Project: Transport: Methods of Improving our Understanding of the Processes by which Transport Choices are made B36
Reference: STP 14/5/15
Last update: 31/12/2009 09:47:28
The objectives of the project are:
to investigate what processes are employed by individuals in making transport choices:
to identify what factors influence the decision process of any given individual:
to examine whether the processes actually employed are consistent with those assumed by the discrete choice models commonly used in transport:
to develop models that offer greater generality in the representation of individual choice behaviour;
to assess the implications, in terms of estimation, forecasting, valuation and appraisal, of mis-specifying the choice process.
How do car drivers choose a route and departure time for travelling to work? How do business or leisure travellers choose what mode to use for a long distance trip? How do car buyers choose between the myriad of cars available on the market? Such questions are fundamental to our understanding of travel behaviour, but it is debatable whether any can be answered with confidence, at least in the longer term. Evidence suggests that the choice processes actually employed by individuals may be inconsistent with those assumed in the models conventionally used in transport. Without an understanding of how people choose and appropriate models for representing choice, however, it is impossible to make reliable forecasts of how policy adjustments will impact upon travel behaviour. This project will address three principal questions. Do the choice processes actually employed by individuals accord with those assumed in disaggregate transport models? If not, does it matter? If it matters, how do we deal with it?
University of Leeds
Institute for Transport Studies, 38 University Road, Leeds, LS2 9JT
+44 (0)113 343 5325
Cost to the Department: £49,997.15
Actual start date: 01 October 2002
Actual completion date: 01 December 2003
Transport: ‘methods of improving our understanding of the processes by which transport choices are made and the factors which affect these’
Author: Dr Richard Batley and Prof. Andrew Daly
Publication date: 28/11/2003
Summary of results
- This report summarises the findings of a study aimed at improving understanding of the processes by which transport choices are made.
The dominant behavioural paradigm underlying transport choice models is currently that of Random Utility Maximisation (RUM), originating from the work of Block and Marschak in 1960. Substantially developed and made rigorous by McFadden (particularly in his 1981 paper), this theory underlies the Generalised Extreme Value (GEV) models which are now used in almost all transport studies. These models are based on a clear assumption of 'rationality', implying utility maximisation by the traveller but a lack of knowledge of the exact choice mechanisms by the analyst who therefore introduces random terms into the model.
In parallel to these developments, Tversky and his associates developed a series of models which are based on a more limited concept of rationality. An important example of these models is the use of elimination strategies which represent choice by the continual elimination of one or more choices which fail to satisfy some criterion until only one alternative remains. A fairly general class of models of this type is the Elimination-By-Aspects (EBA) family. A little-known finding by Tversky is that a simple extension of the assumptions of the model can make it consistent with the RUM paradigm. While Tversky and his associates developed some of the theory for EBA models, there have been very few practical implementations, and procedures for working with the models are not well established.
The EBA family of models can be considered to represent choice as a decision tree, a characteristic they share with all of the GEV models represented in the literature. This commonality and the use of similar probability functions to specify conditional choice has led several distinguished researchers to describe the models as "for all practical purposes indistinguishable" (McFadden), at least for simple cases. The simplest case, which has been studied in particular detail in this study, is the comparison of the hierarchical EBA model (HEBA) with the nested logit model (NL), an important and widely used example of the GEV family.
The initial analysis in this study showed that it is indeed possible to specify HEBA and NL models which imply exactly the same choice probabilities. More generally, EBA models match exactly to a specific class of GEV models, provided normalisation is undertaken to eliminate the differing degrees of over-specification of the two forms. However, these specifications need to be examined in detail to ensure that the fundamental assumptions of EBA and GEV models are maintained.
GEV can be specified to be consistent with a valid EBA, but the requisite specification is invalid as GEV. According to GEV, the utility of an alternative must be derived solely from the attributes of that alternative; otherwise the utility maximisation basis of GEV is destroyed. Consistency with a valid EBA, however, can be achieved only by breaking this rule.
EBA can be specified to be consistent with a valid GEV, but the requisite specification may not be valid as EBA. Choice probability in EBA must be non-negative at each stage in the decision tree, but in achieving consistency with a valid GEV this condition may be violated. (We note, however, that appropriate transformation of EBA/GEV may avoid such violation).
We conclude that the findings of previous researchers concerning this equivalence have been inadequate and that although EBA models may be consistent with RUM, it is not necessary that they are equivalent to the most obviously corresponding GEV model.
Empirical analysis of the two model forms was undertaken on a conveniently available data set which contained features of complexity appearing to make it suitable to illuminate the differences between the model forms. First, assumptions had to be made concerning the practical implementation of the HEBA model and software had to be prepared to estimate its parameters.
The results of this comparison showed that in many cases the HEBA model, with fewer estimated coefficients, outperformed the NL model. Tests were devised for distinguishing between the models, showing that in a number of cases they were indeed significantly different in terms of their parameter values and (therefore) implications for forecasting. Forecasting procedures for EBA models remain difficult to specify. Similarly, it is unclear whether and how attribute valuations can be derived from EBA; though we note that the concept of compensatory choice, which offers a basis for the derivation of valuations, is not clearly precluded by EBA.
Appraisal with EBA models presents a number of problems. Criteria have to be defined that can be extracted from the models - the study suggests some of these but is not able to recommend one strongly over the others. It is difficult to express any evaluation criterion in monetary terms because of the lack of a well-specified cost coefficient in the model. Theoretical issues may also impede the use of EBA models in appraisal.
The strength of the EBA models is in the explanation they give of observed behaviour, which is reported in the literature and has been reproduced in our own findings. This strength makes it worthwhile to continue the investigation of the EBA paradigm, attempting to solve the problems that have been identified and determining whether our results are reproducible on other data.