Modelling Random Taste Variation in the Acceptance of Tolled Roads in the UK



Modelling Random Taste Variation in the Acceptance of Tolled Roads in the UK

Authors

J N Ibáñez, R D Connors, R P Batley, M Wardman, ITS, University of Leeds, UK; G Hyman, Department for Transport, UK

Description

Assessment of route choice behaviour by exploiting different configurations of mixed multinomial logit models to elucidate heteroskedasticity and correlation in the account of the random taste variation present in the users of the UK M6 Toll road.

Abstract

This paper presents results of a modelling exercise aiming to analyse preference data from passengers in relation to their valuation of tolled route choices. In so doing, we seek to gain a better understanding of how toll levels on an interurban trunk road influence travel demands in circumstances where tolled routes compete with free-of-charge alternatives. The context of our investigation is the United Kingdom's M6 Toll (M6T) motorway, the first tolled motorway within the UK, comprising of a 43km three-lane motorway designed to alleviate traffic congestion around Birmingham and used on average by 45k daily motorists.
The research centres upon a large scale market research exercise undertaken as part of a UK Department for Transport funded study and involving questionnaire surveys that incorporate not only a series of Stated Preference (SP) experiments over 3200 motorists but also Revealed Preferences (RP) data describing the route choices and trips currently chosen.
In the main survey, the SP experiments dealt with drivers´ route choice between tolled and untolled motorways and local A-roads. The tolled options considered the use of the existing M6T for part of the journey but also the acceptance for new long and short distance motorway extensions offered as an alternative to current routes. The assessment of these preferences is of particular benefit to the appraisal of new infrastructure schemes.
To conduct this kind of assessment, different models based on random utility maximisation (RUM) have been estimated over the SP and RP data to explain and predict route and time-of-day choices as a function, amongst other factors, of the potential toll charge, fuel consumption and travel times. The analysis also estimates the role of departure time shifts, reliability requirements and congestion conditions on route choices.
The RUM models that we have specified on these data include the basic multinomial logit (MNL) and nested logit (NL) models, the latter as a direct relaxation of the rather limiting properties of MNL. Given that some of the alternatives (e.g. M6T) are common to different available route choices we also exploit the cross nested logit (CNL) model.
Finally, and further relaxing the properties of MNL, NL and CNL, we employ the Mixed Multinomial Logit (MMNL) model to account for the existence of heterogeneity amongst the motorists taking part in the study. The latter is estimated using bespoke software that we have developed ourselves. The development of this code was prompted by the complexity of our model specification, and sheer volume of data involved, which we found to be unmanageable using existing software for estimating MMNL.
The paper describes how the use of MMNL improves the fit of the demand models and, more significantly, leads to a re-interpretation of monetary valuations deriving from the model. In particular, whilst standard MNL and NL, and even CNL, permit simple computation of monetary valuations, we provide insight as to the distributions of these valuations emanating from MMNL. Furthermore, we analyse how these valuations depend upon the heteroskedasticity and correlation prevalent within the mixing distributions of the MMNL parameters; especially the parameters associated with time (time of day choice, late running and the influence of journey time reliability) and cost (toll and fuel). In so doing, and building on current estimates of a significant correlation between distributions for cost and time parameters, we progress beyond previous specifications of MMNL which attributed heterogeneity in preferences entirely to heteroskedasticity.
Additionally, for this particular study there are systematic violations of the assumption of independence between alternatives, arising in particular from the structure of the road network. In many choice situations alternative routes overlap for a significant fraction of their length. We examine how these correlations are manifest in individuals? preferences and show that the MMNL model admits their estimation.
In reporting our results, we underline various issues encountered within the panoply of MMNL models that we have used, including the assumptions made regarding the characteristics of the mixing distributions defining such models. We discuss how MMNL models are better equipped to judge the effect of policies such as tolling, relative to simpler models that do not explicitly admit random taste variation.
In summary, the paper proposes a new approach to the implementation of MMNL models within a transport context, specifying an additional layer of analysis at the interaction between inter-individual taste variation and the socio-economic characteristics of travellers. We apply this specification over a rich set of route preference data, revealing a level of insight on motorists? responses to tolls which has not been possible hitherto.

Publisher

Association for European Transport