Presentation (CTI)

OBJECTIVES: While real-world data (RWD) are increasingly used for causal inference in observational studies, conventional regression methods like propensity score matching can be sensitive to model misspecification, potentially leading to biased estimates of treatment effects. This study evaluated the average treatment effect (ATE) of interventions and comparators on 1-year in-hospital mortality among heart failure (HF) patients using Targeted Maximum Likelihood Estimation (TMLE), a doubly robust causal inference framework designed to address limitations of traditional approaches in administrative claims database studies.
METHODS: Leveraging Japan’s IQVIA Claims Database, which is large-scale and covers all ages, 102,835 adult HF patients initiating two different therapies (drug A vs. drug B) were identified and their observed characteristics including demographics, prevalent comorbidities, procedures, and number of foundational therapies at baseline were included as covariates. The TMLE-based methods, which integrate ensemble of machine learning algorithms and semiparametric efficiency theory, were employed to mitigate bias from model misspecification through targeted bias reduction steps when estimating ATE and relative risk (RR).
RESULTS: Unadjusted 1-year in-hospital mortality was 19.4% in drug A group (n = 1,201) and 8.96% in drug B group (n = 101,634), respectively. SuperLearner (ensemble of generalized linear model and XGboost) was employed to derive efficient and consistent estimates of statistical parameters. As a result, the ATE of drug A versus drug B was 0.074 (95% CI: 0.050, 0.098), and the RR was 1.82 (95% CI: 1.58, 2.11) using a general TMLE method. Further by leveraging doubly robust TMLE method to safeguard against model misspecification, the adjusted ATE of drug A versus drug B was 0.067 (95% CI: 0.046, 0.088), and the adjusted RR was 1.74 (95% CI: 1.38, 2.20).
CONCLUSIONS: TMLE-based methods provide a less biased and robust strategy for causal inference in observational studies, enhancing the reliability of effect estimates when using RWD for rigorous clinical and regulatory decision-making.