A comparison of multilevel ordinal regression models in the analysis of police force ratings

Abstract

In the literature several methods have been developed to model ordinal data while considering their natural ordering. However, this study sought to compare two possible link functions for the multilevel ordinal regression using males’ ratings of the police forces in Uganda as an outcome variable. Variables were obtained from the UNGBS database (Uganda National Governance Baseline Survey). The highest proportion of males rated the police as good (40.9%) followed by fair (24.96%), poor (19.1%), and lastly very good (15.1%). The multilevel ordered logistic regression model with both individual and contextual variables had the lowest AIC compared to other models, fitting the data best. All the likelihood ratio test results indicated that there was significant variation in males' ratings of the police forces across districts. Hence, males from the same district were significantly more similar compared to males from another districts. Researchers using data collected by applying multi-stage sampling or any form of nesting should consider multilevel or mixed-effects models.

Keywords

• logit
• ordinal regression
• police rating
• probit
• Uganda