Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Spatial Heterogeneity in the Nonlinear Impact of Built Environment on Commuting Time of Active Users: A Gradient Boosting Regression Tree Approach

Spatial Heterogeneity in the Nonlinear Impact of Built Environment on Commuting Time of Active... Hindawi Journal of Advanced Transportation Volume 2023, Article ID 6217672, 15 pages https://doi.org/10.1155/2023/6217672 Research Article Spatial Heterogeneity in the Nonlinear Impact of Built Environment on Commuting Time of Active Users: A Gradient Boosting Regression Tree Approach 1,2 2 3 4 Jingxian Wu , Guikong Tang , Huapeng Shen , and Soora Rasouli Key Laboratory of Earthquake Engineering and Engineering Vibration, Institute of Engineering Mechanics, China Earthquake Administration, Beijing, China Department of Transportation Engineering, University of Shanghai for Science and Technology, Jungong Road #516, Shanghai 200093, China Beijing Key Laboratory of Trafc Engineering, Beijing University of Technology, No. 100 Ping Le Yuan, Chaoyang District, Beijing 100124, China Urban Planning and Transportation Group, Eindhoven University of Technology, P.O. Box 513, Eindhoven 5600 MB, Netherlands Correspondence should be addressed to Huapeng Shen; shenhuapeng1997@163.com Received 10 January 2022; Revised 24 January 2023; Accepted 2 February 2023; Published 13 February 2023 Academic Editor: Hongtai Yang Copyright © 2023 Jingxian Wu et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Many studies provided evidence regarding the infuence of built environment (BE) on commuting time. However, few studies have considered the spatial heterogeneity of such impacts. Using data from Nanjing, China, this study employs two-step clustering and gradient boosted regression trees (GBRT) to segment the neighborhoods into diferent types and investigate the efects of BE characteristics on the commuting time of active users. Te results show a strong efect of BE characteristics on commuting time, involving active modes. Te importance of BE characteristics varies among neighborhood types. For active commuters in the internal region of Nanjing, commuting time is afected mostly by the land use mix at the work end. Te lowest impact of BE in internal regions is associated with metro station density. For active commuters in external region of the city, the relative importance of intersection density at the home end is the largest (as high as 5.76%). Moreover, other signifcant diferences are found in the associations between BE characteristics and active commuting time in the two regions. commute distance [4, 5]. Longer trip distances imply that 1. Introduction active commuters will spend more time on the road and thus Active travel mode, referring to walking or cycling, is a viable have less life satisfaction. Despite the fact that active mode alternative to driving in short-to-medium distance trips [1]. commuting generates positive utilities such as a health- Regular active travel such as active commuting to work is enhancing efect, commuters prefer to shorten it due to time thought to beneft both the environment and an individual’s budgets [6]. Active transportation, particularly cycling, can physical health [2]. To encourage active travel, several in- be used for longer distance trips if the time cost of com- terventions such as bike-sharing programs and provisions of muting by active modes is appropriately reduced [4]. As a footpaths and cycle lanes have been implemented. However, result, it is critical to investigate factors that infuence the such promotion does not bring a signifcant increase in the commuting time of active users. share of active commuting, and private cars are still the most Many studies on travel behavior have found that the built widely used mode [3]. environment and sociodemographics are strongly associated Te low prevalence of active commuting can be at- with walking and cycling (e.g., see [7–14]). Individuals’ tributed in part to urbanization’s increasing average active travel choices are infuenced by built-environment 2 Journal of Advanced Transportation characteristics such as walk-bike infrastructure [7], street found that mixing multiple land uses beyond a certain pattern [8], route connectivity [9], street greenery [10], and proportion can have an adverse efect [28]. For commuting, the mix of land use types that specify jobs and houses plays a population density [11]. Changes in the built environment, according to Handy et al. [12], infuence travel mode choice major role. Te appropriate job-to-housing ratio can be an primarily by altering travel time. Furthermore, Eldeeb et al. indication of whether residents are likely to be employed in discovered that improving the built environment does not the neighboring area of their residence. Tis, in turn, in- have a homogeneous impact on the likelihood of using active fuences commute distance and facilitates using active travel mode in diferent parts of the city [13]. However, rare studies to go to work. have focused on the commuting time of active mode and All the above-mentioned studies have confrmed the examined whether the built environment has a spatially contribution of built environment to active travel primarily diferent impact on it. by using regression-based models such as multilevel re- Terefore, this study attempts to advance the literature gression [11], a logit-based model [8, 18, 24], and a structural by investigating the nonlinear associations between the built equation model [23, 25]. However, these models are pri- marily based on a priori, often linear relation between the environment and travel time of active commuters, ac- counting for spatial heterogeneity. First, the two-step built environment and active mobility. A parallel stream of clustering method is used to diferentiate the regions with studies examines the association between built environment various built-environment features. Next, gradient-boosted and travel behavior using machine learning methods. Ding regression trees are generated for predicting the commuting et al. applied decision trees [29] to extract the nonlinear time of active travelers in each region. Travelers’ socio- relationship between a built environment and commute demographics, trip characteristics, and built-environment mode choice. Tao et al. assessed the importance of BE characteristics are considered as conditional variables. components for the energy consumption of active users by Te remainder of this article is structured as follows. applying gradient boosting decision trees [19]. Tey found Section 2 presents a literature review on the association the distance to the nearest park posed the greatest impact. between built environment and active travel. Section 3 Cheng et al. used a random forest to assess BE’s impact on elderly active travel and found that population density had describes the data source used in this study and built-en- vironment characteristics at the trafc analysis zone (TAZ) the highest contribution [30]. Liu et al. in their most recent level, as well as commuters’ personal profles. Section 4 study used the extreme gradient boosting approach to ex- briefy describes the methodology used in this study, while amine the association of built environment and active travel Section 5 discusses the fndings. Finally, we present the choice. Tey found that trip characteristics contributed more summary of the key fndings and discuss their implications than the built environment [31]. for planning practice. 2.2. Spatial Heterogeneity. Spatial heterogeneity refers to the 2. Literature Review varying impact of the same infuential factors at diferent spatial scales or geographical locations. Several studies have 2.1. Association between the Built Environment and Active investigated whether the relationship between the built Travel. Te built environment is the physical setting environment and travel behavior varies across diferent designed to meet people’s need to engage in activities. Te types of neighborhoods [32, 33]. Srinivasan and Ferreira built-environment characteristics that relate to residents’ found the built environment around residential areas to pose travel behavior are defned from “3Ds” into “5Ds,” which are diferent efects on travel mode choice compared to that “density,” “diversity,” block “design,” “destination,” “ac- around workplaces [34]. For instance, land use mix around cessibility,” and “distance” to transit [15, 16]. Some scholars specifcally focused on the impact of the “5Ds,” as refection households’ residence has stronger efects on household travel mode choice and travel distances than that at job of built-environment characteristics, on active travel (e.g., see [7–14, 17–19]). locations [35]. Using the geographically weighted regression (GWR) model, Tu et al. found signifcantly diferent impacts Density, as a key component of built environment, has a of built environment on travel mode choice [36]. Zhang et al. paradoxical efect on active travel choice [20]. Density used a hierarchical linear model to explore the relationship appeared to have a signifcant impact on choosing active between the neighborhood-built environment and trip travel modes in some studies, such as reference [21], but not distance [37]. Tey found signifcant spatial heterogeneity in in other studies (e.g., see [22]). Zhu et al. discovered that the infuence of the built environment on travel behavior. increasing population density increases residents’ possibility of commuting in active mode [11]. Block design such as Neighborhoods located in diferent areas of the city sometimes share similar characteristics, but their efects can street crossing density, road density, and connected side- walks have profound efects on active travel [8, 23, 24]. be diferent. Zhong et al. used a geographically weighted regression model to analyze the spatial heterogeneity of the Accessibility indicators such as employment accessibility efects of an urban built environment on road travel time [18], destination accessibility [25], and transit accessibility and found that spatially varying relationships exist [38]. [26] are found to be positively correlated with residents’ use Ding et al. used spline in a mixed logit model and concluded of active travel. Diversity that measures land use mix in a that nonlinearity exists in the relation between built envi- neighborhood/region is closely associated with the choice of ronment and commuting mode choice [39]. active travel [13, 23, 27]. Furthermore, Raman and Roy Journal of Advanced Transportation 3 In summary, existing studies have empirically done a lot bikes due to the infrastructure limitation. Tus, this study on the measurement of BE on active travel behavior or included it as one of the active travel modes, from home to work in the morning were chosen for this study. spatial heterogeneity of BE on travel behavior. However, none have investigated how the built environment infu- Te built-environment characteristics are measured at ences the commuting time of active modes and compre- the TAZ level using the software of ArcGIS. Data sources, hensively considered the potential impact of spatial including Baidu map POIs, open street maps, and urban heterogeneity. Tus, this study contributes to the existing land use GIS data, are used. Many studies have focused on body of the literature by exploring nonlinear relation the built-environment characteristics surrounding resi- (without any a priori assumptions) between built environ- dences, but others, such as Sun et al. [22] and Ding et al. [29] ment and active commuting time taking into account spatial have emphasized the importance of trip destination char- heterogeneity. It is important to note that the duration of acteristics. Tus, this study measures the TAZ characteristics for both the home and workplace ends. Among all these data active mobility for commuting purpose has never been the subject of examination within the topic of nonlinear rela- sources, POIs provide geographic information about specifc points and are used to calculate transit-related indicators tionship between built environment and active mobility. such as intersection density, bus stop density, and metro station density within the TAZ. Te open street map is used 3. Data to calculate the density of roads in each TAZ. Te urban land 3.1. Study Area and Data Sources. Te data used in this study use GIS data are used to calculate three indicators of land originated from Nanjing, China. Nanjing is a mega-city and use: land use mix and the ratios of residences and working the provincial political and economic center of Jiangsu. It is places to the area of TAZ. Te land use mix is determined by located in the eastern region of China, downstream of the an entropy index of nine land use types around commuters’ Yangtze River. It is an important gateway city for the central residences and work places. Te nine diferent types of land uses include residence, industrial use, public administration, and western regions’ development, which is fueled by ra- diation from the Yangtze River Delta. Nanjing is divided into commercial services, green space and plazas, construction, transportation, public facilities, and warehousing. Te in- 11 administrative regions, covering a total area of 6,587 km and a built-up area of 868 km . Te resident population was dicator is calculated as follows: 9.42 million in 2021, with an urban population of 8.19 −1 LandMix � 􏽘 p ln p , (1) million and an urbanization rate of 86.9%. In 2021, the city’s i i ln n gross regional product reached 163,532 billion yuan. Tis study concentrates on the most urbanized regions, including where p is proportion of the type i land use, and n is the Gulou, Qinhuai, Xuanwu, Jianye, Yuhuatai, Qixia, Jiangn- number of land use types. ing, Pukou, and Luhe (two regions, Gaochun and Lishui, were newly designated as regions during the urbanization 3.2. Data Description and were not involved in this survey). Seven of them are on the southern side of the Yangtze River, while two are on the 3.2.1. Built Environment at the TAZ Level. Table 1 shows the northern side. For transportation census and management, defnitions and descriptive statistics of eight built-envi- the Nanjing transportation planning agency divides the ronment variables obtained at the TAZ level. All 766 TAZs entire study area of Nanjing into 766 trafc analysis zones have the BE characteristics in fve dimensions: density, (TAZ) based on land use and administration boundaries. design, distance to transit, destination accessibility, and Figure 1 depicts the study area of Nanjing, China. diversity. Housing density corresponds to the density of In this study, four data sources are used: Nanjing residential development. Road density and intersection Household Travel Survey data from 2016; Nanjing urban GIS density that describe the characteristics of a street network data; points of interest (POIs) from Baidu map; and an open represent street design. Bus stop density and metro station street map (OSM). Te Nanjing Household Travel Survey is density measure the accessibility of bus and subway services an annual survey conducted by the Nanjing transportation and are related to distance to transit. Job density indicates planning agency. It is carried out through household in- destination accessibility; land use mix represents diversity; terviews in order to learn about the daily mobility patterns of and distance to the CBD refects regional location. urban residents. In 2016, the survey employed the stratifed random sampling technique to guarantee that the sample size was proportional to the population size. In total, 8,387 3.2.2. Statistics of Active Commuters. We matched BE fea- people from 3,015 households were invited to participate in tures for each active commute trip based on both home-end the survey. Te survey collected individual’s sociodemo- and work-end TAZs in the trip records. Table 2 depicts the graphic information (e.g., household income, car ownership, sociodemographics, trip characteristics, and BE character- gender, and age) and their travel diaries (e.g., trip origin, istics of 1,937 commuters by active modes. Males account for destination, purpose, departure time, and travel mode) on a 43.9% of the sample, which is slightly lower than females. given day. Based on the provided trip purpose and travel 56.4% have a bachelor degree or higher the majority are aged mode information, 1,937 commuters used active modes, between 30 and 49 and 52.7% own a driving license. Te such as walking, bicycling, and e-cycling. In China, riding an average household has 0.63 cars, while the average number e-bike does not imply much higher speed than ordinary of children aged six years old or below is 0.12. 