USING MACHINE LEARNING TECHNIQUES TO IDENTIFY AND ASSESS HETEROGENEOUS TREATMENT EFFECTS IN ADULTS WITH TYPE 2 DIABETES

Precisely and efficiently identifying subgroups with heterogeneous treatment effects (HTE) in observational data remains a challenge. Based on the causal forest (CF), we developed iterative causal forest (iCF) to identify HTE in subgroups defined by variables. iCF iteratively grows different depths of CF with important effect modifiers, performs majority votes to obtain decision trees from a family of different depths of CFs, then finds the cross-validated subgroup decision which best predicts the treatment effect as final subgroup decision. Our simulations included of 11 various scenarios including 4-way, 3-way, and 2-way interactions, and no interactions. In empirical study 1, we implemented iCF in sodium-glucose cotransporter-2 inhibitors (SGLT2i) versus glucagon-like peptide-1 receptor agonists (GLP1RA) using a 20% random sample of fee-for-service Medicare beneficiaries aged 65+ 2012-2017 to identify subgroups for hospitalization for Heart Failure. In empirical study 2, we implemented iCF in 2 new-user cohorts of commercially insured patients ages 18-65 from MarketScan data 2007-2019: GLP1RA vesus thiazolidinediones and sulfonylureas, respectively to assess GLP1RA’s effect on incident asthma exacerbation. Follow-up started at the 2nd prescription for a maximum of 2 and 0.5 years for empirical study 1 and 2, respectively, after initiation, ignoring treatment changes. In each identified subgroup, we estimated propensity scores to balance confounders, then estimated adjusted risk difference (aRD) by inverse probability treatment weights.Our simulations showed iCF often outperforms other machine learning methods for interaction/subgroup identification. In empirical study 1, iCF successfully identified heterogeneous subgroups and showed patients with edema (proxy of heart failure) had the largest treatment effect (adjusted RD of -3.92%, 95% CI: -5.98% to -1.87%) from SGLT2i relative to GLP1RA. In empirical study 2, patients with 2+ ED visits had the largest benefit (aRD -3.89%, -5.68% to -2.09%) from GLP1RA relative to sulfonylurea. Our simulations demonstrated iCF often outperforms other methods for subgroup identification and for finding the subgroup with largest CATE. The 2 empirical studies both showed iCF successfully identified subgroups with different treatment effects. The iCF is a promising method to identify subgroups with HTEs in large cohorts where the potential for unmeasured confounding can be limited by study design.