1 Introduction

With increasing urbanization, the urban population, number of motor vehicles, and infrastructure are also increasing rapidly. This series of changes has led to the reorganization and transformation of urban spaces, producing large and complex urban systems. As a result of urban space changes, the built environment is now an important carrier affecting the population’s travel behavior. Therefore, many scholars have advocated changing the built environment through land-use planning and transportation policies to guide travel behaviors, improve the urban traffic environment, and alleviate traffic congestion [1]. Taxis are one of the most critical transportation modes for urban residents and play an indispensable role in travel behavior owing to their flexibility and convenience. The built environment also influences them. In recent years, scholars have also shifted their research focus to the relationship between the built environment and taxi travel behavior. These studies have focused on the mechanism by which the built environment influences behavioral factors, such as the taxi travel distance, travel intensity, and travel time [2–4]. The results have confirmed that the built environment influences taxi travel behavior to varying extents.

The demand for taxi travel is an essential behavioral factor in taxi travel behavior that is coupled with the non-static and varying qualities of built environment elements. As a result, the uneven and complex distribution of built environment elements in urban spaces directly affects the spatiotemporal distribution of residents’ taxi travel demand. In addition, drivers have an inadequate grasp of high-demand areas, which often leads to difficulties in taking taxis and long wait times, which are not conducive to urban development. Therefore, some scholars have begun to pay attention to the influence of built environment variables such as population, land use, and road design on taxi travel demand [5–7]. These studies aim to deeply understand the connection between the built environment and taxi travel demand. Thus, these studies can provide a scientific basis for taxi demand prediction and scheduling. In addition, these studies are important for management services in the taxi industry and for promoting the virtuous cycle of urban transportation.

However, when constructing a model of the influence of the built environment on taxi travel demand, few studies have considered (or have considered insufficiently) the differences in the results caused by the modifiable areal unit problem (MAUP) when aggregating data. The MAUP refers to a phenomenon in which analysis results will vary depending on the definition of the primary research units (as first proposed by Openshaw in 1984 [8]). These differences in definitions mainly involve scale and zoning effects [9, 10]. The scale effect refers to changes in the statistical results of aggregated spatial data that occur when the size of the research unit changes. The zoning effect refers to differences in the statistical results caused by the aggregation of spatial data using different zoning schemes for a fixed research unit size. Some studies have also shown that the MAUP is an essential fundamental issue in many traffic problem studies [11]. For example, Zhao et al. [12] studied the impact of the built environment on online car-hailing travel intensity and found that as the analysis scale increased, the effect of proximity to public transportation on online car-hailing travel intensity increased; however, its effect was not significant at the remaining analysis scales. Therefore, the MAUP cannot be ignored when constructing a model of the influence of built environment on taxi travel demand, and a suitable spatial analysis unit is a prerequisite for such research.

This paper aims to explore the relationship between the built environment and taxi travel demand. Based on point of interest (POI), population, and road network data, the built environment variables were reasonably quantified. The optimal discretization of built environment variables under the MAUP effect was determined through the optimal parameters-based geographical detector (OPGD) model. Then, according to the optimal parameter combination results, the relationship between the built environment and taxi travel demand was analyzed under the MAUP effect, and the optimal spatial analysis unit for the influence of the built environment on taxi travel demand was determined. On this basis, while considering the nonlinear and linear relationships between the built environment and taxi travel demand, the influence of the built environment on taxi travel demand was explored by cross-using the OPGD and multiscale geographically weighted regression (MGWR) models.

The potential academic contributions of this paper are as follows. (1) The optimal discrete parameter combination of the built environment variables is determined based on the OPGD model. This solves the lack of accurate quantitative evaluations of discretization methods and classification numbers when discretizing continuous variables in the OPGD model. (2) At the same time, the optimal scale and zoning scheme for aggregating the built environment and taxi travel demand data are determined using the OPGD model. The selection of spatial unit granularity for this type of problem is thus realized. (3) A “5D” dimension index system of the built environment is constructed to explore its influence on taxi travel demand. This provides a scientific basis for quantifying each index. (4) The OPGD and the MGWR models are cross-used. This method can not only reveal the degree of influence of major built environment variables on taxi travel demand from a global perspective but also reflects the spatial heterogeneity of the influence of the built environment on taxi travel demand from a local perspective. This provides a method for studying the influence of the built environment on taxi travel demand. The reasonable determination of the spatial analysis unit in this study can facilitate a better understanding of the influence of the built environment on taxi travel demand and improve the reliability of the results. At the same time, it also provides a meaningful method for reasonably integrating environmental and transportation policies to guide taxi travel and alleviate problems such as taxi-taking difficulties.

