Do Vehicle Restrictions on Urban Expressways Reduce Carbon Emissions Across the Urban Road Network? Short-Run and Longer-Run Evidence from Shanghai

Abstract

Vehicle restrictions on urban expressways are widely used to relieve traffic congestion and reduce traffic emissions. However, the effects of such restrictions should be assessed over the wider urban road network rather than on expressways alone, and over both short-run and longer-run periods. This study empirically investigates the impacts of vehicle restriction policies on network-level emissions in Shanghai. The network-level vehicle emissions are dynamically estimated using a carbon-emissions macroscopic fundamental diagram (CE-MFD) model based on taxi trajectory data and loop detector data. The effects are then identified using a spatial difference-in-differences (SDID) framework, while geographically weighted regression (GWR) is used to examine spatial heterogeneity in the associated factors. The results show that extending the restriction periods reduced carbon emissions across the urban road network by 9.96% after one month and by 17.93% after one year. The effects are spatially heterogeneous and are associated with population, road-network characteristics, parking supply, and ramp configuration. These findings suggest that the sustainability impacts depend not only on the restrictions themselves, but also on traffic redistribution and local network conditions. Findings provide empirical evidence for designing sustainability-oriented traffic strategies, underscoring the importance of evaluating emissions outcomes across the urban road network over both short-run and longer-run horizons.

1. Introduction

The transportation sector is the second-largest source of global carbon dioxide emissions, accounting for approximately 840 million tons of CO₂ annually [1]. Based on an analysis of 180 cities across 40 countries from 1960 to 2012, the transition toward low-carbon development in most developing countries remains in its early stages [2]. Since the beginning of 21st century, carbon dioxide emissions from China’s road transportation sector have shown a continuous upward trend, with an average annual growth rate of 6.6% [3]. As an effective measure to reduce traffic congestion and emissions, vehicle restriction policies have been widely implemented around the world. For example, traffic emissions have been reduced through license plate-based driving restrictions in Mexico City [4] and Delhi [5], and emissions have been curbed by low-emission zones (LEZs) in Barcelona [6] and Madrid [7].

Table 1 summarizes vehicle restriction policy schemes adopted by various cities and their reported effects. Shanghai was the first city in China to implement such policies. In 2002, Shanghai prohibited non-local vehicles from using urban expressways in city center between 07:30 a.m. and 10:00 a.m. [8]. During the 2008 Beijing Olympic Games, an odd–even license plate restriction policy was introduced. Analysis of traffic volume and speed before and after implementation showed that traffic volume decreased by 20% to 40%, while speed increased 10% to 20% [9]. Similarly, Hangzhou adopted another form of license plate restriction, barring vehicles with specific plate-ending digits from entering restricted zones during designated periods, and carbon emissions were reduced by 6.9% [10]. Such license plate measures can be effective as a short-term restriction strategy during large-scale events. However, they do not fundamentally reduce overall travel demand. In 2010, Beijing introduced a policy to curb the rapid growth in the vehicle fleet. This policy has been shown to effectively constrain vehicle growth and reduce carbon emissions [11]. Nevertheless, some vehicle restriction policies may fail to achieve their expected outcomes, as commuters traveling during peak hours or longer distances are more likely to violate such restrictions [12]. As an alternative form of vehicle restriction, LEZs have also been shown to be effective. Studies across multiple European cities demonstrated that LEZs can significantly reduce emissions, with particularly substantial effects observed in cities with severe pollution [13]. LEZs in Lisbon have been in place since 2009, leading to a noticeable improvement in road traffic pollutant emissions [14]. Political factors and opposition from some communities have made it difficult to enforce such measures effectively [15].

Recently, there has been a growing number of studies using a macroscopic fundamental diagram (MFD) approach to calculate network-level emissions. Emmanouil et al. estimated vehicle emissions by constructing a microscopic emissions model [16]. The systematic relationships, derived from empirical data at the network level, enabled the extension of MFD to an emissions macroscopic fundamental diagram model. For instance, Ji et al. proposed the carbon emission–MFD model, which was designed to evaluate road traffic emissions throughout the entire network [17]. Moreover, recent studies have demonstrated that combining the MFD model with the COPERT model results in relatively fewer errors compared to other models [18]. Therefore, MFD model can serve as an effective tool for calculating traffic emissions across road networks. Furthermore, drawing on the framework proposed by van Jiansong Chen [19], and Yan et al. [20], factors such as driver distraction, fatigue, and risk tolerance can be dynamically quantified and incorporated into the model. Moreover, these driving styles may potentially influence traffic flow and vehicular emissions [21].

When considering the role of policy interventions in reducing carbon emissions, one of the most popular methods is the difference-in-differences (DID) model. For example, to analyze the heterogeneity of factors influencing the effectiveness of vehicle restriction policies, Song et al. employed a Propensity Score Matching–DID model to evaluate the carbon reduction effects of low-carbon city pilot policies. This method not only effectively addresses endogeneity issues but also enhances the reliability of policy evaluation performance [22]. Zhang et al. applied a multi-period DID model to evaluate the effectiveness of license plate restriction policies in controlling urban traffic emissions [23]. In comparison with the DID framework, the propensity score matching method may suffer from estimation bias when unobserved confounding factors are present. Since policy impacts may differ spatially, the spatial DID (SDID) model can also be applied to capture and assess these regional variations more effectively [24].

