Assessing the Impact of Multiple Imputation Algorithms on Pharmacokinetic Model Performance: A Simulation-Based Study

Abstract

This study compared multiple imputation (MI) algorithms in a one-compartment pharmacokinetic (PK) scenario with oral absorption. Four covariates (two continuous, two dichotomous) linked to PK parameters were randomly removed under a missing completely at random (MCAR) mechanism. The aim was to identify which algorithm best preserves covariate distributions and PK parameter estimates. The original dataset included 100 individuals, each with five sampling occasions. Missing data were introduced at 5%, 20%, 50%, and 75% for the four covariates under the MCAR assumption. Five MI algorithms (Mice, Amelia, missForest, rMIDAS, XGBoost) were tested. Absolute and relative errors and concordance metrics were used to assess performance. Population and individual parameter estimates were compared across imputed and original datasets using Monolix2024R1®. MissForest (MF) and Amelia yielded lower errors for continuous covariates whereas dichotomous variables were poorly imputed. Based on objective function values, Mice perform best at 5% and MF at 20% of missingness. Increasing missingness decreased covariate effects and increased the estimated inter-individual variances. Individual parameter estimation accurately captured individual-level variability across all imputed datasets. MI methods appear effective for covariate imputation in PK modeling, offering reliable results up to 20% missingness under an MCAR mechanism. Future research should explore refined strategies, including advanced modeling frameworks and Bayesian approaches for imputation. Enhancing our understanding of missing data processes will be crucial for robust PK analyses across diverse clinical settings.

Keywords: PFIM; monolix; mrgsolve; multiple imputation algorithms; pharmacokinetics.