The remainder of this paper is organized as follows. Section 2 summarizes the related work. Section 3 introduces the main research methods used in this paper, including the method for determining the optimal spatial analysis unit, OPGD model, methods for testing the spatial autocorrelation and multicollinearity, and the MGWR model. Section 4 outlines the study area and data, including the study area and division, data sources, and processing. Section 5 presents an analysis and discussion of the results. Section 6 summarizes the principal conclusions, limitations, and future work.

2 Literature review

The determination of spatial units is the premise and foundation of spatial analysis. The size and shape of the spatial analysis units determine the amount of data to be included, resulting in differences in the data analysis results [13, 14]. Previously, when analyzing the influence of the built environment on taxi travel demand, the research area was often divided into a single space type using the crowd sampling method of “people for their use.” For example, the study area would be divided into a traffic analysis zone (TAZ) [15], census area [16], and grid [17], neglecting the impact of the MAUP on the data aggregation and modeling results. However, scholars have recently begun to focus on this issue. For example, Wang et al. [19] divided a study area into different grid scales and calculated the 90% quantile of built environment variables under different grids using the OPGD model. The 90% quantiles of the built environment variables were found to differ at different grid scales. The scale corresponding to the maximum value was selected as the optimal scale, and the impact of the built environment on ride-hailing travel demand was discussed. However, this study only considered the scale effect in the MAUP and ignored the other substantial zoning effect. Moreover, Cheng et al. [18] divided a study area into two spatial units, TAZ and 1 km × 1 km grids, to study the impact of the built environment, population distribution, and road network structure on taxi travel demand at night. The results showed that the R-squared value of the TAZ-based spatial Durbin model (SDM) was slightly lower than that of the grid-based SDM model; however, the Log-likelihood, Akaike information criterion (AIC), and Bayesian information criterion (BIC) of the former were significantly lower than those of the latter. Finally, the TAZ was selected as the spatial analysis unit. However, this study considered only the zoning effect and ignored the scale effect. Of course, some scholars have considered both the scale effect and zoning effect in the MAUP. Wang et al. [19] divided a study area into TAZ, Thiessen polygons, community units, and 300-1000m (100m interval) grids to explore the impact of the built environment on network passenger volume. The results showed that the road density exhibited a significant spatial clustering distribution in the community units, TAZ, and Thiessen polygons. In contrast, the bus station density exhibited a significant spatial clustering distribution under the 400m grid, 1000m grid, and community units. Gao et al. [20] divided the study area into three partition types: administrative partition, hexagonal grids, and square grids. Spatial units with basically the same corresponding scale were generated according to the street, community, and traffic district scales in the administrative partition, resulting in a total of nine partition schemes. The results showed that the influence of commercial land on taxi commuting demand decreased with an increase in spatial scale. Other studies that considered both the scale and zoning effects of the MAUP have explored the relationship between the built environment and shared bike travel [21] and the built environment and traffic system state [22]. Therefore, the MAUP should be fully considered when investigating the relationship between the built environment and taxi travel demand, which will inevitably affect the modeling results.

Built environment factors that affect taxi travel demand are complex. The built environment was initially constituted by the “3D” dimensions proposed by Cervero and Kockelman [23]: density (such as population density and the density of POIs), diversity (such as the mixed degree of land use), and design (such as the road network density and intersection density). Subsequently, this framework was expanded to the “5D” dimensions [24], which added the characteristics of destination accessibility (such as the distance to the city center and distance to CBD) and distance to transit (such as the distances to bus and subway stations). Currently, this framework has developed into the “7D” dimensions [25], adding demand management and demographic characteristics. However, while these two indicators are closely related to the built environment, they do not directly describe it [26]. The different dimensions of the built environment have led scholars to adopt different built environment variables in their research. For example, based on a review of 29 studies on public transportation and the built environment, Liu et al. [27] found that although many studies were consistent with the ‘5Ds,’ most did not include comprehensive coverage of all domains of the ‘5Ds;’ moreover, some variables used in these studies were inconsistent with the ‘5Ds.’ Chen et al. [28] quantified the built environment using various land-use variables to investigate the relationship between the built environment and taxi travel demand. However, the authors did not divide the dimensions of these variables. Zhu et al. [29] quantified the built environment in terms of road network density, bus coverage, subway coverage, and other variables. They did not explicitly propose the dimension concept of the built environment. Ni and Chen [30] also reported similar situation: only the variables of the built environment were quantified, and these variables were not divided into relevant dimensions. In addition, Xie et al. [31] divided built environment variables into four dimensions: density, design, diversity, and destination accessibility. However, these were not considered comprehensively. In addition, although Zhu et al. [32] mentioned the concept of 5Ds, only built environment variables from the four dimensions of density, design, distance to transit, and diversity were selected. These studies indicate that built environment factors are complex and variable. The mechanism by which the built environment influences taxi travel demand is also complex and needs to be strengthened and further explored.