53.5% of the 4 Journal of Advanced Transportation Legend Administrative regions Gulou Jiangning Jianye Luhe Pukou Qinhuai Qixia Xuanwu Yuhuatai TAZ 40°N CBD 30°N 20°N 120°E 100°E 04 8 16 24 32 km Figure 1: Case study area in Nanjing. Table 1: Built-environment characteristics for 766 TAZs. Names Variable descriptions Means (S.D.) Road density Total road length/TAZ area (km/km ) 6.85 (4.57) Intersection density Intersections/TAZ area (count/km ) 4.39 (11.55) Bus stop density Bus stops/TAZ area (count/km ) 4.85 (4.95) Metro station density Metro stations/TAZ area (count/km ) 0.23 (0.61) House density Residential area/TAZ area 0.19 (0.20) Job density Industrial, public administration, and commercial area/TAZ area 0.24 (0.21) Land use mix An entropy index of nine types of land use 0.41 (0.19) Distance to CBD Euclidean distance from TAZ centroids to CBD (km) 16.40 (9.66) CBD refers to the center of the TAZ in which Xinjiekou business district is located. respondents have an annual household income over 100,000 4.1. Two-Step Clustering Method. Te two-step clustering CNY. Te majority (81.5%) leaves for work between 7:00 am method has been extensively used in the transportation and 8:30 am. Te sample’s average commute time is feld due to its fexibility and capability in data processing 21.52 minutes, and the average trip distance to work is [40, 41]. It has an advantage over other clustering 4.98 km. techniques in that it can handle both continuous and discrete variables simultaneously. In addition, it can determine the optimal number of clusters automatically 4. Methodology and its clustering accuracy is unafected by the size of To examine the impact of built-environment characteristics data [42, 43]. on active commuting time, this study frst divides Nanjing’s Te clustering consists of two procedures. First, it 766 zones into diferent types using a two-step clustering clusters the 766 zones into groups according to their sim- method. Ten, in each region, gradient-boosted regression ilarity in BE characteristics. Ten, the merging algorithm is trees are constructed to investigate the determinants and used to gradually combine these groups until only one group relative importance of the infuential factors on active is left. Te optimal clustering number is determined using commuting time. Te following sections elaborate the Bayesian information criterion (BIC). Interested readers can specifcs of the analysis. refer to Chiu et al. [44]. Journal of Advanced Transportation 5 Table 2: Sample description of active commuters (N � 1937). Names Variable descriptions Means (S.D./percent) Sociodemographics Gender Gender: 1 � male; 0 � female 0 � 56.1%, 1 � 43.9% Education Hold a bachelor degree or above: 1 � yes; 0 � no 0 � 43.6%, 1 � 56.4% Respondent’s age: 1 � 20–29 years old; 2 � 30–39 years old; 3 � 40–49 years old; 1 � 17.6%, 2 � 27.1%, 3 � 36.1%, Age 4 � 50 or more years old 4 � 19.2% License Hold a driving license: 1 � yes; 0 � no 0 � 47.3%, 1 � 52.7% Cars Number of cars owned by a household (count) 0.63 (0.58) Child Number of children at 6 years old or younger (count) 0.12 (0.34) Income Household income per year: 1 � over 100,000 CNY; 0 � other 0 � 46.7%, 1 � 53.3% Trip attributes Te commute trip occurs in morning peak hours from 7:00 am to 8:30 am: 1 � yes; Departure time 0 � 18.5%, 1 � 81.5% 0 � no Euclidean distance from residential TAZ centroid to workplace TAZ centroid Trip distance 4.98 (6.38) (km) Commuting time Commuting time spent on road (min) 21.52 (11.98) Built environment at home end Road density Total road length per TAZ area (km/km ) 9.78 (3.54) Intersection Intersections/TAZ area (count/km ) 10.12 (16.99) density Bus stop density Bus stops/TAZ area (count/km ) 8.32 (5.39) Metro station Metro stations/TAZ area (count/km ) 0.53 (0.92) density House density Residential area/TAZ area 0.31 (0.19) Job density Industrial, public administration, and commercial area/TAZ area 0.30 (0.21) Land use mix An entropy index of nine types of land use 0.52 (0.16) Distance to CBD Euclidean distance from TAZ centroids to CBD (km) 9.39 (8.38) Built environment at work end Road density Total road length per TAZ area (km/km ) 9.77 (3.91) Intersection Intersections/TAZ area (count/km ) 11.48 (19.05) density Bus stop density Bus stops/TAZ area (count/km ) 8.48 (5.47) Metro station Metro stations/TAZ area (count/km ) 0.46 (0.86) density House density Residential area/TAZ area 0.31 (0.19) Job density Industrial, public administration, and commercial area/TAZ area 0.30 (0.20) Land use mix An entropy index of nine types of land use 0.52 (0.16) Distance to CBD Euclidean distance from TAZ centroids to CBD (km) 9.35 (8.59) CBD refers to the centroid of the TAZ in which Xinjiekou business district is located. 4.2. Gradient-Boosting Regression Trees. Gradient-boosted errors of the previous ones. Given the training data regression trees (GBRT) are an ensemble model that (y , x ) , the specifc learning steps are as follows: 􏼈 􏼉 i i 1 combines gradient-boosting and regression trees [45, 46]. It (1) Initialize the base model F (x) to be a constant: has myriad merits over the traditional linear regression methods and has been often used in transportation research F (x) � argmin 􏽘 L y , c􏼁 , (2) [29, 37, 45]. First, GBRT is more efective at data prediction 0 c i�1 and interpretation than general linear regressions or even just a single tree due to its tree-based ensemble feature. where y is the observed value, c is the predicted Second, it accommodates data with missing values and value, and N is the number of observation. Squared avoids multicollinearity of explainable variables. Tird, it error is chosen as the loss function for the regression. can calculate the relative importance of each variable (2) For m � 1 to M (M is the times of iterations or without making assumptions about the variables’ relation- optimal number of trees), compute the residual ships. Fourth, it is adaptable to both continuous and cate- which is mathematically calculated by the negative gorical types and is applicable to small data sets. derivation of loss function with respect to the pre- Furthermore, it avoids the overftting issue that frequently vious model outcome: arises as the number of tree nodes rises by using gradient boosting. zL y , F x􏼁􏼁 i i r � −􏼢 􏼣 , i � {1, . . . , N}, (3) Te GBDT model combines multiple regression trees mi zF x􏼁 F(x)�F (x) m−1 sequentially with each new tree adding up to correct the 6 Journal of Advanced Transportation where r is negative gradient, and F(x ) is the where J is the number of terminal nodes, J-1 is the number of mi i previous model. the nonterminal nodes, υ is the feature associated with the node t, τ is the improvement in squared error after the (3) Fit a regression tree to the residuals r and minimize mi splitting node t, and I[υ � κ] equals to 1 when υ � x , or 0 t t κ the loss function: otherwise. h (x) � 􏽘 c I􏼐x ∈ R 􏼑, m mj mj 5. Results j�1 5.1. Identifcation of the Neighborhood Types. As shown in ⎝ ⎠ ⎛ ⎞ c � argmin 􏽘 L y , F x􏼁 + 􏽘 cI􏼐x ∈ R 􏼑 , mj c i m−1 i mj Table 3, there are 766 TAZs in Nanjing with varying built- x ∈R j�1 i mj environment characteristics. Te two-step clustering (4) method is used to cluster these TAZs with more homoge- neous spatial features. To eliminate the infuence of col- where h (x) is the m th regression tree, J is the tree m linearity on clustering results, the Pearson correlation depth, referring to the number of terminal nodes, coefcient is used to test the association between pairs of BE R is the disjoint region partitioned by the terminal mj variables, and the variance infation factors (VIF) are cal- nodes of m th tree, c is the optimal coefcient for mj culated to measure the degree of collinearity. Except for bus R , and I(x ∈ R ) equals to 1 when (x ∈ R ), or 0 mj mj mj stop density, all BE characteristics have coefcients less than otherwise. 0.6 (0.7–1.0 indicates strongly correlated) and all VIFs (4) Update the model: calculated are less than 3. Both indicate that all BE variables are suitable for clustering. In a stepwise approach, the ratio change in BIC and ratio of distance measures for a variety of F (x) � F (x) + 􏽘 c I􏼐x ∈ R 􏼑. (5) m m−1 mj mj clusters are identifed. A model with two clusters appears to j�1 be optimal, with a silhouette coefcient value of 0.5. Table 3 shows the centroids for the two clustered groups To prevent overftting in the training procedure, hyper- as well as the signifcance of their diferences in each BE parameters including optimal number of trees M, learning characteristic. All 766 TAZs are divided into two groups: rate ], and tree depth J should be estimated by using test data Cluster-1 with 263 TAZs and Cluster-2 with 503 TAZs. Te or cross-validation. Te model is replaced by spatial heterogeneity of TAZs has been interpreted using the centroids for each group. TAZs in Cluster-1 are featured by a F (x) � F (x) + ] 􏽘 c 1􏼐x ∈ R 􏼑, (6) m m−1 higher road density (10.64 km/km ), more intersections mj mj j�1 (11.04 count/km ), more access to metro stations (0.61 count/km ), higher ratio of residential land (0.40), more job where ] is the learning rate that scales the contribution of opportunities (0.33), higher land use mix (0.58), and closer each tree. It has the value range from 0 to 1. Smaller values of proximity to CBD (8.11 km). TAZs in Cluster-2 have a lower learning rate give rise to larger M value and results in minor road density (4.86 km/km ), fewer intersections (0.92 count/ test error. 2 2 km ), less developed metro service (0.04 count/km ), less Te optimal values of these parameters are determined residential land use (0.08), lower job coverage (0.19), a lower by performing the 5-fold cross-validation. Root mean land use mix (0.33), and are located far away from the CBD squared error (RMSE) is chosen as the performance (20.74 km). Given the spatial diference between the cen- measurement. Parameters that result in the lowest cross- troids, we named cluster-1 as the internal region and cluster- validation error are preferred in the fnal model. For the 5 2 as the external region. Te Mann–Whitney U test method test test datasets in cross-validation, RSME is calculated as is used to compare the diferences in BE characteristics follows: between the two groups. Te result demonstrates their 􏽶�������������� spatial diferences. Te two types of TAZs in Nanjing are test shown in Figure 2. (7) RSME � 􏽘 􏽘 y − y 􏽢 􏼁 , t t 5 ′ f�1 f t�1 5.2. Results of GBRT. Using the GBM package in RStudio, where N is the data number in test set f. GBRT models for the commute time of active commuters Meanwhile, the learned regression trees 􏼈T 􏼉 provide living in each region are estimated. Te relative importance interpretative results that show the relative infuence of an of infuential factors is calculated for both identifed regions. explanatory variable x as follows [45]: Te relative importance is measured by comparing the error 2 1 2 reduction of one variable in commute time compared to 􏽢 􏽢 I � 􏽘 I T 􏼁 , κ κ m other variables. All variables included have a total impor- m�1 tance that adds up to 100%. Prior to modeling, hyper-pa- (8) rameters including learning rate, optimal number of J−1 2 ⌢ I T 􏼁 � 􏽘 τ I􏼂υ � κ􏼃, iterations (or the number of trees), and tree depth must be m t κ t t�1 tuned. Ridgeway recommended setting the learning rate for Journal of Advanced Transportation 7 Table 3: Centroids for the TAZ clustering results. Attributes Cluster-1 (263 TAZs) Cluster-2 (503 TAZs) Mann–Whitney U Sig. Road density 10.64 4.86 15990.00 <0.001 Intersection density 11.04 0.92 27695.00 <0.001 Metro station density 0.61 0.04 41884.00 <0.001 House density 0.40 0.08 9588.50 <0.001 Job density 0.33 0.19 34587.00 <0.001 Land use mix 0.58 0.33 16321.00 <0.001 Distance to CBD 8.11 20.74 13653.00 <0.001 Classifcation Administrative regions Internal External 01 5 0 20 30 km Figure 2: Spatial distributions of TAZs in clustered neighborhood types. practice between 0.01 and 0.001 [46]. Te smaller learning order to fnd the best GBRT, we initially developed the rate is thought to improve model performance. We set the model with the depth of the tree ranging from 1 to 49 in learning rate as 0.001 in accordance with Tao et al. [19]. In increments of 1. Te optimal parameters are then 8 Journal of Advanced Transportation the home end. Tis can be explained by the high aggregation determined using the RMSE value of fve-fold cross-vali- dation, which varies as tree depth increases. Figures 3 and 4 of morning commutes at the destination over the origin. Similarly, bus stop density at the work end is as high as visualize the RMSE values versus tree depth and the optimal number of iterations for the internal and external regions, 4.32%, greater than that at the home end. Tis confrms the respectively. Te RMSE in the internal region decreases with roles of transit accessibility on travel behavior [48], as well as increasing tree depth until it reaches 28. In the external the fact that the resultant trip time is more infuenced by the region, however, this indicator becomes stable at a depth of BE feature at the trip end. Metro station density is the least 19. As a result, 28 were set as the tree depth for the model in important BE factor, and its importance at both ends is less the internal region and 19 for the model in the external than 2.00%. Tis could be explained by the least variation in region. Te diference between internal and external regions metro services in the internal region. Geographical locations with respect to the number of iterations is even larger. of home and work ends that are presented by distance to CBD pose the contributions, 3.60% and 3.84%, respectively. According to the results, the commuting time model in the internal region iterated 3,140 times before convergence, In the external region, intersection density at the home end is the most infuential BE factor, with a relative while the model in the external region iterated 2,728 times. Both models ft well, with pseudo-R values of 0.637 and importance of 