A large number of studies have analyzed the relationship between built environment and carbon emissions. For example, machine learning [25], geographically convolutional neural network weighted regression (GCNNWR) model [26], and gradient boosting decision tree [27] have been used to analyze the relationship between carbon emissions and the built environment. While these methods effectively evaluate the relationship between carbon emissions and various influencing factors, this study takes a step further by specifically examining the impact of neighbor-level road network infrastructure on traffic-related carbon emissions. The geographically weighted regression (GWR) model has been widely applied to explore the determinants of urban traffic carbon emissions, covering research scales ranging from the national level [28], urban agglomerations [29,30], and large cities [31,32,33] to small cities [34]. These findings not only contribute to a better understanding of how built environment factors influence motor vehicle emissions but also provide valuable support for the development of policies and strategies aimed at promoting urban energy conservation and emissions reduction.

Table 1. Overview of Vehicle Restrictions and Policies and Their Effects in Various Regions.

Country	City	Restriction Policy	Implementation Date	Measures	Empirical Findings
China	Beijing [9,35]	License plate	July 2008	License plate numbers ending in an odd–even number were permitted to travel on odd–even-numbered dates.	Traffic volume was reduced by 20% to 40%.
		Vehicle ownership	January 2011	Residents must obtain a license plate through a lottery system to purchase a vehicle.	The total stock of cars was reduced by 14%.
	Hangzhou [10]	Vehicle ownership	8 October 2011	Vehicle restriction dates are determined based on the last digit of the license plate.	Total emissions decreased by 6.9%.
	Shanghai [8]	License plate	2002	During the restricted hours, vehicles without local license plates are prohibited from expressways in the city center.	The fraction of congestion time decreased by 20% during morning and evening peaks.
Mexico	Mexico City [4]	License plate	1989	Private vehicles are restricted from driving on specific weekdays based on the last digit of the license plate.	The policy had little impact on reducing overall vehicle travel.
India	Delhi [5]	License plate	2016	Implemented an odd–even license plate restriction scheme (on a trial basis).	The trials generated cleaner air for certain hours of the day.
Spain	Barcelona [6]	Low-emission zones	2020	Only vehicles that meet the required standards are allowed to enter LEZs.	Air quality improved by approximately 30%.
	Madrid [36]	Low-emission zones	2017	The eligibility for on-street parking and access to LEZs is determined based on the vehicle’s environmental label classification.	The policies may have encouraged private car users to shift to more sustainable modes of transportation.
Portugal	Lisbon [14]	Low-emission zones	2011	LEZ 1(2): Only vehicles meeting at least the Euro 3(2) emission standards are permitted to enter.	Pollutant concentrations in both LEZs decreased by varying degrees, ranging from 10% to 30%.

Table 1. Overview of Vehicle Restrictions and Policies and Their Effects in Various Regions.

Country	City	Restriction Policy	Implementation Date	Measures	Empirical Findings
China	Beijing [9,35]	License plate	July 2008	License plate numbers ending in an odd–even number were permitted to travel on odd–even-numbered dates.	Traffic volume was reduced by 20% to 40%.
		Vehicle ownership	January 2011	Residents must obtain a license plate through a lottery system to purchase a vehicle.	The total stock of cars was reduced by 14%.
	Hangzhou [10]	Vehicle ownership	8 October 2011	Vehicle restriction dates are determined based on the last digit of the license plate.	Total emissions decreased by 6.9%.
	Shanghai [8]	License plate	2002	During the restricted hours, vehicles without local license plates are prohibited from expressways in the city center.	The fraction of congestion time decreased by 20% during morning and evening peaks.
Mexico	Mexico City [4]	License plate	1989	Private vehicles are restricted from driving on specific weekdays based on the last digit of the license plate.	The policy had little impact on reducing overall vehicle travel.
India	Delhi [5]	License plate	2016	Implemented an odd–even license plate restriction scheme (on a trial basis).	The trials generated cleaner air for certain hours of the day.
Spain	Barcelona [6]	Low-emission zones	2020	Only vehicles that meet the required standards are allowed to enter LEZs.	Air quality improved by approximately 30%.
	Madrid [36]	Low-emission zones	2017	The eligibility for on-street parking and access to LEZs is determined based on the vehicle’s environmental label classification.	The policies may have encouraged private car users to shift to more sustainable modes of transportation.
Portugal	Lisbon [14]	Low-emission zones	2011	LEZ 1(2): Only vehicles meeting at least the Euro 3(2) emission standards are permitted to enter.	Pollutant concentrations in both LEZs decreased by varying degrees, ranging from 10% to 30%.

Most studies on LEZ policies rely on before–after analyses or use econometric and data-driven methods to identify influencing factors. Existing studies have not adequately captured the heterogeneous spatial effect of policies due to data limitations. Specifically, traffic flows vary significantly across different regions and time periods, leading to heterogeneous effects among restriction policies. In addition, quantitative evidence on changes in short-run and longer-run network-level carbon emissions remains limited. To address these research limitations, this study constructs a CE-MFD model at the road-network level and subsequently employs an SDID model to evaluate the short-term and long-term effects of the expressway traffic restriction policy in Shanghai. Furthermore, the study applies the GWR model to analyze the influence of street-level characteristics on traffic carbon emissions, while also considering the non-motorized travel environment. Figure 1 shows the research framework of this study. This study combines large-scale taxi trajectory data and loop detector data to analyze the heterogeneous impacts of polices. Regional traffic emissions are estimated using a CE-MFD model, and the short-run and longer-run effects of the restriction policy are analyzed.

Figure 1. Research Framework.

