Multivariate Zero-Inflated Proportional Odds Model (MZIPO): A Joint Latent Variable Approach to Scale Development in the Presence of a Preponderance of Zeros

• Activity Number: 48
• Type: Invited
• Date/Time: Sunday, July 29, 2012 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Steffanie Marie Halberstadt*+ and Mary D. Sammel
• Companies: The EMMES Corporation and University of Pennsylvania
• Keywords: Latent class ; Latent variable ; Multiple indicator multiple cause model ; Zero-inflation

Abstract

A salient issue in clinical research is the use of patient reported outcomes (PROs). Often expressed as items in a scale, PROs measure responses to unobservable constructs such as disease severity. Latent variable (LV) methods can simultaneously model the relationship among individual PRO items and compare PROs with clinical measures. It is also necessary to determine which PRO items to include in a scale and whether the scale is best scored as a LV estimate or a sum of items. When many subjects do not endorse particular items, PRO responses exhibit a zero-inflated distribution. Our objectives include concurrently evaluating individual items and validating the scale by comparing it to a clinical measure while accounting for excess zeros. We propose a multivariate zero-inflated proportional odds (MZIPO) model, which incorporates a latent class component into a joint LV framework. The model divides subjects into two unobserved subgroups according to "true" symptom susceptibility. For the susceptible group, the model provides estimates of correlation between items and a latent measure of disease severity as well as association between the latent measure and a clinical measure.