5.76% while its importance at the work end is 4.47%. Land use characteristics at the work end, in- 0.585 in the internal and external regions, respectively. Tese values are greater than those of traditional linear regressions, cluding job density, house density, and land use mix, have which are 0.207 and 0.237. For comparison, we estimated a higher contributions in the external region, ranking third, general GBRT for all active commuters and found that the fourth, and ninth, respectively. In contrast to that in the model has the lower pseudo-R (0.509). Tis indicates that internal region, the density of metro stations at work ends incorporating spatial heterogeneity in creating the GBRT in the external region has a greater impact, accounting for improves the model ft. 4.49% of the total. Tis may be due to the proximity of metro services to the workplace. Te bus stop density at both ends contributes around 3.5%, which is comparable 5.2.1. Relative Importance of Infuential Factors. Relative to the internal region. Te remaining BE variables had importance is commonly used in machine learning to only a minor infuence. For active commuters in the measure how much a factor infuences a dependent variable. external region, the distance from the work end to the All variables in this study have a relative importance that CBD (4.22%) contributes more to their trip duration than sums up to 100%. Te greater the relative importance of the the home end (3.08%) does. factor, the greater it contributes. Table 4 is the calculated relative importance of each infuential factor in determining 5.3. Spatial Heterogeneity in BE Impact. To describe the active commuting time in internal and external regions. Te spatial heterogeneity of BE impact, a more thorough result demonstrates that built-environment characteristics comparison of derived BE importance as well as BE asso- have a higher collective importance than social demo- ciations with active commuting time was made. graphics. Tis is consistent with the fndings of some earlier studies [30, 31]. Te importance of built-environment characteristics at both commute trip ends is 63.29% and 5.3.1. BE Importance. Figure 5 shows the comparison of BE 54.92%, respectively, for internal and external regions. Te importance to active commuting time in two regions. In the roughly 8% gap could be due to the more spatially con- internal region, nearly all BE variables have a relative im- strained nature of active commute trips in the internal re- portance more than 3%, with the exception of metro station gion. In both regions, built-environment features at the density at both ends and intersection density at the home work-end pose higher importance than those at home end, end. In external region, all BE variables at the work end and which is consistent with the fnding of Ding et al. [29]. three out of eight BE variables (intersection density, bus stop Similarly, we fnd diferences in the collective importance of density, and distance to CBD) at the home end have relative sociodemographics in both the regions. Tey are 5.49% more importance over 3%. Te most signifcant disparity is in the important in the external region than in the internal region. roles of street network-related factors, metro station density, Active commuters in the external region have more fexi- and land use-related factors. bility in determining their commuting time than those in the Te road network density at both ends contributes internal region. 1.78%∼2.69% more in the internal region than it does in In the internal region, road network density at both trip the external region. Te intersection density at the home ends contributes signifcantly to active commuting time, end in the internal region, on the other hand, contributes accounting for 5.24% (ranking 3rd) and 5.23% (ranking 4th), half as much as it does in the external region. Although respectively. Te efectiveness of active commuting is closely metro station density at the home end contributes the related to the connectivity of the street network, particularly least in both regions, its importance in the internal region the routes for cycling and walking. Tis is consistent with the is six times that of the external region. At the workplace, fndings of Cao [47]. Te intersections density at the work metro station density is three times as important in the end (4.41%) is shown to have an impact on the trip time of external region as it is in the internal region. When land active commuters. Land use mix, job density, and house use variables related to BE variables at home end, such as density at the work end have higher rankings than those at land use mix, job density, and house density are Journal of Advanced Transportation 9 10.75 5250 10.70 10.65 10.60 10.55 10.50 10.45 10.40 10.35 10.30 10.25 3000 0 5 10 15 20 25 30 35 40 45 50 Depth of the tree RMSE Iterations Figure 3: Result of RMSE in the internal region. 9.65 5250 9.60 9.55 9.50 9.45 9.40 9.35 2500 0 5 10 15 20 25 30 35 40 45 50 Depth of the tree RMSE Iterations Figure 4: Result of RMSE in the external region. compared, their roles range between 3.92%∼4.28% in the 5.3.2. Nonlinear Associations between BE and Active Com- internal region and 1.98%∼2.21% in the external region. muting Time. Partially dependent curves are used to present Notice that at the work end land use mix still holds a more the nonlinear associations between BE characteristics and active commuting time. In GBRT models, partial depen- important role in internal region than in external region (5.49% versus 3.47%). Tese varying efects from region to dence curves are commonly used to visualize the marginal region are closely related to land use characteristics. Te efects of independent variables on the dependent variable. diverse and well-developed land use pattern within the Figure 6 shows the relationships between BE at home internal region implies greater job options for active (columns 1 and 2) and work ends (columns 3 and 4) and commuters. As a result, these factors in the internal region active commuting time in internal (columns 1 and 3) and have great efects. TAZs in the external region, on the external (columns 2 and 4) regions. other hand, are generally less developed in large blocks Figures 6(a) and 6(b) show nonlinear associations be- with homogeneous land use. Distance to the CBD and bus tween street network-related characteristics and active stop density have similar roles in both regions. commute times. Te active commuting time for the internal RMSE RMSE Number of iterations Number of iterations 10 Journal of Advanced Transportation Table 4: Te relative importance of infuential factors in both regions. Internal region External region Variables Rank Relative importance (%) Sum (%) Rank Relative importance (%) Sum (%) Built environment at home end 28.97 21.65 Road density 3 5.24 16 2.55 Intersection density 16 2.68 2 5.76 Bus stop density 15 3.43 8 3.79 Metro station density 17 1.70 25 0.26 House density 11 3.92 18 2.21 Job density 9 4.28 19 2.02 Land use mix 10 4.12 20 1.98 Distance to CBD 13 3.60 12 3.08 Built environment at work end 34.32 33.27 Road density 4 5.23 10 3.45 Intersection density 7 4.41 6 4.47 Bus stop density 8 4.32 11 3.43 Metro station density 18 1.49 5 4.49 House density 6 4.69 4 4.69 Job density 5 4.85 3 5.05 Land use mix 2 5.49 9 3.47 Distance to CBD 12 3.84 7 4.22 Trip attributes 29.68 32.54 Departure time 14 3.50 14 2.72 Trip distance 1 26.18 1 29.82 Sociodemographics 7.05 12.54 Gender 19 1.47 17 2.26 Education 23 0.90 23 1.18 Age 20 1.40 15 2.72 License 22 1.08 22 1.25 Cars 21 1.09 21 1.85 Child 25 0.33 24 0.30 Income 24 0.78 13 2.98 Total 100 100 5.24% Road density 2.55% 2.68% Intersection density 5.76% 3.43% Bus stop density 3.79% 1.70% Metro station density 0.26% 3.92% House density 2.21% 4.28% Job density 2.02% 4.12% Land use mix 1.98% 3.60% Distance to CBD 3.08% 5.23% Road density 3.45% 4.41% Intersection density 4.47% 4.32% Bus stop density 3.43% 1.49% Metro station density 4.49% 4.69% House density 4.69% 4.85% Job density 5.05% 5.49% Land use mix 3.47% 3.84% Distance to CBD 4.22% 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 Relative Importance (%) Internal External Figure 5: Comparison of relative importance in internal and external regions. Work end Home end Journal of Advanced Transportation 11 20.0 20.1 External Internal External Internal 23.5 23.0 19.9 19.9 23.0 19.8 22.5 22.5 19.7 19.7 22.0 22.0 19.5 19.6 510 15 20 25 510 15 0 5 10 15 20 25 30 0 5 10 15 20 25 2 2 2 2 Road density (km/km ) Road density (km/km ) Road density (km/km ) Road density (km/km ) (a) 21.5 22.6 Internal External Internal External 22.6 21.0 22.4 22.4 20.5 22.2 22.2 20.0 22.0 19.5 21.8 22.0 0 20406080 01234 0 20406080 0 20406080 2 2 2 2 Intersection density (count/km ) Intersection density (count/km ) Intersection density (count/km ) Intersection density (count/km ) (b) 21.0 23.0 20.2 Internal External Internal External 22.7 20.5 22.6 20.0 22.5 22.5 19.8 20.0 22.4 22.0 19.6 22.3 19.5 22.2 19.4 21.5 5 10152025 0 5 10 15 0 5 10 15 20 25 0 5 10 15 20 25 2 2 2 2 Bus stop density (count/km ) Bus stop density (count/km ) Bus stop density (count/km ) Bus stop density (count/km ) (c) 22.8 Internal External Internal External 22.4 23 19.730 22.6 22.2 22.4 19.720 21 22.0 22.2 21.8 19.710 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.2 0.4 0.6 0.8 1.0 1.2 2 2 2 Metro station density (count/km ) 2 Metro station density (count/km ) Metro station density (count/km ) Metro station density (count/km ) (d) 21.5 23.5 19.9 23.0 Internal External External Internal 21.0 19.8 23.0 22.5 20.5 19.7 22.5 20.0 22.0 19.6 19.5 22.0 19.5 21.5 0.0 0.2 0.4 0.6 0.8 0.0 0.1 0.2 0.3 0.4 0.5 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 House density House density House density House density (e) 23.0 20.0 Internal External Internal External 23.5 20.0 22.8 19.9 22.6 19.8 23.0 19.8 22.4 19.7 19.6 22.5 22.2 19.6 22.0 19.4 19.5 22.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 Job density Job density Job density Job density (f ) Figure 6: Continued. Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) 12 Journal of Advanced Transportation Internal External Internal External Internal 20.2 19.80 23.5 23.0 19.75 20.0 23.0 22.5 19.70 22.5 19.8 19.65 22.0 19.60 22.0 19.6 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.2 0.4 0.6 0.8 Land use mix Land use mix Land use mix Land use mix (g) 20.6 22.6 21.0 External Internal External Internal 23.5 20.4 22.4 20.5 20.2 23.0 20.0 22.2 20.0 22.5 19.8 22.0 19.6 22.0 19.5 19.4 21.8 0 5 10 15 20 25 30 35 510 15 20 25 30 35 05 10 15 20 25 30 35 0 10 20 30 Distance to CBD (km) Distance to CBD (km) Distance to CBD (km) Distance to CBD (km) BE at the home end BE at the work end (h) 2 2 Figure 6: Nonlinear impact of BE characteristics. (a) Road density (km/km ). (b) Intersection density (count/km ). (c) Bus stop density 2 2 (count/km ). (d) Metro station density (count/km ). (e) House density. (f ) Job density. (g) Land use mix. (h) Distance to CBD (km). region decreases rapidly as the home-end road density in- exhibits the contrast patterns. Te extension of home-end creases from 0 to 11 km/km and bottoms at 22 minutes. It metro stations increases commuting time in the internal then gradually increases. Active commuting in the external region while decreasing them in the external region. Te region decreases until the home-end road density reaches extension of work-end metro stations reduces active com- 7 km/km , and then it goes up sharply. Te curves of work- muting time in the internal region while increasing it in the end road density in both regions also have a nonlinear external region. feature. Within 10 km/km , the active commuting time in Figures 6(e) and 6(g) illustrate the efects of land use- the internal region decreases sharply in a nearly linear related variables, including house density, job density, and pattern. After that, it climbs to 22.8 min and then remains land use mix. In general, active commuting time increases stable. With increasing work-end road density in the ex- with increasing house density at the home end, although there are some fuctuations in the curve of the external ternal region, active commuting time fuctuates in an ap- proximate inverted U shape; when the road density at the region within the ratio of 0.1. Because the data points in work end reaches 5 km/km in the external region, the in- these intervals are too sparse to interpret, the fuctuations creasing trend in commuting time stops. As shown in can be ignored. In both regions, house density at the work columns 1 and 2 of Figure 6, intersection density at home end has a positive relationship with active commuting time. end is positively associated with active commuting time, Work-end house density, as presented by the ratio of res- when it is in the range of 0∼40 per km in internal region and idential land at the work end, is better kept within 0.5 for the 0∼3 per km in external region. Higher intersection density internal region and within 0.3 for the external region. Te is often associated with longer stopping times, which in turn general trend can be explained by the greater sense of safety increases active commute times. Both fndings reinforce the while walking/cycling in areas with higher house density, ambiguous impact of street network design on active mo- which leads people to be willing to walk/cycle longer to their work. Job density at the home end has a U-shaped rela- bility [49]. Improved street network connectivity may demonstrate that there are more alternative shortcuts to tionship with active commuting time in internal region. Te reach destinations, reducing travel time further, but it may threshold of 0.3 indicates the ideal ratio of work-related land. also increase commute time due to more intersections. However, in the external region, the association is generally Figures 6(c) and 6(d) compare the efects of transit- negative. Commuting time stops decreasing when the ratio related variables in both regions. Te density of bus stops at of work-related lands reaches 0.5. Active commuting time the home end has the opposite efects. Active commuting increases in both regions as job density increases at the time increases when home-end bus stops increase in the workplace. external region but decreases when they increase in the For land use mix, its association with active commuting internal region. Both regions see a similar trend in the time varies by region. For active commuters in the internal impact of work-end bus stop density, namely, a reduction in region, the best home-end land use mix is around 0.5. However, in the external region, a land use mix of 0.5 or active commuting time with an increase in bus stop density. A minor diference between the two curves is in the range of higher is preferable for active commuters. Te work-end 10∼18 stops/km in the internal region where a fuctuation land use mix, in contrast to the home end, has a winding, exists. Te impact of metro station density at both ends decreasing association with active commuting time in the Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Journal of Advanced Transportation 13 internal region. Te active commuting time reaches its should be included as part of conditional variables. Second, lowest point (of 22 minutes) when the land use mix is 0.7. In the temporal impact of BE on active commuting time de- serves