(S-069) Evaluation and Mitigation of Time-Dependent Confounding Effects in Conventional Exposure-Response Analyses for Small-Molecule Oncology Drugs

Xuefen Yin – Center for Pharmacometrics and Systems Pharmacology, Department of Pharmaceutics, College of Pharmacy, University of Florida, Orlando, USA.; Ye Xiong – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA.; Youwei Bi – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Xin Wei – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Hong Zhao – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Elimika Fletcher – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Rajanikanth Madabushi – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Amal Ayyoub – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Hao Zhu – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA; Stephan Schmidt – Center for Pharmacometrics and Systems Pharmacology, Department of Pharmaceutics, College of Pharmacy, University of Florida, Orlando, USA; Jiang Liu – Center for Drug Evaluation and Research, U.S. Food and Drug Administration, MD, USA.

Objectives: Conventional exposure-response (E-R) analysis methods for cancer drugs (i.e., logistic regression and time-to-event analysis typically relying on summary exposure metrics from a single dose level) may mischaracterize true E-R relationships and potentially misguide the dose selection decision, particularly when dose modifications are common. This study investigates how time-dependent confounding factors including accumulation of exposure, dose modification patterns, and event onset timing may distort E-R characterizations in such analyses. In addition, we tested several potential strategies to decrease bias and, thus, enable robust E-R evaluations for dose optimization.

Methods: We utilized a simulation-based approach to evaluate two ER scenarios: ER1, where responses generated using Weibull distributions are not affected by the drug exposures (no drug effect or effect plateaued), and ER2, where responses are driven by drug exposures simulated through a joint PK-tumor size model with a sigmoidal Emax effect on tumor shrinkage. For both scenarios, we simulated virtual trials with varying drug half-lives (short, moderate, or long), dosage and dosage ranges, sample sizes, dropout rates, and dosing strategies (constant, dynamic, or empirical dosing). Subsequently, we evaluated the direction and magnitude of E-R mischaracterization across static and time-dependent exposure metrics using conventional E-R analysis methods. In ER1, Type I error (significant E-R slope, p < 0.05) was summarized given no E-R relationship. In ER2, Emax model in logistic regression was assessed considering the typical sigmoid E-R shape. To mitigate E-R mischaracterization, we tested dose range adjustments and a novel adaptive censoring method developed based on identified sources of bias.

Results: Our analyses indicate that when using time-dependent exposure metrics, e.g., average concentration till event/censoring, PK accumulation tends to induce an inverse E-R trend, while dose modifications likely induce a positive E-R trend. Employing static exposure metrics (e.g., 1st cycle or steady state) can minimize these biases. Additionally, adopting an Emax model aligned with the underground truth in ER2 could reduce bias. When significant dose modifications are present, incorporating a second lower dose arm and employing an adaptive censoring method may mitigate bias, though determining the optimized dose range remains challenging.

Conclusions: Our results suggest using multiple exposure metrics, including static ones, to assess E-R consistency; when uncertainty is high, then validating causal effects through a randomized trial with multiple dose levels can improve the robustness of the E-R analyses.

Citations: N/A

Keywords: exposure-response, time-dependent confounding factors, mitigating E-R mischaracterization