Structural Limitations of Early–Late Treatment Comparisons Under Target Trial Emulation: A Simulation Study

Le résumé fourni par la source

Early versus Late (E-L) treatment initiation comparisons, which are prone to immortal time bias, are widely considered amenable to causal inference through target trial emulation (TTE) using clone-censor-weighting (CCW). However, whether standard CCW yields a well-defined causal estimand for the E-L comparison under ideal implementation conditions has not been formally examined. We conducted a structural simulation study to evaluate CCW-based TTE for E-L comparisons. The design varied along three dimensions: treatment-initiation mechanisms (random versus baseline-confounded timing), benefit-harm structures governing delayed-treatment harm and post-treatment benefit, and five distinct E-L initiation-window configurations. We implemented CCW as conventionally recommended and summarized estimated 12-month marginal risk ratios across settings. Depending on the placement and structure of the initiation windows, estimated risk ratios exhibited systematic window-dependent variation, including direction reversals across window specifications, even in settings where the true marginal causal contrast was 1 by design. These patterns were not consistently explained by weight instability or lack of support overlap. Our findings indicate that, in E-L comparisons, the window-defined strategy contrast estimated using CCW-based TTE can be strongly shaped by risk accumulated before treatment initiation under the specified windows, underscoring the need for caution when interpreting E-L analyses.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.