attention if new data sources become available. Fi- external region, the work-end land use mix has an inverted U-shaped relationship with active commuting time. It nally, while the study’s fndings are applicable to Nanjing suggests that increasing the land use mix in the external and may provide basic references for cities in similar con- region at the work end is only recommended to a certain texts to Nanjing, the BE impact varies from city to city [50]. extent in order to promote active commuting. More case studies are recommended in the future. Figure 6(h) shows the relationship between distance to the CBD and active commute time. Te residence distance to Data Availability the CBD has a U-shaped relationship with active commute time in both regions. For active commuters in the internal Te household travel survey data used to support the fndings of this study are not available because of data region, with the distance of their residence to the CBD around 5 km, active commuting time reaches its lowest value privacy and protection. (of 21.5 minutes); however, for commuters in the external region, when the distance is in the range of 15∼20 km, the Conflicts of Interest time keeps short. At the workplace, its average distance to Te authors declare that they have no conficts of interest. the CBD is 8.8 km for the internal region and 16.9 km for the external region. We focus more on curve intervals of less than 10 km for internal region and over 15 km for external Authors’ Contributions region. Active commuting time increases with the increasing Jingxian Wu and Huapeng Shen were in charge of con- distance of the workplace from the CBD in both intervals. ceptualization, methodology, formal analysis, and initial Te rise in active commuting time as a consequence of draft preparation. Jingxian Wu, Guikong Tang, and Soora increasing distance from home to CBD is monotonic. Rasouli were responsible for review and editing. Jingxian Wu was in charge of funding acquisition, investigation, and 6. Conclusions and Discussion supervision. Guikong Tang was in charge of data visuali- zation. Soora Rasouli was responsible for language checks With the evidence from Nanjing, China, this paper inves- and supervision. tigates the spatial heterogeneity in the BE impact on active commuting time. It uses the two-step clustering method to cluster 766 TAZs according to their BE features. Gradient- Acknowledgments boosted regression trees are then constructed for each Tis research work was supported by the Scientifc Research distinguished cluster to examine the heterogeneity in the Fund of the Institute of Engineering Mechanics, China importance of BE for active commuting time. It has been Earthquake Administration (grant no. 2020D23), National concluded that built-environment characteristics have more Natural Science Foundation of China (grant no. 52122215), importance than sociodemographics because they contrib- and the Soft Science Key Project of Shanghai (grant no. ute 63.29% to active commuting time in the internal region 21692105100). and 54.92% in the external region. Tis confrms the con- clusions of Cheng et al. that despite of the minor impact of single BE factor, their total impacts were larger than that of References sociodemographics [30]. Te spatial heterogeneity of BE’s [1] S. Cook, L. Stevenson, R. Aldred, M. Kendall, and T. Cohen, role to active commuting time is further proved by com- “More than walking and cycling: what is ‘active travel’?” paring the nonlinear impact of BE on active commuting time Transport Policy, vol. 126, pp. 151–161, 2022. in two regions. Such heterogeneity is so pronounced that in [2] R. Buehler, T. Kuhnimhof, A. Bauman, and C. Eisenmann, many cases the impact of BE on active commuting time is “Active travel as stable source of physical activity for one third opposite in two regions. of German adults: evidence from longitudinal data,” Trans- Te implications of this study for practice are two-fold. portation Research Part A: Policy and Practice, vol. 123, First, if the goal of urban planners and practitioners is to pp. 105–118, 2019. encourage the use of active modes, they need to prioritize [3] R. Buehler, J. Pucher, and A. Bauman, “Physical activity from diferent policies in each region depending on their im- walking and cycling for daily travel in the United States, 2001–2017: demographic, socioeconomic, and geographic portance in reaching the ultimate goal. Tis is especially variation,” Journal of Transport & Health, vol. 16, Article ID important when there is a limited budget that can be spared 100811, 2020. for improving one or a few built-environment character- [4] A. Banerjee, M. Łukawska, A. F. Jensen, and S. Haustein, istics. Second, it is very important for policymakers to be “Facilitating bicycle commuting beyond short distances: in- aware of the adverse efects of certain policies in some re- sights from existing literature,” Transport Reviews, vol. 42, gions if those policies are developed on the basis of average no. 4, pp. 526–550, 2022. efects. [5] R. Goel, A. Goodman, R. Aldred et al., “Cycling behaviour in Te current work comes with some limitations, which 17 countries across 6 continents: levels of cycling, who cycles, leave space for future improvements. First, if the data allows, for what purpose, and how far?” Transport Reviews, vol. 42, walkability measures such as sidewalk presence and width no. 1, pp. 58–81, 2022. 14 Journal of Advanced Transportation [6] A. Paez ´ and K. Whalen, “Enjoyment of commute: a com- factors,” American Journal of Preventive Medicine, vol. 42, parison of diferent transportation modes,” Transportation no. 6, pp. 610–615, 2012. [22] B. Sun, A. Ermagun, and B. Dan, “Built environmental im- Research Part A: Policy and Practice, vol. 44, no. 7, pp. 537– pacts on commuting mode choice and distance: evidence 549, 2010. from Shanghai,” Transportation Research Part D: Transport [7] H. M. A. Aziz, N. N. Nagle, A. M. Morton, M. R. Hilliard, and Environment, vol. 52, pp. 441–453, 2017. D. A. White, and R. N. Stewart, “Exploring the impact of [23] K. M. Leung, P. K. Chung, D. Wang, and J. D. Liu, “Impact of walk–bike infrastructure, safety perception, and built-envi- physical and social environments on the walking behaviour of ronment on active transportation mode choice: a random Hong Kong’s older adults,” Journal of Transport & Health, parameter model using New York City commuter data,” vol. 9, pp. 299–308, 2018. Transportation, vol. 45, no. 5, pp. 1207–1229, 2018. [24] M. Moniruzzaman and A. Paez, “An investigation of the [8] J. Feng, “Te built environment and active travel: evidence attributes of walkable environments from the perspective of from Nanjing, China,” International Journal of Environmental seniors in Montreal,” Journal of Transport Geography, vol. 51, Research and Public Health, vol. 13, no. 3, p. 301, 2016. pp. 85–96, 2016. [9] R. Cervero, O. L. Sarmiento, E. Jacoby, L. F. Gomez, and [25] C. Yin, J. Zhang, and C. Shao, “Relationships of the multi- A. Neiman, “Infuences of built environments on walking and scale built environment with active commuting, body mass cycling: lessons from Bogota,” ´ International journal of sus- index, and life satisfaction in China: a GSEM-based analysis,” tainable transportation, vol. 3, no. 4, pp. 203–226, 2009. Travel Behavior and Society, vol. 21, pp. 69–78, 2020. [10] J. Wu, B. Wang, N. Ta, K. Zhou, and Y. Chai, “Does street [26] C. Chen, H. Gong, and R. Paaswell, “Role of the built envi- greenery always promote active travel? Evidence from Bei- ronment on mode choice decisions: additional evidence on jing,” Urban Forestry and Urban Greening, vol. 56, Article ID the impact of density,” Transportation, vol. 35, no. 3, 126886, 2020. pp. 285–299, 2008. [11] Z. Zhu, Z. Li, Y. Liu, H. Chen, and J. Zeng, “Te impact of [27] M. Moniruzzaman, A. Paez, K. M. Nurul Habib, and urban characteristics and residents’ income on commuting in C. Morency, “Mode use and trip length of seniors in Mon- China,” Transportation Research Part D: Transport and En- treal,” Journal of Transport Geography, vol. 30, pp. 89–99, vironment, vol. 57, pp. 474–483, 2017. [12] S. L. Handy, M. G. Boarnet, R. Ewing, and R. E. Killingsworth, [28] R. Raman and U. K. Roy, “Taxonomy of urban mixed land use “How the built environment afects physical activity: views planning,” Land Use Policy, vol. 88, Article ID 104102, 2019. from urban planning,” American Journal of Preventive [29] C. Ding, X. Cao, and Y. Wang, “Synergistic efects of the built Medicine, vol. 23, no. 2, pp. 64–73, 2002. environment and commuting programs on commute mode [13] G. Eldeeb, M. Mohamed, and A. Paez, ´ “Built for active travel? choice,” Transportation Research Part A: Policy and Practice, Investigating the contextual efects of the built environment vol. 118, pp. 104–118, 2018. on transportation mode choice,” Journal of Transport Geog- [30] L. Cheng, J. D. Vos, P. Zhao, M. Yang, and F. Witlox, “Ex- raphy, vol. 96, Article ID 103158, 2021. amining non-linear built environment efects on elderly’s [14] Y. Wang, C. K. Chau, W. Y. Ng, and T. M. Leung, “A review walking: a random forest approach,” Transportation Research on the efects of physical built environment attributes on Part D: Transport and Environment, vol. 88, Article ID enhancing walking and cycling activity levels within resi- 102552, 2020. dential neighborhoods,” Cities, vol. 50, pp. 1–15, 2016. [31] J. Liu, B. Wang, and L. Xiao, “Non-linear associations between [15] R. Cervero and K. Kockelman, “Travel demand and the 3Ds: built environment and active travel for working and shopping: density, diversity, and design,” Transportation Research Part an extreme gradient boosting approach,” Journal of Transport D: Transport and Environment, vol. 2, no. 3, pp. 199–219, Geography, vol. 92, Article ID 103034, 2021. [32] H. Yang, Z. Zhao, C. Jiang, Y. Wen, and M. Muneeb Abid, [16] R. Ewing and R. Cervero, “Travel and the built environment,” “Spatially varying relation between built environment and Journal of the American Planning Association, vol. 76, no. 3, station-level subway passenger-distance,” Journal of Ad- pp. 265–294, 2010. vanced Transportation, vol. 2022, Article ID 7542560, 18 [17] L. Cheng, X. Chen, S. Yang, Z. Cao, J. De Vos, and F. Witlox, pages, 2022. “Active travel for active ageing in China: the role of built [33] L. Yang, J. Liu, Y. Liang, Y. Lu, and H. Yang, “Spatially varying environment,” Journal of Transport Geography, vol. 76, efects of street greenery on walking time of older adults,” pp. 142–152, 2019. ISPRS International Journal of Geo-Information, vol. 10, no. 9, [18] L. Yang, S. Grifn, K. T. Khaw, N. Wareham, and J. Panter, p. 596, 2021. “Longitudinal associations between built environment char- [34] S. Srinivasan and J. Ferreira, “Travel behavior at the household acteristics and changes in active commuting,” BMC Public level: understanding linkages with residential choice,” Health, vol. 17, no. 1, Article ID 839, 458 pages, 2017. Transportation Research Part D: Transport and Environment, [19] T. Tao, X. Wu, J. Cao, Y. Fan, K. Das, and A. Ramaswami, vol. 7, no. 3, pp. 225–242, 2002. “Exploring the nonlinear relationship between the built en- [35] F. M. Dieleman, M. Dijst, and G. Burghouwt, “Urban form vironment and active travel in the Twin Cities,” Journal of and travel behaviour: micro-level household attributes and Planning Education and Research, Article ID residential context,” Urban Studies, vol. 39, no. 3, pp. 507–527, 0739456X2091576, 2020. [20] X. Ma, J. Yang, C. Ding, J. Liu, and Q. Zhu, “Joint analysis of [36] W. Tu, R. Cao, Y. Yue, B. Zhou, Q. Li, and Q. Li, “Spatial the commuting departure time and travel mode choice: role of variations in urban public ridership derived from GPS tra- the built environment,” Journal of Advanced Transportation, jectories and smart card data,” Journal of Transport Geogra- vol. 2018, Article ID 4540832, 13 pages, 2018. phy, vol. 69, pp. 45–57, 2018. [21] M. A. Beenackers, S. Foster, C. B. Kamphuis et al., “Taking up [37] W. Zhang, Y. Zhao, X. Cao, D. Lu, and Y. Chai, “Nonlinear cycling after residential relocation: built environment efect of accessibility on car ownership in Beijing: pedestrian- Journal of Advanced Transportation 15 scale neighborhood planning,” Transportation Research Part D: Transport and Environment, vol. 86, Article ID 102445, [38] S. Zhong, Z. Wang, Q. Wang, A. Liu, and J. Cui, “Exploring the spatially heterogeneous efects of urban built environment on road travel time variability,” Journal of Transportation Engineering Part A Systems, vol. 147, no. 1, Article ID 04020142, 2021. [39] C. Ding, X. Cao, B. Yu, and Y. Ju, “Non-linear associations between zonal built environment attributes and transit commuting mode choice accounting for spatial heterogene- ity,” Transportation Research Part A: Policy and Practice, vol. 148, pp. 22–35, 2021. [40] M. A. Abas, S. Rajoo, and S. F. Zainal Abidin, “Development of Malaysian urban drive cycle using vehicle and engine parameters,” Transportation Research Part D: Transport and Environment, vol. 63, pp. 388–403, 2018. [41] J. Große, A. S. Olafsson, T. A. Carstensen, and C. Fertner, “Exploring the role of daily “modality styles” and urban structure in holidays and longer weekend trips: travel be- haviour of urban and peri-urban residents in Greater Copenhagen,” Journal of Transport Geography, vol. 69, pp. 138–149, 2018. [42] M. T. Ballestar, P. Grau-Carles, and J. Sainz, “Customer segmentation in e-commerce: applications to the cashback business model,” Journal of Business Research, vol. 88, pp. 407–414, 2018. [43] A. Tkaczynski, S. Rundle-Tiele, and N. Beaumont, “Desti- nation segmentation: a recommended two-step approach,” Journal of Travel Research, vol. 49, no. 2, pp. 139–152, 2010. [44] T. Chiu, D. Fang, J. Chen, Y. Wang, and C. Jeris, “A robust and scalable clustering algorithm for mixed type attributes in large database environment,” in Proceedings of the 7th ACM SIGKDDD International Conference on Knowledge Discovery and Data Mining, pp. 263–268, San Francisco, August 2001. [45] J. H. Friedman, “Greedy function approximation: a gradient boosting machine,” Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001. [46] G. Ridgeway, Generalized Boosted Models: A Guide to the GBM Package, 2005. [47] X. J. Cao, “Examining the impacts of neighborhood design and residential self-selection on active travel: a methodo- logical assessment,” Urban Geography, vol. 36, no. 2, pp. 236–255, 2015. [48] Y. Fan, K. V. Das, and Q. Chen, “Neighborhood green, social support, physical activity, and stress: assessing the cumulative impact,” Health & Place, vol. 17, no. 6, pp. 1202–1211, 2011. [49] J. Wu, B. Wang, R. Wang, T. Na, and Y. Chai, “Active travel and the built environment: a theoretical model and multi- dimensional evidence,” Transportation Research Part D: Transport and Environment, vol. 100, Article ID 103029, 2021. [50] P. Naess, “Meta-analyses of built environment efects on travel: no new platinum standard,” Journal of Planning Ed- ucation and Research, vol. 42, no. 2, pp. 199–205, 2019. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Advanced Transportation Hindawi Publishing Corporation

Spatial Heterogeneity in the Nonlinear Impact of Built Environment on Commuting Time of Active Users: A Gradient Boosting Regression Tree Approach

Download PDF
15 pages

Loading next page...
