1 Introduction

Transport CO2 emissions contribute substantially to global greenhouse gas (GHG) emission increases, which will cause climate change and increase the vulnerability of ecosystems and human societies (Ostad-Ali-Askar et al., 2018). It is projected by the International Energy Agency’s World Energy Outlook 2013 that transport fuel demand will increase nearly 40% globally by 2035 (ADB, 2016), and there will be large increases of transport CO2 emissions from develo** Asian countries (Timilsina & Shrestha, 2009).

Understanding transport CO2 emission factors’ effects comprehensively is important for making rational planning strategies. Previous studies have examined and modeled the relationships between transport CO2 emissions and their significant factors, including socio-economic characteristics, urban form factors, and metro accessibility. It is found that car availability, high income, and location in the outer and sprawling areas will notably cause larger transport CO2 emissions (Brand & Boardman, 2008; Brand & Preston, 2010; Brand et al., 2013; Büchs & Schnepf, 2013; Ko et al., 2011; Shuai et al., 2018; Wang et al., 2017; Yang et al., 2020). Fostering polycentric and satellite cities can decrease transport CO2 emissions significantly (Cirilli & Veneri, 2014; Grunfelder et al., 2015; Knaap et al., 2016; Modarres, 2011; Sun, Zacharias, et al., 2016; Sun, Zhou, et al., 2016; Veneri, 2010; Yang et al., 2020). Metro or light rail constructions can reduce driving distance and transport CO2 emissions (Cao, 2019; Huang et al., 2017; Yang et al., 2020). Also, develo** satellite cities can make individuals commute shorter distances, make more local trips, and use more non-motorized modes of transportation, thus producing significantly less transport CO2 emissions (Yang et al., 2020). Numerous studies also find that household locations can influence transport CO2 emissions significantly. It is found that in the Greater Toronto Area, a 1 km increase of the straight-line distance from the city center will cause a 0.25 km increase of vehicle miles travelled (Miller & Ibrahim, 1998). In the Minneapolis-St. Paul Twin Cities Area, a 1 mile increase of the distance to the downtown area is connected to approximately 0.08–0.1 kg increase of transport CO2 emissions per day (Wu et al., 2019). Households located near the radial road and outer ring road areas produce much more transport CO2 emissions than those located in the city center in Chinese cities (Wang et al., 2017; Yang et al., 2020).

Presently, numerous cities have begun metro or light rail constructions to provide mass public transportation services, with fast speeds, for the urban residents. Plenty of scholars have started to focus on the impacts of rail transit on travel behaviors and transport CO2 emissions. Moving into metro neighborhoods is found to be positively related to decreased driving distances by autos and more transit uses (Cao & Ermagun, 2016; Huang et al., 2020). In recent years, some new forms of transfer modes greatly promote the first-and-last mile access trips to metros, such as shared bicycles/cars, feeder buses/customized shuttle buses, electric bicycles/motors, and electric scooters (Baek et al., 2021; Chen et al., 2021; Zuo et al., 2020). These convenient transfers have attracted more residents to use transit and metro modes, and have reduced transport CO2 emissions to some extent.

Also, recent study results show that nature-based solutions can improve and protect ecosystem services, and can bring about changes in land use and land cover in urban areas (Zwierzchowska et al., 2021; Pan et al., 2021). Greenways can reduce pedestrian exposure to air pollution (Ahn, et al., 2021), and residents living near the greenway drive less, and thus, reduce their transport emissions (Ngo et al., 2018). Pedestrian and environmentally-friendly designed streets can promote walkability and access to metro trips (Sun, Zacharias, et al., 2016; Sun, Zhou, et al., 2016). All these nature-based solutions will be beneficial for reducing transport emissions and mitigating climate change.

The above predictors of transport CO2 emissions have been discussed extensively, but limitations still exist. At present, few studies describe the varied effects of the impact factors at different locations of transport CO2 emission distribution. It is vital to obtain heterogeneous emission factor characteristics comprehensively to effectively make planning strategies and implement specific emission reduction policies to address the current develo** situations of Chinese and Indian cities. These development situations include urban growths, motorization and economic increases, metro and rail network constructions, urban agglomeration developments, energy technology improvements, and new energy vehicle promotions. In order to identify factors’ heterogenous effects, models based on the conditional mean method, which are frequently used in the previous studies (Huang et al.,

3 Data and methodology

3.1 Data collection and description

Simple random samplings were carried out in the urban areas of Bei**g, **, Huairou, Shunyi, Miyun, **gu, Fangshan, and Daxing. More detailed information of the four case cities can be obtained in the literature of Yang et al. (2020). The three Chinese cities’ metro lines have been in operation, but Bangalore’s metro line was under construction during the survey year.

