Quantitative modelling of cyclists' route choice behaviour

With the rapid growth of the urban sprawl, there arises a question about a proper allocation of the existing public urban space. The current evidence suggests that active transport modes have been underprioritised in national models and planning, with the bicycle being the most disadvantaged mode. On the other hand, there is enough knowledge support stating that bicycle travel contributes to positive impacts on the environment, such as reduction in air and noise pollution; it makes cities more liveable, improves individual health and increases physical activity levels.

One of the aspects to consider when encouraging other road users to shift to bicycles is to understand the behaviour of the current cyclists in terms of their choices of routes. There is a considerable interest in this topic and corresponding strong motivation for quantifying the above-mentioned route choice preferences. This thesis addresses the wide topic of cyclists’ route choice preferences, focusing on quantitative methods to model the problem based on revealed preference data.

With the general aim of the PhD study being to understand the route choice behaviour of cyclists, various approaches used in the existing literature are described in Chapter 2 of this thesis. The content of this chapter is supported by Appendix A which includes a systematic summary of the existing bicycle route choice literature.

The main core of this thesis consists of five self-contained studies.

As a basis to facilitate cycling for longer distances, the literature review in Paper 1 aims to synthesise and improve knowledge on bicycle commuting beyond 5 km, by analysing both socio-psychological and physical factors. Perceived trip benefits, cycling habits, bicycle-friendly infrastructure, and e-bike usage were identified as key factors. This review also emphasises the relevance of encouraging people to cycle longer distances and discusses different tailored intervention strategies. It states that further research is required to fully understand the dynamics of bicycle commuting beyond short distances, for example a more detailed comparison of preferences between short- and long-distance cyclists.

Addressing the latter gap outlined in Paper 1, Paper 2 estimates several bicycle route choice models for various cycling frequencies and trip distances. The models are estimated as a joint path-size logit model and account for a wide range of bicycle network attributes, such as bicycle infrastructure type, land use, surface type or cycle superhighways. The model performs very well on a hold-out sample, also when considering the similarity between the observed and predicted route, not only their binary consistency. Finally, this paper formulates several policy measures relevant to promote cycling.

A bicycle route choice model, applying a novel method that takes the complete network into account and avoids generating choice sets, is estimated in Paper 3. This study provides valuable insights about the route choice behaviour of cyclists. It uses the estimated preferences to compute a generalised cost of bicycle travel, which is correlated with a large number of bicycle trips across many origin-destination pairs. Counterfactual simulations suggest that the extensive Copenhagen bicycle lane network has caused the number of bicycle trips and the bicycle kilometres travelled to increase significantly, which translates into a huge annual benefit, owing to changes in generalised travel cost, health, and accidents.

Extending the traditional mixed multinomial logit model by relaxing the condition of intraindividual homogeneity, Paper 4 proposes a new modelling approach. In this method, a neural network maps contextual information to interpretable shifts in the preference parameters of each individual in each choice occasion. The model can leverage information from other decision-makers to infer the effect of a particular context on a particular decision-maker. It supports both continuous and discrete variables, as well as complex non-linear interactions between them, and each context specification is considered jointly as a whole by the neural network. The concept and interpretation of the proposed model is illustrated in a simulation study, and a bicycle route choice model on a real-life dataset is estimated to showcase its scalability and application.

Modelling individual behaviour based on large datasets is a challenging task, and the results are not straightforward to interpret without considering the sample structure. Paper 5 addresses the issue of panel datasets that are imbalanced in terms of number of observations per individual. It proposes a strategy that ensures equal contribution of each individual to the estimation results, while reducing the computational burden related to estimating models on large datasets. The study includes a simulation experiment to study how the application of bias reduction strategies influences model results, and finds that combining subsampling and weighting techniques helps to find an optimal tradeoff between bias and variance of the estimates. Considering the conclusions from the simulation study, a large-scale case study estimates bicycle route choice models with different correction strategies. 

Analyses and studies included in this thesis are based on a large-scale crowdsourced dataset of GPS trajectories from users of a head protection airbag for cyclists, with a sample size unprecedented in the bicycle route choice literature. Paper 2 provides a general description of the data source, reports how the dataset was processed for modelling purposes, and includes insights about cycling frequency and trip length in the dataset. Paper 3 includes a description of socio-demographic characteristics of the individuals. In Paper 4, the GPS dataset is enriched with data from further sources and with information inferred from the dataset itself, providing information about trip context. Finally, Paper 5 discusses the imbalance in the dataset in terms of number of trips per individual.

The contribution of this PhD study is discussed in the light of three research directions: i) utilising large-scale cycling GPS data, ii) quantifying cyclists’ route choice preferences and formulating policy implications, and iii) capturing heterogeneity in behaviour.

Route choice models estimated in this thesis are based on a large-scale dataset of observed bicycle trajectories and those models account for a wide range of bicycle network attributes, determined at a very fine level of resolution, which allows distinguishing between numerous infrastructure types. Model estimation procedures in this thesis apply both state-of-the-art and novel methods from the existing literature, as well as methods and approaches developed during this PhD study. Moreover, recent implementations and modern hardware resources are utilised, so that models for several hundred thousands of observations are estimated within minutes. 

This thesis advances the bicycle route choice modelling literature by quantifying preferences for a detailed set of attributes in behaviourally realistic route choice models, and by formulating policy recommendations based on the results. Route choice models in this thesis suggest, for example, that painted bicycle lanes alone do not provide sufficient protection on large roads. Moreover, the findings indicate that cyclists avoid interaction with other modes of transport and value the continuity of their ride and the connectivity of the route. Analysis of counterfactual scenarios suggest that such route-level investments can lead to large societal benefits. Furthermore, findings from the literature review indicate that the right intervention strategies, together with the help of technological advancements, can make a bicycle an attractive alternative to car for longer distances.

Studies in this thesis recognise the heterogeneity in route choice behaviour of cyclists and apply various approaches to account for this heterogeneity in the models. For example, models in Paper 2 reveal differences in preferences for infrequent versus frequent cyclists, and cyclists on short versus long trips. Moreover, route choice models in this thesis account for the unobserved heterogeneity either by including the land-use attributes as random parameters (Papers 4 and 5), by assuming that contextual trip attributes have influence on the preferences for some land-use attributes (Paper 4), or by including parameters accounting for interactions between the type of bicycle infrastructure and the neighbouring land-use (Paper 3). 

From the methodological perspective, this thesis makes a twofold contribution to the choice modelling field, which is not restricted to the route choice task. Firstly, Paper 4 introduces an efficient approach to model context-dependent intra-respondent heterogeneity, which allows for systematic variations in preference parameters as a function of contextual variables. Furthermore, Paper 5 proposes strategies for modelling behaviour on imbalanced datasets, which assure that the model is equitable, i.e. each of the individuals contributes equally to the estimation results, regardless of their representation in the sample. The outcomes of these studies are transferable to other choice situations in various research fields applying choice modelling to understand individual behaviour.