When exploring the relationship between the built environment and taxi travel demand, some scholars have typically assumed a linear relationship between the built environment and taxi travel demand, and impact models considering spatial correlation and heterogeneity have been established, such as the SDM [33], spatial error model (SEM) [1], and geographically weighted regression (GWR) model [34]. Because all of the independent variables in the spatially heterogeneous GWR model have the same bandwidth, and the changes in the relationships between independent and dependent variables at different spatial action scales are ignored, the MGWR model has been developed. However, studies using this model to explore the impact of the built environment on taxi travel demand are rare [32]. This type of research can not only fit the relationship between the built environment and taxi travel demand but also analyze the spatial heterogeneity of the built environment effect at small scales. However, owing to the complexity and changeability of travel purposes and the limitations of urban activity spaces, taxi travel demand sometimes varies linearly with the built environment. Therefore, some scholars have proposed that nonlinear relationships should also be considered [35]. This research is generally conducted by introducing machine learning methods (such as the gradient boosting decision tree (GBDT) [36] and random forest (RF) models [37]). However, these studies have largely focused on the nonlinear relationship between the built environment and shared-bike usage [38–42], bus travel [43, 44], carpool usage [45, 46], and trail transit travel [47]. Research on the nonlinear relationship between the built environment and taxi travel demand remains lacking. Simultaneously considering the nonlinear and linear relationships between the built environment and taxi travel demand can allow for a more accurate exploration of the correlation between the two and determination of the hidden complex relationship.

In summary, previous studies have rarely considered the MAUP in the process of data aggregation and modeling when discussing the mechanism by which the built environment influences taxi travel demand. When constructing a built environment index system, the influence of the “5D” dimensions of built environment variables on taxi travel demand has rarely been considered. Regarding research methods, the nonlinear and linear relationships between the built environment and taxi travel demand have rarely been recognized. To solve these problems, this paper considered both the scale and zoning effects in the MAUP and divided the study area into multiple spatial units. Based on multi-source data and the characteristics of each dimension of the built environment, a “5D” dimension index system of built environment variables was constructed. Referring to the method reported by Wang et al. for selecting the optimal spatial analysis unit using the OPGD model [48], a more reasonable spatial unit division scheme was determined. The nonlinear and linear relationships between the built environment and taxi travel demand were thoroughly discussed by cross-using the OPGD and MGWR models [32]. Based on these existing methods, a more detailed and comprehensive analysis was realized than in previous studies.

3 Methods

The MAUP and lack of nonlinear and linear relationships were considered to analyze the influence of the built environment on taxi travel demand. First, optimal discretization of the built environment variables under each scale and zoning was performed using the OPGD model. We then calculated the q-value of each built environment variable that affects taxi travel demand as well as the q-value ranking, 90% quantile of the q-values, and their growth rates to select the optimal spatial analysis unit. Furthermore, based on the optimal spatial analysis unit, the nonlinear relationship between the built environment and taxi travel demand was characterized using the factor detection module of the OPGD model. Finally, through spatial autocorrelation and multicollinearity tests of the built environment and taxi travel demand variables, the linear relationship between the built environment and taxi travel demand was characterized using the MGWR model. The overall research framework of the paper is shown in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Research framework.

https://doi.org/10.1371/journal.pone.0292363.g001

4 Study area and data

4.1 Study area and division

As the most populous megacity in northwest China, Xi’an has a longitude range of [107.4, 109.49] and a latitude range of [33.42, 34.45] [55]. At the end of 2022, Xi’an covered an area of 10,108 square kilometers [55] and had a permanent population of 12.9959 million [56]. The city is well-equipped with public transport, and residents’ taxi travel demand is mainly concentrated in the central urban area of Xi’an. Therefore, this study considered the central urban area of Xi’an as the study area (Fig 2(A)), including the districts of Weiyang, Lianhu, Xincheng, Beilin, Yanta, and Baqiao (Fig 2(B)). The study area is approximately 831.87 square kilometers [57].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Study area.

(a) Study area location; (b) overview of the study area.

https://doi.org/10.1371/journal.pone.0292363.g002

ArcGIS software was used to divide the study area into grids according to its scope. This classification method was selected because it has strong operability, a wide range of applications, and is suitable for intensive data research. To explore the scale effect of the MAUP, it was necessary to choose the side length of the grid. Considering the scope of the study area and spatial resolution of the multisource data, 100m was chosen as the minimum side length of the grid. A review of the literature on the influence of the built environment on taxi travel demand revealed that most previous studies used 500m or 1000m as the side length of the grid [28, 29, 58, 59]. Therefore, 1000m was selected as the maximum side length, with a side length interval of 100m. To explore the zoning effect of the MAUP, two common grid types were used: regular hexagonal grids and square grids.

The specific operation method created regular hexagonal grids with ten scales and side lengths ranging from 100m to 1000m (with an interval of 100m). The areas of the regular hexagonal grids were then fixed at each scale, and corresponding square grids were generated (a total of 10). Ultimately, 20 types of research units were generated. Here, only the 300m, 600m, and 900m scales are taken as examples, and the study area is divided into regular hexagonal and square grids, as shown in Fig 3.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Study area division.