Commuting CO2 emissions are equal to the CO2 emission factor (by mode, fuel type, and occupancy) multiplied by the commuting trip distance (IPCC 1997). Well-To-Wheel (WTW) CO2 emission intensities for different fuel types were calculated to obtain CO2 emission factors. The calculation method is described in greater detail in Wang et al. (2017) and Yang et al., (2020). Table 1 reports the statistics and percentiles of individual commuting CO2 emissions in the four city samples. Bangalore in India has larger transport CO2 emissions than the Chinese city of **’an, despite their similar economic levels and urban forms. Bei**g’s top 25% emitters produce the most transport CO2 emissions. Wuhan has the smallest transport CO2 emissions, and **’an’s emissions are in the middle level.

Table 1 Summary statistics of individual commuting CO2 emissions
Full size table

3.2 The quantile regression model

The quantile regression model was first introduced by Koenker and Bassett (1978). This method can depict a more complete picture about the relationship between the outcome \(y\) and the regressors \(x\) at different points in the conditional distribution of \(y\) (Cameron & Trivedi, 2009). Compared to the conditional mean regression method used in the previous studies, such as the frequently used ordinary least square (OLS) method, or other methods based on the OLS method, quantile regression has several advantages. First, OLS regression is sensitive to outliers; however, quantile regression estimates are more robust to address this problem. Second, quantile regression could estimate the covariate effects on any percentile of the distribution, not only obtaining the conditional mean estimation on the entire distribution. Third, the quantile regression method avoids assumptions about the regression error distribution. This means that if the regression error is not normally distributed and is heteroscedastic, using OLS method would produce biased estimation results. The quantile regression method does not need to be applied under the above assumptions. (Cameron & Trivedi, 2009; Wang et al., 2019; Xu & Lin, 2016).

The standard quantile approach is used to specify the conditional quantile function to be linear, and parameters of the intercept and slope may vary with each quantile. The \(q\) th conditional quantile function of \(y\) given \(x\) is denoted as \(Q_{q} \left( {y|x} \right)\). The standard linear conditional quantile function is

$$y_{i} = {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q} + \varepsilon_{qi} ,\quad 0 < q < 1$$
(1)
$$Q_{q} \left( {y_{i} |{\varvec{x}}_{i} } \right) = {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q}$$
(2)

where \(Q_{q} \left( {y_{i} |x_{i} } \right)\) means the \(q\) th quantile of the dependent variable \(y_{i}\); \({\varvec{x}}_{i}^{\prime }\) indicates the vector of independent variables; \({\varvec{\beta}}_{q}\) is the vector of estimated coefficients; \(\varepsilon_{qi}\) indicates a random error term.

The \(q\) th quantile regression estimator \(\widehat{{\varvec{\beta}}}_{q}\) is the solution of minimizing the following function

$$Q\left( {{\varvec{\beta}}_{q} } \right) = \sum\limits_{{i:y_{i} \ge {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q} }}^{N} q \left| {y_{i} - {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q} } \right| + \sum\limits_{{i:y_{i} < {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q} }}^{N} {(1 - q)} \left| {y_{i} - {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q} } \right|$$
(3)

From Eq. (3) it can be seen that quantile regression could be considered as a weighted regression. As Cameron and Trivedi (2009, pp. 207) point out, “If \(q = 0.9\), for example, then much more weight is placed on prediction for observations with \(y_{i} \ge {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q}\) than for observations with \(y_{i} < {\varvec{x}}_{i}^{\prime } {\varvec{\beta}}_{q}\)”.

