Like It or Not, You Are Using One Value Representation

Do we use the same information to decide what we like and what we do not like? Best-worst scaling-where respondents select their most and their least preferred option from a set of options-is an efficient method for obtaining information of direct relevance to this question. Many best-worst scaling applications use multinomial logit (MNL) models to predict such best and worst choice data, explicitly or implicitly assuming that best and worst choices are driven by the same parameters for utility information. Some recent literature, however, has criticized this common practice as an overly simplistic representation of the choice process. We tested this assumption by applying three MNL-type models of increasing complexity in their parameterization to the stated best-worst choices from a total of 1,200 individuals drawn from five data sets. Our Bayesian latent mixture modeling found clear evidence that the same utility parameters drive individuals' best and worst choices, although usually with an additional scale parameter leading to more variable worst choices. These conclusions also held for stated best-worst choices of the same individuals for the same alternatives after a 6-, 12-, and 18-month delay. We argue that the conclusion of several recent papers that best and worst choices are driven by different utility information or reflect different decision processes is based on inadequate data and/or data analyses.