Spatial heterogeneity of built environment impacts on metro–bus transfer ridership: evidence from Shanghai, China

Wenxiang Li; Shuxia Pang; Tianyi Li; Yin Han; Wenxiang Li; Shuxia Pang; Tianyi Li; Yin Han

doi:10.48130/dts-0026-0023

First- and last-mile connectivity in dense megacities depends heavily on metro–bus transfers, yet the two directions, bus-to-metro (access) and metro-to-bus (egress), are usually treated as one and modeled with global regressions that assume spatially uniform effects. Using 1 month of smart card records from 306 Shanghai metro stations, we estimate directional transfer rates with Ordinary Least Squares (OLS) and Geographically Weighted Regression (GWR), and mask statistically insignificant local coefficients with pseudo t-statistics. GWR substantially improves model fit over OLS and reveals a clear core–periphery gradient. For bus-to-metro flows, metro ridership and distance to the central business district (CBD) strengthen feeder demand in peripheral districts but lose significance near the urban core, while parking density suppresses transfers network-wide. For metro-to-bus flows, egress is more sensitive to CBD distance, and parking density operates as a localized rather than uniform deterrent. The same built-environment factor can therefore push transfer behavior in opposite directions depending on station context, which argues for station-specific rather than citywide integration policies.

HTML

Introduction

Metro systems in global megacities move billions of passengers annually, yet their effectiveness hinges critically on integration with feeder services. In Shanghai alone, millions of daily metro trips depend on bus connections for first- and last-mile access, making transfer efficiency a pivotal determinant of system-wide performance. The well-known 'first-mile/last-mile' problem constrains the effectiveness of public transport, and metro–bus transfers are central to bridging these gaps^[1]. Understanding what drives transfer ridership is therefore crucial for designing seamless multimodal systems in dense urban contexts such as Shanghai.

Extensive research has examined determinants of public transport ridership, with particular attention to built environment factors summarized in the 5Ds framework: density, diversity, design, distance to transit, and destination accessibility^[2,3]. Empirical studies consistently highlight how urban form and land-use composition shape access to and egress from metro systems. For instance, population and employment density tend to increase metro usage, while parking supply or low street connectivity may discourage feeder trips. These findings suggest that built environment characteristics shape not only overall transit demand but also the propensity to combine modes—yet their effects on transfer behavior specifically remain underexplored.

The increasing availability of smart card data has enabled richer insights into multimodal behavior. Such data provide detailed spatiotemporal records of passenger trips, allowing researchers to detect transfer patterns with much higher resolution than surveys or aggregate counts^[4−6]. Studies have used smart card datasets to analyze metro access modes, characterize transfer behavior, and monitor temporal variations in flows. Yet most of this work has treated transfers in aggregate, often overlooking the directional distinction between bus-to-metro (B$ \rightarrow $M) and metro-to-bus (M$ \rightarrow $B) transfers, which serve different functions in daily commuting. B$ \rightarrow $M transfers function primarily as access trips, concentrated in morning peaks and shaped by bus stop accessibility and station proximity. M$ \rightarrow $B transfers, by contrast, serve as egress trips linked to employment density, land-use mix, and activity centers, often spread across the day. Treating these flows as a single category conceals their distinct drivers and limits the effectiveness of targeted policies for multimodal integration.

A second limitation is methodological: many prior studies employ global models such as Ordinary Least Squares (OLS)^[7], which assume spatially invariant effects of explanatory variables^[8]. However, urban contexts vary widely: bus stop proximity may be highly predictive of transfers in peripheral districts but less influential in city centers, where networks are denser. Similarly, parking availability may suppress transfers in suburban locations but play little role downtown. Recent transport research increasingly recognizes that such relationships are spatially heterogeneous^[9]. Ignoring this heterogeneity risks obscuring localized dynamics and misinforming policy.

To address these gaps, this study integrates large-scale smart card data with built environment measures to investigate metro–bus transfers in Shanghai. We distinguish between B$ \rightarrow $M and M$ \rightarrow $B transfer rates and apply both global (OLS) and local (GWR) modeling approaches to capture spatially varying relationships. The objectives are threefold: (1) to identify key built environment and service-related factors affecting metro–bus transfer ridership; (2) to examine how the influence of these factors varies across space, highlighting differences between central and peripheral areas; (3) to provide empirical evidence that supports differentiated strategies for enhancing metro–bus integration in megacities.