The contribution of this study to the topic of sustainability has three aspects. First, it provides a data-driven framework for measuring the network-level carbon outcomes of an empirical traffic-control policy. Second, it compares short-run and longer-run effects, which is essential for judging whether a policy supports sustained emission reductions rather than temporary traffic redistribution. Third, it identifies the street-level conditions associated with heterogeneous carbon outcomes, thereby offering evidence for more targeted, equitable, and sustainable transport-management strategies. The remainder of this paper is organized as follows. Section 2 introduces the study area and data. Section 3 develops the network-level traffic emissions models and evaluates the effects of the vehicle restriction policy. Section 4 presents the short-run and long-run policy effects and the spatial heterogeneity of carbon emissions. Finally, Section 5 provides conclusions and limitations.

2. Study Area and Research Data

Shanghai ranks among the largest metropolitan areas in China. It has a total land area of about 6340 km², which is divided into 16 administrative districts and 108 sub-districts. With the rapid development of Shanghai, traffic flow on expressways in the city center has increased dramatically. The government has begun to gradually implement stricter traffic restriction policies, making the estimation of carbon emissions and the analysis of these policies increasingly important. Due to limitations in detection equipment, the focus of this research is a 105 km² section in the central urban zone (see Figure 2). In the downtown core, the expressway system mainly includes the Inner Ring Elevated Road, the North–South Elevated Road, and the Yan’an Elevated Road. These facilities generally provide two to three lanes for each direction, with design speeds ranging from 60 to 80 km/h. Major arterial corridors, such as Zhongshan Road and Yan’an Road, usually have two to four lanes per direction, where the speed is between 30 and 50 km/h. It should be noted that local residential streets and minor branch roads are excluded from the present analysis. Shanghai’s restriction stages are shown in Table 2.

Figure 2. Study area and road network map.

Table 2. Basic information about the four restriction periods.

Investigation Period	Number of Effective Days	Restricted Time
Stage I: 1–14 April 2015	8	a.m.: 07:30–09:30 p.m.: 16:30–18:30
Stage II: 15–30 April 2015	7	a.m.: 07:00–10:00 p.m.: 16:00–19:00
Stage III: 18–30 March 2016	8	a.m.: 07:00–10:00 p.m.: 16:00–19:00
Stage IV: 1–14 August 2016	8	a.m.: 07:00–10:00 p.m.: 15:00–20:00

In China’s administrative hierarchy, the “street” functions as a township-level administrative unit, operating under the authority of a district and being administered by a sub-district office. In the study area, each street has an average resident population of approximately 78,000 individuals and covers an average area of about 2.4 km².

This research data integrate taxi floating car data (FCD) and loop detector data (LDD) to empirically assess the effects of the restriction policy. The taxi floating car data used in this study was provided by Shanghai Qiangsheng taxi company and spans from April 2015 to August 2016. It contains sequential trajectory information and operational states of around 13,000 taxis. Each trajectory records the vehicle identifier, travel date, travel timestamp, local information, and travel speed. An operational status indicator was assigned to each record, where a value of 1 denotes that the taxi is vacant, and a value of 0 indicates that it is occupied. Detailed explanations of previous research work can be found in the works of Sun, Zheng and Duan [37], Huang and Sun [38], and Sun and Ding [39].

Loop detector data were obtained between 2015 and 2016 from 326 sensors distributed along urban expressways and their connecting ramps. The records contain details on detector position (for instance, locations spanning from one ramp to another), road type (including expressways, on-ramps, and off-ramps), as well as traffic flow data aggregated at both hourly and daily levels. Detectors are distributed along each expressway, covering both on-ramps and off-ramps to capture vehicle flow dynamics.

In addition, the study collected data on the built environment of the streets, including the population and the number of facilities. To further analyze whether the configuration of non-motorized vehicle lanes affects residents’ choices of travel modes and thus traffic carbon emissions, we included data such as the percentage of roads with physical barriers and the percentage of roads with non-motorized vehicle restrictions as control variables.

3. Methods

In this study, the COPERT (Computer Programme to Calculate Emissions from Road Transport) model and the MFD model were combined to assess the carbon emissions under different traffic restriction policies, as well as under the same policy, across different periods. After that, a spatial difference-in-differences model and a geographically weighted regression model are employed to examine factors associated with traffic emissions and their spatial heterogeneity.

3.1. Network-Level Carbon Emissions Model

The CE-MFD framework integrates the empirical MFD with the COPERT emissions model. In this study, the empirical MFD was derived using ten-minute aggregation intervals:

$${\widehat{q}}_{FCD}={\displaystyle \frac{d_{FCD}^{tot}}{L{T}_{10}\widehat{p}}}$$

(1)

where $L$ denotes the total length of the neighborhoods network, while $d_{FCD}^{tot}$ refers to the cumulative travel distance of the floating taxis. $T_{10}$ denotes the division of daily data into ten-minute intervals, and $\widehat{p}$ stands for the penetration rate of taxis serving as probe vehicles. Considering the differences in taxi penetration rates across different areas and time periods, taxi penetration rates were calculated separately for each street at ten-minute intervals.

