The Role of Tangible Attributes in Hybrid Discrete Choice Models

The Role of Tangible Attributes in Hybrid Discrete Choice Models


M Francisca Yañez, F Bahamonde-Birke, S Raveau, J de Dios Ortuzar, Pontificia Universidad Catolica de Chile, CL


This paper studies the inclusion of tangible attributes on a latent variable model. Tangible attributes have effects: (i) directly on the utility through the choice model, and (ii) indirectly through the explanatory variables of the latent variables.


Modelling the choice of transport mode is a key element in travel demand estimation. The traditional approach has used mainly tangible and objective modal attributes as explanatory variables. However, attitudes and perceptions may also influence the users? decision process (Ashok et al., 2002; Ben-Akiva et al., 2002). In the last 30 years, the literature has reported different attempts to include the effect of both traditional variables such as time and cost, but also psychometric and non-measurable (or latent) variables in choice models:

- Greene (1984) suggested the inclusion of perception indicators directly in the utility function. Nevertheless, this method may lead to bias, as indicators are a manifestation of perceptions rather than their cause.
- Keane (1997) proposed the inclusion of latent attributes in the utility function. As these attributes are inferred from individual choices, this method does not use information beyond the observed choice and the traditional objective alternative attributes. The problem is that the method demands to assume specific latent variables over alternatives, which are constant over individuals.
- Bollen (1989) suggested to treat the latent variables through a MIMIC (multiple indicator multiple cause) model, where the latent variables are explained by characteristics from the users and the alternatives through structural equations; at the same time, the latent variables explain the perception indicators through measurement equations. Nowadays, the use of hybrid choice models including a latent variable (MIMIC) model and a discrete choice model, is the most popular approach.
Now, despite the fact that using hybrid choice models has become more popular, up to now most applications have included only socioeconomic characteristics from the users in the structural equations of the latent variable model. Consequently, these hybrid models are not better than the traditional discrete choice models in terms of their sensitivity to evaluate policies that modify the transport system (Yáñez et al., 2009). Hence, even though the utility functions of the choice models do capture the effects of changes in the transport system through changes in the tangible modal attributes, there is still a problem when a certain policy affects individual perceptions. For example, new seats or air conditioning in the underground may increase the perception of comfort by this mode; also, electronic boards indicating expected waiting times at bus stops may increase the perception of reliability.

The above discussion clearly suggests the inclusion of tangible attributes related to the alternatives not only in the choice model, but also in the latent variable model. However, it is important to note that we are not referring to well known tangible attributes such as travel cost and time. The tangible attributes required must be related to the latent variables. For instance, attributes such as the density of passengers or the standard deviation of waiting times could explain, respectively, latent variables such as comfort and reliability.

This paper studies the issues related to the inclusion/omission of tangible attributes on the latent variable model. Using both simulated and real data, we provide empirical evidence on the effect that tangible attributes have: (i) directly on the utility value through the choice model, and (ii) indirectly through the explanatory variables of the latent variables.

Our results show that the omission of tangible attributes in the latent variable model has two undesirable consequences: the significance of the latent variable(s) associated may fall artificially, and, as for traditional models (Ortúzar, 1999), this omission could lead to inconsistent parameters being estimated. Complementary, the experiments with simulated data show the importance of finding the correct hybrid model specification. In particular, we analyze where to allocate the explanatory variables associated with the alternatives (i.e. in the latent variable model, in the choice model, or in both models). We show that the most natural technique, i..e trying to include any such attribute directly in the choice model, is not always correct.


Association for European Transport