A reduction in between subject variability is not mandatory for selecting a new covariate

Area of Science:

• Pharmacometrics
• Pharmacokinetics
• Pharmacodynamics

Background:

• Population pharmacokinetic-pharmacodynamic (PK/PD) analysis models population mean response and variability.
• Between-subject variance (BSV) quantifies parameter variability across individuals.
• BSV comprises predictable (BSVP) and random (BSVR) components; BSVP is explained by covariates.

Purpose of the Study:

• To investigate if significant covariates always decrease BSVR.
• To explore conditions where BSVR may increase despite covariate inclusion.
• To assess if specific covariate models can resolve BSVR anomalies.

Main Methods:

• Nonlinear hierarchical modeling using NONMEM (ver 7.2).
• Simulations performed in MATLAB (2011a) with a 1-compartment IV bolus PK model.
• Evaluated base models vs. six covariate models (nested, non-nested, interaction) with varying covariate-CL correlation (0-100%).

Main Results:

• BSVR can increase when a correlated covariate is added due to model misspecification.
• Non-nested covariate models (NNCM) showed negative BSVP (inflated BSVR) up to 50% correlation.
• Nested covariate models (NCM) showed biased BSVP; covariate-eta interaction models improved BSVP estimation.
• Perfect correlation (100%) led to similar performance across models.

Conclusions:

• Statistically misspecified covariate models can inflate BSVR, even with significant covariate correlation.
• Covariate-eta interaction models can improve the diagnostic assessment of covariate effects on variability.
• Careful model selection is crucial for accurate PK/PD analysis and covariate effect evaluation.