The value of $\widehat{p}$ is derived from the LDD corresponding to the same time interval. The overall network flow is then computed based on the hourly aggregated traffic volume together with the segment length $l_i$:

$${\widehat{q}}_{LDD}={\displaystyle \frac{{\displaystyle {\sum }_i{q}_i{l}_i}}{{\displaystyle {\sum }_i{l}_i}}}$$

(2)

For every one-hour period, the overall network flow ${\widehat{q}}_{LDD}$ obtained from LDD can be alternatively represented as the mean value of the network flow ${\widehat{q}}_{FCD}(j)$ estimated from FCD across the j-th ten-minute interval:

$${\widehat{q}}_{LDD}={\widehat{q}}_{FCD}\cong {\displaystyle {\sum }_{j=1}^6}{\widehat{q}}_{FCD}(j)/6={\displaystyle {\sum }_{j=1}^6\frac{d_{FCD}^{tot}(j)}{L{T}_{10}\widehat{\rho }(j)}}/6$$

(3)

The COPERT model provides a tool for assessing the energy consumption of both internal combustion engine vehicles (ICEV) and battery electric vehicles (BEV). By incorporating the gasoline emissions factor, the model is able to estimate the per-lane, per-kilometer CO₂ emissions associated with gasoline-powered vehicles as follows:

$$\begin{array}{ll}{EF}_{ICEV,C{O}_2}& ={EF}_{C{O}_2, ICEV}\times {10}^{-3}\times {EF}_{ICEV}\\ & =2.9251\times {10}^{-3}\times (5.862\times {10}^{-13}\times V^8-3.051\times {10}^{-10}\times V^7+8.71\times {10}^{-8}\times V^6\\ & -1.17\times {10}^{-5}\times V^5+0.0009218\times V^4-0.04334\times V^3+1.195\times V^2\\ & -18.44\times V+171.2)\end{array}$$

(4)

${EF}_{ICEV}$ (g/km-ln) quantifies the fuel usage of a vehicle for each kilometer traveled across a single lane. V (km/h) characterizes the average driving speed observed during the trip.

Similarly, the carbon emissions of BEVs are calculated from electricity consumption, and then converted into the carbon emissions per kilometer per lane for BEVs through the grid emissions coefficient ${EF}_{ICEV,C{O}_2}$ (g/km-ln):

$$\begin{array}{ll}{EF}_{BEV,C{O}_2}=& {CF}_{C{O}_2, BEV}\times {10}^{-3}\times {EF}_{BEV}\\ & =0.581\times {10}^{-3}\times (1.134\times {10}^{-12}\times V^8-6.972\times {10}^{-10}\times V^7+1.786\times {10}^{-7}\times V^6\\ & -2.472\times {10}^{-5}\times V^5+0.002008\times V^4-0.09751\times V^3+2.783\times V^2\\ & -43.56\times V+424.1)\end{array}$$

(5)

${CF}_{C{O}_2, BEV}$ means grid emissions factor, which is set to 0.581.

Let $p$ denote the penetration rate of vehicles on the road, and let $p_{ICEV}$ and $p_{BEV}$ represent the proportions of internal combustion engine vehicles and battery electric vehicles, respectively. Then, the average carbon emissions of ICEVs in the network per unit length per lane over a 10 min interval are denoted as $E_{ICEV}(t)$, and the corresponding average carbon emissions of BEVs are denoted as $E_{BEV}(t)$:

$$E_{ICEV}(t)={\displaystyle {\sum }_{i=1}^I{EF}_{ICEV,C{O}_2}(t)}\times {\widehat{q}}_{LDD}\times p_{ICEV}\times {\displaystyle \frac{l_i}{{\displaystyle {\sum }_{i=1}^I{l}_i}}}$$

(6)

$$E_{BEV}(t)={\displaystyle {\sum }_{i=1}^I{EF}_{BEV,C{O}_2}(t)}\times {\widehat{q}}_{LDD}\times p_{BEV}\times {\displaystyle \frac{l_i}{{\displaystyle {\sum }_{i=1}^I{l}_i}}}$$

(7)

Taking into account the total length of lanes L of the street, the carbon emissions of the street per 10 min are estimated as follows:

$$E_{total}={\displaystyle {\sum }_{t\in T}E(t)L}={\displaystyle {\sum }_{t\in T}\left[E_{ICEV}(t)+E_{BEV}(t)\right]}L$$

(8)

At present, Chinese emissions standards align with the European emissions framework, and existing studies have validated the consistency of the COPERT model under this standard [40]. Therefore, the relevant parameters within the COPERT model employed in this study are applicable to traffic conditions in Shanghai.

3.2. Spatial Difference-in-Difference Model

The DID model is a commonly used non-experimental model in policy evaluation. In the panel data of the DID model’s usage scenarios, data individuals can be divided into two categories: the “treatment group” is subject to policy shocks, and the “control group” is not subject to policy shocks. Whether or not to be affected by policy shocks is achieved by introducing dummy variables, in other words, setting the pre-shock to 0 and 1, and the control shock to 0 and 1. The double difference model is shown below:

$$y_{it}={\alpha }_0+{\beta }_1{post}_t+{\beta }_2{treat}_i+{\beta }_3{post}_t·{treat}_i+{\epsilon }_{it}$$

(9)

where

$${treat}_i=\left\{\begin{array}{l}1, if i\in treatment group\hfill \\ 0,if i\in control group\hfill \end{array}\right.$$

(10)

$${post}_t=\left\{\begin{array}{l}1, if t\in post treatment\hfill \\ 0, if t\in before treatment\hfill \end{array}\right.$$

(11)

where $y_{it}$ is the value of the dependent variable. ${\beta }_1$ represents the impact of the treatment group on the control group. ${\beta }_2$ represents the overall impact of policy shocks on the dependent variable. ${post}_t·{treat}_i$ is the interaction term between the grouping dummy variable and the policy implementation dummy variable. The coefficient of the interaction term ${\beta }_3$ is the policy effect. ${\alpha }_0$ represents the intercept of the model. ${\epsilon }_{it}$ is the error term.