 
/lp/hindawi-publishing-corporation/spatial-heterogeneity-in-the-nonlinear-impact-of-built-environment-on-Va9XznhYU0

References (49)

Publisher
Hindawi Publishing Corporation
ISSN
0197-6729
eISSN
2042-3195
DOI
10.1155/2023/6217672
Publisher site
See Article on Publisher Site

Abstract

Hindawi Journal of Advanced Transportation Volume 2023, Article ID 6217672, 15 pages https://doi.org/10.1155/2023/6217672 Research Article Spatial Heterogeneity in the Nonlinear Impact of Built Environment on Commuting Time of Active Users: A Gradient Boosting Regression Tree Approach 1,2 2 3 4 Jingxian Wu , Guikong Tang , Huapeng Shen , and Soora Rasouli Key Laboratory of Earthquake Engineering and Engineering Vibration, Institute of Engineering Mechanics, China Earthquake Administration, Beijing, China Department of Transportation Engineering, University of Shanghai for Science and Technology, Jungong Road #516, Shanghai 200093, China Beijing Key Laboratory of Trafc Engineering, Beijing University of Technology, No. 100 Ping Le Yuan, Chaoyang District, Beijing 100124, China Urban Planning and Transportation Group, Eindhoven University of Technology, P.O. Box 513, Eindhoven 5600 MB, Netherlands Correspondence should be addressed to Huapeng Shen; shenhuapeng1997@163.com Received 10 January 2022; Revised 24 January 2023; Accepted 2 February 2023; Published 13 February 2023 Academic Editor: Hongtai Yang Copyright © 2023 Jingxian Wu et al. Tis is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Many studies provided evidence regarding the infuence of built environment (BE) on commuting time. However, few studies have considered the spatial heterogeneity of such impacts. Using data from Nanjing, China, this study employs two-step clustering and gradient boosted regression trees (GBRT) to segment the neighborhoods into diferent types and investigate the efects of BE characteristics on the commuting time of active users. Te results show a strong efect of BE characteristics on commuting time, involving active modes. Te importance of BE characteristics varies among neighborhood types. For active commuters in the internal region of Nanjing, commuting time is afected mostly by the land use mix at the work end. Te lowest impact of BE in internal regions is associated with metro station density. For active commuters in external region of the city, the relative importance of intersection density at the home end is the largest (as high as 5.76%). Moreover, other signifcant diferences are found in the associations between BE characteristics and active commuting time in the two regions. commute distance [4, 5]. Longer trip distances imply that 1. Introduction active commuters will spend more time on the road and thus Active travel mode, referring to walking or cycling, is a viable have less life satisfaction. Despite the fact that active mode alternative to driving in short-to-medium distance trips [1]. commuting generates positive utilities such as a health- Regular active travel such as active commuting to work is enhancing efect, commuters prefer to shorten it due to time thought to beneft both the environment and an individual’s budgets [6]. Active transportation, particularly cycling, can physical health [2]. To encourage active travel, several in- be used for longer distance trips if the time cost of com- terventions such as bike-sharing programs and provisions of muting by active modes is appropriately reduced [4]. As a footpaths and cycle lanes have been implemented. However, result, it is critical to investigate factors that infuence the such promotion does not bring a signifcant increase in the commuting time of active users. share of active commuting, and private cars are still the most Many studies on travel behavior have found that the built widely used mode [3]. environment and sociodemographics are strongly associated Te low prevalence of active commuting can be at- with walking and cycling (e.g., see [7–14]). Individuals’ tributed in part to urbanization’s increasing average active travel choices are infuenced by built-environment 2 Journal of Advanced Transportation characteristics such as walk-bike infrastructure [7], street found that mixing multiple land uses beyond a certain pattern [8], route connectivity [9], street greenery [10], and proportion can have an adverse efect [28]. For commuting, the mix of land use types that specify jobs and houses plays a population density [11]. Changes in the built environment, according to Handy et al. [12], infuence travel mode choice major role. Te appropriate job-to-housing ratio can be an primarily by altering travel time. Furthermore, Eldeeb et al. indication of whether residents are likely to be employed in discovered that improving the built environment does not the neighboring area of their residence. Tis, in turn, in- have a homogeneous impact on the likelihood of using active fuences commute distance and facilitates using active travel mode in diferent parts of the city [13]. However, rare studies to go to work. have focused on the commuting time of active mode and All the above-mentioned studies have confrmed the examined whether the built environment has a spatially contribution of built environment to active travel primarily diferent impact on it. by using regression-based models such as multilevel re- Terefore, this study attempts to advance the literature gression [11], a logit-based model [8, 18, 24], and a structural by investigating the nonlinear associations between the built equation model [23, 25]. However, these models are pri- marily based on a priori, often linear relation between the environment and travel time of active commuters, ac- counting for spatial heterogeneity. First, the two-step built environment and active mobility. A parallel stream of clustering method is used to diferentiate the regions with studies examines the association between built environment various built-environment features. Next, gradient-boosted and travel behavior using machine learning methods. Ding regression trees are generated for predicting the commuting et al. applied decision trees [29] to extract the nonlinear time of active travelers in each region. Travelers’ socio- relationship between a built environment and commute demographics, trip characteristics, and built-environment mode choice. Tao et al. assessed the importance of BE characteristics are considered as conditional variables. components for the energy consumption of active users by Te remainder of this article is structured as follows. applying gradient boosting decision trees [19]. Tey found Section 2 presents a literature review on the association the distance to the nearest park posed the greatest impact. between built environment and active travel. Section 3 Cheng et al. used a random forest to assess BE’s impact on elderly active travel and found that population density had describes the data source used in this study and built-en- vironment characteristics at the trafc analysis zone (TAZ) the highest contribution [30]. Liu et al. in their most recent level, as well as commuters’ personal profles. Section 4 study used the extreme gradient boosting approach to ex- briefy describes the methodology used in this study, while amine the association of built environment and active travel Section 5 discusses the fndings. Finally, we present the choice. Tey found that trip characteristics contributed more summary of the key fndings and discuss their implications than the built environment [31]. for planning practice. 2.2. Spatial Heterogeneity. Spatial heterogeneity refers to the 2. Literature Review varying impact of the same infuential factors at diferent spatial scales or geographical locations. Several studies have 2.1. Association between the Built Environment and Active investigated whether the relationship between the built Travel. Te built environment is the physical setting environment and travel behavior varies across diferent designed to meet people’s need to engage in activities. Te types of neighborhoods [32, 33]. Srinivasan and Ferreira built-environment characteristics that relate to residents’ found the built environment around residential areas to pose travel behavior are defned from “3Ds” into “5Ds,” which are diferent efects on travel mode choice compared to that “density,” “diversity,” block “design,” “destination,” “ac- around workplaces [34]. For instance, land use mix around cessibility,” and “distance” to transit [15, 16]. Some scholars specifcally focused on the impact of the “5Ds,” as refection households’ residence has stronger efects on household travel mode choice and travel distances than that at job of built-environment characteristics, on active travel (e.g., see [7–14, 17–19]). locations [35]. Using the geographically weighted regression (GWR) model, Tu et al. found signifcantly diferent impacts Density, as a key component of built environment, has a of built environment on travel mode choice [36]. Zhang et al. paradoxical efect on active travel choice [20]. Density used a hierarchical linear model to explore the relationship appeared to have a signifcant impact on choosing active between the neighborhood-built environment and trip travel modes in some studies, such as reference [21], but not distance [37]. Tey found signifcant spatial heterogeneity in in other studies (e.g., see [22]). Zhu et al. discovered that the infuence of the built environment on travel behavior. increasing population density increases residents’ possibility of commuting in active mode [11]. Block design such as Neighborhoods located in diferent areas of the city sometimes share similar characteristics, but their efects can street crossing density, road density, and connected side- walks have profound efects on active travel [8, 23, 24]. be diferent. Zhong et al. used a geographically weighted regression model to analyze the spatial heterogeneity of the Accessibility indicators such as employment accessibility efects of an urban built environment on road travel time [18], destination accessibility [25], and transit accessibility and found that spatially varying relationships exist [38]. [26] are found to be positively correlated with residents’ use Ding et al. used spline in a mixed logit model and concluded of active travel. Diversity that measures land use mix in a that nonlinearity exists in the relation between built envi- neighborhood/region is closely associated with the choice of ronment and commuting mode choice [39]. active travel [13, 23, 27]. Furthermore, Raman and Roy Journal of Advanced Transportation 3 In summary, existing studies have empirically done a lot bikes due to the infrastructure limitation. Tus, this study on the measurement of BE on active travel behavior or included it as one of the active travel modes, from home to work in the morning were chosen for this study. spatial heterogeneity of BE on travel behavior. However, none have investigated how the built environment infu- Te built-environment characteristics are measured at ences the commuting time of active modes and compre- the TAZ level using the software of ArcGIS. Data sources, hensively considered the potential impact of spatial including Baidu map POIs, open street maps, and urban heterogeneity. Tus, this study contributes to the existing land use GIS data, are used. Many studies have focused on body of the literature by exploring nonlinear relation the built-environment characteristics surrounding resi- (without any a priori assumptions) between built environ- dences, but others, such as Sun et al. [22] and Ding et al. [29] ment and active commuting time taking into account spatial have emphasized the importance of trip destination char- heterogeneity. It is important to note that the duration of acteristics. Tus, this study measures the TAZ characteristics for both the home and workplace ends. Among all these data active mobility for commuting purpose has never been the subject of examination within the topic of nonlinear rela- sources, POIs provide geographic information about specifc points and are used to calculate transit-related indicators tionship between built environment and active mobility. such as intersection density, bus stop density, and metro station density within the TAZ. Te open street map is used 3. Data to calculate the density of roads in each TAZ. Te urban land 3.1. Study Area and Data Sources. Te data used in this study use GIS data are used to calculate three indicators of land originated from Nanjing, China. Nanjing is a mega-city and use: land use mix and the ratios of residences and working the provincial political and economic center of Jiangsu. It is places to the area of TAZ. Te land use mix is determined by located in the eastern region of China, downstream of the an entropy index of nine land use types around commuters’ Yangtze River. It is an important gateway city for the central residences and work places. Te nine diferent types of land uses include residence, industrial use, public administration, and western regions’ development, which is fueled by ra- diation from the Yangtze River Delta. Nanjing is divided into commercial services, green space and plazas, construction, transportation, public facilities, and warehousing. Te in- 11 administrative regions, covering a total area of 6,587 km and a built-up area of 868 km . Te resident population was dicator is calculated as follows: 9.42 million in 2021, with an urban population of 8.19 −1 LandMix � 􏽘 p ln p , (1) million and an urbanization rate of 86.9%. In 2021, the city’s i i ln n gross regional product reached 163,532 billion yuan. Tis study concentrates on the most urbanized regions, including where p is proportion of the type i land use, and n is the Gulou, Qinhuai, Xuanwu, Jianye, Yuhuatai, Qixia, Jiangn- number of land use types. ing, Pukou, and Luhe (two regions, Gaochun and Lishui, were newly designated as regions during the urbanization 3.2. Data Description and were not involved in this survey). Seven of them are on the southern side of the Yangtze River, while two are on the 3.2.1. Built Environment at the TAZ Level. Table 1 shows the northern side. For transportation census and management, defnitions and descriptive statistics of eight built-envi- the Nanjing transportation planning agency divides the ronment variables obtained at the TAZ level. All 766 TAZs entire study area of Nanjing into 766 trafc analysis zones have the BE characteristics in fve dimensions: density, (TAZ) based on land use and administration boundaries. design, distance to transit, destination accessibility, and Figure 1 depicts the study area of Nanjing, China. diversity. Housing density corresponds to the density of In this study, four data sources are used: Nanjing residential development. Road density and intersection Household Travel Survey data from 2016; Nanjing urban GIS density that describe the characteristics of a street network data; points of interest (POIs) from Baidu map; and an open represent street design. Bus stop density and metro station street map (OSM). Te Nanjing Household Travel Survey is density measure the accessibility of bus and subway services an annual survey conducted by the Nanjing transportation and are related to distance to transit. Job density indicates planning agency. It is carried out through household in- destination accessibility; land use mix represents diversity; terviews in order to learn about the daily mobility patterns of and distance to the CBD refects regional location. urban residents. In 2016, the survey employed the stratifed random sampling technique to guarantee that the sample size was proportional to the population size. In total, 8,387 3.2.2. Statistics of Active Commuters. We matched BE fea- people from 3,015 households were invited to participate in tures for each active commute trip based on both home-end the survey. Te survey collected individual’s sociodemo- and work-end TAZs in the trip records. Table 2 depicts the graphic information (e.g., household income, car ownership, sociodemographics, trip characteristics, and BE character- gender, and age) and their travel diaries (e.g., trip origin, istics of 1,937 commuters by active modes. Males account for destination, purpose, departure time, and travel mode) on a 43.9% of the sample, which is slightly lower than females. given day. Based on the provided trip purpose and travel 56.4% have a bachelor degree or higher the majority are aged mode information, 1,937 commuters used active modes, between 30 and 49 and 52.7% own a driving license. Te such as walking, bicycling, and e-cycling. In China, riding an average household has 0.63 cars, while the average number e-bike does not imply much higher speed than ordinary of children aged six years old or below is 0.12. 