Evaluating a data-driven approach for choice set identification

The specification of the choice set for travel behaviour analysis is a non-trivial task, as its size and composition are known to influence the results of model estimation and prediction. Most studies specify the choice set using choice set generation algorithms. These methods can introduce two severe errors to the specified choice set: false negative (not generating observed routes) and false positive (including irrelevant alternatives) errors. Due to increased availability of revealed preference data, like GPS, it is possible to identify the choice set in different way: data-driven. The data-driven path identification approach (DDPI), introduced in this paper, combines all unique routes that are observed for one origin-destination pair into the choice set. This paper evaluates this DDPI approach, by comparing it to two choice set generation methods (breadth-first search on link elimination and labelling). The evaluation is based on three main purposes of choice sets: analysis of alternatives, model estimation and prediction. The conclusion is that the DDPI approach is a useful alternative for choice set identification. The findings indicate that in analysing alternatives, the DDPI approach is most suitable, as it is equal to the observed behaviour. For model estimation the DDPI approach provides a useful alternative to choice set generation methods, as it provides insights into the preferences of individuals. In terms of prediction, the DDPI approach is suitable on a network level, but not on the individual level. The average performance over all alternatives is similar for all choice sets, but on individual level the DDPI method does not predict well.