Since the objective function (3) is not differentiable, it is not applicable to use gradient optimization methods. However, the linear programming method can be used and can provide relatively fast computation of \(\widehat{{\varvec{\beta}}}_{q}\) (Cameron & Trivedi, 2005). The following concisely introduces this method, quoted from StataCorp (2017, pp. 2158):

“Define \(\uptau\) as the quantile to be estimated; the median is \(\tau = 0.5\). For each observation \(i\), let \(\varepsilon_{i}\) be the residual

$$\varepsilon_{i} = y_{i} - {\mathbf{x}}_{i}^{^{\prime}} \widehat{{\varvec{\beta}}}_{\tau }$$

The objective function to be minimized is

$$\begin{aligned} c_{\tau } (\varepsilon _{i} ) = & \left( {\tau {\mathbf{1}}\{ \varepsilon _{i} \ge 0\} + \left( {1 - \tau } \right){\mathbf{1}}\{ \varepsilon _{i} < 0\} } \right)\left| {\varepsilon _{i} } \right| \\ = & \left( {\tau {\mathbf{1}}\{ \varepsilon _{i} \ge 0\} - \left( {1 - \tau } \right){\mathbf{1}}\{ \varepsilon _{i} < 0\} } \right)\varepsilon _{i} \\ = & \left( {\tau - {\mathbf{1}}\{ \varepsilon _{i} < 0\} } \right)\varepsilon _{i} \\ \end{aligned}$$
(4)

where \({\mathbf{1}}\left\{ \cdot \right\}\) is the indicator function. This function is referred to as the check function; the slope of \(c_{\tau } (\varepsilon_{i} )\) is \(\uptau\) when \({\upvarepsilon }_{\mathrm{i}}>0\) \(\varepsilon_{i} > 0\) and is \(\tau - 1\) when \({\upvarepsilon }_{\mathrm{i}}<0\) \(\varepsilon_{i} < 0\), but is undefined for \({\upvarepsilon }_{\mathrm{i}}=0\) \(\varepsilon_{i} = 0\). Choosing the \({\widehat{\upbeta }}_{\uptau }\) \(\widehat{{\varvec{\beta}}}_{\tau }\) that minimize \(c_{\tau } (\varepsilon_{i} )\) is equivalent to finding the \(\widehat{{\varvec{\beta}}}_{\tau }\) that make \(\boldsymbol{x\beta }_{\tau }\) best fit the quantiles of the distribution of \(y\) conditional on \({\varvec{x}}\).

This minimization problem is set up as a linear programming problem and is solved with linear programming techniques. Here \(2n\) slack variable, \(u_{n \times 1}\) and \(v_{n \times 1}\), are introduced, where \(u_{i} \ge 0,v_{i} \ge 0\) and \(u_{i} \times v_{i} = 0\), reformulating the problem as

$$\min_{{{\varvec{\beta}}_{\tau } ,u,v}} \{ \tau {\mathbf{1}}_{n}^{^{\prime}} u + \left( {1 - \tau } \right){\mathbf{1}}_{n}^{^{\prime}} v|y - \boldsymbol{X\beta }_{\tau } = u - v\}$$
(5)

where \({1}_{\mathrm{n}}\) \({\mathbf{1}}_{n}\) is a vector of \(1s\). This is a linear objective function on a polyhedral constraint set with \(\left(\begin{array}{c}\mathrm{n}\\ \mathrm{k}\end{array}\right)\) \(\left( {\begin{array}{*{20}c} n \\ k \\ \end{array} } \right)\) vertices, and the goal is to find the vertex that minimizes (4). Each step in the search is described by a set of \(k\) observations through which the regression plane passes, called the basis. A step is taken by replacing a point in the basis if the linear objective function can be improved. If this occurs, a line is printed in the iteration log. The definition of convergence is exact in the sense that no amount of added iterations could improve the objective function. A series of weighted least-squares (WLS) regression is used to identify a set of observations as a starting basis. The WLS algorithm for \(\tau = 0.5\) is taken from Schlossmacher (1973) with a generalization for \(0<\uptau <1\) \(0 < \tau < 1\) implied from Hunter and Lange (2000).”

The bootstrap method is applied to acquire the standard errors and confidence intervals of the estimated coefficients in the quantile regression model. In order to obtain a sound estimation, the bootstrap replications should be more than 500 or 1,000 times (Efron & Tibshirani, 1993). In our estimations, the bootstrap replications are 1,000 times.

The independent variables in the quantile regression model consist of the following impact factors: household car availability, household annual income, dummy variable of whether the city form is polycentric or has a strong center, dummy variable of whether a commuter is located in the satellite cities, dummy variable of whether a city has metro services, and distance from home to the city center/subcenter (HCD/HSD). The dependent variable is individual commuting CO2 emissions in the form of a natural logarithm. Table 2 presents the definitions and the summary statistics of the variables in the quantile regression models.

