Comparison of Regression Models for Ordinal Outcome Data

• Activity Number: 215
• Type: Contributed
• Date/Time: Tuesday, August 13, 2002 : 12:00 PM to 1:50 PM
• Sponsor: Section on Statistics in Epidemiology*
• Author(s): Barbara Neas*+ and Binitha Kunnel
• Affiliation(s): University of Oklahoma Health Sciences Center and University of Oklahoma Health Sciences Center
• Keywords: ordinal outcomes ; proportional odds ; continuation ratio ; power

Abstract

Although ordinal scales are common in medical research, ordinal logistic regression has not been widely used. Ordinal scales are often treated as nominal or the ordinal outcome reduced to a binary one. Methods that reduce ordinal scales to nominal or dichotomous ones or that assume ordinal scales have the properties of interval scales may lead to erroneous statistical inferences. The purpose of this research was to investigate two ordinal models, the proportional odds model (POM) and the continuation ratio (CR) model and to compare their properties using simulated data. The specific aims were to: 1.) create simulated data representing a mixture of continuous and nominal covariates; 2.) compare the coefficients produced by the two models; 3.) compare the power of the respective models; and 4.) compare the predicted probabilities of the two models. The CR model coefficients were about 25% smaller than the POM coefficients. Overall, the CR model had slightly higher power to detect a significant coefficient, 5%-8%. The predicted probabilities were nearly identical for the two models under all conditions tested. The POM performed slightly better than the CR model for small sample sizes.