To analyze the impact of the vehicle restriction policy on carbon emissions, we defined a control group that is not affected by the policy and a treatment group that is affected by the policy. We divided the carbon emissions data for each stage into four groups: the control group before policy implementation, the treatment group after policy implementation, the treatment group before policy implementation, and the control group after policy implementation. The model analyzes the carbon emissions of four groups based on the time points of changes in policy restrictions in order to determine the impact of the restriction policies. In this study, neighborhoods with expressways passing through them were designated as the treatment group, while the remaining neighborhoods without expressways were designated as the control group.

Based on the DID model, the spatial difference-in-difference model is obtained by introducing a spatial weight matrix and fully considering the avoidance demand of multicollinearity. As the carbon dioxide emissions of one street will have a certain impact on the surrounding streets, the spatial lag model was selected to analyze the impact of other streets with carbon emissions. The spatial lag model is as follows:

$$Y=\rho WY+X\beta +\epsilon$$

(12)

where $W$ is the spatial weight matrix, representing the adjacency relationship between regions, and $WY$ represents the impact of carbon emissions from adjacent regions on the current region. $\rho$ is the spatial lag coefficient, $\beta$ is the coefficient vector of the explanatory variable, and $\epsilon$ is the error term.

3.3. Geographically Weighted Regression Model

This study uses the GWR model to analyze the impact of environmental factors on traffic carbon emissions. The GWR model [41] is a local linear regression model based on spatial change relationship modeling. It produces a regression model describing the local relationship in every part of the study area, which can well explain the local spatial relationship and heterogeneity of variables. The model is as follows:

$$y_i={\beta }_0(u_i,v_i)+{\displaystyle \sum _{k=1}^m{\beta }_k}(u_i,v_i)x_{ik}+{\epsilon }_i$$

(13)

where $y_i$ is the dependent variable at position $i$. $x_{ik}$ is the independent variable at position $i$. $(u_i,v_i)$ is the coordinate of regression analysis point $i$. ${\beta }_0(u_i,v_i)$ is the intercept term. ${\beta }_k(u_i,v_i)(k=1,2,\cdots m)$ is the regression analysis coefficient. ${\epsilon }_i$ is the error term. Moran’s I was used to test for spatial correlation and determine whether spatial clustering exists.

4. Results

4.1. Distribution of Carbon Emissions in Shanghai

Based on Table 3, we conducted t-tests at half-hour intervals during the morning and evening peaks to determine whether there was a significant reduction in carbon dioxide emissions compared to the previous stage. For the first stage of vehicle restrictions, only a limited number of periods in stage 2 exhibited a significant decline in carbon emissions. In contrast, stage 3 exhibited a notable decrease in all periods when compared to the preceding stage. Following the second extension of the restricted hours, the data indicate that carbon dioxide emissions between 08:00–09:00 and 17:00–19:00 within the original restricted periods did not show significant reductions. Instead, the most significant reductions during this phase were concentrated at the beginning and end of the restriction periods.

Table 3. Carbon emissions of the morning and evening peaks during the four vehicle restriction periods.

Investigation Period		Stage I (n = 8) (t)	Stage II (n = 7) (t)	Stage III (n = 8) (t)	Stage IV (n = 8) (t)
Morning peak	06:30–07:00	14.200	14.657	12.845	10.518 **
	07:00–07:30	21.244	20.603	18.739 *	14.591 **
	07:30–08:00	36.323	34.108	29.888 **	25.700 ***
	08:00–08:30	37.591	36.575	31.891 ***	29.901
	08:30–09:00	34.078	32.502	27.773 ***	26.406
	09:00–09:30	37.727	35.557	31.077 ***	28.725 **
	09:30–10:00	36.792	34.160 **	32.089 *	27.711 ***
Evening peak	15:00–15:30	51.514	52.206	46.828 **	32.997 ***
	15:30–16:00	48.318	46.507	41.663 **	28.170 ***
	16:00–16:30	37.527	35.733	30.523 ***	23.484 ***
	16:30–17:00	43.494	40.093	33.255 ***	30.449 *
	17:00–17:30	43.355	41.384	34.600 ***	33.365
	17:30–18:00	41.733	39.758 *	34.487 ***	35.524
	18:00–18:30	48.892	43.663 ***	39.438 *	39.790
	18:30–19:00	41.582	35.834 ***	32.531	32.253
	19:00–19:30	30.746	28.371 *	24.789 ***	21.610 **
	19:30–20:00	29.574	29.098	25.959 **	20.881 ***

Note: p-value represents the statistical significance of indicator changes compared with the previous stage: * p < 0.05; ** p < 0.01; *** p < 0.001. The underlying data indicate carbon emissions during the restricted traffic periods.

Figure 3 shows the road network carbon emissions across the four stages, revealing the general trends and changes over time. The greatest reductions in carbon emissions were concentrated in the northern and central parts of the study area, particularly those that had the highest emissions during stage 1, such as Changfeng Xincun Street, Jing’an Temple Street, Hunan Road Street, and Tianmu West Road Street.

Figure 3. Daily carbon emissions of the neighborhoods(t).

