Informative Dropout and Visual Predictive Check of Exposure-Response Modeling of Ordered Categorical Data
C. Hu(1), P. Szapary(1), N. Yeilding(1), H. Zhou(1)
(1)Centocor Research & Development, Inc., Malvern, PA, USA
Objectives: Physician's Global Assessment (PGA) score is a 6-point measure of psoriasis severity (0=cleared and 5=severe). PGA scores were collected from two Phase 3 studies (PHOENIX 1 and PHOENIX 2) of ustekinumab in patients with moderate-to-severe psoriasis [1]. The objective of this study was to develop an exposure-response model for PGA scores and to (1) account for potential dropout influence on model predictions, (2) conduct accurate visual predictive check (VPC) in light of uncertainties in future dosing regimens, and (3) understand the predictive ability of the model.
Methods: A novel joint longitudinal-dropout model was developed. The longitudinal component extended a previous approach using a latent variable semi-mechanistic drug model and placebo effect under the mixed-effect logistic regression framework to model prob(PGA≤k), k=0, 1, ..., by incorporating disease progression. The dropout component extended a previous approach [2] to categorical data, and used the flexible Weibull hazard function. Sequential PK/PD and limited simultaneous estimations were implemented in NONMEM. VPC needs to account for informative dropout. However, simulating dropouts requires the knowledge of future doses beyond the last observation, and assuming the nominal dosing regimen may create biases especially if dose titration is present. A conditional approach, treating the observed data trend as conditional on observed dropouts, was developed to overcome this obstacle. It was then extended to external model validation of the longitudinal component of the joint model. External validation of the dropout component was assessed qualitatively using Kaplan-Meier plots.
Results: 1,995 Patients contributed to 19,340 ustekinumab concentration measurements and 41,668 PGA scores from studies PHOENIX 1 and 2 collected over 2 years, with 373 dropouts. An informative dropout model with Weibull hazard best fitted the data. Conditional VPC results confirmed the differences between the informative dropout model and other models. External validation showed that the prediction errors were small (<3%) in the treatment optimization period but larger (6%) for an extrapolation period.
Conclusions: An informative dropout exposure-response approach was developed to model PGA scores as ordered categorical data. The conditional VPC approach, with no dependence on unknown future dosing regimens, is useful for accurately evaluating informative dropout models and model validation.
References:
[1] C. Hu, P. Szapary, N. Yeilding and H. Zhou, Informative Dropout Modeling of Longitudinal Ordered Categorical Data and Model Validation: Application to Psoriatic Patients Treated with Ustekinumab. Journal of Pharmacokinetics and Pharmacodynamics, in press.
[2] C. Hu and M. Sale, A joint model for nonlinear longitudinal data with informative dropout, Journal of Pharmacokinetics and Pharmacodynamics, 30(1): 83-103, 2003.
