Non-linear associations between built environment and active travel for working and shopping: An extreme gradient boosting approach

Abstract

Active travel has environmental, social, and public health-related benefits. Researchers from diverse domains have extensively studied built-environment associations with active travel. However, limited attention has been paid to distinguishing the associations between built environment characteristics at both the origins and destinations and active travel for working and shopping. Scholars have started to examine non-linear associations of built environment with travel behaviour, but active travel has seldom been a focus. Therefore, this study, selecting Xiamen, China, as the case, utilises a state-of-the-art machine learning method (i.e., extreme gradient boosting) to explore the non-linear associations between built environment and active travel for working and shopping. Our findings are as follows. (1) For both purposes, trip characteristics contribute the greatest, and the built environment is also quite important and has larger collective contributions for active travel than does socioeconomics. (2) The relative importance of built environment on active travel for shopping is evidently larger than that for working. (3) All built-environment variables have non-linear associations with active travel, and associations with active travel for working are generally in inverted U or V shapes, while those with shopping trips have much more complex patterns. (4) Differences in the threshold value and gradient exist between built-environment associations with active travel for working and shopping and between variables at origins and destinations. Decision makers are recommended to meticulously disentangle the complex influences of built environment on active travel and distinguish between diverse purposes to make informed and targeted interventions.

Introduction

Active travel, i.e., making journeys by physically active means like walking and cycling (Litman, 2003), is widely recognised to have multiple benefits. As environmentally friendly transport, active travel can act as an efficient alternative for motorised transport modes, especially private driving, and reduce a variety of pains induced by increasing motorisation, such as traffic congestion, air pollution, and energy consumption (Frank et al., 2010; Woodcock et al., 2009; Maizlish et al., 2017). Active travel is also found to be associated with higher transportation satisfaction (De Vos et al., 2016). Furthermore, by integrating physical activity into people's daily lives (Sahlqvist et al., 2012), active travel is confirmed capable of preventing or mitigating diverse chronic diseases, such as obesity (Bassett et al., 2008; Flint and Cummins, 2016), type 2 diabetes (Millett et al., 2013), cardiovascular diseases (Celis-Morales et al., 2017), coronary heart disease, hypertension (Saunders et al., 2013), and mental disorder (Singleton, 2019), potentially saving billions each year in health-related costs (Jarrett et al., 2012). Additionally, as one of the key agreeable outcomes of compact development that can, in turn, demonstrate the effectiveness of investment for sustainability (Koszowski et al., 2019), active travel is oftentimes associated with higher spatial vibrancy and sense of community (Wood et al., 2010) and provides high social capital by promoting social integration and interaction (van den Berg et al., 2017).

Therefore, scholars from diverse domains such as transportation, urban planning, and public health have extensively studied active travel, in particular, on its socio-economic and built-environment factors. For example, in their seminal work, Cervero and Kockelman (1997) proposed the ‘3Ds’ model (i.e., density, diversity, and design) to delineate and quantify the essential characteristics of the built environment. They found that higher density, land use diversity, and pedestrian-oriented design are statistically significantly associated with a higher probability of non-motorised travel (Cervero and Kockelman, 1997). The ‘3Ds’ model has then become the most frequently used model to portray the built environment in the exploration of its relationship with travel behaviours. Later, the ‘3Ds’ model was advanced to the ‘5Ds’ model by Ewing and Cervero (2001), with the addition of distance to transit and destination accessibility. Ewing and Cervero (2010) further expanded the ‘5Ds’ model to the ‘6Ds’ and ‘7Ds’, with demand management and demographics added. They also confirmed that such variables as land use diversity, intersection density, and destination accessibility can significantly encourage walking (Ewing and Cervero, 2010). The ‘Ds’ model has triggered a multitude of studies on associations between the built environment and travel behaviours including active travel (e.g., (Handy et al., 2005, Hong et al., 2014, Clark et al., 2014, Grow et al., 2008, Aziz et al., 2018), providing policy makers with rich insights for making informed interventions.

However, despite innumerable empirical findings, existing research on active travel is insufficient in the following two aspects.

First, scholars have predominantly focused on work travel and its correlates/determinants, and limited research attention has been given to active travel for other purposes, such as shopping, and even fewer studies delve into the differences in the correlates between active travel for working and shopping purposes. Although working travel is often the most important and regular travel for most adults, which is why it draws the heaviest research attention, travel for shopping has taken up an increasingly significant role in people's daily life (Loo and Wang, 2018), only second to working travel for most people (Meena et al., 2019). Sustaining active travel for shopping or diverting the motorised mode choice for shopping toward active travel may be more feasible than that for working, because shopping trips are much less spatially and temporally constrained than working trips (Popovich and Handy, 2015). The fact that working and shopping trips may be distributed geographically unevenly also entails that urban planners make targeted priorities for different trips in different places. Nevertheless, given the potential differences in the associations between built environment and travel for working and shopping, policymakers basing their decisions toward travel for shopping on the existing knowledge and evidence that are mainly derived from travel for working might be problematic. Thus, examining whether and how the built-environment factors influencing active travel for such two purposes differ with each other is highly necessary.

Second, existing findings are mainly based on the hypothesis that the built environment has linear or pre-determined non-linear (e.g., logarithmic and nth power) effects on travel behaviour. However, such presumptions may oversimplify and miscalculate the complex associations between built environment and travel behaviour. In recent years, availing themselves of state-of-the-art methods (e.g., machine learning methods), scholars have attempted to escape such assumptions and examine the non-linear associations, of which the shapes are not pre-determined, between the built environment and travel behaviour (e.g., (Hagenauer and Helbich, 2017, Ding et al., 2018a, Ding et al., 2018b, Wang and Ross, 2018, Cheng et al., 2019, Ding et al., 2019b, Wu et al., 2019, Zhou et al., 2019, Zhao et al., 2020, Cheng et al., 2020)). Nevertheless, analogical efforts on active travel remain limited. The only two exceptions, to the extent of our knowledge, are the studies by Christiansen et al. (2016) and Tao et al. (2020b). Therefore, additional explorations in the non-linear associations between built environment and active travel are warranted. The present study advances the above two by distinguishing different travel purposes.

Against this backdrop, this study, taking Xiamen, China, as the case, uses a state-of-the-art machine learning method, i.e., extreme gradient boosting (XGBoost) method, to explore the non-linear associations of built environment with the probability of active travel (walking and cycling) for working and shopping. It contributes to the literature from the following aspects. First, it depicts the built environment characteristics at both the origins and destinations of each working and shopping trip, advancing most of its fellow studies that delineate the built environment at only the origins of trips (usually residential neighbourhood). Second, this study is among the first ones to model the non-linear associations between the built environment and active travel, and it interprets the models by presenting the relative contribution of each variable and illustrating the non-linear built environment-active travel associations with partial dependence plots (pdp). Third, it elucidates the non-linear associations of built environment with active travel for both working and shopping and conducts comparisons wherever feasible, providing fruitful implications for decision makers.

The remainder of the article is organized as follows. Section 2 reviews the relevant studies. Section 3 describes the study area, data source, and method. Section 4 describes, analyses, and discusses the modelling results. Section 5 concludes.