Building upon the integration of large-scale smart card data with detailed built-environment and socio-demographic indicators, and employing both global (OLS) and local (GWR) regression approaches, this study makes three major contributions. Methodologically, it applies GWR alongside OLS to capture spatial heterogeneity overlooked by global models. Empirically, it combines multimodal travel data and contextual variables to analyze directional transfer rates in Shanghai, distinguishing between B$ \rightarrow $M and M$ \rightarrow $B flows. Practically, it reveals how the influence of key factors varies between central and peripheral areas, offering evidence for targeted strategies to enhance metro–bus connectivity in megacities.

The remainder of the paper is organized as follows. The literature review synthesizes prior work and identifies the research gap; the data preparation section introduces the study area and datasets; the methodology section outlines the modeling framework; the results and discussion section presents the findings and examines spatial heterogeneity; and the conclusions summarize key insights and policy implications.

Literature review

Studies on multimodal integration have examined metro–bus transfers in terms of ridership, built-environment influences, and spatial heterogeneity, yet these strands are rarely integrated: ridership analyses overlook spatial variation, built-environment studies assume uniform effects, and spatial models seldom distinguish directional flows. To situate this study, the review covers three areas—empirical evidence on transfers, built-environment correlates, and spatial methods—together with a synthesis of the remaining research gap.

Empirical evidence on metro–bus transfers
Research on multimodal integration has long emphasized the role of transfers in sustaining public transport ridership. Early work examined transfer penalties, showing that additional walking, waiting, and uncertainty can generate substantial perceived disutility for passengers^[10]. Subsequent studies used passenger survey data to measure transfer inconvenience and its impact on mode choice, finding that poor transfer design significantly suppresses public transport demand. More recent work has applied operational data to evaluate transfer efficiency at interchange hubs, highlighting the importance of minimizing access distance and coordinating bus arrivals with metro schedules. Other studies have employed smart card data to quantify temporal dynamics of transfers, offering more precise operational evidence^[4,6,11]. Similar approaches have been extended internationally, such as in Seoul and Singapore, where automatic fare collection (AFC) data have revealed how transfer penalties vary across user groups and times of day^[5]. Taken together, this body of research demonstrates that metro–bus transfers are pivotal to ridership outcomes, with their efficiency and convenience directly conditioning the overall performance of urban public transport systems.

Built-environment effects on metro–bus transfers
Beyond service coordination, the built environment^[12] shapes how easily passengers can reach and leave metro stations. The '3D/5D' paradigm has been widely used to explain station-level ridership^[12,13]. Applying these principles to transfers, scholars have shown that higher residential density and mixed land use around stations tend to increase feeder bus demand, while large parking supply can reduce bus reliance by encouraging park-and-ride. Station-area pedestrian and bicycle facilities also influence transfer choice, as safe and connected networks make non-motorized options more competitive. More recently, researchers have employed points-of-interest (POI) data and fine-grained street network measures to capture the urban design characteristics that encourage or discourage transfers. Case studies in cities such as Guangzhou^[14] confirm that the relationship between built environment and transfer propensity is context-dependent and may differ substantially between access and egress transfers. However, many of these studies remain geographically limited and often treat metro ridership as a single outcome, without distinguishing directional transfers between bus-to-metro (access) and metro-to-bus (egress).

Spatial analysis methods and heterogeneity
Methodologically, most existing studies estimate global regression models that assume spatially uniform effects across a network. Yet mounting evidence suggests that transfer determinants vary geographically. For example, bus stop proximity may be decisive in suburban areas but negligible in dense city centers, while parking supply may strongly discourage bus feeders in outer districts but play little role downtown^[15]. GWR and other local models provide a means to capture such heterogeneity, offering location-specific parameter estimates^[16,17]. Recent applications in transport geography confirm that built-environment impacts on ridership can differ significantly across stations^[14]. Spatiotemporal extensions of GWR also demonstrate that built-environment effects may change across both location and time of day, underscoring the importance of accounting for contextual variation. Despite these advances, relatively few studies have explicitly applied spatially varying coefficient models to metro–bus transfer ridership, leaving a gap in understanding the localized dynamics of access and egress transfers.

