The use of pooled RP-SP choice data to simultaneously identify alternative attributes and random coefficients on those attributes

Abstract

Random utility maximization-based discrete choice models involve utility functions that are typically specified with explanatory variables representing alternative-specific attributes. It may be useful to specify some alternative-specific attributes as stochastic in situations when the analyst cannot accurately measure the attribute values considered by the decision maker. In addition, the parameters representing decision makers’ response to the attributes may have to be specified as stochastic to recognize response heterogeneity in the population. Ignoring either of these two sources of stochasticity can lead to biased parameter estimates and distorted willingness-to-pay estimates. Further, in some situations the analyst may not even have access to measurements of important alternative-specific attributes to include them in the utility specification. In this study, we explore the feasibility of simultaneously inferring alternative attributes and the corresponding coefficients, as well as stochasticity in both – without the help of external measurement data on alternative attributes – using mixed logit models on pooled revealed preference (RP) and stated preference (SP) choice datasets. To do so, we first theoretically examine parameter identifiability for different specifications and distributional forms of alternative attributes and their coefficients. Next, we illustrate this through simulation experiments in a travel mode choice setting and demonstrate the conditions under which pooled RP-SP data can help disentangle stochastic alternative attributes from random coefficients. In addition, an empirical application is presented in the context of commute mode choice in Bengaluru, India, to demonstrate the importance of recognizing stochasticity in mode-specific in-vehicle travel times along with the random coefficient on in-vehicle travel times.