4.2. The Impact of Restriction Policy on Carbon Emissions

4.2.1. Short-Run and Longer-Run Policy Effect

Table 4 shows the regression results of the DID model. Overall, the impact of the vehicle restriction policy on carbon emissions is in line with expectations. The results show that the vehicle restriction policy has a significant impact on reducing carbon emissions. After the first change in the vehicle restriction policy, during 07:00–10:00 and 16:00–19:00, the network carbon emissions decreased 3.819 t (9.96%) after one month, and decreased 6.878 t (17.93%) after one year. This indicates that longer-run traffic restrictions have stronger effects than shorter-run ones. After the second extension of the time restriction, the network carbon emissions decreased 6.908 t (17.19%) during 07:00–10:00 and 15:00–20:00 compared with stage 1.

Table 4. The emission reduction effect of the vehicle restriction policy.

Stage	Treat Stage	ATT (t)	Standard Error (t)	Control Group			Treat Group
				Min (t)	Mean (t)	Max (t)	Min (t)	Mean (t)	Max (t)
2	2	−3.819 *	1.068	15.427	35.831	63.351	4.832	26.559	72.902
3	2	−6.878 *	1.080	11.438	31.358	57.911	1.305	25.145	70.021
4	2	−3.739 *	0.957	19.179	32.478	55.260	4.357	23.125	53.968
2	4	−2.150	1.861	21.184	39.325	79.317	4.832	26.559	72.902
3	4	−2.771	1.781	14.652	35.139	58.928	1.305	25.145	70.021
4	4	−6.908 *	1.366	15.724	26.211	40.233	4.357	23.125	53.968

Note: ATT denotes the average treatment effect on the treated. * represents significance at the 5% level.

Table 5 shows Moran’s I for carbon emissions across the four stages. From stage 1 to stage 4, the Moran index showed a downward trend, indicating that the spatial correlation of regional carbon emissions was weakening. The main reason for this change can be attributed to the implementation of the vehicle restriction policy. With the vehicle restriction measures restricting the passage of high-emission vehicles in a specific period, the traffic pressure on the elevated expressway was relieved, resulting in a reduction in carbon emissions. At the same time, the reduced traffic flow in the restricted area did not completely disappear. It was forced to transfer to other ground roads. This change led to an increase in the traffic pressure on ground roads. Although the emissions of elevated expressways decreased, the carbon emissions of ground roads increased and became more dispersed. In addition, the carbon dioxide spillover from the streets with high carbon emissions to the surrounding streets was also weakened. Under the influence of the vehicle restriction policy, the streets with high carbon emissions no longer significantly affect the emission level of the surrounding areas, and the spatial carbon emissions spillover effect was restrained. This means that the regional concentration of carbon emissions was alleviated to some extent, which makes the distribution of carbon emissions in the whole region more balanced.

Table 5. Moran’s I of the four stages of emissions.

Stage	Moran’s I	Z-Score	p-Value
1	0.3408	3.8516	0.0001 ***
2	0.3405	3.8487	0.0001 ***
3	0.3349	3.8181	0.0001 ***
4	0.3076	3.5287	0.0002 ***

Note: *** represents p-value < 0.001.

Observations show that only a small number of drivers choose public transportation, likely because the original commute, including the expressway was already long, and public transportation further increases commuting time. In the long run, residents’ travel demands will not decrease due to the traffic restrictions. Most non-local license plate vehicles restricted to the inner ring expressway have two ways to avoid the restrictions: either use the outer ring expressway or surface roads, or adjust their departure times to avoid the restricted hours. This is consistent with the public transport data, which show no significant change in passenger flow before and after the change in the restriction policy. Secondly, before and after the restricted hours, the number of vehicles per kilometer per lane on expressway ramps increased by 50–100 vehicles. The average speed on surface roads decreased significantly.

4.2.2. Parallel Trend Test

Figure 4 shows the results of the parallel trend test. The stage of policy implementation is taken as the benchmark, and the data in the first stage were removed to avoid collinearity. In the two stages before the implementation of the vehicle restriction policy, the confidence interval of the carbon dioxide emissions reduction intersected with the horizontal axis, indicating that there was no significant difference between the treatment group and the control group before the change in the vehicle restriction policy. At the same time, after the change in vehicle restriction policy, carbon dioxide emissions decreased significantly, indicating that there was a significant difference between the control group and the treatment group, which was in line with the hypothesis of the parallel trend test.

Figure 4. Results of parallel trend test. Note: The X-axis denotes the stage for restriction. The Y-axis represents the variation in restriction periods. The stage before restriction is implemented as the base period.

4.2.3. Placebo Test

In the process of analyzing the impact of policies on carbon emissions, other unobservable urban characteristics may affect the estimation results over time. Therefore, the placebo test was used to exclude the impact of other factors. Figure 5 shows the results of the placebo test. The view of the placebo test is that if a hypothetical causal relationship is true, then no significant effect should be observed in the treatment of “false” or “unrelated”. In this study, we randomly scrambled the carbon emissions data and reprocessed it as “false” data with the DID model. After repeated processing for ten thousand times, we calculated the “false” carbon reduction effect and its probability after each processing. It can be seen from Figure 5 that it roughly follows the normal distribution, and the average value of the “false” carbon reduction effect is 0, which represents no significant effect. The analysis result is reliable. This shows that the observed reduction in carbon emissions can be attributed to the vehicle restriction policy.

Figure 5. Results of the placebo test. Note: The x-axis is a virtual carbon reduction effect. The y-axis is density.

