Improve the Prediction of Landmark Event Time by Incorporating Actual Enrollment Data

• Activity Number: 45
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Yong Zhang* and Zhichao Sun and Kalyanee Appanna and Cheng Zheng and Kaushal Mishra and Feng Tai and Can Cai
• Companies: Novartis Pharmaceuticals and University of Michigan and Novartis Pharmaceuticals and Novartis Pharmaceuticals and Novartis Pharmaceuticals and Novartis Pharmaceuticals and Novartis Oncology
• Keywords: prediction of landmark event time ; Bayesian prediction model ; auto-regressive model ; moving average model

Abstract

In event-based clinical trials, it is important to predict early and accurately a landmark event time. A targeted event time is determined by three factors: the accrual process, the event process and the loss to follow-up process. Bagiella and Heitjan (2001) proposed a Bayesian prediction model for landmark event times by assuming exponential distributions for the above three processes when treatment arms are known. Donovan et al. (2006) developed a parametric mixture model with a known mixture proportion under the Bayesian framework when the treatment arms are masked. Both methods assumed a simple exponential distribution and constant enrollment rate for the accrual process, however often times these assumptions do not hold. For example, many trials may have gradually increasing enrollment rates and some trials may have U-shape enrollment rates. In this work, we try to improve predictive performance by using an auto-regressive model or moving average model for the accrual process. Simulation studies and data examples show that our proposed methods outperform the simple average model if the enrollment rate changes over time, especially when enrollment is still actively ongoing.