Three gaps therefore remain. First, most studies aggregate directional flows, ignoring functional differences between bus-to-metro (access) and metro-to-bus (egress) transfers. A related body of work has approached metro–bus integration from environmental, demand-forecasting, and bus-operation perspectives. Wu et al.^[18] combined phase-space reconstruction with a CNN–LSTM architecture to forecast short-term inbound passenger flows at urban rail stations; Yuan et al.^[19] and Sui et al.^[20] emphasized bus frequency, timetable coordination, and facility planning in multimodal systems. These studies offer useful context, but they leave open a more specific question: how do station-area built-environment and service attributes affect directional B$ \rightarrow $M and M$ \rightarrow $B transfer rates, and do these effects vary across space? This study addresses that question using station-level smart card data and GWR. Second, built environment research assumes spatially uniform effects despite evidence that relationships vary across urban contexts. Third, limited integration of smart card and built environment data constrains station-level analysis. This study addresses these gaps through directional disaggregation, geographically weighted regression, and integrated multi-source data.

Methodology

This study employs GWR and OLS models to examine how built-environment variables influence metro–bus transfer rates. To ensure model robustness, the VIF test was first applied to the 14 explanatory variables, thereby addressing potential multicollinearity and improving estimation reliability. Subsequently, a stepwise regression procedure was conducted to retain only statistically significant variables. This process reduced redundancy among predictors and ensured that the final models focused on the most relevant determinants of $ R_{{\mathrm{B}} \rightarrow {\mathrm{M}}} $ and $ R_{{\mathrm{M}} \rightarrow {\mathrm{B}}} $, providing a consistent basis for both global (OLS) and local (GWR) analyses.

Multicollinearity test
Before estimating the regression models, it is essential to ensure that explanatory variables are not strongly correlated with one another, as multicollinearity may bias coefficient estimates and reduce the reliability of statistical inference. To address this issue, the variance inflation factor ($ VIF $) was employed to diagnose multicollinearity among the selected variables. The $ VIF $ is calculated for each independent variable based on the coefficient of determination ($ R^{2} $) obtained from an auxiliary regression, as shown in Eq. (3):

$ VIF = \dfrac{1}{1 - R^{2}} $ (3)

A commonly used threshold is $ VIF \gt 10 $, which indicates a high level of multicollinearity and suggests that the variable should be reconsidered or removed^[28]. In this study, the VIF test was applied to the fourteen explanatory variables described in the variable descriptions subsection to ensure that multicollinearity did not compromise the accuracy of the subsequent OLS and GWR models.

Ordinary least squares

To examine the global determinants of bus–metro transfer rates, we employed the OLS regression model as a baseline analytical framework^[29]. The dependent variables were defined as the $ R_{{\mathrm{M}} \rightarrow {\mathrm{B}}} $ and $ R_{{\mathrm{B}} \rightarrow {\mathrm{M}}} $ at each metro station, while explanatory variables were identified using the multicollinearity and stepwise screening procedures described above.

The OLS model is specified as:

$ Y_i = \beta_0 + \sum_{k=1}^{p} \beta_k X_{ik} + \varepsilon_i, \quad i=1,2,\cdots,n, $

(4)

where $ Y_i $ denotes the $ R_{{\mathrm{M}} \rightarrow {\mathrm{B}}} $ and $ R_{{\mathrm{B}} \rightarrow {\mathrm{M}}} $ at station $ i $, $ X_{ik} $ represents the $ k $^th explanatory variable for station $ i $, $ \beta_k $ is the corresponding regression coefficient, and $ \varepsilon_i $ is the random error term.

Stepwise regression

To identify the most influential explanatory variables, stepwise regression was employed prior to model estimation. This procedure iteratively introduces or eliminates candidate variables based on the $ F $-test, retaining only those with significant explanatory power^[28]. Separate regressions were conducted for $ R_{{\mathrm{B}} \rightarrow {\mathrm{M}}} $ and $ R_{{\mathrm{M}} \rightarrow {\mathrm{B}}} $, and the selected variables were subsequently used in the OLS and GWR models.

The decision criterion for stepwise selection is based on the $ F $-test statistic:

$ F = \dfrac{(RSS_{r} - RSS_{ur})\,/\,(p_{ur} - p_{r})}{RSS_{ur}\,/\,(n - p_{ur} - 1)} $

(5)

where, $ RSS_{r} $ denotes the residual sum of squares (RSS) of the restricted model, $ RSS_{ur} $ is the RSS of the unrestricted model, $ p_{r} $ and $ p_{ur} $ represent the number of explanatory variables in the restricted and unrestricted models, respectively, and $ n $ is the sample size. When the $ F $ statistic exceeds the critical value, the newly added variable is considered to significantly improve model fit and is retained. Through repeated testing of variable significance, stepwise regression ensures that only robust predictors of transfer rates are preserved, providing reliable inputs for subsequent OLS and GWR analysis. A lenient entry threshold was adopted at this stage so that borderline predictors would not be discarded prematurely; final variable retention for the GWR specifications was therefore determined by the subsequent OLS t-tests rather than by the stepwise procedure alone.

