Chapter 7: A Comparison of Conjoint, Multi-Criteria, Conditional Logit and Neural Network Analyses for Rank-Ordered Preference Data

The chapter presents a systematic comparison of four different methodologies for separately analysing individuals’ stated preferences relative to the choice of alternative solutions of freight transport. These are defined by five ‘qualitative’ attributes: frequency of service, transport time, reliability, carrier’s flexibility and transport losses, plus the monetary cost. As the data consist of alternatives’ rankings, the models applied here are somewhat unusual in the field, at least in transportation analysis: conjoint analysis, UTA-type multi-criteria analysis, rank-ordered conditional logit, and neural network analysis. In an earlier paper Beuthe and Bouffioux (2008) applied the ordered logit model to a preference survey of 103 firms at an aggregate level. With linear utility functions, this model’s results indicated a wide heterogeneity in decision making. In order to go deeper into the process of decision making, the present chapter applies the above four models, some of them with non-linear utility functions, to the individual rankings of nine firms’ transport managers selected from the survey in diverse industrial sectors. The alternatives submitted to their judgement are designed according to an orthogonal fractional factorial design. Each estimation methodology is adjusted and specified to suit the specific data. Over this small set of individual firms, each applied method shows that the cost factor is the most important one in the individual choice making of seven out of nine transport managers. Reliability is often more important than transport time, even though its relative importance varies from case to case. The other factors may play a significant role in some cases. These outcomes are mostly coherent with the earlier results obtained at an aggregate level. The two better performing methods are the multi-criteria and the neural network analyses, which both involve non-linear partial utility functions.