4.2.4. Results of the SDID Model

The carbon emissions data cover the period from 7:00 to 20:00, and the control and explanatory variables were adjusted. Table 6 shows the regression results of the SDID model. Column (1) shows the regression results without control variables, and column (2) shows the results with control variables added.

Table 6. Regression results of the SDID model.

Variables	(1) PCe		(2) PCe
Stage	−8.200 ** (3.466)		−8.016 *** (2.424)
RL			−1.350 ** (0.639)
PB			−6.175 (11.708)
PCr			192.718 *** (36.459)
PPr			71.725 ** (36.339)
PSr			348.562 *** (93.041)
PD			−1089.163 *** (165.300)
Parking			−1.021 *** (0.362)
Bus station			0.313 ** (0.131)
On-ramp			27.364 *** (4.674)
Off-ramp			−12.194 *** (3.933)
Intercept	23.876 ** (9.660)		107.411 *** (24.930)
Observations	3348		3348
R-squared	0.542		0.776
Stage	√		√
Neighborhoods	√		√
Rho = 0.520
test		p-value
LR teat = 25.892		0.000 ***
Wald statistic = 39.115		0.000 ***
z-value = 6.2542		0.000 ***
LM test value = 0.876		0.349

Note: The p-value represents the statistical significance of the spatial DID model: *** p < 0.001, ** p < 0.01. PCe means per capita carbon dioxide emissions. RL means road length. PB means primary road length per km. PCr means percentage of car-only road, where car-only road refers to road sections restricted to motor vehicles, with the use of non-motorized vehicles strictly prohibited. PPr means percentage of roads with physical separation. PSr means percentage of shared roads. PD means the number of populations.

The results of the SDID are shown in Table 6. The results from both columns indicate that the policy has a significant effect in reducing emissions. When the control variables are considered in the emissions reduction effect of the vehicle restriction policy, they show different effects. From the perspective of the road network and traffic flow characteristics, car-only roads (PCrs) can increase the number of cars, thereby increasing the carbon emissions per unit of population. In contrast, roads with a high percentage of sharing roads (PSr) and percentage of physical separation (PPr) will also increase per capita carbon emissions. Such roads usually have a large number of bicycles. The increased interactions between motor vehicles and bicycles may reduce traffic speeds and increase carbon emissions. In practice, some bicycles may aggravate traffic congestion for convenience by violating traffic rules. The increase in road length will reduce carbon emissions per capita, which is related to the structural characteristics of Shanghai’s road network. A well-designed road network is conducive to traffic diversion and improving traffic efficiency. From the perspective of street facilities, the increase in parking spaces will help reduce carbon emissions. It may be that convenient parking reduces the time for vehicles to find parking spaces, thus reducing carbon emissions.

4.3. Result of the GWR Model

To further align with the expressway traffic restriction policy and achieve the goal of emissions reduction, this study employs a GWR model to conduct an in-depth analysis of the heterogeneous effects of the built environment on traffic carbon emissions within the road network.

4.3.1. Results of GWR Model

Spatial autocorrelation measures the relationship between spatial data. The Moran’s I is one of the most commonly used indexes to measure spatial correlation. The range of the Moran index is [0, 1]. When the Moran index is close to 1, it is a positive spatial autocorrelation, reflecting aggregation. When the Moran index is close to −1, it is a negative spatial autocorrelation and an abnormal value. When it is close to 0, it shows randomness. In GWR model, Z-Score is a standardized index used to measure the significance of the Moran index, which indicates the standard deviation of the expected value of statistics under the assumption of random distribution (Z-Score > |1.96| for 95% confidence). Table 7 presents the Moran’s I values of the relevant variables. The Moran’s I value for carbon emissions is 0.3935, indicating that the variable has a strong positive spatial autocorrelation.

Table 7. Moran’s I of carbon emission factors.

Variable	Description	Mean (SD)	Moran’s I	Z-Score	p-Value
CE	Carbon emission(t)	29.878 (27.761)	0.3935	18.269	0.000 ***
PD	Population(millions)	0.079 (0.026)	0.4751	22.36	0.000 ***
RL	Road length in research area(km)	14.572 (7.741)	0.2231	10.665	0.000 ***
PB	Primary road length per km	0.597 (0.30)	0.4676	22.01	0.000 ***
Non-r	Number of on-ramps per km	0.052 (0.060)	0.2185	10.559	0.000 ***
Noff-r	Number of off-ramps per km	0.058 (0.069)	0.2381	11.431	0.000 ***
PCr	Percentage of car-only road (%)	0.114 (0.120)	0.4384	20.627	0.000 ***
PPr	Percentage of roads with physical separation (%)	0.187 (0.111)	0.4236	19.952	0.000 ***
PSr	Percentage of shared roads (%)	0.018 (0.031)	0.1869	9.1258	0.000 ***
NHs	Number of hospitals per km	1.026 (0.748)	0.1137	5.6073	0.000 ***
NSs	Number of schools per km	1.776 (0.942)	0.1292	6.3257	0.000 ***
NPs	Number of parking lots per km	0.779 (0.522)	0.3950	18.724	0.000 ***
NBs	Number of bus stations per km	3.855 (1.939)	0.1972	9.5838	0.000 ***

Note: The p-value represents the statistical significance of the GWR model: *** p < 0.001.

4.3.2. The Impact of Built Environment on Carbon Emissions

The variation in the local coefficients estimated by the GWR model is shown in Table 8. The local coefficient reflects the spatial heterogeneity of the variable, and the global coefficient reflects the average impact of the variable on all streets.

