Intra-respondent Taste Heterogeneity in Instantaneous Panel Surveys
S Hess, Imperial College London, UK; J Rose, University of Sydney, AU
This paper discusses the estimation of models allowing for intra-respondent taste heterogeneity alongside inter-respondent heterogeneity
Most discrete choice applications recognise the fact that the data used in modelling is highly unlikely to come from a homogeneous sample of respondents, such that it is important to allow for taste heterogeneity. While some effort generally still goes into attempting to explain at least some of these variations in tastes in a deterministic fashion, an increasing number of applications rely almost exclusively on a random approach to modelling taste heterogeneity, using random coefficients models such as Mixed Logit.
The presence of taste heterogeneity, be it modelled in a random or deterministic fashion, is often especially prominent in the case of Stated Preference (SP) data. Here, it can be argued that the retrieval of differences across respondents is made easier by the fact that multiple observations exist for each respondent. In modelling, an assumption of within-individual taste homogeneity is generally made, such that tastes vary across respondents, but not across observations for the same respondent. In the case of deterministic taste heterogeneity, this means that any variation is linked only to variables that stay constant across replications for the same individual, such as socio-demographic information like income or age. In the case of random taste heterogeneity, this assumption manifests itself in the positioning of the integration over the distribution of taste coefficients outside rather than inside the product over choice probabilities for the same individual.
In this paper, we argue that this assumption of within-individual taste homogeneity is potentially not justified in all cases. While it is easier to make a case for within-agent heterogeneity in the case of long-term panel data, such as typically used in Revealed Preference (RP) studies, such variation in tastes potentially also arises in the case of instantaneous panels. As an example, it is well known that respondents do, over the course of an SP experiment, experience both learning and cognitive burden effects, potentially resulting in variations in their behaviour. A further argument against the assumption of within respondent preference homogeneity derives from the very nature of SP data. Unlike most RP surveys where market data is largely correlated in terms of the attribute levels experienced over time, SP surveys are designed to induce trading off by markedly changing the attribute levels experienced from one choice situation to another. Thus, why would a respondent's preferences not vary when faced by two very different choice situations? For example, compare a situation in which a respondent faces a 10 minute travel time relative to a travel cost of £3 to a situation where the same respondent faces an alternative with a travel time of 20 minutes at a cost of £1.
The paper then looks in detail at ways of modelling intra-respondent heterogeneity in addition to inter-respondent heterogeneity. First, we investigate the potential of linking taste heterogeneity to variables that vary across observations for the same respondent, i.e. the levels of the attributes presented in SP surveys. Next, we show how the inclusion of terms that represent learning and fatigue effects can similarly retrieve taste heterogeneity across observations for the same respondent. Finally, we discuss an approach that allows for two layers of random taste heterogeneity within a random coefficients framework, with one level of variation limited to the level of individual respondents, and one level allowing for variation within respondents. Mathematically, this corresponds to the positioning of one level of integration outside the product across choice probabilities, and one level of integration inside the product. The results of a preliminary analysis, primarily relying on a random coefficients approach, suggest the presence of significant levels of within-individual heterogeneity in the case of a standard SP dataset, such as widely used in practice. Crucially, this leads to substantively different results for trade-offs when compared to a model making an assumption of within-individual homogeneity in tastes.
Association for European Transport