A COMPARISON OF REGRESSION-MODELS FOR THE ANALYSIS OF ORDERED CATEGORICAL-DATA

In a clinical trial on the treatment of lung cancer two chemotherapeutic strategies were compared with respect to tumour response, which was assessed under four ordinal categories. Two regression models for the analysis of ordered categorical data, the proportional odds model and the stereotype model, are used to analyse the data. Differences between the models with respect to estimation and significance tests are further investigated using simulation studies. The results indicate that estimates are less biased in the proportional odds model, but in the stereotype model an improvement may be achieved if adjacent categories are amalgamated. Similar behaviour is observed with respect to tests concerning regression coefficients which are of primary interest in most clinical trials. While in the proportional odds model no relevant deviation from the nominal level of significance is observed, comparable results are obtained in the stereotype model only after amalgamation of categories.