Enhancing Pharmacometric Modeling with Full Bayesian Inference and Students t-Based M3 Censoring: A Simulated Population PK Study on Robust Handling of Outliers and Censored Data

Abstract

Pharmacometric modeling and simulation, with population pharmacokinetic (PopPK) modeling as a key component, play a crucial role in characterizing drug disposition and variability, supporting dose optimization and regulatory decision-making. Traditional maximum likelihood estimation (MLE) methods, although efficient, are sensitive to outliers, particularly when datasets contain observations below the limit of quantification (BLQ). As clinical datasets become increasingly complex, more robust modeling approaches are needed. In this study, a simulation-based evaluation was conducted to compare four modeling strategies combining different residual error structures (normal vs. Students t) and censoring methods (M1 vs. M3). Two-compartment pharmacokinetic profiles were simulated for fifty subjects, incorporating varying degrees of outlier contamination and BLQ data. Full Bayesian inference using Markov Chain Monte Carlo (MCMC) methods was employed to estimate posterior distributions of PK parameters at both the population and individual levels. Notably, the combination of Students t-distributed residuals with M3 censoring (Students t_M3) consistently produced the most accurate and precise parameter estimates across all simulation scenarios, even under extreme outlier contamination and substantial BLQ presence. The combination of Students t-distributed residuals with M3 censoring within a Bayesian framework offers a robust and resilient strategy for PopPK modeling, effectively addressing both outlier contamination and data censoring challenges. These findings support the broader adoption of robust Bayesian modeling techniques in pharmacometric practice, particularly for complex and irregular clinical datasets such as cell and gene therapies.