53.5% of the 4 Journal of Advanced Transportation Legend Administrative regions Gulou Jiangning Jianye Luhe Pukou Qinhuai Qixia Xuanwu Yuhuatai TAZ 40°N CBD 30°N 20°N 120°E 100°E 04 8 16 24 32 km Figure 1: Case study area in Nanjing. Table 1: Built-environment characteristics for 766 TAZs. Names Variable descriptions Means (S.D.) Road density Total road length/TAZ area (km/km ) 6.85 (4.57) Intersection density Intersections/TAZ area (count/km ) 4.39 (11.55) Bus stop density Bus stops/TAZ area (count/km ) 4.85 (4.95) Metro station density Metro stations/TAZ area (count/km ) 0.23 (0.61) House density Residential area/TAZ area 0.19 (0.20) Job density Industrial, public administration, and commercial area/TAZ area 0.24 (0.21) Land use mix An entropy index of nine types of land use 0.41 (0.19) Distance to CBD Euclidean distance from TAZ centroids to CBD (km) 16.40 (9.66) CBD refers to the center of the TAZ in which Xinjiekou business district is located. respondents have an annual household income over 100,000 4.1. Two-Step Clustering Method. Te two-step clustering CNY. Te majority (81.5%) leaves for work between 7:00 am method has been extensively used in the transportation and 8:30 am. Te sample’s average commute time is feld due to its fexibility and capability in data processing 21.52 minutes, and the average trip distance to work is [40, 41]. It has an advantage over other clustering 4.98 km. techniques in that it can handle both continuous and discrete variables simultaneously. In addition, it can determine the optimal number of clusters automatically 4. Methodology and its clustering accuracy is unafected by the size of To examine the impact of built-environment characteristics data [42, 43]. on active commuting time, this study frst divides Nanjing’s Te clustering consists of two procedures. First, it 766 zones into diferent types using a two-step clustering clusters the 766 zones into groups according to their sim- method. Ten, in each region, gradient-boosted regression ilarity in BE characteristics. Ten, the merging algorithm is trees are constructed to investigate the determinants and used to gradually combine these groups until only one group relative importance of the infuential factors on active is left. Te optimal clustering number is determined using commuting time. Te following sections elaborate the Bayesian information criterion (BIC). Interested readers can specifcs of the analysis. refer to Chiu et al. [44]. Journal of Advanced Transportation 5 Table 2: Sample description of active commuters (N � 1937). Names Variable descriptions Means (S.D./percent) Sociodemographics Gender Gender: 1 � male; 0 � female 0 � 56.1%, 1 � 43.9% Education Hold a bachelor degree or above: 1 � yes; 0 � no 0 � 43.6%, 1 � 56.4% Respondent’s age: 1 � 20–29 years old; 2 � 30–39 years old; 3 � 40–49 years old; 1 � 17.6%, 2 � 27.1%, 3 � 36.1%, Age 4 � 50 or more years old 4 � 19.2% License Hold a driving license: 1 � yes; 0 � no 0 � 47.3%, 1 � 52.7% Cars Number of cars owned by a household (count) 0.63 (0.58) Child Number of children at 6 years old or younger (count) 0.12 (0.34) Income Household income per year: 1 � over 100,000 CNY; 0 � other 0 � 46.7%, 1 � 53.3% Trip attributes Te commute trip occurs in morning peak hours from 7:00 am to 8:30 am: 1 � yes; Departure time 0 � 18.5%, 1 � 81.5% 0 � no Euclidean distance from residential TAZ centroid to workplace TAZ centroid Trip distance 4.98 (6.38) (km) Commuting time Commuting time spent on road (min) 21.52 (11.98) Built environment at home end Road density Total road length per TAZ area (km/km ) 9.78 (3.54) Intersection Intersections/TAZ area (count/km ) 10.12 (16.99) density Bus stop density Bus stops/TAZ area (count/km ) 8.32 (5.39) Metro station Metro stations/TAZ area (count/km ) 0.53 (0.92) density House density Residential area/TAZ area 0.31 (0.19) Job density Industrial, public administration, and commercial area/TAZ area 0.30 (0.21) Land use mix An entropy index of nine types of land use 0.52 (0.16) Distance to CBD Euclidean distance from TAZ centroids to CBD (km) 9.39 (8.38) Built environment at work end Road density Total road length per TAZ area (km/km ) 9.77 (3.91) Intersection Intersections/TAZ area (count/km ) 11.48 (19.05) density Bus stop density Bus stops/TAZ area (count/km ) 8.48 (5.47) Metro station Metro stations/TAZ area (count/km ) 0.46 (0.86) density House density Residential area/TAZ area 0.31 (0.19) Job density Industrial, public administration, and commercial area/TAZ area 0.30 (0.20) Land use mix An entropy index of nine types of land use 0.52 (0.16) Distance to CBD Euclidean distance from TAZ centroids to CBD (km) 9.35 (8.59) CBD refers to the centroid of the TAZ in which Xinjiekou business district is located. 4.2. Gradient-Boosting Regression Trees. Gradient-boosted errors of the previous ones. Given the training data regression trees (GBRT) are an ensemble model that (y , x ) , the specifc learning steps are as follows: 􏼈 􏼉 i i 1 combines gradient-boosting and regression trees [45, 46]. It (1) Initialize the base model F (x) to be a constant: has myriad merits over the traditional linear regression methods and has been often used in transportation research F (x) � argmin 􏽘 L y , c􏼁 , (2) [29, 37, 45]. First, GBRT is more efective at data prediction 0 c i�1 and interpretation than general linear regressions or even just a single tree due to its tree-based ensemble feature. where y is the observed value, c is the predicted Second, it accommodates data with missing values and value, and N is the number of observation. Squared avoids multicollinearity of explainable variables. Tird, it error is chosen as the loss function for the regression. can calculate the relative importance of each variable (2) For m � 1 to M (M is the times of iterations or without making assumptions about the variables’ relation- optimal number of trees), compute the residual ships. Fourth, it is adaptable to both continuous and cate- which is mathematically calculated by the negative gorical types and is applicable to small data sets. derivation of loss function with respect to the pre- Furthermore, it avoids the overftting issue that frequently vious model outcome: arises as the number of tree nodes rises by using gradient boosting. zL y , F x􏼁􏼁 i i r � −􏼢 􏼣 , i � {1, . . . , N}, (3) Te GBDT model combines multiple regression trees mi zF x􏼁 F(x)�F (x) m−1 sequentially with each new tree adding up to correct the 6 Journal of Advanced Transportation where r is negative gradient, and F(x ) is the where J is the number of terminal nodes, J-1 is the number of mi i previous model. the nonterminal nodes, υ is the feature associated with the node t, τ is the improvement in squared error after the (3) Fit a regression tree to the residuals r and minimize mi splitting node t, and I[υ � κ] equals to 1 when υ � x , or 0 t t κ the loss function: otherwise. h (x) � 􏽘 c I􏼐x ∈ R 􏼑, m mj mj 5. Results j�1 5.1. Identifcation of the Neighborhood Types. As shown in ⎝ ⎠ ⎛ ⎞ c � argmin 􏽘 L y , F x􏼁 + 􏽘 cI􏼐x ∈ R 􏼑 , mj c i m−1 i mj Table 3, there are 766 TAZs in Nanjing with varying built- x ∈R j�1 i mj environment characteristics. Te two-step clustering (4) method is used to cluster these TAZs with more homoge- neous spatial features. To eliminate the infuence of col- where h (x) is the m th regression tree, J is the tree m linearity on clustering results, the Pearson correlation depth, referring to the number of terminal nodes, coefcient is used to test the association between pairs of BE R is the disjoint region partitioned by the terminal mj variables, and the variance infation factors (VIF) are cal- nodes of m th tree, c is the optimal coefcient for mj culated to measure the degree of collinearity. Except for bus R , and I(x ∈ R ) equals to 1 when (x ∈ R ), or 0 mj mj mj stop density, all BE characteristics have coefcients less than otherwise. 0.6 (0.7–1.0 indicates strongly correlated) and all VIFs (4) Update the model: calculated are less than 3. Both indicate that all BE variables are suitable for clustering. In a stepwise approach, the ratio change in BIC and ratio of distance measures for a variety of F (x) � F (x) + 􏽘 c I􏼐x ∈ R 􏼑. (5) m m−1 mj mj clusters are identifed. A model with two clusters appears to j�1 be optimal, with a silhouette coefcient value of 0.5. Table 3 shows the centroids for the two clustered groups To prevent overftting in the training procedure, hyper- as well as the signifcance of their diferences in each BE parameters including optimal number of trees M, learning characteristic. All 766 TAZs are divided into two groups: rate ], and tree depth J should be estimated by using test data Cluster-1 with 263 TAZs and Cluster-2 with 503 TAZs. Te or cross-validation. Te model is replaced by spatial heterogeneity of TAZs has been interpreted using the centroids for each group. TAZs in Cluster-1 are featured by a F (x) � F (x) + ] 􏽘 c 1􏼐x ∈ R 􏼑, (6) m m−1 higher road density (10.64 km/km ), more intersections mj mj j�1 (11.04 count/km ), more access to metro stations (0.61 count/km ), higher ratio of residential land (0.40), more job where ] is the learning rate that scales the contribution of opportunities (0.33), higher land use mix (0.58), and closer each tree. It has the value range from 0 to 1. Smaller values of proximity to CBD (8.11 km). TAZs in Cluster-2 have a lower learning rate give rise to larger M value and results in minor road density (4.86 km/km ), fewer intersections (0.92 count/ test error. 2 2 km ), less developed metro service (0.04 count/km ), less Te optimal values of these parameters are determined residential land use (0.08), lower job coverage (0.19), a lower by performing the 5-fold cross-validation. Root mean land use mix (0.33), and are located far away from the CBD squared error (RMSE) is chosen as the performance (20.74 km). Given the spatial diference between the cen- measurement. Parameters that result in the lowest cross- troids, we named cluster-1 as the internal region and cluster- validation error are preferred in the fnal model. For the 5 2 as the external region. Te Mann–Whitney U test method test test datasets in cross-validation, RSME is calculated as is used to compare the diferences in BE characteristics follows: between the two groups. Te result demonstrates their 􏽶�������������� spatial diferences. Te two types of TAZs in Nanjing are test shown in Figure 2. (7) RSME � 􏽘 􏽘 y − y 􏽢 􏼁 , t t 5 ′ f�1 f t�1 5.2. Results of GBRT. Using the GBM package in RStudio, where N is the data number in test set f. GBRT models for the commute time of active commuters Meanwhile, the learned regression trees 􏼈T 􏼉 provide living in each region are estimated. Te relative importance interpretative results that show the relative infuence of an of infuential factors is calculated for both identifed regions. explanatory variable x as follows [45]: Te relative importance is measured by comparing the error 2 1 2 reduction of one variable in commute time compared to 􏽢 􏽢 I � 􏽘 I T 􏼁 , κ κ m other variables. All variables included have a total impor- m�1 tance that adds up to 100%. Prior to modeling, hyper-pa- (8) rameters including learning rate, optimal number of J−1 2 ⌢ I T 􏼁 � 􏽘 τ I􏼂υ � κ􏼃, iterations (or the number of trees), and tree depth must be m t κ t t�1 tuned. Ridgeway recommended setting the learning rate for Journal of Advanced Transportation 7 Table 3: Centroids for the TAZ clustering results. Attributes Cluster-1 (263 TAZs) Cluster-2 (503 TAZs) Mann–Whitney U Sig. Road density 10.64 4.86 15990.00 <0.001 Intersection density 11.04 0.92 27695.00 <0.001 Metro station density 0.61 0.04 41884.00 <0.001 House density 0.40 0.08 9588.50 <0.001 Job density 0.33 0.19 34587.00 <0.001 Land use mix 0.58 0.33 16321.00 <0.001 Distance to CBD 8.11 20.74 13653.00 <0.001 Classifcation Administrative regions Internal External 01 5 0 20 30 km Figure 2: Spatial distributions of TAZs in clustered neighborhood types. practice between 0.01 and 0.001 [46]. Te smaller learning order to fnd the best GBRT, we initially developed the rate is thought to improve model performance. We set the model with the depth of the tree ranging from 1 to 49 in learning rate as 0.001 in accordance with Tao et al. [19]. In increments of 1. Te optimal parameters are then 8 Journal of Advanced Transportation the home end. Tis can be explained by the high aggregation determined using the RMSE value of fve-fold cross-vali- dation, which varies as tree depth increases. Figures 3 and 4 of morning commutes at the destination over the origin. Similarly, bus stop density at the work end is as high as visualize the RMSE values versus tree depth and the optimal number of iterations for the internal and external regions, 4.32%, greater than that at the home end. Tis confrms the respectively. Te RMSE in the internal region decreases with roles of transit accessibility on travel behavior [48], as well as increasing tree depth until it reaches 28. In the external the fact that the resultant trip time is more infuenced by the region, however, this indicator becomes stable at a depth of BE feature at the trip end. Metro station density is the least 19. As a result, 28 were set as the tree depth for the model in important BE factor, and its importance at both ends is less the internal region and 19 for the model in the external than 2.00%. Tis could be explained by the least variation in region. Te diference between internal and external regions metro services in the internal region. Geographical locations with respect to the number of iterations is even larger. of home and work ends that are presented by distance to CBD pose the contributions, 3.60% and 3.84%, respectively. According to the results, the commuting time model in the internal region iterated 3,140 times before convergence, In the external region, intersection density at the home end is the most infuential BE factor, with a relative while the model in the external region iterated 2,728 times. Both models ft well, with