Modelling Multimodal Mode Choice Behaviour with Spatial Variability in Level of Service Attributes

Abstract

Cities worldwide are grappling with declining public transit usage, primarily due to insufficient spatial coverage of mass transit corridors in sprawling urban regions and inadequate first- and last-mile connectivity options for transit users. These limitations have compelled travellers to rely heavily on personal vehicles, exacerbating traffic congestion and related issues. To tackle these challenges, many cities are now focusing on developing integrated multimodal public transit systems. This approach combines the strengths of various travel modes, including mass transit and personalized access and egress mode options, to improve transit system accessibility and enhance door-to-door connectivity for transit users. As cities are increasingly moving toward multimodal transit systems, the need for analytical tools has increased to: (a) enhance the current understanding of traveller behaviour and mode choices in the context of multimodal transit systems, and (b) accurately forecast travel demand patterns under various configurations of transit connectivity with other modes of travel. Random utility maximisation (RUM)-based mode choice models are commonplace for the analysis of traveller preferences and forecasting mode shares in cities with public transit systems. Such models involve the characterization of an indirect utility function for each mode, which is specified as a function of the traveller’s characteristics and the level of service (LOS) attributes, such as travel times and travel costs of the mode. It is a common practice to treat these LOS variables as free of errors. However, it has long been recognised that the LOS attributes used in mode choice models may be associated with measurement errors. A well-recognized source of such errors is spatial aggregation of travel origin and destination locations into centroids of aggregate spatial units called traffic analysis zones (TAZs). Ignoring such aggregation-induced imprecision in LOS attributes can potentially lead to inferior fit, biased parameter estimates, and distorted forecasts. The first objective of this thesis is to develop a multimodal mode choice model that captures the multimodal nature of transit trips. This model is structured as a two-level mixed multinomial logit model, where the primary mode choice is modelled at the top level, and the access and egress mode choices are modelled at the bottom level. To reflect the influence of the LOS variables of access and egress modes on the preference for a primary transit mode, two logsum variables – one from the access modes available for that transit mode and another from the egress modes – are incorporated as additional covariates in the utility function of the transit mode. The second objective is to extend the above formulation to recognize the spatial variability in LOS attributes of primary and access/egress modes due to imprecise location data of trip destinations. To do so, the LOS attributes are specified as stochastic explanatory variables, with their distributions empirically characterized beforehand. The third objective is to apply the above modelling frameworks to analyse commute mode choice in Bengaluru, India. Central to this empirical work is a multimodal mode choice analysis that not only accounts for traditional modes but also considers emerging shared modes, such as ride-hailing as options for first/last-mile connectivity to transit modes. The fourth objective is to apply the proposed multimodal mode choice models considering different levels of spatial aggregation – Census wards and Census block groups – to evaluate a suitable aggregation for defining TAZs for mode choice modelling in Indian cities. The empirical analysis of multimodal commute mode choice in Bengaluru suggests that making public transit options more accessible to travel origins and destinations would be among the most effective ways of enhancing public transit mode share in the city. Further, models that consider the measurement error in LOS attributes (due to spatial aggregation) offered better data fit, parameter estimates, and willingness to pay measures than the models that ignore measurement errors. If the analyst is unable to consider the measurement error due to computational or other reasons, empirical models with Census block groups (with an average area of 0.27 km2) as spatial units yielded better results than those with Census wards (with an average area of 3.5 km2). Therefore, the study recommends adopting Census block groups as TAZs instead of Census wards for modelling travel demand in Indian cities, as this approach strikes a balance between model performance and computational effort for large scale travel demand modelling applications.