(a) 300m regular hexagonal grid; (b) 600m regular hexagonal grid; (c) 900m regular hexagonal grid; (d) 300m square grid; (e) 600m square grid; (f) 900m square grid.

https://doi.org/10.1371/journal.pone.0292363.g003

4.2 Data sources and processing

4.2.1 Taxi travel demand data.

The taxi travel demand data used in this paper were obtained from taxi GPS trajectory data provided by the taxi management office of Xi’an city. The selected time range was 30 days, from March 1, 2019 to March 30, 2019. Owing to the extensive period of the selected data, the data volume was also large. Moreover, there were no special holidays during this period; therefore, the data was assumed to be objective and representative. The original trajectory data included the Vehicle ID, Intime, Time, Longitude, and Latitude. An example of the taxi trajectory data structure is presented in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. An instance of taxi trajectory data structure.

https://doi.org/10.1371/journal.pone.0292363.t001

The total original trajectory data obtained in this paper was approximately 11.6 million. Python software was then used to remove the driving records with outliers and those occurring outside the central urban area of Xi’an. Data corresponding to pick-up points (the points at which the vehicle state changed from “4” to “5”) and drop-off points (the points at which the vehicle state changed from “5” to “4”) were then extracted as the taxi travel demand data for the central urban area of Xi’an. Finally, a total of 10.3 million effective taxi trajectory data were obtained. In other words, 88.79% of the valid data was obtained. It was considered that displaying all of the taxi pick-up and drop-off points for 30 days was prone to data being tightly distributed and unattractive. Therefore, the simple random sampling method was adopted to display the density distribution based on 5% of the taxi pick-up and drop-off points in the central urban area of Xi’an after cleaning and treatment (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Density of taxi pick-up and drop-off points.

(a) Pick-up points; (b) drop-off points.

https://doi.org/10.1371/journal.pone.0292363.g004

4.2.2 Built environment data.

Based on the “5D” dimensions of built environment characteristics, this paper selected built environment variables from the five dimensions of density, design, diversity, destination accessibility, and distance to transit. There were three primary sources of data. First, population data were obtained from the Seventh National Population Census of Xi’an City, China (http://tjj.xa.gov.cn). Data were collected at the street scale. Then, it was processed on a grid scale as the built environment variable for the density dimension according to the research scope. Second, POI data representing various facilities were obtained from the 2019 Xi’an data on the Amap open platform (https://lbs.amap.com/). Reptiles were collected to obtain 13 first-level POI categories, including shopping service POI, catering service POI, accommodation services POI, etc., which were used to quantify the built environment variables of the density dimension. Based on the POI data, the mixed degree of urban functions was calculated as the built environment variable for the diversity dimension. In addition, bus stop POI and subway station POI were extracted to quantify the built environment variables in the distance-to-transit dimension. In addition, road data was obtained from Xi’an in 2019 on OpenStreetMap (the OSM data download address in China is https://download.geofabrik.de/). According to the study area and classification of the road grade, roads were quantified into four levels: primary roads, secondary roads, tertiary roads, and urban express roads, which were used to represent the built environment variables of the design dimension. Finally, the Euclidean distance from each grid centroid to the city center was used to quantify the built environment variables in the destination accessibility dimension. The main variables for each dimension and their calculation methods are shown in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. The “5D” dimension of built environment variables.

https://doi.org/10.1371/journal.pone.0292363.t002

5 Results and discussion

5.1 Determination result of the optimal spatial analysis unit

According to the above description of the OPGD model, this paper discretized the built environment variables at different scales and zones using five methods: equal breaks, natural breaks, quantile breaks, geometric breaks, and SD breaks, and the classification levels were divided into 3–10 categories. The optimal discretization of built environment variables was obtained using R Programming Language (see Supporting information for the results of the optimal discretization of the built environment variables).

Furthermore, the q-values of the built environment variables were calculated according to the optimal discretization results for the built environment variables at different scales and zones. The q-values of each built environment variable were then ranked at different scales and zones to represent the influence stability of each variable (the results are shown in Figs 5 and 6). A smaller fluctuation in the ranking of the q-values for each variable with the change in scale indicates that the influence is more stable, and the scale of the corresponding spatial analysis unit is more appropriate.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. q-values of built environment variables at different scales and their ranking (regular hexagonal grid).

https://doi.org/10.1371/journal.pone.0292363.g005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. q-values of built environment variables at different scales and their ranking (square grid).

