and 776 had poor adherence (4 h/night); 1,217 were usual care and never used CPAP, yielding 60.3% non-adherence among CPAP recipients. We incorporate adherence in our previous causal survival forest (CSF) specifications for CPAP vs usual care: as (1) adherence as a predictor, and (2) adherence-adjusted via inverse-probability weighting (IPW), where the probability of adherence to CPAP is modeled via logistic regression with patient characteristics and incorporated as weights in the CSF training, adjusting for confounders between adherence and CVE. Strategies were compared based on model performance for HTE (Area Under the Target Operator Characteristic -AUTOC) and agreement of ITE estimates. Results All models show significant HTE model exist and can be explained by baseline variables (AUTOC0, p 10-10). Adding adherence as predictor slightly improved the ability to capture HTE as compared to the ITT model 1 while adherence via IPW produced a modest improvement most visible in the top 40-60% ITE quantiles. Overall, ITE rankings were highly concordant, indicating strong agreement in ITT estimates. Nonetheless, adherence modeling meaningfully reclassified patients near decision boundaries: with IPW, ∼12% moved from neutral in ITT to benefit or harm in the IPW, with higher reclassification among highly adherent CPAP users. Conclusions Accounting for adherence in CSF models produced only modest changes in overall ITE magnitude and patient ranking. However, it meaningfully reclassified individuals near the Benefit/Neutral/Harm boundaries—particularly among patients with high CPAP adherence. These findings demonstrate that adherence has a measurable impact on individualized treatment-effect estimation and should be incorporated into causal modeling frameworks to better support personalized decision-making for CPAP therapy. This abstract is funded by: NIH
Suárez-Farin∼as et al. (Fri,) studied this question.