Hierarchical Elimination-by-aspects and Nested Logit Models of Stated Preferences for Alternative-fuel Vehicles
R Batley, J Toner, ITS, University of Leeds, UK
Concerns regarding both the environmental impacts of petrol and diesel cars, and the security of oil supplies, have stimulated the development of alternative-fuel vehicle (AFV) technology. Examples of AFVs include vehicles powered by batteries, liquefied petroleum gas or fuel cell. Since AFVs are yet to achieve mass-market penetration, particularly in the UK, purchase costs are relatively high. Furthermore, current AFV technology typically offers inferior performance, over at least some characteristics, compared with conventional vehicles. The pertinent policy question, therefore, is perhaps as follows. ?Are consumers willing to pay more for a vehicle that offers environmental benefits but is inferior in terms of some performance criteria??
Government and commercial interest notwithstanding, there has been a relative paucity of academic literature on the subject of AFV demand, and almost all of the published work has been US-based. In this work, we have sought to research the behavioural issues associated with estimating the demand for AFVs in a UK context. More specifically, we have addressed the following questions:
1. Can we design and apply a survey instrument to investigate how people process complex choice tasks?
2. Are the processes employed consistent with those assumed in conventional choice models?
3. If not, can and should we fit alternative decision process models?
4. What might be the take-up of AFVs?
In terms of methodological innovation, this paper makes two principal contributions; first, in developing and applying alternative decision process models, and second, in developing and applying a complex stated preference experiment.
The principal survey instrument was a mail-back stated preference (SP) experiment analysing households? choices between hypothetical vehicle alternatives. This was supplemented by a focus group analysis, the aim of which was to explore the decision processes involved in vehicle ownership, purchase and use and the interaction between the ownership and use decisions. SP questionnaires were also administered to the attendees of the focus groups. Although concentrating on the SP experiment, this paper notes the influence of the focus group findings in informing model development.
Developing and applying alternative decision process models
Throughout the last thirty years, researchers have striven continuously, to achieve greater generality in the representation of the discrete choice behaviour of individuals. The vast majority of the research effort has, however, adhered rigidly to the conventional framework of random utility maximisation (RUM), which is based fundamentally on the micro-economic axioms of completeness, transitivity and continuity. This is despite continual assault from the decision sciences, which have identified, through experimental investigation, behavioural phenomena producing choices that violate these axioms. Of particular relevance to this paper is the observation that as the volume of information presented to an individual is increased, a decision task can quickly become unmanageable, and lead to the adoption of simplifying decision heuristics. Such heuristics may achieve a satisfactory, but not necessarily utility-maximising, solution. This creates a dichotomy between RUM and non-RUM models of discrete choice.
Some authors have identified structural similarities between RUM and non-RUM models. In particular, the nested logit (NL) and cross-nested logit (CNL) models appear to offer RUM analogies of the (seemingly) non-RUM elimination-by-tree (EBT) and elimination-by-aspects (EBA) models, respectively. This paper develops, and applies, specifications of all of these models, before extending RUM to the limits of current practice in the form of mixed logit.
The basic specifications of EBA and EBT characterise alternatives as collections of ?aspects? that are unique to the alternatives or common to subsets of the alternatives. These aspects are not, however, decomposed into their component attributes. For example, the unique characteristics of an AFV might include, say, battery power, limited range and lengthy recharging time. The basic specification combines all such characteristics of the AFV into an aspect (?AFV-ness?) yet, for practical purposes, it is necessary to know which are the more important attributes associated with AFV-ness. If the range can be improved, will that increase the uptake of AFVs, the other limitations notwithstanding? Or is the recharging time a bigger deterrent? For forecasting purposes, it is essential to disaggregate AFV-ness into tangible attributes. We have done this.
Developing and applying complex stated preference experiments
Conventional design practice in the UK is to avoid ?overloading? the respondent, in terms of the numbers of alternatives and attributes presented. If, from the analyst?s perspective, there is an interest in a large number of attributes, conventional practice is to develop a series of small designs featuring one common attribute, and to merge the designs at the modelling stage. Moreover, a typical design in UK applications might feature, say, two alternatives and four attributes.
The experiment developed in this paper was much larger in dimension, featuring three alternatives and eight attributes. A car purchase choice task was presented involving three alternatives; Car A was broadly consistent with a conventional petrol or diesel car, Car C was broadly consistent with a near-term AFV, and Car B was a compromise option, which might represent some form of future ?clean? petrol or diesel vehicle or a future AFV with performance features more comparable with a petrol or diesel vehicle. Each car was described in terms of the following attributes: on-the-road price, running costs, range on a full refuel or recharge, time for a full refuel or recharge, top speed, time taken to accelerate from 0-60 mph, retained value after 3 years or 36,000 miles, and emissions as a percentage of a year 2000 petrol car. Each attribute was specified at four levels.
The large numbers of attributes and levels not only precluded the use of a full factorial design, since this would have required 424 replications, but were beyond the scope of conventional fractional factorial design plans. Design was instead carried out using the OPTEX and FACTEX procedures in the SAS software, and this yielded a D-optimal design on 25 replications.
How do people make complex choices?
Table 1 presents some indicative results from our analysis, where we have applied the stated preference data to the estimation of MNL, NL and EBT models. Our analysis of EBT was motivated, in particular, by an interest in the existence of a minimum range threshold, which was identified as a significant driver of AFV choice in both the literature and our own focus group research. Several alternative specifications of EBT were estimated. Our preferred specification imposed a tree structure with the conventional vehicle and the AFV nested together and the ?compromise? vehicle (i.e. vehicle with highest range) in a single alternative nest. NL was specified as having an identical nesting.
Perusal of Table 1 reveals that EBT offers the best fit, followed by NL and MNL. Interestingly, the three models show a slightly different pattern of parameter significance. Comparing NL and EBT, for example, NL reports an insignificant retained value parameter but a significant emissions parameter, while EBT reports the converse. As regards range, it can be seen that the estimated parameter is more precise in EBT than in NL. This offers some support to the hypothesis of a threshold effect on range.
These are indicative results; the paper extends the analysis to consider CNL and EBA specifications.
In seeking to gain an understanding of what factors might influence the existence and nature of thresholds, the models were developed further to include segmentation variables. For example, in MNL, we identified significant positive interactions between the range variable and dummy variables representing both households using at least one car to travel 9001-11000 miles per year and households not using any cars to take children to a school or nursery. Extending the investigation of taste variation, we developed a random parameters logit specification, estimating significant standard deviations for running costs, range on full refuel or recharge and time for full refuel or recharge.
Potential take-up of AFVs
Using standard sample enumeration techniques, the market share of AFVs was forecast under a range of scenarios. Our headline result is that, if the representation of vehicle types in the experiment is anything like realistic, the market share for AFVs would be at most 8% or 9%, and this would increase to 9% to 11% in the event of a 20% reduction in on-the-road price. Furthermore, relatively small changes in AFV performance (improvements in running costs, range, refuelling time of up to 20%) would still, in most cases, fail to increase the AFV share substantially above 10%.
Association for European Transport