https://doi.org/10.1371/journal.pone.0292363.g006

As shown in Figs 5 and 6, regardless of whether the regular hexagonal grid or square grid, the influence of most built environment variables increased with an increase in scale. This phenomenon was particularly evident when the scale was less than 900m. In addition, regardless of the regular hexagonal or square grid used, the influence of the ranking of built environment variables differed at different scales. This indicates that different scales had different effects on taxi travel demand. For instance, the rankings of the public facility POI density (Pf_d), urban express road network density (Uern_d), distance to the city center (Dcc), and subway station POI density (Ss_d) were relatively stable across all scales, indicating that they had low scale sensitivity for the influence of taxi travel demand. However, the q-values of the other built environment variables and their rankings were sensitive to scale. Urban planners and managers should thus pay attention to the spatial scale when considering the influence of these built environment variables. When the scale was greater than 900m, the fluctuations in most of the built environment variables decreased. Based on this analysis, the 900m scale was selected as the initial candidate for the optimal spatial unit for the regular hexagonal and square grids.

Furthermore, the 90% quantiles of the q-values and their growth rates were calculated for the built environment variables at different scales and zones. The results are shown in Fig 7.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. The 90% quantile of the q-values of built environment variables and their growth rates at each scale.

(a) Regular hexagonal grid; (b) square grid.

https://doi.org/10.1371/journal.pone.0292363.g007

Fig 7 shows that for both the regular hexagonal grid and the square grid, the 90% quantile of the q-values of the built environment variables gradually increased with increasing scale. However, when the scale exceeded 900m, the 90% quantile of q-values decreases. That is, when the scale was 900m, the 90% quantile of the q-values of the built environment variables reached its maximum value. The growth rate of the 90% quantile of the q-values showed a trend of decreasing, followed by a slight increase and then a further decrease. The growth rate of the 90% quantile of the q-values of built environment variables was relatively large before the 900m scale. When the scale exceeded 900m, the growth rate decreased. Therefore, it is possible to select the 900m scale as the optimal spatial analysis unit for regular hexagonal and square grids. Finally, the q-values of each built environment variable corresponding to the regular hexagonal grid and square grid at the 900m scale were compared. The results are presented in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. q-values of each built environment variable at 900m scale.

https://doi.org/10.1371/journal.pone.0292363.t003

As indicated in Table 3, at the 900m scale, the q-values of each built environment variable based on the regular hexagonal grid were generally larger than those based on the square grid. This indicates that the model performance of the regular hexagonal grid division scheme was slightly better than that of the square grid division scheme. Correlation analysis has also shown that using a regular hexagonal grid as the spatial analysis unit can reduce the visual field deviation compared with a square grid [60].

In summary, based on the scope of the study area, the optimal spatial analysis unit was selected as a regular hexagonal grid with a side length of 900m.

5.2 Result of factor detection

This paper analyzed the influence of the built environment on the total demand for taxi pick-up and drop-off on weekdays and weekends and obtained results for the difference between weekdays and weekends. Prior to the analysis, an optimal discrete selection of built environment variables corresponding to taxi travel demand variables was performed. The results are summarized in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. The optimal discretization results of built environment variables on weekdays and weekends.

https://doi.org/10.1371/journal.pone.0292363.t004

Furthermore, according to the optimal discretization results for the built environment variables, the factor detector was used to explore the independent influence of the built environment variables on the total demand for taxi pick-up and drop-off on weekdays and weekends. The results are provided in Table 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. The independent effect of built environment variables on the total demand for taxi pick-up and drop-off.

https://doi.org/10.1371/journal.pone.0292363.t005

Table 5 indicates that on both weekdays and weekends, only the Uern_d variable in the design dimension did not pass the significance test, i.e., the P-value was more significant than 5%. All of the other built environment variables passed the significance test and thus had statistical significance. In general, the built environment variables were ranked as destination accessibility > density > diversity > distance to transit > design according to the “5D” dimensions and influence. Specifically, the top variables of the built environment in terms of influence were Dcc, Commercial residence POI density (Cr_d), Transportation facility POI density (Tf_d), Population density (P_d), Catering service POI density (Cs_d), Life service POI density (Ls_d), and Financial and insurance service POI density (Fis_d). At the bottom of the rankings were Uern_d and Ss_d. Moreover, the influence of built environment variables was generally higher on weekdays than on weekends. Only the Tourist attraction POI density (Ta_d), Government agency and social organization POI density (Gaso_d), Dcc, and Ss_d had a lower influence on weekdays than on weekends. In particular, the influence of Ta_d on weekdays was lower than that on weekends compared with the other variables. In contrast, the influence of the Public facility POI density (Pf_d) and Dcc was similar on weekdays and weekends.

