Accommodating taste heterogeneity and desired substitution pattern in exit choices of pedestrian crowd evacuees using a mixed nested logit model

Mixed logit has been recognised and widely practised by researchers as a highly flexible modelling tool that can address the main shortcomings of the standard logit Despite the potential to be generalised, the random-coefficient modelling has rarely been integrated with more advanced GEV-type models, possibly due to the unavailability of such estimation options in most econometric software. This particular generalisation has been recommended by a number of econometricians for analysing choice problems in which capturing taste variation and specific non-IIA patterns of substitution are both of modeller's concern. This way, the analyst will be able to limit the number of explanatory variables to the ones whose distributions of coefficients offer behavioural interpretations about taste variation, and leave the imposition of the desired substitution pattern to the GEV core. In this paper, we report on a mixed nested logit application aimed at understanding evacuation behaviour of human crowd, particularly the choice of exit. Evacuees' trade-off between contributing factors are elicited through a data set of stated exit choices. We attempt to improve the realism of the choice experiments by conducting face-to-face interviews with pedestrians (respondents) as they exit a building; and designing scenarios that refer to their recent actual choice of exit. The primary analysis strongly suggests the presence of taste heterogeneity as well as the violation of the IIA assumption (i.e. the presence of a particular pattern of correlation between exit utilities). We propose and estimate a mixed nested logit model through which we accommodate these two econometric aspects of the problem in a unified model. Compared to the counterpart mixed multinomial logit model in which the desired substitution pattern is embodied through specification of the explanatory variables, the proposed model reduces the dimension of simulation in the likelihood maximisation procedure, and also yields a slightly better statistical fit. (C) 2015 Elsevier Ltd. All rights reserved.