Table 2 Definitions and summary statistics of the variables in the quantile regression models
Full size table

4 Quantile regression model results

4.1 Pooled samples of four case cities

Columns 1 to 5 in Table 3 shows the quantile regression model results of the pooled samples at the 10th, 25th, 50th, 75th, and 90th quantiles of commuting CO2 emissions. Column 6 in Table 3 shows the estimation results using the OLS regression method. Figure 1 illustrates the varied coefficients at different quantiles of the emissions, the coefficients’ mean levels using the OLS method, and the distribution of commuting CO2 emissions. It is found that the signs of the estimated coefficients are consistent between the quantile regression and OLS methods. The OLS coefficients are approximately at the middle level of the upper and lower limits of the quantile regression method Fig. 2.

Table 3 Quantile regression model results of four case city pooled samples
Full size table
Fig. 1
figure 1

Quantile regression coefficients and commuting CO2 emission distribution. Note: (a). The lines in bold black refer to the coefficients of OLS regression; (b). The lines in dark green refer to the varied coefficients in the quantile regression; (c). The gray bands refer to the confidential intervals of the coefficients in the quantile regression; (d). The bar graph shows the distribution of commuting CO2 emissions

Full size image
Fig. 2
figure 2

Percentages of mode choice and car availability below and above the 10th, 25th, 50th, 75th, and 90th quantiles in the pooled samples

Full size image

Results in Table 3 show that the changing tendency of the coefficient of car availability turns out to be an inverted U-shape. As commuting CO2 emissions increase from the lower level to the middle level, the positive coefficient of car availability displays a tendency to increase rapidly. At the 10th quantile, car availability could increase 69.5% of the emissions, while this percentage will amount to 118% at the 50th quantile. When the commuting CO2 emissions are larger than the middle level, the positive coefficient of car availability has a slight decrease, increasing 93.2–115.8% of the emissions. Car availability’s effect on increasing the emissions rises by approximately 30% as emissions rise from the 10th quantile to the 50th quantile. As shown in Fig. 3, the marginal effects of car availability are between 33.36 g and 72.38 g of CO2 per trip. At the 75th quantile, the marginal effect turns out to be the largest, registering at 2.2 times that of the minimum. These results are mainly due to the fact that, shown in Fig. 2 and Fig. 4, longer commute distances (averaging 11 km), as well as higher car mode share (66.3%), exist among the commuters with emissions larger than the middle level. Among the low emitters, with emissions smaller than the middle level, the average commute distance is only 4.2 km, and the car mode share is only 17.9%. Conversely, the bus mode share for these individuals is as high as 62.1%.

Fig. 3
figure 3

Average marginal effects on commuting CO2 emissions below and above the 10th, 25th, 50th, 75th, and 90th quantiles in the pooled samples

Full size image
Fig. 4
figure 4

Average distance below and above the 10th, 25th, 50th, 75th, and 90th percentiles in the pooled samples

Full size image

The changing tendency of the coefficient of HCD/HSD also turns out to be an inverted U-shape. The positive coefficient of HCD/HSD has a slight increasing trend as commuting CO2 emissions increase to the 75th quantile. At the 25th quantile, a 1 km increase of HCD/HSD causes the emissions to increase by 2.16%; at the 75th quantile, this quantity increases to 3.77%. HCD/HSD’s effect on increasing the emissions rises by 32% as emissions rise from the 25th quantile to the 75th quantile. At the 90th quantile, the positive coefficient of HCD/HSD has a slight decrease, increasing 2.46% of the emissions. The marginal effects of HCD/HSD also increase as emissions increase to the 75th quantile. Figure 3b shows that a 1 km increase of HCD/HSD could cause 1.1 g to 1.67 g increases of emissions per trip. At the 75th quantile, the marginal effect becomes the biggest, at 2.1 times that of the minimum. These results are caused by the increased commute distances as quantiles increase, from the smallest average level of 1.47 km, increasing to the largest average level of 15.75 km, shown in Fig. 4.