In summary, the influence of built environment variables on the total demand for taxi pick-up and drop-off on weekdays and weekends reveals the following: (1) In the density dimension, the influence of Ta_d on weekends was higher than that on weekdays, which confirms the weekend effect of tourism. (2) In the diversity dimension, the influence of the Mixed degree of urban functions (Mduf) was moderate in the overall built environment variable. (3) In the design dimension, the influence of the Secondary road network density (Srn_d) was stronger than those of the Primary road network density (Prn_d), Tertiary road network density (Trn_d), and Uern_d. This indicates that Srn_d in road design is the main reason for the difference in the distribution of the total demand for taxi pick-up and drop-off. In addition, the influence of Uern_d was the lowest among all of the built environment variables. (4) In the destination accessibility dimension, Dcc had the greatest influence on the overall built environment variables. This shows that different urban locations can lead to large differences in the total demand for taxi pick-up and drop-off. (5) In the distance to transit dimension, the influence of the Bus stop POI density (Bs_d) was much higher than that of Subway station POI density (Ss_d). This indicates that the Bs_d distribution is the most important transportation factor affecting the total demand for taxi pick-up and drop-off.

5.3 Results of spatial autocorrelation and multicollinearity test

5.3.1 Results of the spatial autocorrelation test.

Moran’s I was used to measure the global spatial autocorrelation of all taxi travel demand and built environment variables, as listed in Table 6.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. The global Moran’s I of taxi travel demand variable and built environment variable.

https://doi.org/10.1371/journal.pone.0292363.t006

As presented in Table 6, the global Moran’s I of all taxi travel demand variables and built environment variables were greater than zero, and the P-value was less than 0.05. This indicates a significant positive spatial correlation at the 95% confidence level. In addition, the Z-scores were all greater than 1.96, indicating that all variables had spatial agglomeration characteristics. These are consistent with the preconditions for the MGWR.

5.3.2 Results of the multicollinearity test.

The VIF was used to conduct a multicollinearity test on the built environment variables; the results are provided in Table 7.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Results of multicollinearity test of built environment variables on weekdays and weekends.

https://doi.org/10.1371/journal.pone.0292363.t007

Table 7 reveals that only the VIF of Cs_d was greater than 10. This indicates that there is multicollinearity among the built environment variables. Furthermore, we used SPSS software to screen the important variables with a significant influence using a stepwise regression method. The variable screening results are shown in Table 8.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Stepwise regression screening results for significant variables on weekdays and weekends.

https://doi.org/10.1371/journal.pone.0292363.t008

The results in Table 8 demonstrate that after stepwise regression, the selected built environment variables did not have multicollinearity and were all significant. Thus, these variables could be used as the input variables for the MGWR model.

5.4 Results of spatial heterogeneity analysis

5.4.1 Results of model evaluation.

We established MGWR models based on the selected built environment and taxi travel demand variables; GWR Models were simultaneously constructed as controls. A comparison of the results is presented in Table 9.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Comparison results of the GWR and the MGWR models on weekdays and weekends.

https://doi.org/10.1371/journal.pone.0292363.t009

Table 9 reveals that the R² values and values of the MGWR model were higher than those of the GWR model for both weekdays and weekends. Moreover, the values of AIC, AICc, and RSS were also lower than those of the GWR model. Thus, it can be concluded that the MGWR model has a better fitting effect and explanatory power than the GWR model. In addition, the difference between the AICc values in the MGWR model and the GWR model was greater than three, indicating that the MGWR model is superior to the GWR model. A bandwidth comparison between the GWR and MGWR models is presented in Table 10.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Comparison of bandwidth between the GWR and MGWR models on weekdays and weekdays.

https://doi.org/10.1371/journal.pone.0292363.t010

Table 10 indicates that the GWR model and MGWR model have vastly different results in terms of bandwidth. The bandwidths of each variable in the GWR model are fixed, whereas those in the MGWR model vary significantly. In other words, the action scales of different built environment variables are quite different. The P_d, Srn_d, Medical service POI density (Ms_d), and Cs_d variables have the same action scales of 1700, 51, 838, and 1661, respectively, accounting for 99.94%, 3.00%, 49.27%, and 97.65% of the total samples, respectively, for both weekdays and weekends. The results show that P_d, Ms_d, and Cs_d have small spatial heterogeneity. In contrast, Srn_d has large spatial heterogeneity, and the effects of these variables are the same on weekdays and weekends. In addition, the action scales of Tf_d on weekdays and weekends are 1309 and 1700, respectively, accounting for 76.95% and 99.94% of the total samples, respectively. The results show little spatial heterogeneity and a certain difference between weekdays and weekends. The action scales of Dcc on weekdays and weekends are 896 and 1179, respectively, accounting for 52.67% and 69.31% of the total samples, respectively, similar to the results for Tf_d.

In summary, the MGWR model was superior to the GWR model in terms of fit. Moreover, the spatial scale of the relationship between different built environment variables was considered. The MGWR model could fit the taxi travel demand well and further reveal the spatial differences in the influence of the built environment on taxi travel demand. Therefore, the MGWR model was used to elaborate the spatial heterogeneity of the degree of impact of the built environment on taxi travel demand.

