Identification and Estimation of Latent Attitudes and their Behavioral Implications

Abstract

This paper (i) formalizes conditions under which a population distribution of categorical responses to attitudinal questions (‘items’) has a scale representation; (ii) develops tests for whether a particular sample of item responses is consistent with a scale representation; (iii) develops methods for nonparametrically estimating the relation between an outcome and a scale value; and (iv) generalizes the foregoing to the multi-scale case. An implication of these results is that the effect of multiple latent attitudes on behaviour can be identified, even though the attitudes of an individual can never be precisely observed. We illustrate our methods using survey data from the 1992 U.S. Presidential election, where the ‘outcome’ is an individual’s vote and the ‘items’ are expressions of social and policy preferences