Evaluation and Mitigation of Time‐Dependent Confounding Effects in Conventional Exposure‐Response Analyses for Oncology Drugs

Conventional exposure-response (E-R) analyses, such as logistic regression and time-to-event analysis using summary exposure metrics, are often conducted in oncology using data from the pivotal trial(s) with a single dose level to support dosing decisions. However, these E-R analyses, affected by multiple confounding factors, may mischaracterize true E-R relationships, potentially limiting their utility in dosing decisions. This study investigates potential mischaracterization in such analyses influenced by the following time-dependent confounding factors: exposure accumulation, dose modification patterns, and event onset time. We used a simulation-based approach to evaluate two E-R scenarios: ER1, where event time generated with a Weibull distribution is not affected by drug exposure, and ER2, where the response is driven by drug exposure via a joint PK-tumor size model. Our analyses indicate that when using time-dependent exposure metrics (e.g., average concentration till event/censoring), exposure accumulation tends to induce an inverse E-R trend, while dose modifications (interruptions/reductions) likely induce a positive E-R trend. Simulations suggest that employing static exposure metrics (e.g., first-cycle or steady-state) minimizes these biases. Additionally, adopting an Emax model aligned with the underground truth in ER2 in the E-R analyses could reduce bias. When significant dose modifications are present, including relevant data from a dose-range study and employing modified methods for time-dependent exposure derivation may help mitigate bias. Overall, using multiple exposure metrics (including static ones) to assess E-R consistency and interpreting the potential causal effects with totality of evidence (including dose-response results) should be implemented to better inform dosing decisions.