Search

This paper analyzes the choice behavior of the high-speed rail Italian passengers, with a random utility latent class structure for modelling travelers' tastes. Latent class structures are widely used for addressing inter/intra-respondent taste variation in transport field. The simple Multinomial Logit (MNL) is commonly used as a kernel model (in the sense it will be clarified in the paper) into such latent class structures. The resulting random utility model (RUM) is well known in the literature as latent class Multinomial Logit (LC-MNL). In this study, we propose to test kernel models with different inherent distributional assumptions for random residuals of the utilities than that of the simple Multinomial Logit (MNL), namely the Gumbel distribution assumption (Extreme Value type I). Such different kernel RUMs are obtained by assuming random residuals of the utilities distributed as: Weibull (i.e. Extreme Value-type III), Exponential, Rayleigh, Pareto and q-Gumbel. Results on an estimation exercise, based on a dataset coming from a stated preference (SP) survey, show that some of these assumptions allow the resulting LC-RUM to outperform the well-known LC-MNL in terms of goodness of fit (GOF).

Background: Route choice set definition is a very sensitive phase of the route choice simulation. Several heuristics, generally based on shortest path algorithm repetition, give as output choice sets that are very large, lading to questions about their behavioural consistency. Objective: This paper proposes a comparison of the main route choice set generation methods, contrasting the results of the commonly implemented heuristics with the revealed choice sets of a sample of employees and students moving within the Metropolitan Area of Naples. Methods: We described the data collection process and provided a statistical analysis of the sample data. In addition, since coverage measures and performance indicators, usually applied in the literature, do not take into account any possible biases related to the generated choice set cardinality. The current work proposes an analysis of the coverage of routes that are generated by the heuristics towards the revealed routes. Results: We observed that when the heuristics did not provide overlapped routes, although giving higher network coverage, they introduced a higher number of links not belonging to any observed route. In general, this may cause significant network loading errors. Therefore, the quality of a method for choice set generation should be measured as a function of the trade-off amongst network coverage and network loading bias due to excessive cardinality of the generated choice-sets. Conclusion: We found the randomization method, which is also less computational demanding, provided the best trade-off amongst network coverage and network loading bias