Model results also indicate that, among the high emitters with emissions larger than the middle level, satellite city forms could reduce 56.4–92.4% of the commuting CO2 emissions. These effects are not statistically significant among the low emitters at the 10th or 25th quantiles. Figure 3(d) shows that the marginal effects of satellite cities are between − 38.35 g and − 57.75 g of CO2 per trip. At the 75th percentile, the marginal effect turns out to be the largest, at 1.5 times that of the minimum. Additionally, polycentric urban forms could generally reduce commuting CO2 emissions. At the 10th, 25th, and 90th quantiles, polycentricity could reduce 39.9%, 14.8%, and 11.7% of the emissions, respectively. Figure 3c reports that the marginal effects of polycentric forms are between − 7.55 g to − 19.15 g of CO2 per trip. At the 10th percentile, the marginal effect becomes the largest, calculated as being 2.5 times that of the minimum level.

The significant effects polycentric or satellite city forms have on reducing emissions can be illustrated from the following characteristics. High emitters are characterized by much longer commute distances and more frequent car usage, especially for the high emitters in the top 10th percentile. These commuters have an average travel distance of 15.75 km, with a car mode share of 89.6%, shown in Fig. 2e and Fig. 4b. However, polycentric and satellite city urban forms can promote job-to-housing balances, shorten commute distances, and, thus, reduce driving frequency, which will substantially lower transport CO2 emissions. Among the samples taken from the polycentric city of Wuhan, fewer driving trips, more non-motorized trips, and more intra-trips inside the three subcenters existed, resulting in lower levels of transport CO2 emissions. Among the samples taken from the satellite cities in Bei**g, more local trips inside the satellite cities were observed, and, thus, the transport CO2 emissions were lower, compared with the sprawling areas.

Based on the above results, it can be concluded that controlling the percentage of oil-fueled cars, shortening the HCD/HSD, and develo** satellite cities have the largest effects for reducing transport CO2 emissions from the emitters of the middle and higher levels. Moreover, these middle and higher levels of emitters account for about 50% of the total emitters. Therefore, the above policies will have enormous effects on transport CO2 emission reductions.

Another notable finding reveals that the effects metro services have on reducing the emissions decrease as the emissions increase from the 25th quantile to the 90th quantile. At the 10th quantile, cities with metro services could reduce 34.3% of the emissions, and at the 25th quantile, cities with metro services could reduce 39.1% of the emissions; while at the 75th quantile, this effect drops to 13.5%, and at the 90th quantile, this effect becomes the lowest, at only 9.41%. Calculated from the model results, it can be obtained that the effects of metro services reducing the emissions will suffer a 19% decrease when commuting CO2 emissions increase by 1 kg. Figure 3e and Fig. 5 indicate that the marginal effects of this impact factor are between − 19.94 g to − 6.4 g of CO2 per trip. At the 25th quantile, the marginal effect is the greatest, at 3.1 times that of the minimum. On the one hand, these results are due to the increasing tendencies of car availability, household income, driving frequencies, and commute distances, and the decreasing tendency of metro mode usage as transport CO2 emissions increase, as reflected in Fig. 2 and Fig. 4. Despite the advanced metro services and metro networks in the sprawling areas in Bei**g, commuters will still heavily rely on driving when car availability and long-distance commutes are present. Another possible reason lies in the long travel times that exist while transferring to other modes when using buses and metros. For instance, among the bus or metro users of Bei**g, half of the commuters take more than a 1/3 of their total travel time just in transferring; among the bus users in Bangalore, half of the commuters spend more than 1/2 of their total travel time before and after riding the bus.

Fig. 5
figure 5

Commuters’ percentages, transport CO2 emissions per trip, and metro’s marginal effects at different quantiles. Note: HWD refers to home-work distance in kilometers; CA refers to car availability

