Interpreting discrete choice models based on best-worst data: A matter of framing

Abstract

Best worst choice response tasks have become increasingly popular as a means of increasing the amount of information captured from respondents undertaking stated preference experiments. In analysis, best worst data is often exploded to provide additional pseudo observations which may aid in model estimation. Recent studies however have questioned many of the underlying assumptions which typically accompany best worst studies, such as the symmetry of preferences across the best and worst responses as well as assumptions about equal error variances across the two response types. This paper first provides a detailed description of the various best worst tasks that have appeared within the literature before arguing that violations of preference symmetry and homogeneity of error variance should be the norm. This is because in asking respondent to choose their most and least preferred option out of a set of alternatives reflects different response frames, one positive and one negative, and behaviourally there exists no reason why one would assume that the preferences (and error variances) obtained from one type of question should precisely mirror that of the other. Using an empirical case study, the impact of the framing of these questions is examined. Finally, the argument put forward is that best-worst data should be treated in a similar to data fusion methods, where one combines two different sources of discrete choice data.