2F-01 - Validation and Extension of IPW-Adjusted Cox Regression Using Super Learning: Insights from Antihypertensive and Antihyperlipidemic Therapy Comparisons

Background: Super Learning (SL) is an advanced ensemble machine learning method increasingly applied in medicine. Inverse probability weighting (IPW) is widely used for causal effect estimation. Previously, we used IPW-adjusted Cox regression to compare the effectiveness of different antihypertensive and antihyperlipidemic monotherapies in primary CVD prevention, considering only a limited variable set without SL.

Objectives: This study aims to validate the findings from IPW-adjusted Cox regression analyses using existing advanced R packages and SuperLearner R package, and to investigate the impact of incorporating a broader set of variables on the results.

Methods: A retrospective cohort study using the IADB.nl dispensing database (1996–2020, Netherlands) compared antihypertensive and antihyperlipidemic monotherapies. Exposures included beta-blockers (BBs), calcium channel blockers (CCBs), thiazides, ACEIs, and statins. The outcome was the time to first acute cardiac drug therapy (CDT) prescription, serving as a proxy for major CVD events. Cox proportional hazards models, adjusted with inverse probability weighting (IPW), were applied using R packages (twang, ipw, SuperLearner). Two covariate sets were analyzed: a limited set (age, gender, calendar year, and history of diabetes, rheumatoid arthritis, and asthma/COPD medications) and an expanded set (all drugs starting with ATC codes A, B, C, D, G, H, J, L, M, N, P, R, S, V at the 4th level). Hazard ratios (HRs) with 95% confidence intervals (CIs) were calculated.

Results: Among 33,427 antihypertensive users, SL-IPW-adjusted analyses (limited variables) showed CCBs had higher CDT hazards than BBs (HR: 1.27, 95% CI: 1.13–1.42), while thiazides had lower hazards (HR: 0.91, 95% CI: 0.85–0.98), consistent with traditional IPW trends. With expanded variables, CCBs remained associated with higher hazards (HR: 1.35, 95% CI: 1.20–1.51), and thiazides remained lower (HR: 0.92, 95% CI: 0.85–0.99). ACEIs were also linked to increased hazards.
Among 18,375 antihyperlipidemic users, SL-IPW-adjusted analyses (limited variables) indicated higher hazards for atorvastatin and rosuvastatin vs. simvastatin (HRs: 1.26, 95% CI: 1.14–1.40; and 1.22, 95% CI: 1.03–1.44). Findings remained consistent with expanded variables (HRs: 1.22, 95% CI: 1.10–1.35; and 1.22, 95% CI: 1.02–1.45), with minor statistical variations.

Conclusions: The findings from the SuperLearner package, whether using limited or extensive variables, closely aligned with results from advanced IPW methods. This consistency reinforces the robustness of previous findings and highlights SuperLearner’s reliability in producing high-quality causal effect estimates.