A Multiple Imputation Workflow for Handling Missing Covariate Data in Pharmacometrics Modeling

Covariate missingness is a prevalent issue in pharmacometrics modeling. Incorrect handling of missing covariates can lead to biased parameter estimates, adversely affecting clinical practice and drug development dosing decisions. Single imputation is usually favored by pharmacometricians for its simplicity, but it ignores the uncertainty about imputed values, potentially leading to biased estimates and standard errors. Multiple imputation, in contrast, generates multiple plausible values from a predictive distribution, addressing this uncertainty and thus is a preferable approach over single imputation to handle covariate missingness. Yet, its application in pharmacometrics remains limited due to perceived complexity. To address this, we developed a multiple imputation workflow specifically tailored for pharmacometricians, encouraging wider adoption of this more reliable method in pharmacometrics modeling. We compared single imputation and multiple imputation in estimating covariate effects using a publicly available dataset on warfarin pharmacokinetics in healthy volunteers. A one-compartment population pharmacokinetic model with baseline body weight as the only covariate was used to describe the warfarin pharmacokinetics. We simulated five scenarios in which 6.25%, 12.5%, 25%, 50%, and 75% of the subjects had their body weight missing under a missing at random mechanism conditioned on age and sex. We confirm that multiple imputation better reflects uncertainty estimates than single imputation, regardless of the degree of missingness. This confirms multiple imputation as a superior alternative to single imputation for handling missing covariate data in pharmacometrics.