A Mixture Model for Scaling Preferences Under Response-Style Heterogeneity

• In Event: Comparative Politics Applications of Latent Variable Models

Abstract

Preference items ask survey respondents to rate objects such as parties or candidates, often on thermometer or propensity-to-vote scales. Multidimensional scaling summarizes such responses with low-dimensional spaces that are sometimes interpreted in ideological or policy terms. Existing methods assume that (1) all the respondents use the scale in the same way, and that (2) all the responses are informative. Neither may always be the case. The former assumption is violated for example when some respondents ignore some of the points on the scale. The latter assumption is violated for example when some respondents use the mid-point of the scale to give “don’t know” answers. This paper proposes a mixture unfolding model that relaxes both assumptions. The model sorts respondents into groups with different response styles and identifies and downweights non-informative responses. Furthermore, a generalization of the model that allows for between-respondent multidimensionality provides an alternative method for the measurement of shared cultural schemas. Compared to Relational Class Analysis (RCA) (Goldberg 2011) and Correlational Class Analysis (CCA) (Boutyline 2017), the proposed approach more closely fits the theory of cultural schemas, provides uncertainty estimates, and allows to incorporate covariates and extra-data information. The approach is demonstrated on propensity-to-vote batteries from the European Election Studies.

Author

• Juraj Medzihorsky, London School of Economics & Political Science