Identifying Taste Variation in Choice Models

Among the many attractive features of the mixed logit model is its ability to take account of taste variation among decision-makers by allowing coefficients to follow pre-specified distributions (usually normal or lognormal). Whilst accounting for heterogeneity in the population, simple applications of the technique fail to identify valuable information on differences in behaviour between market segments. This information is likely to be of use to those involved in policy and investment analysis, product design and marketing. The ‘standard’ approach to overcome this problem when working with the mixed logit model is to identify segments prior to modelling and either specify a set of constant coefficients for each market segment together with an additional error term to ‘mop-up’ any residual variation, or by allowing separate distributions for each market segment. An alternative approach is to adapt an exciting new methodology that offers the ability to estimate reliable individual specific parameters (Revelt and Train, 1999). This approach is documented in Section 3 and involves three key stages: • First use maximum simulated likelihood to estimate distributions of tastes across the population. • Next examine individual’s choices to arrive at estimates of their parameters, conditional on know distributions across the population (including accounting for uncertainty in the population estimates). This process again involves the use of maximum simulated likelihood. • Finally, differences in behaviour between market segments are identified by regressing individual ‘part-worths’ against the characteristics of the decision-maker or attributes of the choice alternatives. In the first instance the technique is validated under ‘controlled’ circumstances on a simulated data set with know taste distributions. This simulation involves a binary choice situation in which the alternatives are described in terms of time and cost. The choices of a group of decision-makers are simulated with each with a value of time drawn from a known distribution. The resulting choices are then analysed and individual values recovered with a surprisingly high degree of precision. The findings of this validation are set out in Section 4. Following a successful validation of the technique on simulated data, the methodology is applied to data from two stated preference experiments in which 326 respondents were asked to choose between alternate motor vehicle 1 specifications defined by purchase price, running costs, engine size, emissions and safety features. The results of this analysis are set out in Section 5 and are compared to the findings of previously calibrated models that identified significant differences in tastes across market segments.