Spatial autocorrelation test

Before applying the GWR, the spatial dependence of the global model residuals is examined, since residual spatial autocorrelation indicates that a global specification has failed to capture spatially varying relationships^[30]. Moran's $ I $ is additionally computed for the individual explanatory variables to characterize their spatial clustering, although variable-level clustering is descriptive and does not by itself establish coefficient non-stationarity. Spatial autocorrelation refers to the degree of similarity between values of a variable across spatial locations, where positive autocorrelation indicates clustering of similar values and negative autocorrelation implies spatial dispersion.

In this study, Moran's $ I $ statistic was employed to evaluate the global spatial autocorrelation of the explanatory variables, including built-environment measures, transport service indicators, and demographic attributes. Moran's $ I $ is defined as:

$ I = \dfrac{n}{W} \cdot \dfrac{ \displaystyle\sum_{i=1}^{n}\displaystyle\sum_{j=1}^{n} w_{ij}(x_i - \bar{x})(x_j - \bar{x})} { \displaystyle\sum_{i=1}^{n}(x_i - \bar{x})^2} $

(6)

where, $ n $ is the number of spatial units; $ x_i $ and $ x_j $ are the observed values of variable $ x $ at locations $ i $ and $ j $; $ \bar{x} $ is the mean of $ x $; $ w_{ij} $ is the spatial weight matrix element between $ i $ and $ j $; and $ W = \displaystyle\sum_i\displaystyle\sum_j w_{ij} $ is the sum of all spatial weights.

If the test results indicate significant positive spatial autocorrelation, global OLS estimates may fail to capture localized spatial processes. In such cases, local models such as GWR are necessary to account for spatial heterogeneity and to provide more accurate inference.

Geographically weighted regression model

In regression analysis, if linear regression models are used without accounting for spatial context, the estimated coefficients may ignore spatial non-stationarity and fail to capture the true underlying relationships^[30]. Specifically, the association between independent variables and dependent outcomes often varies across space, reflecting localized dynamics rather than global uniformity. Ignoring this heterogeneity can obscure important insights and lead to biased inferences. In the context of this study, spatial non-stationarity is particularly relevant for analyzing metro–bus transfer rates, where built environment factors may exert different effects in central versus peripheral districts^[14]. To address this issue, we adopt GWR, which extends the traditional OLS model by allowing coefficients to vary across geographic space.

The GWR model can be expressed as:

$ Y_i = \beta_0(u_i,v_i) + \sum_{k=1}^{p}\beta_k(u_i,v_i)\,X_{ik} + \varepsilon_i, \quad i=1,2,\cdots,n $

(7)

where, $ (u_i,v_i) $ denotes the spatial coordinates of observation $ i $; $ \beta_k(u_i,v_i) $ represents the location-specific coefficient for variable $ k $; and $ \varepsilon_i $ is the error term. This specification captures the spatial heterogeneity of the relationship between explanatory variables and metro–bus transfer rates.

The coefficients are estimated using a weighted least squares procedure:

$ \hat{\beta}(u_i,v_i) = \bigl(X^{\top} W(u_i,v_i)\,X\bigr)^{-1} X^{\top} W(u_i,v_i)\,Y $

(8)

where, $ X $ is the design matrix:

$ \begin{array}{l} X = \begin{bmatrix} 1 & x_{11} & \cdots & x_{1k} \\ 1 & x_{21} & \cdots & x_{2k} \\ \vdots & \vdots & \ddots & \vdots \\ 1 & x_{n1} & \cdots & x_{nk} \end{bmatrix}, \end{array} $

(9)

$ W(u_i,v_i) $ is the spatial weight matrix:

$ W(u_i,v_i) = \operatorname{diag}\bigl\{ w_{1}(u_i,v_i),\, w_{2}(u_i,v_i),\,\ldots,\, w_{n}(u_i,v_i) \bigr\}. $

(10)

In this study, an adaptive bisquare kernel function was employed to construct the spatial weight matrix. The weight assigned to observation $ j $ at location $ i $ is defined as:

