Local Odds Ratio Is More Efficient Than Correlation Coefficient For Modeling Longitudinal Ordinal Data

• Activity Number: 243 - Contributed Poster Presentations: Biopharmaceutical Section
• Type: Contributed
• Date/Time: Monday, July 31, 2017 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Xinkai Zhou* and Ronghui Xu and David Elashoff
• Companies: Statistics Core@UCLA and University of California, San Diego and University of California, Los Angeles
• Keywords: Longitudinal Ordinal Data ; Proportional Odds Cumulative Logit Model ; Generalized Estimating Equations ; Local Odds Ratio

Abstract

While correlation coefficient is commonly used for parameterizing the "working" correlation structure in generalized estimating equations (GEE) for modeling longitudinal ordinal data using the proportional odds cumulative logit model, it is well known that its range is severely constrained as a result of the Frechet bound. Although alternative parameterizations have been proposed, a direct comparison between them is lacking. Consequently, analysts usually fall back to the correlation coefficient method as the default option even though they are aware of its potential problems. To inform modeling choice, this paper conducted a simulation study and found that the correlation coefficient approach is not optimal in a wide range of scenarios. In fact, we found that the local odds ratio approach can achieve up to 30% efficiency gains (in a sense that will be defined in the article) compared to the correlation coefficient approach.