Full size image

Figure 5 shows that, between the 25th quantile and the 50th quantile of the emissions, metro’s effects on reducing the emissions have a great extent of decrease. We analyze the samples’ characteristics between the 25th to 50th quantile, between the 50th to 75th quantile, and between the 75th to 90th quantile. It is found that, when commuting distance reaches and exceeds 5.8 km, and the percentage of car availability reaches 41.2% or more, metro’s effects in reducing the emissions drop continuously at a rate of 37.8%. Thus, for low-carbon transportation development, it is suggested to form employment and life circles within a 5–6 km radius. Meanwhile, in order to attract more high emitters to use metros, better transit resources, such as feeder buses or customized shuttle buses, need to be allocated around the metro stations. For the first-and-last mile access to metros, shared bicycles and electric scooters are now popularly used (Baek, 2021; Zuo et al., 2020; Chen et al., 2021), which can expand the access distance to metros, compared with walking, making transfer to metros more convenient for those residents located farther away from the station. These non-motorized travel modes need to be encouraged. Also, urban roads with better walkability and greenways can encourage more travelers to walk or use non-motorized traffic modes in their transfers to metros (Ahn, et al., 2021; Pan et al., 2021; Ngo et al., 2018; Sun, Zhou, et al., 2016; Sun, Zacharias, et al., 2016). Therefore, bicycle lanes, pedestrian streets, and greenways need to be constructed with high service levels. These above suggested measures will be beneficial for the first-and-last mile access to metros and will make metros more attractive to those commuters with cars and high income. In addition, it is necessary to combine proper traffic demand management policies and metro network construction together to control the driving frequencies. The percentage of oil-fueled cars in ownership needs to be controlled, and newer energy-operated vehicles need to be encouraged through the use of transport policies. Car ownership restriction or congestion pricing policymaking need to consider the travel behaviors and sensitivity of travel consumptions among the high-income commuters.

By using the model equations at the 50th, 75th, and 90th quantiles, the increasing rates of transport CO2 emissions produced by the high emitters were calculated when the impact factors’ values varied. These impact factors include distance to the city center, metro provision, mono/polycentric form, and whether satellite cities exist. The high emitters refer to those with car availability and higher household annual incomes between US$ 20,000–40,000. The results are shown in Fig. 6. As the distances to the city center increase, the increasing rates of transport CO2 emissions increase continuously. At the 50th quantile, the increasing rates change between 0.01 and 0.07, at the 75th quantile, the increasing rates change between 0.02 and 0.18, and at the 90th quantile, the increasing rates change between 0.03 and 0.14. Generally, the increasing rates of transport CO2 emissions become larger as the emissions increase. When there are metro provisions, polycentric forms, or satellite cities, the increasing rates of transport CO2 emissions will decrease, especially for the scenarios with satellite city developments. It is noteworthy that, under the scenarios with polycentric and satellite city developments, as the distances to the city center increase, though, the increasing rates of transport CO2 emissions will have more declines. At the 75th quantile, there exist larger extents of the decreased emissions when the distance to the city center is more than about 8–10 km, and at the 50th and 90th quantiles, the emissions will have more decreases when the distance to the city center is more than about 15–18 km. These results indicate that when the urban radius is more than approximately 10–15 km, transport CO2 emissions in the outer areas will increase significantly, while develo** polycentric and satellite cities can greatly decrease the increasing extents in the outer areas. These results suggest that, for transport emission reductions, it is necessary to control the urban development radius within 10–15 km under the monocentric pattern, and when the city continues to sprawl, polycentric structures and satellite cities need to be formed.

Fig. 6
figure 6figure 6

Increasing rates of transport CO2 emissions and distances to the city center under different scenarios

Full size image

4.2 Samples of four single cities, car availability, high income, and polycentric/satellite cities

In order to test the robustness of the heterogeneous effects of the impact factors, different quantile regression models were established by using different subsamples. These subsamples include samples of the four single cities, a sample of the commuters with car availability, a sample of the commuters with high household annual income (> US$ 40,000), and a sample of the commuters located in the polycentric city or satellite city.

Table 4 shows the model results in the samples of the four single cities. In the cities of Bei**g, ** Chinese and Indian cities.

The quantile regression method was applied, because it can overcome the shortcomings of the conditional mean method, frequently used in the previous literature. The conditional mean method can only provide the mean level of the impact factors’ effects, and the regression error has normal distribution and homoscedastic assumptions, while the quantile regression method can estimate the covariate effects on any percentile of transport emission’s distribution and is not limited to the conditional mean model’s assumptions. The quantile regression model is estimated by linear programming techniques, and the standard errors and confidence intervals of the estimated coefficients are calculated by the bootstrap method.

The significant impact factors of transport CO2 emissions include household car availability, mono/polycentric urban form, satellite city development, metro service, and distance from home to the city center/subcenter. The statistical test and robust test results indicate the reliability of the model results. According to the study results, specific urban and transport planning strategies are proposed to reduce transport CO2 emissions under Chinese and Indian cities’ development situations. These development situations include urban growths, metro and rail network constructions, urban agglomeration formations, motorization and economic increases, energy technology improvements, and new energy vehicle promotions. The above development situations are equally important issues for other global cities. Therefore, this study’s findings will be beneficial to transport CO2 emission reductions globally.