$ w_{ij} = \left\{\begin{aligned} &\left[1 - \left(\dfrac{d_{ij}}{b_i}\right)^2\right]^2, && \text{if } d_{ij} \lt b_i &\\ &0, && \text{otherwise}& \end{aligned}\right. $

(11)

where, $ d_{ij} $ is the Euclidean distance between locations $ i $ and $ j $, and $ b_i $ is the adaptive bandwidth at location $ i $. The adaptive specification was selected to account for the uneven spatial distribution of metro stations across Shanghai, where station density is substantially higher in the inner city than in suburban corridors. The optimal bandwidth was determined by minimizing the corrected Akaike Information Criterion (AICc), yielding bandwidths of 191 neighboring features for the B$ \rightarrow $M model and 112 neighboring features for the M$ \rightarrow $B model. These values indicate that each local regression was calibrated using the respective number of nearest observations, ensuring stable local estimates regardless of local data density.

$ \beta $ is the matrix of estimated coefficients:

$ \begin{array}{l} \beta = \begin{bmatrix} \beta_0(u_1,v_1) & \cdots & \beta_k(u_1,v_1) \\ \beta_0(u_2,v_2) & \cdots & \beta_k(u_2,v_2) \\ \vdots & \ddots & \vdots \\ \beta_0(u_n,v_n) & \cdots & \beta_k(u_n,v_n) \end{bmatrix}, \end{array} $

(12)

and $ Y $ is the response vector:

$ \begin{array}{l} Y = \begin{bmatrix} y_1 \\ y_2 \\ \vdots \\ y_n \end{bmatrix}. \end{array} $

(13)

By systematically comparing OLS and GWR, we evaluate both global and spatially varying determinants of transfer ridership.

Conclusions

This study examined the spatially heterogeneous determinants of directional metro–bus transfer rates in Shanghai by integrating 1 month of smart card transactions with station-area built-environment, transit-service, and socio-demographic indicators. By distinguishing B$ \rightarrow $M access transfers from M$ \rightarrow $B egress transfers, the analysis shows that metro–bus integration is not a single uniform process, but a direction-specific and location-dependent travel behavior.

The first major finding is methodological. The significant spatial autocorrelation remaining in the OLS residuals indicates that global coefficients are insufficient for explaining station-level transfer behavior. After applying GWR, residual spatial autocorrelation becomes statistically insignificant, and model fit improves substantially: the adjusted $ R^2 $ increases from 0.374 to 0.490 for B$ \rightarrow $M transfers and from 0.291 to 0.418 for M$ \rightarrow $B transfers. This confirms that a considerable share of the variation unexplained by global models is spatially structured rather than random.

Second, the results reveal a clear core–periphery gradient. In peripheral corridors, metro ridership and distance to the CBD generally strengthen reliance on bus feeders or bus egress services, while longer distance to the nearest bus stop creates stronger deterrent effects. In central areas, many local coefficients are weak or statistically insignificant, reflecting dense bus coverage, higher walkability, and more diversified access or egress options. Thus, the same built environment factor may differ in magnitude and statistical significance across urban contexts, and variables such as CBD distance may even exhibit sign changes between central and peripheral areas.

Third, the two transfer directions are governed by related but asymmetric mechanisms. For B$ \rightarrow $M transfers, metro ridership and CBD distance mainly increase feeder-bus dependence in outer areas, while parking density, road network density, and bus-stop distance generally suppress bus-based metro access where statistically significant. Tourist POI proportion is positively associated with B$ \rightarrow $M transfers in the global model, but its local interpretation is limited because only a small number of stations show statistically significant GWR coefficients. For M$ \rightarrow $B transfers, metro ridership and CBD distance also show stronger positive effects in peripheral corridors, but the egress direction is more sensitive to CBD distance, with significant local coefficients ranging from −0.922 to 1.270.

These findings clarify the contributions of the study. Methodologically, this study combines OLS, residual spatial autocorrelation diagnostics, GWR, and pseudo t-statistic masking to distinguish statistically reliable local effects from insignificant coefficient variation. Empirically, it provides station-level evidence that directional metro–bus transfer rates respond differently to built-environment and service-related factors, demonstrating that access and egress transfers should not be treated as a single aggregated outcome. Practically, the results translate spatially varying coefficients into differentiated planning priorities: improving bus-stop proximity and pedestrian transfer conditions at suburban stations, aligning feeder-bus frequency and capacity with metro ridership in peripheral corridors, and coordinating parking management with bus-service planning in automobile-oriented station areas.

