ABSTRACT

If the categories of the response variable form a natural order, the researcher has the option of fitting an ordinal logistic regression model. Ordinal variables, as they are known, arise in different contexts. Some may simply be less fine gradations of an underlying continuous variable, such as ten-year age bands. A very common way of generating ordered variables is to elicit people’s attitudes, beliefs or levels of satisfaction about a particular issue or product where views may be ranked from strongly negative to strongly positive. In such situations, as with many latent variables, there may be an underlying continuous variable but we do not know the metric, that is, the distance between adjacent levels. Other variables have a natural order but no clear metric. Examples here included educational attainment, which can range from no educational qualifications to obtaining a higher degree. Sentence awarded (and the subject of analysis in sections 3.1 and 6.2) might be considered to reflect a scale of severity of punishment, with discharge as the least severe and custody as the most severe penalty. Although subsequent analysis (not reported here) indicated that sentence should not be regarded as an ordered but a multinomial variable.