Effect of correlation on covariate selection in linear and nonlinear mixed effect models

Area of Science:

• Pharmacometrics
• Population Pharmacokinetics
• Statistical Modeling

Background:

• Covariate selection is crucial in population pharmacokinetic (PopPK) analyses for understanding drug behavior.
• High correlation between covariates can complicate accurate selection in mixed-effects models.
• Previous studies have not fully elucidated the impact of correlated covariates on model selection.

Purpose of the Study:

• To investigate the effect of correlated covariates on covariate selection in linear and nonlinear mixed-effects models.
• To determine how correlated covariates can be mistakenly identified as significant predictors.
• To explain variability in covariate identification across different PopPK analyses.

Main Methods:

• Monte Carlo simulations were used to generate concentration-time profiles with a single true covariate affecting oral clearance (CL/F).
• Demographic data from the National Health and Nutrition Examination Survey III database were utilized.
• Population pharmacokinetic models were built using univariate covariate analysis, comparing models with and without covariates, and using likelihood ratio tests (LRT) or AIC for selection.

Main Results:

• Highly correlated covariates (e.g., weight and body surface area, r=0.98) led to the selection of the incorrect covariate (body surface area) in up to 20% of simulations.
• In a second simulation involving drug metabolites and QTc interval prolongation, a metabolite was incorrectly selected as a predictor when only parent drug concentrations influenced the outcome.
• These findings demonstrate that strong correlation alone can lead to the selection of a non-causal covariate.

Conclusions:

• Correlated covariates can be erroneously selected as significant predictors in population pharmacokinetic and pharmacodynamic (PopPK/PD) analyses due to statistical artifacts.
• This phenomenon can explain discrepancies in covariate identification for the same drug across different studies.
• Careful consideration of covariate correlations is essential for robust and accurate PopPK/PD model building.