Overall, the evidence suggests that metro–bus integration in megacities should be planned as a direction-specific and station-context-specific system. Uniform citywide interventions are unlikely to address the uneven geography of transfer demand; instead, policies should prioritize peripheral and suburban corridors where dependence on bus-based first- and last-mile connections is strongest.

shu — Limitations, challenges, and future work

Limitations, challenges, and future work

[1]	Venter CJ. 2020. Measuring the quality of the first/last mile connection to public transport. Research in Transportation Economics 83:100949 doi: 10.1016/j.retrec.2020.100949 CrossRef Google Scholar
[2]	Liu W, Pang S, Li W, Han Y. 2025. Assessing the CO₂ emission reduction potential of metro-bus combined travel through interpretable machine learning. Transportmetrica A: Transport Science 1−24 doi: 10.1080/23249935.2025.2472869 CrossRef Google Scholar
[3]	Ewing R, Cervero R. 2001. Travel and the built environment: a synthesis. Transportation Research Record: Journal of the Transportation Research Board 1780:87−114 doi: 10.3141/1780-10 CrossRef Google Scholar
[4]	Seaborn C, Attanucci J, Wilson NHM. 2009. Analyzing multimodal public transport journeys in London with smart card fare payment data. Transportation Research Record: Journal of the Transportation Research Board 2121(1):55−62 doi: 10.3141/2121-06 CrossRef Google Scholar
[5]	Pelletier MP, Trepanier M, Morency C. 2011. Smart card data use in public transit: a literature review. Transportation Research Part C: Emerging Technologies 19(4):557−568 doi: 10.1016/j.trc.2010.12.003 CrossRef Google Scholar
[6]	Ma X, Wu Y, Wang Y, Chen F, Liu J. 2013. Mining smart card data for transit riders' travel patterns. Transportation Research Part C: Emerging Technologies 36:1−12 doi: 10.1016/j.trc.2013.07.010 CrossRef Google Scholar
[7]	He Y, Zhao Y, Tsui KL. 2018. An analysis of factors influencing metro station ridership: insights from Taipei Metro. arXiv Preprint:1904.01280 doi: 10.48550/arXiv.1904.01280 CrossRef Google Scholar
[8]	Gilbert B, Ogburn EL, Datta A. 2023. Consistency of common spatial estimators under spatial confounding. arXiv Preprint:230812181 doi: 10.48550/arXiv.2308.12181 CrossRef Google Scholar
[9]	Fotheringham AS, Brunsdon C, Charlton M. 1998. Geographically weighted regression: a natural evolution of the expansion method for spatial data analysis. Environment and Planning A 30(11):1905−1927 doi: 10.1068/a301905 CrossRef Google Scholar
[10]	Wardman M. 2001. A review of British evidence on time and service quality valuations. Transportation Research Part E: Logistics and Transportation Review 37(2–3):107−128 doi: 10.1016/s1366-5545(00)00012-0 CrossRef Google Scholar
[11]	Zhao D, Wang W, Woodburn A, Ryerson MS. 2017. Isolating high-priority metro and feeder bus transfers using smart card data. Transportation 44(6):1535−1554 doi: 10.1007/s11116-016-9713-7 CrossRef Google Scholar
[12]	Ewing R, Cervero R. 2010. Travel and the built environment: a meta-analysis. Journal of the American Planning Association 76(3):265−294 doi: 10.1080/01944361003766766 CrossRef Google Scholar
[13]	Cervero R, Kockelman K. 1997. Travel demand and the 3Ds: density, diversity, and design. Transportation Research Part D: Transport and Environment 2(3):199−219 doi: 10.1016/S1361-9209(97)00009-6 CrossRef Google Scholar
[14]	Li S, Lyu D, Huang G, Zhang X, Gao F, Chen Y, et al. 2020. Spatially varying impacts of built environment factors on rail transit ridership at station level: a case study in Guangzhou, China. Journal of Transport Geography 82:102631 doi: 10.1016/j.jtrangeo.2019.102631 CrossRef Google Scholar
