Order Matters (?): Alternatives to Conventional Practices for Ordinal Categorical Response Variables

Social scientists, particularly political scientists, frequently use ordinal survey items as dependent variables in models of political attitudes. Commonly, normal-theory modeling strategies like ordinary least squares regression are applied to these items. Additionally, workers also make frequent use of the proportional odds (ordinal logit) model or cumulative probit model when working with such items. Often, scholars opt to report linear regression estimates in lieu of alternative estimates on the grounds that regression results are “easier to interpret.” In this paper, we consider the implications of this strategy and present several alternative modeling strategies in the case when a researcher is working with ordinal kinds of survey items. In so-doing, we consider models that relax certain key assumptions of the ordinal logit model and models for nominal data that can be usefully applied to seemingly ordinal response variables.