pseudo-R values of 0.637 and importance of 5.76% while its importance at the work end is 4.47%. Land use characteristics at the work end, in- 0.585 in the internal and external regions, respectively. Tese values are greater than those of traditional linear regressions, cluding job density, house density, and land use mix, have which are 0.207 and 0.237. For comparison, we estimated a higher contributions in the external region, ranking third, general GBRT for all active commuters and found that the fourth, and ninth, respectively. In contrast to that in the model has the lower pseudo-R (0.509). Tis indicates that internal region, the density of metro stations at work ends incorporating spatial heterogeneity in creating the GBRT in the external region has a greater impact, accounting for improves the model ft. 4.49% of the total. Tis may be due to the proximity of metro services to the workplace. Te bus stop density at both ends contributes around 3.5%, which is comparable 5.2.1. Relative Importance of Infuential Factors. Relative to the internal region. Te remaining BE variables had importance is commonly used in machine learning to only a minor infuence. For active commuters in the measure how much a factor infuences a dependent variable. external region, the distance from the work end to the All variables in this study have a relative importance that CBD (4.22%) contributes more to their trip duration than sums up to 100%. Te greater the relative importance of the the home end (3.08%) does. factor, the greater it contributes. Table 4 is the calculated relative importance of each infuential factor in determining 5.3. Spatial Heterogeneity in BE Impact. To describe the active commuting time in internal and external regions. Te spatial heterogeneity of BE impact, a more thorough result demonstrates that built-environment characteristics comparison of derived BE importance as well as BE asso- have a higher collective importance than social demo- ciations with active commuting time was made. graphics. Tis is consistent with the fndings of some earlier studies [30, 31]. Te importance of built-environment characteristics at both commute trip ends is 63.29% and 5.3.1. BE Importance. Figure 5 shows the comparison of BE 54.92%, respectively, for internal and external regions. Te importance to active commuting time in two regions. In the roughly 8% gap could be due to the more spatially con- internal region, nearly all BE variables have a relative im- strained nature of active commute trips in the internal re- portance more than 3%, with the exception of metro station gion. In both regions, built-environment features at the density at both ends and intersection density at the home work-end pose higher importance than those at home end, end. In external region, all BE variables at the work end and which is consistent with the fnding of Ding et al. [29]. three out of eight BE variables (intersection density, bus stop Similarly, we fnd diferences in the collective importance of density, and distance to CBD) at the home end have relative sociodemographics in both the regions. Tey are 5.49% more importance over 3%. Te most signifcant disparity is in the important in the external region than in the internal region. roles of street network-related factors, metro station density, Active commuters in the external region have more fexi- and land use-related factors. bility in determining their commuting time than those in the Te road network density at both ends contributes internal region. 1.78%∼2.69% more in the internal region than it does in In the internal region, road network density at both trip the external region. Te intersection density at the home ends contributes signifcantly to active commuting time, end in the internal region, on the other hand, contributes accounting for 5.24% (ranking 3rd) and 5.23% (ranking 4th), half as much as it does in the external region. Although respectively. Te efectiveness of active commuting is closely metro station density at the home end contributes the related to the connectivity of the street network, particularly least in both regions, its importance in the internal region the routes for cycling and walking. Tis is consistent with the is six times that of the external region. At the workplace, fndings of Cao [47]. Te intersections density at the work metro station density is three times as important in the end (4.41%) is shown to have an impact on the trip time of external region as it is in the internal region. When land active commuters. Land use mix, job density, and house use variables related to BE variables at home end, such as density at the work end have higher rankings than those at land use mix, job density, and house density are Journal of Advanced Transportation 9 10.75 5250 10.70 10.65 10.60 10.55 10.50 10.45 10.40 10.35 10.30 10.25 3000 0 5 10 15 20 25 30 35 40 45 50 Depth of the tree RMSE Iterations Figure 3: Result of RMSE in the internal region. 9.65 5250 9.60 9.55 9.50 9.45 9.40 9.35 2500 0 5 10 15 20 25 30 35 40 45 50 Depth of the tree RMSE Iterations Figure 4: Result of RMSE in the external region. compared, their roles range between 3.92%∼4.28% in the 5.3.2. Nonlinear Associations between BE and Active Com- internal region and 1.98%∼2.21% in the external region. muting Time. Partially dependent curves are used to present Notice that at the work end land use mix still holds a more the nonlinear associations between BE characteristics and active commuting time. In GBRT models, partial depen- important role in internal region than in external region (5.49% versus 3.47%). Tese varying efects from region to dence curves are commonly used to visualize the marginal region are closely related to land use characteristics. Te efects of independent variables on the dependent variable. diverse and well-developed land use pattern within the Figure 6 shows the relationships between BE at home internal region implies greater job options for active (columns 1 and 2) and work ends (columns 3 and 4) and commuters. As a result, these factors in the internal region active commuting time in internal (columns 1 and 3) and have great efects. TAZs in the external region, on the external (columns 2 and 4) regions. other hand, are generally less developed in large blocks Figures 6(a) and 6(b) show nonlinear associations be- with homogeneous land use. Distance to the CBD and bus tween street network-related characteristics and active stop density have similar roles in both regions. commute times. Te active commuting time for the internal RMSE RMSE Number of iterations Number of iterations 10 Journal of Advanced Transportation Table 4: Te relative importance of infuential factors in both regions. Internal region External region Variables Rank Relative importance (%) Sum (%) Rank Relative importance (%) Sum (%) Built environment at home end 28.97 21.65 Road density 3 5.24 16 2.55 Intersection density 16 2.68 2 5.76 Bus stop density 15 3.43 8 3.79 Metro station density 17 1.70 25 0.26 House density 11 3.92 18 2.21 Job density 9 4.28 19 2.02 Land use mix 10 4.12 20 1.98 Distance to CBD 13 3.60 12 3.08 Built environment at work end 34.32 33.27 Road density 4 5.23 10 3.45 Intersection density 7 4.41 6 4.47 Bus stop density 8 4.32 11 3.43 Metro station density 18 1.49 5 4.49 House density 6 4.69 4 4.69 Job density 5 4.85 3 5.05 Land use mix 2 5.49 9 3.47 Distance to CBD 12 3.84 7 4.22 Trip attributes 29.68 32.54 Departure time 14 3.50 14 2.72 Trip distance 1 26.18 1 29.82 Sociodemographics 7.05 12.54 Gender 19 1.47 17 2.26 Education 23 0.90 23 1.18 Age 20 1.40 15 2.72 License 22 1.08 22 1.25 Cars 21 1.09 21 1.85 Child 25 0.33 24 0.30 Income 24 0.78 13 2.98 Total 100 100 5.24% Road density 2.55% 2.68% Intersection density 5.76% 3.43% Bus stop density 3.79% 1.70% Metro station density 0.26% 3.92% House density 2.21% 4.28% Job density 2.02% 4.12% Land use mix 1.98% 3.60% Distance to CBD 3.08% 5.23% Road density 3.45% 4.41% Intersection density 4.47% 4.32% Bus stop density 3.43% 1.49% Metro station density 4.49% 4.69% House density 4.69% 4.85% Job density 5.05% 5.49% Land use mix 3.47% 3.84% Distance to CBD 4.22% 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 8.00 Relative Importance (%) Internal External Figure 5: Comparison of relative importance in internal and external regions. Work end Home end Journal of Advanced Transportation 11 20.0 20.1 External Internal External Internal 23.5 23.0 19.9 19.9 23.0 19.8 22.5 22.5 19.7 19.7 22.0 22.0 19.5 19.6 510 15 20 25 510 15 0 5 10 15 20 25 30 0 5 10 15 20 25 2 2 2 2 Road density (km/km ) Road density (km/km ) Road density (km/km ) Road density (km/km ) (a) 21.5 22.6 Internal External Internal External 22.6 21.0 22.4 22.4 20.5 22.2 22.2 20.0 22.0 19.5 21.8 22.0 0 20406080 01234 0 20406080 0 20406080 2 2 2 2 Intersection density (count/km ) Intersection density (count/km ) Intersection density (count/km ) Intersection density (count/km ) (b) 21.0 23.0 20.2 Internal External Internal External 22.7 20.5 22.6 20.0 22.5 22.5 19.8 20.0 22.4 22.0 19.6 22.3 19.5 22.2 19.4 21.5 5 10152025 0 5 10 15 0 5 10 15 20 25 0 5 10 15 20 25 2 2 2 2 Bus stop density (count/km ) Bus stop density (count/km ) Bus stop density (count/km ) Bus stop density (count/km ) (c) 22.8 Internal External Internal External 22.4 23 19.730 22.6 22.2 22.4 19.720 21 22.0 22.2 21.8 19.710 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0 0.0 0.2 0.4 0.6 0.8 1.0 1.2 2 2 2 Metro station density (count/km ) 2 Metro station density (count/km ) Metro station density (count/km ) Metro station density (count/km ) (d) 21.5 23.5 19.9 23.0 Internal External External Internal 21.0 19.8 23.0 22.5 20.5 19.7 22.5 20.0 22.0 19.6 19.5 22.0 19.5 21.5 0.0 0.2 0.4 0.6 0.8 0.0 0.1 0.2 0.3 0.4 0.5 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 House density House density House density House density (e) 23.0 20.0 Internal External Internal External 23.5 20.0 22.8 19.9 22.6 19.8 23.0 19.8 22.4 19.7 19.6 22.5 22.2 19.6 22.0 19.4 19.5 22.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 0.0 Job density Job density Job density Job density (f ) Figure 6: Continued. Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) 12 Journal of Advanced Transportation Internal External Internal External Internal 20.2 19.80 23.5 23.0 19.75 20.0 23.0 22.5 19.70 22.5 19.8 19.65 22.0 19.60 22.0 19.6 0.3 0.4 0.5 0.6 0.7 0.8 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.2 0.4 0.6 0.8 Land use mix Land use mix Land use mix Land use mix (g) 20.6 22.6 21.0 External Internal External Internal 23.5 20.4 22.4 20.5 20.2 23.0 20.0 22.2 20.0 22.5 19.8 22.0 19.6 22.0 19.5 19.4 21.8 0 5 10 15 20 25 30 35 510 15 20 25 30 35 05 10 15 20 25 30 35 0 10 20 30 Distance to CBD (km) Distance to CBD (km) Distance to CBD (km) Distance to CBD (km) BE at the home end BE at the work end (h) 2 2 Figure 6: Nonlinear impact of BE characteristics. (a) Road density (km/km ). (b) Intersection density (count/km ). (c) Bus stop density 2 2 (count/km ). (d) Metro station density (count/km ). (e) House density. (f ) Job density. (g) Land use mix. (h) Distance to CBD (km). region decreases rapidly as the home-end road density in- exhibits the contrast patterns. Te extension of home-end creases from 0 to 11 km/km and bottoms at 22 minutes. It metro stations increases commuting time in the internal then gradually increases. Active commuting in the external region while decreasing them in the external region. Te region decreases until the home-end road density reaches extension of work-end metro stations reduces active com- 7 km/km , and then it goes up sharply. Te curves of work- muting time in the internal region while increasing it in the end road density in both regions also have a nonlinear external region. feature. Within 10 km/km , the active commuting time in Figures 6(e) and 6(g) illustrate the efects of land use- the internal region decreases sharply in a nearly linear related variables, including house density, job density, and pattern. After that, it climbs to 22.8 min and then remains land use mix. In general, active commuting time increases stable. With increasing work-end road density in the ex- with increasing house density at the home end, although there are some fuctuations in the curve of the external ternal region, active commuting time fuctuates in an ap- proximate inverted U shape; when the road density at the region within the ratio of 0.1. Because the data points in work end reaches 5 km/km in the external region, the in- these intervals are too sparse to interpret, the fuctuations creasing trend in commuting time stops. As shown in can be ignored. In both regions, house density at the work columns 1 and 2 of Figure 6, intersection density at home end has a positive relationship with active commuting time. end is positively associated with active commuting time, Work-end house density, as presented by the ratio of res- when it is in the range of 0∼40 per km in internal region and idential land at the work end, is better kept within 0.5 for the 0∼3 per km in external region. Higher intersection density internal region and within 0.3 for the external region. Te is often associated with longer stopping times, which in turn general trend can be explained by the greater sense of safety increases active commute times. Both fndings reinforce the while walking/cycling in areas with higher house density, ambiguous impact of street network design on active mo- which leads people to be willing to walk/cycle longer to their work. Job density at the home end has a U-shaped rela- bility [49]. Improved street network connectivity may demonstrate that there are more alternative shortcuts to tionship with active commuting time in internal region. Te reach destinations, reducing travel time further, but it may threshold of 0.3 indicates the ideal ratio of work-related land. also increase commute time due to more intersections. However, in the external region, the association is generally Figures 6(c) and 6(d) compare the efects of transit- negative. Commuting time stops decreasing when the ratio related variables in both regions. Te density of bus stops at of work-related lands reaches 0.5. Active commuting time the home end has the opposite efects. Active commuting increases in both regions as job density increases at the time increases when home-end bus stops increase in the workplace. external region but decreases when they increase in the For land use mix, its association with active commuting internal region. Both regions see a similar trend in the time varies by region. For active commuters in the internal impact of work-end bus stop density, namely, a reduction in region, the best home-end land use mix is around 0.5. However, in the external region, a land use mix of 0.5 or active commuting time with an increase in bus stop density. A minor diference between the two curves is in the range of higher is preferable for active commuters. Te work-end 10∼18 stops/km in the internal region where a fuctuation land use mix, in contrast to the home end, has a winding, exists. Te impact of metro station density at both ends decreasing association with active commuting time in the Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Commuting time (min) Journal of Advanced Transportation 13 internal region. Te active commuting time reaches its should be included as part of conditional