[15]	Diao M. 2019. Towards sustainable urban transport in Singapore: policy instruments and mobility trends. Transport Policy 81:320−330 doi: 10.1016/j.tranpol.2018.05.005 CrossRef Google Scholar
[16]	Fotheringham AS, Brunsdon C, Charlton M. 2002. Geographically weighted regression: the analysis of spatially varying relationships. US: John Wiley & Sons. 269 pp.
[17]	Cardozo OD, García-Palomares JC, Gutiérrez J. 2012. Application of geographically weighted regression to the direct forecasting of transit ridership at station-level. Applied Geography 34:548−558 doi: 10.1016/j.apgeog.2012.01.005 CrossRef Google Scholar
[18]	Wu J, Xu J, Xu D. 2025. Short-term inbound passenger flow forecasting for urban rail transit based on phase space reconstruction and deep learning. Digital Transportation and Safety 4(3):177−187 doi: 10.48130/dts-0025-0015 CrossRef Google Scholar
[19]	Yuan K, Cui D, Long J. 2023. Bus frequency optimization in a large-scale multi-modal transportation system: integrating 3D-MFD and dynamic traffic assignment. Digital Transportation and Safety 2(4):241−252 doi: 10.48130/DTS-2023-0020 CrossRef Google Scholar
[20]	Sui X, Yan H, Pan S, Li X, Gu X. 2024. Bus system optimization for timetables, routes, charging, and facilities: a summary. Digital Transportation and Safety 4(1):1−9 doi: 10.48130/dts-0024-0024 CrossRef Google Scholar
[21]	Zhou X, Gao Y. 2017. "Shanghai 2035" Green transport development plans and strategies. UTC, China. Available from: www.chinautc.com/upload/fckeditor/%E5%91%A8%E7%BF%94.pdf
[22]	Li W, Chen S, Dong J, Wu J. 2021. Exploring the spatial variations of transfer distances between dockless bike-sharing systems and metros. Journal of Transport Geography 92:103032 doi: 10.1016/j.jtrangeo.2021.103032 CrossRef Google Scholar
[23]	National Bureau of Statistics. 2012. Tabulation on the 2010 Population Census of the People's Republic of China. Beijing: China Statistics Press. www.stats.gov.cn
[24]	OpenStreetMap contributors. 2015. OpenStreetMap data. www.openstreetmap.org (Accessed: 2025-02-16)
[25]	Gaode Maps. 2015. Gaode maps open platform. https://lbs.amap.com/. (Accessed: 2025-02-16)
[26]	Shi Y, Zeng L. 2025. How do built environment characteristics influence metro-bus transfer patterns across metro station types in Shanghai? Journal of Transport Geography 123:104137 doi: 10.1016/j.jtrangeo.2025.104137 CrossRef Google Scholar
[27]	Bagchi M, White PR. 2005. The potential of public transport smart card data. Transport Policy 12(5):464−474 doi: 10.1016/j.tranpol.2005.06.008 CrossRef Google Scholar
[28]	Draper NR, Smith H. 1998. Applied regression analysis, 3rd edition. New York: Wiley. doi: 10.1002/9781118625590
[29]	Han D, Choi CG. 2025. The spatial dynamics of urban vegetation and housing prices: insights from pre- and post-pandemic Chicago using OLS and MGWR models. PLoS One 20(9):e0330932 doi: 10.1371/journal.pone.0330932 CrossRef Google Scholar
[30]	Klar R, Rubensson I. 2024. Spatio-temporal investigation of public transport demand using smart card data. Applied Spatial Analysis and Policy 17(1):241−268 doi: 10.1007/s12061-023-09542-x CrossRef Google Scholar
[31]	He Y, Zhao Y, Tsui KL. 2021. An adapted geographically weighted LASSO (Ada-GWL) model for predicting subway ridership. Transportation 48(3):1185−1216 doi: 10.1007/s11116-020-10091-2 CrossRef Google Scholar
[32]	Yang H, Xu T, Chen D, Yang H, Pu L. 2020. Direct modeling of subway ridership at the station level: a study based on mixed geographically weighted regression. Canadian Journal of Civil Engineering 47(5):534−545 doi: 10.1139/cjce-2018-0727 CrossRef Google Scholar

