Using Proportional Odds Model of Ordinal Logistic Regression to Rate National Reporter Panel

• Activity Number: 259
• Type: Contributed
• Date/Time: Monday, August 5, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Xuemei Pan*+
• Keywords: proportional odds model ; ordinal logistic regression ; model fit ; Maximum Likelihood Estimates ; Bootstrap resampling ; SAS 9

Abstract

Human behavior related performance outcomes often possess an intrinsic ordering. In this application, a sample of expert rated IBM maintained nationally dispersed reporters' behaviors and past performance was analyzed using the Proportion Odds Model. The parallel logit surfaces were tested. Validity of interpreting significant independent variables as predicting the outcomes were assessed. Bootstrap resampling technique was used to ensure the validity of model estimates. The predicted probability of each response category was utilized to assign cases to categories. The results revealed that the classification rate from the model matched well, with a nearly 90% match rate. An optimization method was used to transform the ordinal rating bucket to continuous score. This predicting model has been working smoothly and has been applied in the large national reporters' panel (more than 10,000 individuals) performance evaluation for over 4 years.