Choice models estimated on datasets with large numbers of alternatives present significant challenges leading to rapidly expanding computational cost, as well as potential behavioural realism issues. Sampling of alternatives has been a well-established method for overcoming the computational limitations, mostly applied to models of residential location. Nonetheless, destination choice models of discretionary activities require a different type of analysis, since the choice can be governed by time–space constraints and familiarity regarding the alternatives. Observing the general areas of travel for a period of days using high resolution GPS tracking can provide important information of the individuals’ whereabouts. The present study, taking advantage of such a dataset, proposes a more behaviourally realistic sampling protocol to reduce the choice set utilising the geography-based concepts of activity spaces. Differential importance sampling rates are applied depending on the individual's activity space and trip chain making the resulting sampled choice set a function of person-specific spatial awareness and mode-specific time–space constraints. The performance of the sampling protocol developed is assessed using a model estimated with the full choice set and compared with random sampling and several other importance sampling protocols. The modelling outputs suggest that random sampling should be used with care, since it can result in highly biased estimates, but with low standard errors, as well. The proposed approach incorporates both time–space constraints and individual spatial awareness and is able to produce less biased estimates, achieve higher sampling stability and statistical efficiency, while also avoiding overfitting.
CORE (COnnecting REpositories)
CORE (COnnecting REpositories)