Category	Variable	Symbol	Min	Mean	Max	SD
Dependent variables
	Bus$ \rightarrow $Metro transfer rate	$ R_{{{\mathrm{B}}\to {\mathrm{M}}}} $ (ratio)	0.000296	0.122	0.696	0.123
	Metro$ \rightarrow $Bus transfer rate	$ R_{{{\mathrm{M}}\to {\mathrm{B}}}} $ (ratio)	0.000132	0.113	0.803	0.118
Transit service and station network
	Metro ridership (monthly)	$ MR $ (#)	1,171	436,923	2,567,207	387,063
	Number of metro lines	$ N_{\text{line}} $ (lines)	1	1.310	4	0.621
	Number of station exits	$ N_{\text{exit}} $ (#)	1	5.143	20	2.826
	Distance to nearest bus stop	$ D_{\text{bus\_nn}} $ (km)	0.003	0.256	1.872	0.326
	Bus stop density	$ D_{\text{bus}} $ (#/km²)	0	7.065	15.924	3.153
Built environment
	Parking density	$ D_{\text{par}} $ (#/km²)	0	5.118	13.113	3.771
	Bicycle lane density	$ D_{\text{bld}} $ (km/km²)	0	0.452	4.023	0.863
	Road network density	$ D_{\text{road}} $ (km/km²)	0.036	12.791	32.331	5.540
	Distance to CBD	$ D_{\text{CBD}} $ (km)	0.552	13.735	65.984	10.324
	Residential POI share	$ P_{\text{res}} $ (ratio)	0.215	0.472	0.798	0.096
	Office POI share	$ P_{\text{off}} $ (ratio)	0	0.101	0.288	0.029
	Tourism POI share	$ P_{\text{tour}} $ (ratio)	0	0.328	0.667	0.060
	Population density (buffer)	$ P_{\text{pop}} $ (pers/km²)	56	18,851.271	44,682	12,126.491
	Employment density (buffer)	$ P_{\text{job}} $ (jobs/km²)	0.080	285.476	905.573	246.295
SD, standard deviation; POI, point of interest; CBD, central business district. #, number (count). Transfer rates and POI shares are expressed as ratios (0–1), not percentages. Metro ridership is the total number of entries recorded at the station over the 1-month study period (April 2015). All built-environment and socio-demographic indicators are measured within a 1 km buffer around each metro station.

Variable	Coefficient	t-statistic	p-value	Significance
Road network density	−0.139395	−2.187172	0.029495	**
Distance to nearest bus stop	−0.074204	−1.864957	0.063174	*
Distance to CBD	0.230025	3.210951	0.001468	***
Population density^a	−0.029297	−0.622181	0.534300	ns
Proportion of tourist POIs	0.318515	2.044617	0.041763	**
Parking density	−0.136822	−3.438710	0.000681	***
Metro ridership	0.154115	3.837729	0.000161	***
*, , and * indicate that the coefficient is statistically significant at the 1%, 5%, and 10% levels, respectively (i.e., p < 0.01, p < 0.05, and p < 0.10); ns indicates not significant (p ≥ 0.10). ^a Population density was retained in the OLS results for transparency, but excluded from the GWR estimation due to its insignificant OLS t-statistic.

Variable	Coefficient	t-statistic	p-value	Significance
Distance to CBD	0.336207	2.407780	0.016641	**
Population density	−0.047775	−0.761604	0.446886	ns
Distance to nearest bus stop	−0.122878	−2.490553	0.013285	**
Parking density	−0.160087	−3.049310	0.002507	***
Metro ridership	0.266319	4.689708	0.000006	***
*, , and * indicate that the coefficient is statistically significant at the 1%, 5%, and 10% levels, respectively (i.e., p < 0.01, p < 0.05, and p < 0.10); ns indicates not significant (p ≥ 0.10).

Model	Moran's $ I $	$ Z $-score	p-value	Significance
OLS (B$ \rightarrow $M)	0.113058	4.451014	< 0.00001	***
OLS (M$ \rightarrow $B)	0.123467	4.858613	< 0.00001	***
*, , and * indicate statistical significance at the 1%, 5%, and 10% levels, respectively; ns indicates not significant (p ≥ 0.10).

{{lists.name}}

Spatial heterogeneity of built environment impacts on metro–bus transfer ridership: evidence from Shanghai, China

Abstract

Rights and permissions

References

About this article

Cite this article

Article Metrics

Access History

Other Articles By Authors

Spatial heterogeneity of built environment impacts on metro–bus transfer ridership: evidence from Shanghai, China

HTML

Empirical evidence on metro–bus transfers

Built-environment effects on metro–bus transfers

Spatial analysis methods and heterogeneity

Study area and data

Data processing

Variable descriptions

Transportation and travel characteristics

Built-environment and land-use characteristics

Dependent variables

Multicollinearity test

Ordinary least squares