Zero-Inflated and Zero-Altered (Hurdle) Models in HIV Therapy and Behavior Adherence Research

• Activity Number: 657
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Biopharmaceutical Section
• Author(s): Charles Rose*+
• Companies: CDC
• Keywords: Zero-inflated Poisson (ZIP) ; Zero-altered Poisson (ZAP) ; Hurdle model ; HIV ; Adherence data

Abstract

HIV therapy and behavior adherence data are usually continuous, non-negative, and left-skewed with a high percentage of values at 100% but can be right-skewed transformed by subtracting percent adherence from 100%. This transformation typically results in a high proportion of zeroes (0% non-adherence). Zero-inflated and zero-altered (hurdle) models have been developed in many fields to model count and binomial data with excess zeroes but have been little used in HIV therapy research. Zeroes generally are classified as structural (subject not "at-risk") and sample (subject "at-risk"). Zero-inflated models are increasingly common but the focus is usually on the model fit compared to conventional models and not on the interpretation and adequacy of these models. Scant research has focused on the adequacy of the zero-inflated Poisson (ZIP) model for the estimated structural and sample zeroes or on parameter interpretation. We use HIV therapy and behavior adherence simulated data to fit and compare ZIP, zero-altered Poisson (ZAP), and standard Poisson models. We focus on interpretation of the ZIP and ZAP models and adequacy of the ZIP model for describing structural and sample zeroes.