Handling missing data in a duloxetine population pharmacokinetic/pharmacodynamic model - imputation methods and selection models.

Abstract

In pharmacokinetic (PK)/pharmacodynamic (PD) modelling and simulations (M&S), omitting dropouts can cause inaccuracies in parameter estimation and clinical trial simulations (CTS). This study examines the impact of different imputation methods for missing data on the interpretation of model results, as well as develops a selection model (where dropout and efficacy are jointly modelled) for use in CTS. Missing data were imputed using single and multiple imputation and pattern mixtures methods for a previously reported duloxetine PK/PD model. The probability of dropout was described in the selection model and CTS was conducted with a hypothetical drug to examine the impact of dropout on trial results. The study completion rate was 75% and dropouts were not random. Model parameters obtained with different imputation methods were mostly within 40% (range 0 to 63%) compared to the model without dropouts. CTS showed 0.3 points lower median pain scores and 3% lower coefficient of variation over the 12-week simulations when dropout was included. Missing data had little impact on the original population PK/PD analyses. Sensitivity analyses for dropouts should be conducted in M&S exercises. The utility of selection models in CTS was explored via a hypothetical case study.