A Joint Model for Nonlinear Longitudinal Data with Informative Dropout

Abstract

Subject withdrawal from a study (also called dropout, or right censoring), is common in late phase clinical trials. A number of methods dealing with dropouts have been used in practice, the most common being “last observation carried forward” (LOCF). Many of these methods, including LOCF, can result in biased estimates of the efficacy or potency of the drug, especially in the modeling context. If the likelihood of dropout is correlated to the underlying unobserved data, the dropout is informative and should not be ignored in the modeling process. The topic of informative dropout in the context of longitudinal data has received much attention in the statistical literature, in the setting of linear and generalized linear models. We extend the approach to nonlinear models. The dropout hazard, as well as the longitudinal data, is modeled parametrically. Parameters are estimated by maximizing the approximate joint likelihood as implemented in the software NONMEM. Using data from actual clinical trials, we explore the impact of the dropout model on the ability of the joint model to predict observed longitudinal data patterns.