variables. Second, lowest point (of 22 minutes) when the land use mix is 0.7. In the temporal impact of BE on active commuting time de- serves attention if new data sources become available. Fi- external region, the work-end land use mix has an inverted U-shaped relationship with active commuting time. It nally, while the study’s fndings are applicable to Nanjing suggests that increasing the land use mix in the external and may provide basic references for cities in similar con- region at the work end is only recommended to a certain texts to Nanjing, the BE impact varies from city to city [50]. extent in order to promote active commuting. More case studies are recommended in the future. Figure 6(h) shows the relationship between distance to the CBD and active commute time. Te residence distance to Data Availability the CBD has a U-shaped relationship with active commute time in both regions. For active commuters in the internal Te household travel survey data used to support the fndings of this study are not available because of data region, with the distance of their residence to the CBD around 5 km, active commuting time reaches its lowest value privacy and protection. (of 21.5 minutes); however, for commuters in the external region, when the distance is in the range of 15∼20 km, the Conflicts of Interest time keeps short. At the workplace, its average distance to Te authors declare that they have no conficts of interest. the CBD is 8.8 km for the internal region and 16.9 km for the external region. We focus more on curve intervals of less than 10 km for internal region and over 15 km for external Authors’ Contributions region. Active commuting time increases with the increasing Jingxian Wu and Huapeng Shen were in charge of con- distance of the workplace from the CBD in both intervals. ceptualization, methodology, formal analysis, and initial Te rise in active commuting time as a consequence of draft preparation. Jingxian Wu, Guikong Tang, and Soora increasing distance from home to CBD is monotonic. Rasouli were responsible for review and editing. Jingxian Wu was in charge of funding acquisition, investigation, and 6. Conclusions and Discussion supervision. Guikong Tang was in charge of data visuali- zation. Soora Rasouli was responsible for language checks With the evidence from Nanjing, China, this paper inves- and supervision. tigates the spatial heterogeneity in the BE impact on active commuting time. It uses the two-step clustering method to cluster 766 TAZs according to their BE features. Gradient- Acknowledgments boosted regression trees are then constructed for each Tis research work was supported by the Scientifc Research distinguished cluster to examine the heterogeneity in the Fund of the Institute of Engineering Mechanics, China importance of BE for active commuting time. It has been Earthquake Administration (grant no. 2020D23), National concluded that built-environment characteristics have more Natural Science Foundation of China (grant no. 52122215), importance than sociodemographics because they contrib- and the Soft Science Key Project of Shanghai (grant no. ute 63.29% to active commuting time in the internal region 21692105100). and 54.92% in the external region. Tis confrms the con- clusions of Cheng et al. that despite of the minor impact of single BE factor, their total impacts were larger than that of References sociodemographics [30]. Te spatial heterogeneity of BE’s [1] S. Cook, L. Stevenson, R. Aldred, M. Kendall, and T. Cohen, role to active commuting time is further proved by com- “More than walking and cycling: what is ‘active travel’?” paring the nonlinear impact of BE on active commuting time Transport Policy, vol. 126, pp. 151–161, 2022. in two regions. Such heterogeneity is so pronounced that in [2] R. Buehler, T. Kuhnimhof, A. Bauman, and C. Eisenmann, many cases the impact of BE on active commuting time is “Active travel as stable source of physical activity for one third opposite in two regions. of German adults: evidence from longitudinal data,” Trans- Te implications of this study for practice are two-fold. portation Research Part A: Policy and Practice, vol. 123, First, if the goal of urban planners and practitioners is to pp. 105–118, 2019. encourage the use of active modes, they need to prioritize [3] R. Buehler, J. Pucher, and A. Bauman, “Physical activity from diferent policies in each region depending on their im- walking and cycling for daily travel in the United States, 2001–2017: demographic, socioeconomic, and geographic portance in reaching the ultimate goal. Tis is especially variation,” Journal of Transport & Health, vol. 16, Article ID important when there is a limited budget that can be spared 100811, 2020. for improving one or a few built-environment character- [4] A. Banerjee, M. Łukawska, A. F. Jensen, and S. Haustein, istics. Second, it is very important for policymakers to be “Facilitating bicycle commuting beyond short distances: in- aware of the adverse efects of certain policies in some re- sights from existing literature,” Transport Reviews, vol. 42, gions if those policies are developed on the basis of average no. 4, pp. 526–550, 2022. efects. [5] R. Goel, A. Goodman, R. Aldred et al., “Cycling behaviour in Te current work comes with some limitations, which 17 countries across 6 continents: levels of cycling, who cycles, leave space for future improvements. First, if the data allows, for what purpose, and how far?” Transport Reviews, vol. 42, walkability measures such as sidewalk presence and width no. 1, pp. 58–81, 2022. 14 Journal of Advanced Transportation [6] A. Paez ´ and K. Whalen, “Enjoyment of commute: a com- factors,” American Journal of Preventive Medicine, vol. 42, parison of diferent transportation modes,” Transportation no. 6, pp. 610–615, 2012. [22] B. Sun, A. Ermagun, and B. Dan, “Built environmental im- Research Part A: Policy and Practice, vol. 44, no. 7, pp. 537– pacts on commuting mode choice and distance: evidence 549, 2010. from Shanghai,” Transportation Research Part D: Transport [7] H. M. A. Aziz, N. N. Nagle, A. M. Morton, M. R. Hilliard, and Environment, vol. 52, pp. 441–453, 2017. D. A. White, and R. N. Stewart, “Exploring the impact of [23] K. M. Leung, P. K. Chung, D. Wang, and J. D. Liu, “Impact of walk–bike infrastructure, safety perception, and built-envi- physical and social environments on the walking behaviour of ronment on active transportation mode choice: a random Hong Kong’s older adults,” Journal of Transport & Health, parameter model using New York City commuter data,” vol. 9, pp. 299–308, 2018. Transportation, vol. 45, no. 5, pp. 1207–1229, 2018. [24] M. Moniruzzaman and A. Paez, “An investigation of the [8] J. Feng, “Te built environment and active travel: evidence attributes of walkable environments from the perspective of from Nanjing, China,” International Journal of Environmental seniors in Montreal,” Journal of Transport Geography, vol. 51, Research and Public Health, vol. 13, no. 3, p. 301, 2016. pp. 85–96, 2016. [9] R. Cervero, O. L. Sarmiento, E. Jacoby, L. F. Gomez, and [25] C. Yin, J. Zhang, and C. Shao, “Relationships of the multi- A. Neiman, “Infuences of built environments on walking and scale built environment with active commuting, body mass cycling: lessons from Bogota,” ´ International journal of sus- index, and life satisfaction in China: a GSEM-based analysis,” tainable transportation, vol. 3, no. 4, pp. 203–226, 2009. Travel Behavior and Society, vol. 21, pp. 69–78, 2020. [10] J. Wu, B. Wang, N. Ta, K. Zhou, and Y. Chai, “Does street [26] C. Chen, H. Gong, and R. Paaswell, “Role of the built envi- greenery always promote active travel? Evidence from Bei- ronment on mode choice decisions: additional evidence on jing,” Urban Forestry and Urban Greening, vol. 56, Article ID the impact of density,” Transportation, vol. 35, no. 3, 126886, 2020. pp. 285–299, 2008. [11] Z. Zhu, Z. Li, Y. Liu, H. Chen, and J. Zeng, “Te impact of [27] M. Moniruzzaman, A. Paez, K. M. Nurul Habib, and urban characteristics and residents’ income on commuting in C. Morency, “Mode use and trip length of seniors in Mon- China,” Transportation Research Part D: Transport and En- treal,” Journal of Transport Geography, vol. 30, pp. 89–99, vironment, vol. 57, pp. 474–483, 2017. [12] S. L. Handy, M. G. Boarnet, R. Ewing, and R. E. Killingsworth, [28] R. Raman and U. K. Roy, “Taxonomy of urban mixed land use “How the built environment afects physical activity: views planning,” Land Use Policy, vol. 88, Article ID 104102, 2019. from urban planning,” American Journal of Preventive [29] C. Ding, X. Cao, and Y. Wang, “Synergistic efects of the built Medicine, vol. 23, no. 2, pp. 64–73, 2002. environment and commuting programs on commute mode [13] G. Eldeeb, M. Mohamed, and A. Paez, ´ “Built for active travel? choice,” Transportation Research Part A: Policy and Practice, Investigating the contextual efects of the built environment vol. 118, pp. 104–118, 2018. on transportation mode choice,” Journal of Transport Geog- [30] L. Cheng, J. D. Vos, P. Zhao, M. Yang, and F. Witlox, “Ex- raphy, vol. 96, Article ID 103158, 2021. amining non-linear built environment efects on elderly’s [14] Y. Wang, C. K. Chau, W. Y. Ng, and T. M. Leung, “A review walking: a random forest approach,” Transportation Research on the efects of physical built environment attributes on Part D: Transport and Environment, vol. 88, Article ID enhancing walking and cycling activity levels within resi- 102552, 2020. dential neighborhoods,” Cities, vol. 50, pp. 1–15, 2016. [31] J. Liu, B. Wang, and L. Xiao, “Non-linear associations between [15] R. Cervero and K. Kockelman, “Travel demand and the 3Ds: built environment and active travel for working and shopping: density, diversity, and design,” Transportation Research Part an extreme gradient boosting approach,” Journal of Transport D: Transport and Environment, vol. 2, no. 3, pp. 199–219, Geography, vol. 92, Article ID 103034, 2021. [32] H. Yang, Z. Zhao, C. Jiang, Y. Wen, and M. Muneeb Abid, [16] R. Ewing and R. Cervero, “Travel and the built environment,” “Spatially varying relation between built environment and Journal of the American Planning Association, vol. 76, no. 3, station-level subway passenger-distance,” Journal of Ad- pp. 265–294, 2010. vanced Transportation, vol. 2022, Article ID 7542560, 18 [17] L. Cheng, X. Chen, S. Yang, Z. Cao, J. De Vos, and F. Witlox, pages, 2022. “Active travel for active ageing in China: the role of built [33] L. Yang, J. Liu, Y. Liang, Y. Lu, and H. Yang, “Spatially varying environment,” Journal of Transport Geography, vol. 76, efects of street greenery on walking time of older adults,” pp. 142–152, 2019. ISPRS International Journal of Geo-Information, vol. 10, no. 9, [18] L. Yang, S. Grifn, K. T. Khaw, N. Wareham, and J. Panter, p. 596, 2021. “Longitudinal associations between built environment char- [34] S. Srinivasan and J. Ferreira, “Travel behavior at the household acteristics and changes in active commuting,” BMC Public level: understanding linkages with residential choice,” Health, vol. 17, no. 1, Article ID 839, 458 pages, 2017. Transportation Research Part D: Transport and Environment, [19] T. Tao, X. Wu, J. Cao, Y. Fan, K. Das, and A. Ramaswami, vol. 7, no. 3, pp. 225–242, 2002. “Exploring the nonlinear relationship between the built en- [35] F. M. Dieleman, M. Dijst, and G. Burghouwt, “Urban form vironment and active travel in the Twin Cities,” Journal of and travel behaviour: micro-level household attributes and Planning Education and Research, Article ID residential context,” Urban Studies, vol. 39, no. 3, pp. 507–527, 0739456X2091576, 2020. [20] X. Ma, J. Yang, C. Ding, J. Liu, and Q. Zhu, “Joint analysis of [36] W. Tu, R. Cao, Y. Yue, B. Zhou, Q. Li, and Q. Li, “Spatial the commuting departure time and travel mode choice: role of variations in urban public ridership derived from GPS tra- the built environment,” Journal of Advanced Transportation, jectories and smart card data,” Journal of Transport Geogra- vol. 2018, Article ID 4540832, 13 pages, 2018. phy, vol. 69, pp. 45–57, 2018. [21] M. A. Beenackers, S. Foster, C. B. Kamphuis et al., “Taking up [37] W. Zhang, Y. Zhao, X. Cao, D. Lu, and Y. Chai, “Nonlinear cycling after residential relocation: built environment efect of accessibility on car ownership in Beijing: pedestrian- Journal of Advanced Transportation 15 scale neighborhood planning,” Transportation Research Part D: Transport and Environment, vol. 86, Article ID 102445, [38] S. Zhong, Z. Wang, Q. Wang, A. Liu, and J. Cui, “Exploring the spatially heterogeneous efects of urban built environment on road travel time variability,” Journal of Transportation Engineering Part A Systems, vol. 147, no. 1, Article ID 04020142, 2021. [39] C. Ding, X. Cao, B. Yu, and Y. Ju, “Non-linear associations between zonal built environment attributes and transit commuting mode choice accounting for spatial heterogene- ity,” Transportation Research Part A: Policy and Practice, vol. 148, pp. 22–35, 2021. [40] M. A. Abas, S. Rajoo, and S. F. Zainal Abidin, “Development of Malaysian urban drive cycle using vehicle and engine parameters,” Transportation Research Part D: Transport and Environment, vol. 63, pp. 388–403, 2018. [41] J. Große, A. S. Olafsson, T. A. Carstensen, and C. Fertner, “Exploring the role of daily “modality styles” and urban structure in holidays and longer weekend trips: travel be- haviour of urban and peri-urban residents in Greater Copenhagen,” Journal of Transport Geography, vol. 69, pp. 138–149, 2018. [42] M. T. Ballestar, P. Grau-Carles, and J. Sainz, “Customer segmentation in e-commerce: applications to the cashback business model,” Journal of Business Research, vol. 88, pp. 407–414, 2018. [43] A. Tkaczynski, S. Rundle-Tiele, and N. Beaumont, “Desti- nation segmentation: a recommended two-step approach,” Journal of Travel Research, vol. 49, no. 2, pp. 139–152, 2010. [44] T. Chiu, D. Fang, J. Chen, Y. Wang, and C. Jeris, “A robust and scalable clustering algorithm for mixed type attributes in large database environment,” in Proceedings of the 7th ACM SIGKDDD International Conference on Knowledge Discovery and Data Mining, pp. 263–268, San Francisco, August 2001. [45] J. H. Friedman, “Greedy function approximation: a gradient boosting machine,” Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001. [46] G. Ridgeway, Generalized Boosted Models: A Guide to the GBM Package, 2005. [47] X. J. Cao, “Examining the impacts of neighborhood design and residential self-selection on active travel: a methodo- logical assessment,” Urban Geography, vol. 36, no. 2, pp. 236–255, 2015. [48] Y. Fan, K. V. Das, and Q. Chen, “Neighborhood green, social support, physical activity, and stress: assessing the cumulative impact,” Health & Place, vol. 17, no. 6, pp. 1202–1211, 2011. [49] J. Wu, B. Wang, R. Wang, T. Na, and Y. Chai, “Active travel and the built environment: a theoretical model and multi- dimensional evidence,” Transportation Research Part D: Transport and Environment, vol. 100, Article ID 103029, 2021. [50] P. Naess, “Meta-analyses of built environment efects on travel: no new platinum standard,” Journal of Planning Ed- ucation and Research, vol. 42, no. 2, pp. 199–205, 2019.

Journal

Journal of Advanced Transportation – Hindawi Publishing Corporation

Published: Feb 13, 2023

There are no references for this article.