Comparing heavy-tailed residual error models for outlier handling in population PK modeling

Area of Science:

• Pharmacokinetics
• Statistical Modeling
• Computational Biology

Background:

• Population pharmacokinetic (PopPK) estimation is challenged by outliers, particularly under Gaussian error models.
• Conditional weighted residuals (CWRES) filtering is a common but often insensitive method due to variance inflation and model masking.

Purpose of the Study:

• To evaluate the performance of different residual error distributions in handling outliers in PopPK models.
• To compare the robustness of Normal, Laplace, Generalized Error Distribution (GED), and Student's t distributions against outlier contamination.

Main Methods:

• Implemented a one-compartment model in Monolix with a custom likelihood workaround.
• Assessed CWRES sensitivity under extreme contamination.
• Utilized theoretical tail-behavior analysis, simulation studies, and a real-world caffeine PK case study.

Main Results:

• CWRES diagnostics proved unreliable, with extreme outliers often masked by inflated residual variance in the Normal model.
• Exponential-tail models (Laplace, GED) showed limited robustness against extreme outliers.
• The Student's t distribution demonstrated stable and minimally biased parameter estimates across various contamination levels.

Conclusions:

• CWRES-based outlier screening alone is insufficient for reliable PopPK analysis.
• The Student's t distribution offers superior stability and robustness in the presence of influential outliers compared to Normal, Laplace, and GED models.