Inflated discrete Beta regression models for Likert and discrete rating scale outcomes

Discrete ordinal responses such as Likert scales are regularly proposed in questionnaires and used as dependent variable in modeling. The response distribution for such scales is always discrete, with bounded support and often skewed. In addition, one particular level of the scale is frequently inflated as it cumulates respondents who invari- ably choose that particular level (typically the middle or one extreme of the scale) without hesitation with those who chose that alternative but might have selected a neighboring one. The inflated discrete beta regression (IDBR) model addresses those four critical characteristics that have never been taken into account simultaneously by existing models. The mean and the dispersion of rates are jointly regressed on covariates using an underlying beta distribution. The probability that choosers of the inflated level invariably make that choice is also regressed on covariates. Simulation studies used to evaluate the statistical properties of the IDBR model suggest that it produces more precise predictions than competing models. The ability to jointly model the location and dispersion of (the distribution of) an ordinal response, as well as to characterize the profile of subject selecting an ”inflated” alternative are the most relevant features of the IDBR model. It is illustrated with the analysis of the political positioning on a ”left-right” scale of the Belgian respondents in the 2012 European Social Survey.