Missing Data in Model‐Based Pharmacometric Applications: Points to Consider

63 2010 50 63S-74S 63S M data are a ubiquitous complication associated with the repeated-measures longitudinal studies that are typically analyzed in pharmacometric applications. The extent of missing data differs across diseases and end points and can be manifested in many forms, such as missing response values (eg, due to dropout), censored observations (time to event data, data below analytical quantitation limits), and missing covariate values. While developing mathematical models of longitudinal pharmacokinetic, pharmacodynamic, disease progression, and clinical outcome end points, pharmacometricians are faced with the challenges of: (1) identifying missing data, (2) developing quantitative modeling strategies to appropriately accommodate the missing data, (3) checking model performance in the presence of missing data, and (4) investigating the sensitivity of model-based conclusions to missing data assumptions. In pharmacometrics, models are used for a variety of purposes, including but not limited to estimation of parameters, population description, model checking, prediction, and clinical trial simulation. For example, one might ask what the typical clearance, EC50, and Emax of a drug are. What effects do weight, age, and genotype have on clearance and associated unexplained random variability? Is Emax dependent upon severity of disease at baseline? Extrapolation of response to other dosing regimens might be simulated. Models are also used to optimize and inform study designs for purposes of parameter estimation and other trial operating characteristics. The extent to which one needs to address missing data depends on the planned use of the model and the likely missing data mechanism (MDM). Development of an appropriate modeling strategy in the face of missing data is essential for the reliable use of the derived models for any pharmacometric modeling/ simulation goal.