Sparse Bayesian Semiparametric Predictive Modeling with Applications in Unobserved Dose-Response Prediction

• Activity Number: 122
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Ben Haaland*+
• Companies: Duke University
• Keywords: Bayesian predictive modeling ; functional data analysis ; low-dimensional representation ; prior specification ; sparsity ; Gibbs sampling

Abstract

Statistical scientists are often confronted with the task of making informed predictions which will potentially have substantial financial and ethical impact in situations which are simultaneously novel and expected to be similar to a few relatively well-studied scenarios. Many of these situations are distinguished from typical regression-based prediction by the importance of expert information and the combination of only a few similar scenarios with high dimensional inputs and outputs. Here, we focus on a particular version of this problem that presents itself in the pharmaceutical industry. Predicting both side effect and endpoint dose-responses before the initiation of a clinical trial has enormous ethical and financial importance in the pharmaceutical industry. A sparse Bayesian semi-parametric model for predicting unobserved clinical dose-response curves conditional on preclinical data, data from similar compounds, and prior knowledge is proposed. Posterior sampling is achieved through a computationally efficient Gibbs sampler, allowing straightforward incorporation into a risk assessment model. The model is applied to actual data from the pharmaceutical industry.