In the Break or in the Peloton - Factors Affecting Cyclists Route Choice

In the Break or in the Peloton - Factors Affecting Cyclists Route Choice


Jon Crockett, Mott MacDonald, Steven O'Hare, Mott MacDonald, Adrian Barritt,


Research into cyclists’ route choice in London. This combined three waves of data collection, including observed paths via a bespoke App, and choice modelling to derive utility parameters suitable for implementation in a new network model.


‘The Mayor’s Vision for Cycling in London’ (Greater London Authority, 2013) sets out an ambitious programme of investment in cycling. In order to support this, Transport for London has been developing its analytical capability to better appraise the impacts of cycling infrastructure. This will help to understand better where investment in cycling should be prioritised, The study which is the focus of this paper aimed to provide key insight into how cyclists choose their route for implementation in TfL’s new cycling route choice model. The key objectives of this study were to:
 Provision of parameters or values for a Generalised Cost (GC) function in a network-based cycle route choice model, segmented as appropriate to best explain observed behaviour; and
 Provide outputs which are compatible for incorporation into TfL’s network based model of cyclist route choice (CYNEMON).
The study was undertaken in six stages:
 A Literature Review of past studies, including two Revealed Preference (RP) studies in Zurich and San Francisco, plus a number of Stated Preference (SP) studies;
 A Stage 1 survey to recruit respondents and gain detailed socio-demographic and travel pattern characteristics for later segmentation analysis;
 A Stage 2 data collection exercise, where recruited respondents recorded their cycle journeys in London;
 A Stage 3 survey, to help validate the findings of the choice modelling, where respondents were asked to explain their observed choices;
 Build of a detailed GIS network of cycle links in London, including the mapping of multiple possible explanatory variables to a spatially detailed network; and
 Route choice models for the observed journeys, involving the generation of alternative routes to create a ‘choice set’, examination of different combinations of explanatory variable, and the transformation of outputs suitable for implementation in CYNEMON.
In comparison to assignment of motor vehicles in highway assignment modelling, which typically focuses on time and cost [based on distance] attributes, cycle route choice decision-making is perceived to be influenced by both different, and a greater number of attributes. In the UK it is a relatively under-researched field, reflecting the relatively low mode share for cycling in recent decades and a historic, relative to the present day, lack of policy focus during that period. This environment has now shifted markedly, with an increased recognition of the economic, health, and social benefits associated with increasing the proportion of trips made by cycling.
In Stage 2, 8,663 observed routes were collected from 774 unique users of the App. These were then subject to a set of automated cleaning processes with the aim of identifying routes with a significant volume of signal loss, missing links (e.g. where there is no network), potential walking (slow average speed), potential transit use (high average speed), and deviation between the observed route and that created using best path analysis.
The resulting choice models explaining a high degree of the variance in the choice set, with adjusted rho-squared statistics of 0.89, and showed the following factors to have a statistically significant influence on route choice:
 Distance;
 Proportion of route on different classifications of link, e.g. A Road, Minor Road, Canal or Park route etc.;
 Proportion of the route with different types of cycle infrastructure, including on and off highway provision and the availability of bus lanes;
 One-way restrictions in the absence of a dedicated contraflow for cyclists;
 Proportion of route which is signed;
 Average annual 12 hour, bi-directional, traffic volume; and
 Total metres gained or lost per 100m.
The absence of any junction/node/turn specific attributes ran against prior hypotheses, although the final set of attributes did accord with the findings of the two user surveys. We believe that this finding is a result of collinearity in the dataset, the relative quality/spatial accuracy of the datasets available, and also the choice set generation process – where prior assumptions on the attributes of interest are required to create alternative routes.
Finally, for implementation in an assignment model, the parameters from the choice model had to be reconciled from a route based level to a link and node level in a network with over 470,000 unique components. We conclude by reporting on calibration and validation stages in CYNEMON, and the recommended next steps for London and in this field more generally.


Association for European Transport