Recursive Partitioning with a Latent Response: Application to HIV Viral Genetic Data

• Activity Number: 585
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Shannon Stock*+ and Victor DeGruttola
• Companies: Harvard School of Public Health and Harvard School of Public Health

Abstract

Clinical studies evaluating the efficacy of antiretroviral (ARV) therapy in suppressing Human immunodeficiency virus (HIV) replication frequently use viral rebound to clinically indicate treatment failure. However, since HIV RNA measurements are obtained with measurement error, there is uncertainty whether an observed increase in viral load is due to measurement error rather than viral rebound. Therefore, analyses of such studies must accommodate a latent response. In this paper we present a recursive partitioning method for a latent event that uses weights estimated via maximum likelihood to estimate the probability an event occurred between successive (in time) pairs of observations, given an observed change in outcome measurements. The only parametric assumptions our method require relate to the estimation of the outcome measurements' measurement error distribution. We evaluate the performance of our method using simulation studies, and provide an application based on HIV viral genetic data using the AIDS Clinical Trials Group (ACTG) 398 clinical study.