5.4.2 Spatial heterogeneity analysis of regression coefficients.

The local regression coefficients were visualized to demonstrate the influence of built environment variables on taxi travel demand in different spatial positions on weekdays and weekends (Figs 8 and 9). Positive and negative symbols indicate a positive or negative influence of each built environment variable on taxi travel demand. The absolute values indicate the degree of influence.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Spatial distribution of local regression coefficients of built environment variables on weekdays.

(a) P_d; (b) Cs_d; (c) Tf_d; (d) Ms_d; (e) Srn_d; (f) Dcc.

https://doi.org/10.1371/journal.pone.0292363.g008

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Spatial distribution of local regression coefficients of built environment variables on weekends.

(a) P_d; (b) Cs_d; (c) Tf_d; (d) Ms_d; (e) Srn_d; (f) Dcc.

https://doi.org/10.1371/journal.pone.0292363.g009

Density dimension. The regression coefficients for P_d on weekdays and weekends are both negative (Figs 8(A) and 9(A)), indicating that P_d has a negative inhibitory effect on the total demand for taxi pick-up and drop-off. The overall negative inhibitory effect gradually increases from the southwest to the northeast. The areas where P_d has a large negative inhibitory effect on the total demand for taxi pick-up and drop-off on weekdays and weekends are located in the northeast and east of the Baqiao District. Because of the low level of economic development and resident income in these areas, and because most of the area is under development and construction, residents are more willing to choose buses and subways as their main means of transportation. Therefore, compared with the less strongly negative inhibitory effect in the southeast of Yanta District, which is relatively developed and has a high resident income, P_d has a greater influence on taxi travel demand.

The regression coefficients of Cs_d on weekdays and weekends are both positive and negative (Figs 8(B) and 9(B)). This indicates that in some regions, Cs_d has a positive promoting effect on the total demand for taxi pick-up and drop-off. In contrast, in some regions, Cs_d has a negative inhibitory effect on the total demand for taxi pick-up and drop-off. The areas where Cs_d has a significant negative inhibitory effect on the total demand for taxi pick-up and drop-off on weekdays and weekends are all located in the east of the Baqiao District. Cs_d on weekdays has a significant positive influence on the total demand for taxi pick-up and drop-off, particularly in most areas of the Lianhu District and at the junction of the Lianhu District, Xincheng District, and Beilin District. The areas with a significant positive promotion effect on weekends include most areas of the Xincheng District and the southeast of the Weiyang District. This change indicates a difference between weekdays and weekends in areas where Cs_d has a greater positive influence on the total demand for taxi pick-up and drop-off.

The regression coefficients of Tf_d on weekdays and weekends are both positive and negative (Figs 8(C) and 9(C)), and the overall spatial distribution of the regression coefficients is not significantly different. The overall trend is from east to west; in this direction, the negative inhibitory effect gradually weakens and the positive promoting effect gradually increases. The negative inhibitory effect of Tf_d on the total demand for taxi pick-up and drop-off on weekdays and weekends is strongest in the east of the Baqiao District. The positive promoting effect is strongest in the northwest of the Weiyang District. Closer to the central urban area, from south to north, both the negative inhibitory and positive promotion effects are weakened. Because many transportation facilities such as subway stations and bus stops are available for residents to travel along this section, the increase or decrease in transportation facilities has less influence on the total demand for taxi pick-up and drop-off compared with other areas.

The regression coefficients of Ms_d on weekdays and weekends are both positive and negative (Figs 8(D) and 9(D)), and the overall spatial distribution of the regression coefficients is not significantly different. The areas with a significant negative inhibitory effect of Ms_d on the total demand for taxi pick-up and drop-off on weekdays and weekends are located in the east, northeast, and southeast of the Baqiao District. Residents in these areas have relatively low incomes and relatively few medical service resources, and they usually choose buses, subways, or other modes of travel. Therefore, increased medical resources have reduced the demand for taxi pick-up and drop-off among nearby residents. Residents who are further away usually choose to seek medical treatment nearby, and thus will not take taxis. In addition, the positive promotion effect is greater in central Weiyang District. Owing to the relatively abundant medical resources in the region, people often choose convenient and cost-effective transportation such as taxis and private cars when going to buy medicine or see a doctor; a reasonable addition of medical services in this region can increase the total demand for taxi pick-up and drop-off.

Design dimension. The regression coefficients of Srn_d on weekdays and weekends are both positive and negative (Figs 8(E) and 9(E)). The overall spatial distributions of the regression coefficients do not differ significantly. Srn_d on weekdays and weekends has a significant positive impact on the total demand for taxi pick-up and drop-off in south-central of Yanta District, most areas of the Beilin District, most areas of the Xincheng District, and south-central and north-central of Weiyang District. These areas have a high population density, and a moderate increase in secondary roads will increase residents’ demand for taxi pick-up and drop-off. The negative inhibitory effect is greater along the West Third Ring Road, Taibai Interchange, and Jinhua South Road. These roads are mostly connected to urban elevated expressways, and subway stations have less radiation; thus, residents are more inclined to travel by private cars.