Table 8. Geographically weighted regression results (n = 180).

	Min	1st Quarter	Median	3rd Quarter	Max	Global
Intercept	−5.3067	−1.0783	−0.3873	0.0524	4.9066	0.2194
PD	−61.2196	−33.0918	−4.8906	1.9652	73.4467	−4.6240
RL	−0.3629	−0.0055	0.0269	0.0536	0.1991	−0.0196
PB	−1.6089	−0.0430	0.5108	1.0264	2.6462	0.2451
Non-r	−26.8694	0.3169	6.4822	12.0741	23.3020	1.2028
Noff-r	−23.8925	−9.8537	−5.7750	−3.3608	19.5555	0.6683
PCr	−8.7173	−1.2916	0.3356	2.6399	6.9589	−0.2169
PPr	−8.9075	−1.0247	0.1598	2.1969	7.0630	−1.2402
PSr	−29.1883	−5.3543	2.2642	14.5882	33.8521	−0.8914
NHs	−2.2779	0.0964	0.5129	0.7520	1.4272	−0.0720
NSs	−2.3844	−0.4796	−0.0761	0.3206	1.4037	−0.1039
NPs	−1.3416	−0.3330	0.3470	0.7221	4.1151	0.2099
NBs	−0.5117	−0.0030	0.1878	0.3858	1.1045	0.1071

Figure 6 shows the relationship between various factors and street carbon emissions, and the change in local coefficient reflects the spatial heterogeneity of each effect. The results show that the positive coefficient of population (PD) is mainly concentrated in the northern streets of the study area, and the streets with larger populations are also clustered on the north side of the study area. Therefore, the impact of the northern part of the study area on carbon emissions is higher than that of the southern part. Road length (RL), number of hospitals per km (NHs), and number of schools per km (NSs) have a positive impact on the carbon emissions of almost all streets. A higher density of schools and hospitals as traffic attraction points is positively correlated with the carbon emissions of the streets. The number of parking lots per km (NPS) has little impact on carbon emissions in most areas. In areas with a higher density of parking lots, the driver has to drive a shorter path to park, and the corresponding carbon emissions will be lower. According to the COPERT formula, the relationship between average speed and carbon emissions is a convex function. When BEVs operate at 30–40 km/h and gasoline vehicles are operating at 50–60 km/h, carbon emissions are lowest. Interactions between cars and buses lead to frequent acceleration and deceleration, thereby increasing CO₂ emissions. The percentage of car-only roads (PCr) shows a positive impact in the middle of the study area. The higher the proportion of car-only roads, the fewer the interactions between cars and bicycles. Drivers are more willing to choose such roads for travel, and such roads make a greater contribution to carbon emissions. The on-ramps and off-ramps of the expressway show the opposite situation. On-ramps may help reduce vehicle accumulation on surface roads, whereas off-ramps may increase vehicle accumulation on surface roads.

Figure 6. GWR results of each coefficient.

5. Conclusions

This paper analyzes the short-run and longer-run impacts of vehicle restriction policies on carbon emissions in Shanghai. First, the network-level carbon emissions data for 45 streets in Shanghai were calculated using the MFD and COPERT model, and then the impact of vehicle restriction policies and influencing factors on carbon emissions was analyzed using the GWR model and SDID model. At the street level, the longer-run (17.93%) policy effect is better than the short-run (9.96%) policy effect. After that, the emission reduction effect gradually weakens. A similar phenomenon was also observed in the low-emission zone of central Madrid [7]. This suggests that the policy has a stronger effect during the initial stage of implementation. However, travel demand among residents is not necessarily reduced by vehicle restriction policies [9]. Long-term effect depends on the continuous optimization of follow-up mechanisms. This result is similar to the results of Shanghai’s 2021 vehicle restriction policy [42]. Government agencies may adopt complementary policies, including vehicle purchase restrictions, license plate regulations, fuel taxation, congestion pricing and adjustments to vehicle restriction policies based on expressway restriction conditions. With the implementation of a vehicle restriction policy, more vehicles are transferred to the ground roads. The road characteristics, parking lots, traffic attraction points, and amount of non-motor vehicles on the streets will have different effects on carbon emissions, which in turn lead to heterogeneity among the carbon emissions of the streets. Based on the results of the SDID and GWR models, for the northern streets in the study area, where population and traffic attractions are dense, green-wave traffic management can be used to control vehicle speeds within the lowest carbon emissions range, and public transportation routes can be adjusted to encourage residents to use public transportation. For the central and southern streets, the conflict between motorized and non-motorized vehicles should be balanced, establishing continuous, non-interfering, dedicated, non-motorized vehicle lanes, and strengthening the supervision of non-motorized vehicle behavior. In Shanghai, LEZs can be implemented on some streets. For expressway ramp areas, adaptive traffic lights should be introduced to control merging flow to address the weaving and congestion caused by the ramps. For off-ramp areas, a smooth connection between the off-ramp and the surface road network should be maintained to accelerate the dispersal of vehicles flowing down from elevated roads and shorten the vehicle stay time.

This study has several limitations. The available data are not continuous. We cannot observe the effects of vehicle restriction policies over a continuous horizon. Therefore, we evaluated restriction effects within a single year only. Longer-term impacts require further investigation. Due to the limited number of traffic restriction days collected in each phase, we are unable to adequately account for variations in weather, seasonal patterns, and travel demand. In future work, we will collect more comprehensive data, reassess the restriction policies, identify the duration of policy impacts, and provide evidence-based recommendations for policymakers.