Time to Event modeling of dropout event in clinical trials
Andrew C. Hooker (1), Gomeni R (2), Stefano Zamuner (2)
(1) Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden; (2) Clinical Pharmacology Modeling & Simulation, GlaxoSmithKline, Verona, Italy.
Background: Dropouts can greatly affect the outcome of clinical trials and a proper understanding of the dropping-out mechanisms is critical for the correct interpretation of the outcomes of clinical trials. Among the different reasons for dropout are the lack of efficacy, the occurrence of unwanted effects and that the exposure-response parameter estimates can be inappropriate in presence of non-random dropouts.
Objectives: 1) To develop a parametric time-to-event model for dropout investigating the potential effect of both efficacy and PK as significant covariates; 2) To develop a joint model accounting for PK/efficacy and dropout [1]
Methods: Drop out data from two clinical trials (in the neuroscience therapeutic area) were analyzed using a parametric time to event model. Several parametric descriptions of the hazard function were evaluated (exponential, Gompertz, Weibull including a cure rate term). Treatments (i.e., placebo, test drug, and active comparator), PK and clinical endpoints were investigated as potential covariates. The parameters were estimated maximizing the joint likelihood using the laplacian approximation as implemented in NONMEM VI. Model selection was based on the log-likehood ratio test; in addition visual predictive checks (VPC) were used to evaluate model performance using simulated vs non parametric estimates of survival (Kaplan-Maier plots). Finally, a joint model including efficacy and/or PK and time to event data was attempted.
Results: The Weibull model including a cure rate term was the best model to describe the dropout data. Time to Event analysis on dropouts showed a significantly higher drop-out rate for the test drug (approximately 40%) compared to both placebo and active comparator (24 and 22% , respectively). Overall PK exposure (AUC) during treatment was not found to be a significant covariate to explain the probability of drop-out in the test drug arm. Results including the efficacy as covariate showed a great improvement in the objective function suggesting that lack of efficacy was one of the main reason of drop out.
Conclusions: The effect of drop out event is critical and needs to be properly considered in the development of PK and PK-PD models. Parametric time to event models are suitable for this description. All this should be properly implemented before embarking in clinical trial simulation work.
References:
[1] Hu C, Sale M. A joint model for nonlinear longitudinal data with informative dropout. J Pharmacokinetic Pharmacodynamic 2003; 30:82-103.