Hierarchical Hidden Markov Models for HIV-Transmission Behavior Outcomes
Date(s) - Oct 14, 2008
3:15 PM - 4:15 PM
Center for Community Health
Li-Jung Liang, Ph.D.
Longitudinal assessments of sexual behaviors are common in studies of cognitive behavioral interventions designed to reduce HIV-transmission behaviors. These studies elicit frequencies of various sexual behaviors over a previous unit of time such as three months. Study participants may not engage in some transmission or transmission-associated behaviors at several time points, leading to zero-inflated data that is inadequately described by a unimodal count distribution. This data is commonly modeled by a zero-inflated Poisson distribution. However, for sexual behaviors, the distribution of non-zero counts is not well-described by a single Poisson distribution. For example, for three month recall of the number of sex partners, many study participants reported either zero or one sexual partner (abstinent or monogamous), some reported counts in the low single digits (multiple partners), but there were several reports of partner counts in the high 2 digits, or even in the 100s (extreme). We propose a Bayesian approach where the sexual behavior states (abstinent, monogamous, multiple partners, extreme) are modeled as the hidden states of a continuous time hidden Markov model to describe the history of sexual behavior. The effects of covariates such as intervention and HIV serostatus on transitions between sexual behavior states are incorporated using a generalized regression framework. Models are derived for use on a motivating data set from a study designed to reduce HIV-transmission and improve the health of a large HIV positive cohort recruited from metropolitan cities.
*The research being presented is a joint work with Rob Weiss and Scott Comulada.
Li-Jung Liang, Ph.D. (UCLA Biostatistics, 2005), is Assistant Professor of Medicine Statistics Core at the UCLA School of Medicine and Senior Statistician at the UCLA Semel Institute Center for Community Health. Prior to her doctoral studies, she worked in the pharmaceutical and biotechnology industries for over 10 years. Her dissertation was on Bayesian methodology for hierarchical regression models, and she has experience applying Bayesian statistics, longitudinal and repeated measures methodologies, and multi-level modeling to a broad range of health related research problems, including behavioral and social sciences, clinical trials, health services research, HIV/AIDS, and basic sciences.