Making Use of Respondent Reported Processing Information to Understand Attribute Importance: a Latent Variable Scaling Approach



Making Use of Respondent Reported Processing Information to Understand Attribute Importance: a Latent Variable Scaling Approach

Authors

S Hess, ITS, University of Leeds, UK; D Hensher, University of Sydney, AU

Description

This paper explores the use of a latent variable approach in an information processing context

Abstract

Stated choice surveys routinely collect information from respondents on their ?stated? way of processing the data presented to them in choice scenarios. Typically, this includes indications as to whether a given attribute was ignored and also the relative ranking of the different attributes. It has for a while been recognised that direct incorporation of such respondent-provided information not only exposes the analyst to the potential risk of endogeneity bias, but there has also been some evidence to suggest that there may not necessarily a perfect correspondence between the responses to such processing questions and the behaviour in the stated choice component. As an example, it has been observed that respondents who claim to have ignored a certain attribute still exhibit a significant albeit lower sensitivity to that attribute. A number of modelling approaches have been put forward towards using such information without the risk of producing biased results. In the present paper, we present a novel way of looking at this issue, based on a latent variable approach. In particular, we hypothesise that for each attribute and each respondent, the unobserved sensitivity to that attribute drives both the responses in the stated choice component, and the answers to the information processing questions. We represent this through a latent variable which we use in jointly modelling the stated choice component, the response to the ?ignoring? questions, and the reported importance ranking of attributes. The results from a route choice experiment show a clear link between the two sets of responses. More importantly, we show that this set of latent variables, which we can also relate to a number of socio-demographic characteristics, helps to explain a large part of inter-respondent heterogeneity in sensitivities, significantly reducing the remaining random heterogeneity component, in contrast to more simple mixed logit models estimated on the stated choice data only (i.e. also without the latent variables).

Publisher

Association for European Transport