Model results indicate that the marginal effects of a city having metro services vary from − 19.94 to − 6.4 g of CO2 per trip. From the lowest emitters to the highest emitters, the effects of metros reducing the emissions drop at a rate of 27%, averagely. That means metro service provisions could not bring about substantial emission reductions among the high emitters, while the marginal effects of car availability change from 33.4 to 340.4 g of CO2 per trip. From the low emitters to the high emitters, the effects of car availability increasing the emissions have a general rising tendency. The largest marginal effect of car availability increasing the emissions is seen within the high-income sample (254.5 g to 340.4 g of CO2 per trip). It is also found that, after the commute distance reaches 5.8 km or more and the car availability’s percentage amounts to 41.2% or greater, metro’s effects on reducing the emissions decrease continuously at a rate of 37.8%.

Therefore, for low-carbon transportation development, it is recommended to form employment and life circles within a 5–6 km radius. Dependence only on metro construction cannot bring about desired emission reductions in future. It is necessary to combine traffic demand management policies together to reduce driving, including controlling the percentage of oil-fueled cars owned, car use restriction, and congestion pricing. In the policymaking of car restriction or congestion pricing, it is necessary to consider the travel behaviors and sensitivity of travel consumptions among the high-income commuters. In addition, to attract high emitters toward using public transit and metro services more often, better public transit resources (feeder buses or customized shuttle buses) need to be allocated around the metro stations. Additionally, high service levels of bike lane facilities, pedestrian streets, and greenways need to be constructed. These could attract more travelers to use low-carbon traffic modes when transferring to metros, and attract more residents located farther from metro stations to access metros.

Model results also indicate that the marginal effects of HCD/HSD are between 0.5 to 18.5 g of CO2 per trip. The effects of HCD/HSD contributing to the increase of emissions grow at a rate of approximately 32.1%. In the sample of the commuters with car availability, marginal effects of HCD/HSD are a little larger (1.7–2.6 g of CO2 per trip) than those in the pooled samples (1.1–2.3 g of CO2 per trip), and in the high-income sample, the marginal effects turn out to be the largest, at 12.4–18.5 g of CO2 per trip. Under future urban expansion and urban agglomeration development, HCD/HSD will continue to become longer. This reality, paired with the higher marginal effects of HCD/HSD among the high-income and car availability commuters, will inevitably create surges in transport CO2 emissions. In order to mitigate these rapid increases in future, it is of great importance to foster polycentric forms inside the main city and develop satellite cities in the outer areas, since the model results report that polycentric and satellite city forms could reduce transport CO2 emissions significantly, with marginal effects of − 145.9 g to − 7.6 g of CO2 per trip and − 1202.5 g to − 38.4 g of CO2 per trip, respectively. Additionally, these two factors’ effects become larger in the high-income and car availability samples. Furthermore, car availability’s and HCD/HSD’s effects on increasing the emissions decline under the scenarios of develo** polycentric and satellite city forms. Under the scenarios with polycentric and satellite city developments, as HCD increases farther, the increasing rates of transport CO2 emissions have significant declines. When the distance to the city center is more than 8–10 km and 15–18 km, develo** polycentric and satellite cities can greatly decrease the increasing extents of the emissions caused by the increase of HCD. These above results manifest that, under the situation of the urban form with one strong center, there will be smaller transport CO2 emissions if the urban radius is within approximately 10–15 km. When the city sprawls beyond this range, transport CO2 emissions will substantially increase. Therefore, it is necessary to control the urban radius within 10–15 km under the monocentric urban pattern and to foster polycentric structures and satellite cities in the outer areas.

In summary, for transport CO2 emission reductions, it is necessary to combine the following planning strategies together, including controlling the percentages of oil-fueled cars, metro and rail network construction, and providing better public transit services around metro stations. Meanwhile, bicycle lanes, pedestrian streets, and greenways need to be constructed with high service levels for the first-and-last mile transfer to public transit or metros. Also, employment and life circles are suggested to be within a 5–6 km radius, and urban radius is recommended to be 10–15 km under the urban form with one strong center. When the city continuously sprawls, polycentric structures and satellite cities need to be formed.