Destination accessibility dimension. The regression coefficients of Dcc on both weekdays and weekends are negative (Figs 8(F) and 9(F)). This shows that with an increase in Dcc, the total demand for taxi pick-up and drop-off decreases. Overall, the magnitude of this decrease gradually increases from east to west and north to south. This is because the southwest part of the central urban area (southwest of Yanta District) has better economic development than the eastern region (northeast, east, and southeast of Baqiao District) and has more shopping centers and office spaces. Residents living in the southwest can choose nearby areas to meet their daily living and work needs. Therefore, Dcc has a greater impact on the choice of taxi travel.

6 Conclusions

It is important to investigate the influence of the built environment on taxi travel demand. Previously, little attention has been paid to the MAUP in data aggregation, and the impact of the “5D” dimensions of built environment variables on taxi travel demand was considered less commonly. At the same time, there was a lack of consideration for the nonlinear and linear relationships between the two. This study analyzed the central urban area of Xi’an as an example. A “5D” dimension index system of built environment variables was constructed using multi-source data, which provided a scientific basis for quantifying each index. Based on the OPGD model, an optimal spatial analysis unit was selected to study the effects of the built environment on taxi travel demand, which enriched the quantitative basis for selecting the research units. The OPGD and MGWR models were then cross-used to reveal the influence of the built environment on taxi travel demand from global and local perspectives, thereby providing a method for studying this problem. The main conclusions are summarized as follows:

1. It is necessary to consider the MAUP when discussing the impact of the built environment on taxi travel demand. It is preferable to divide the space units into regular hexagonal grids rather than into square grids. A regular hexagonal grid with a side length of 900m yielded the best effect and was selected as the optimal spatial analysis unit.
2. The factor detection results show that on weekdays or weekends, Dcc, Cr_d, Tf_d, P_d, Cs_d, Ls_d, and Fis_d have a greater influence on the taxi travel demand.
3. Comparing the results of the GWR and MGWR models, the MGWR model considering the spatial heterogeneity scale difference has a better fitting effect. On both weekdays and weekends, the spatial action scale of the Srn_d variable is small. This shows that taxi travel is susceptible to spatial location, and the influence of spatial location on taxi travel should be prioritized in urban construction.
4. The spatial heterogeneity of the regression coefficient shows that Srn_d, Dcc, and Ms_d have a significant influence on local taxi travel demand. Cs_d is significantly influenced by weekdays and weekends, and there is an apparent spatial difference.

These results provide an important reference for urban planners and traffic managers. For urban planners, the optimal spatial analysis unit for taxi travel demand is determined in combination with the built environment, and a demand forecast is conducted on this basis. Centralized passenger carrying points can be set in areas with high demand forecasts. When adjusting the demand for taxis in the entire region, urban planners and traffic managers can prioritize updating and adjusting the built environment variables with greater influence, such as Cr_d, Tf_d, P_d, and Dcc. When making local adjustments to taxi demand, priority can be given to optimizing Srn_d, Dcc, and Ms_d. Improvement strategies can be proposed according to local conditions, optimizing the urban layout, guiding taxi travel behavior, and promoting the virtuous cycle of urban traffic development. For taxi industry management, an optimal spatial analysis unit can provide a reference for traffic zoning in taxi demand forecasting. At the same time, it can guide drivers to areas with a high demand for transportation and promote the scheduling and coordinated development of taxi vehicles.

7 Limitations and future work

The above conclusions are expected to help decision-makers formulate targeted urban planning and traffic management strategies. Nevertheless, there are some limitations of this study, and the next steps need to be improved and studied.

1. The variables of an urban built environment are complex and changeable. Although the main elements of the “5D” dimensions are included in the selection of built environment variables, common indicators such as employment density and number of intersections have not been taken into account. In the future, we plan to incorporate these indicators to reveal their relationship with taxi travel demand.
2. The conclusion of this study regarding the selection of an optimal spatial analysis unit is based only on the optimal solution for a specific research area and data set. When changing the study area and data set, the spatial analysis unit must be redefined. However, the determination method is equally applicable to other areas. In the future, we will focus on the calculation of new data in a new study area under various partition schemes, and the effect of the MAUP also needs to be further researched.
3. Although this study discussed the influence of the built environment on taxi travel demand, it did not provide specific suggestions for vehicle scheduling. In the future, a taxi travel demand prediction model will be constructed to provide a reasonable scheme for the dynamic spatial scheduling of taxis.

Supporting information

Acknowledgments

The authors would like to thank the anonymous reviewers for their constructive comments, as well as the editors for their helpful suggestions. We also acknowledge the administrative and technical support from Engineering Research Center of Highway Infrastructure Digitalization, Ministry of Education, Chang’an University.