Comparing the Effectiveness of a Longitudinal Model With Logistic Regression for Binary Endpoints to Support Better Clinical Decision Making

Exposure-Response (ER) analyses are key to optimizing dosing regimens for the safe and effective use of drugs. The choice of modeling approach should depend on the questions being asked and on the method's ability to appropriately characterize the ER relationship. While logistic regression (LR) is a common choice to evaluate ER relationships for binary endpoints, limitations can affect its reliability. Specifically, it cannot explicitly account for time-varying exposure, necessitating the use of summary exposure metrics that may bias its results. Longitudinal models can address some time-dependent factors known to affect LR analyses, but these require greater resource investment. As a result, clearer guidance is needed for selecting appropriate exposure metrics in LR and for determining when the additional complexities of longitudinal models are warranted. In this study, we compared Markov modeling with LR using systematically varied simulations designed to evaluate multiple strengths of the ER relationship, representative pharmacokinetics for small and large molecules, early and late event onset, high and low dose modification rates, different total sample sizes, and multiple exposure metrics. Overall, Markov modeling demonstrated a superior ability to identify ER relationships, especially those of intermediate strength and in smaller sample sizes, and tended to yield less biased estimates of the drug effect. Results did not suggest advantages of event-based metrics for LR. This study is the first to demonstrate the effectiveness of a longitudinal model relative to LR for ER analyses, and it introduces a simulation framework that guides method and exposure metric selection across common drug development scenarios.