Abstract
Commonly used discrete choice model analyses (e.g., probit, logit and multinomial logit models) draw on the estimation of importance weights that apply to different attribute levels. But directly estimating the importance weights of the attribute as a whole, rather than of distinct attribute levels, is challenging. This article substantiates the usefulness of partial least squares structural equation modeling (PLS-SEM) for the analysis of stated preference data generated through choice experiments in discrete choice modeling. This ability of PLS-SEM to directly estimate the importance weights for attributes as a whole, rather than for the attribute’s levels, and to compute determinant respondent-specific latent variable scores applicable to attributes, can more effectively model and distinguish between rational (i.e., optimizing) decisions and pragmatic (i.e., heuristic) ones, when parameter estimations for attributes as a whole are crucial to understanding choice decisions.
Introduction
Understanding why individuals make certain decisions that entail a discrete choice—such as purchasing from a particular retailer, while not purchasing from an alternative retailer; buying one brand rather than another; or accepting one employment offer but not another—is crucially important in business (Hensher et al. 2015; Louviere et al. 2008). Irrespective of whether an individual engages in rational, optimizing decision-making or in pragmatic, heuristic decision-making, the ensuing intention concerning, or choice that pertains to, a particular alternative is based on an assessment of the attributes of each alternative and the individuals’ preferences for such attributes. A variety of analytical approaches have been applied; aimed at trying to specify the relative impact of these attributes empirically so that to understand and predict choices based on these attributes. For example, discrete choice modeling (DCM; Louviere et al. 2000) has been applied to understand the impact of revenue management and loyalty program attributes on travelers’ purchasing choices (Mathies et al. 2013), of value creation and value appropriation attributes on managers’ outsourcing choices (Lin et al. 2016) and of location attributes on foreign direct investment choices (Buckley et al. 2007). Then, as one of the alternative approaches, partial least squares structural equation modeling (PLS-SEM; Lohmöller 1989; Sarstedt et al. 2017a; Wold 1982) has been employed to explain and predict the impact of attributes such as expected return and asset familiarity on choice of investment portfolio (Seetharaman et al. 2017), ease of use and trustworthiness on intentions to use consumer-generated media for travel planning (Ayeh et al. 2013) and price and convenience on, ultimately, intentions to purchase travel online (Amaro and Duarte 2015). Notwithstanding other analytical approaches, DCM, however, remains the commonly referred to analytical approach to explain discrete choices.
Early work involving DCM rested on the assumption that decision-making is a rational and optimizing (i.e., utility maximizing) process (McFadden 1974), but more recently DCM has been applied to assess pragmatic, heuristic decision-making (Bateman et al. 2017). The approach draws on different types of data such as revealed preference data to explain choices pertaining to actual alternatives, stated preference data to explain choices related to hypothetical alternatives, or both (Louviere et al. 2000, Chapter 1). But analyzing any data to empirically determine the importance weights that individuals place on different attributes, which in turn shape their preference for a particular alternative, remains challenging (Kamargianni et al. 2014). More specifically, the estimated parameters in traditional DCM analyses are the marginal utility associated with a change in the attribute level in moving from one alternative to another (e.g., changes in payment terms from 30 to 60 days). But directly estimating the importance weights of the attribute as a whole (e.g., payment terms relative to opening hours), rather than of distinct attribute levels, is less straightforward and requires additional calculations, since commonly used DCM analyses draw on the estimation of importance weights that apply to different attribute levels (for example, see Louviere et al. 2000, Chapters 11 and 12). Traditional DCM estimations apply commonly used multivariate analysis methods (e.g., probit, logit and multinomial logit models) when analyzing stated preference data generated through discrete choice experiments (DCE), but focus on the attribute level rather than the attribute itself (Louviere et al. 2000, Chapter 1). Indeed, “despite common practice, relative attribute impacts in DCEs cannot be inferred directly from parameter estimates” (Lancsar et al. 2007, p. 1752).