Accommodating Heterogeneity for Reducing Traffic Pollution: a ?mixed? Latent Class Approach



Accommodating Heterogeneity for Reducing Traffic Pollution: a ?mixed? Latent Class Approach

Authors

D Campbell, M Boeri, A Longo, Queen's University Belfast, UK

Description

In this paper we use a model that assumes a latent class structure of unobserved heterogeneity whilst allowing the parameters within each of the latent classes to take on a continuous distribution.

Abstract

Discrete choice models which accommodate unobserved preference heterogeneity are now widely used. These models include continuous representation of tastes, e.g., random parameter models, and finite representations of taste, e.g., latent class models. Both the random parameters and latent class models have their strengths and weaknesses. Under the random parameters specification, it is necessary to define a functional form for the parameter. Parameters of this distribution are then obtained (e.g., the mean and variance). However, there remains an ongoing debate and uncertainty regarding the most appropriate function form to represent the unobserved heterogeneity. In contrast, latent class models do not require the functional form to be specified. Instead, the number of finite classes is defined. For each class a set of estimated parameters is derived along with a probability mass representing the share of respondents associated with these parameters. Within each latent class, preference homogeneity is assumed.
In this paper we use a model that assumes a latent class structure of unobserved heterogeneity. However, due to the quite stringent assumption of preference homogeneity within each class we use also allow heterogeneity within each class. This is achieved by allowing the parameters within each of the latent classes to take on a continuous distribution. This is similar to a mixtures of distributions approach, but instead of retrieving separate probabilities for the mixture for every parameter, we constrain the mixing probabilities to be the same across all parameters.
To demonstrate this approach we use a stated choice dataset collected to elicit willingness to pay estimates for the reduction of pollution arising from traffic in three European cities, namely Warsaw, The Hague and Leipzig. We compare this approach against the standard random parameters and latent class models. Our results show significant gains in model fit and a richer description of unobserved heterogeneity.

Publisher

Association for European Transport