Understanding the Heterogeneity in the Effect of Driving Restriction Policies on Air Quality: Evidence from Chinese Cities

Abstract

Many cities around the world have adopted driving restriction policies to reduce vehicle emissions. However, evidence on the effectiveness of these policies is mixed. I exploit detailed and comprehensive data on Chinese cities to conduct a large-scale study of the effectiveness of a variety of driving restriction policies in a variety of locations. I estimate the monitor-specific short-run treatment effects of each driving restriction policy using a regression discontinuity in time approach, and the average treatment effect using a panel fixed-effect approach. The regression discontinuity in time estimation results show strong heterogeneity. Among the eight measures of air quality used, CO, $NO_2$, PM2.5, PM10, and AQI respond most to driving restriction policies. The average reduction of CO and $NO_2$ are consistent with back-of-the-envelope calculations for policies that effectively limit vehicle use. Using detailed information on the design of each policy, I show that policy details and pollution concentration are the major factors that affect the actual and estimated effects of driving restriction policies, and also a potential explanation for earlier studies that showed driving restriction policies had little effect.

Appendices

Appendix

A: Back-of-the-Envelope Calculations for Pollutants Shares

The mobile sources of CO emissions in mainland China from 2015 to 2018 are 34.61, 34.19, 33.27, 30.89 million tons, respectively (MEE 2019). The detailed emission inventory list is not accessible, but we can find some evidence in the literature. In 2001, CO emissions from mobile sources were 32.63 million tons in mainland China, which is similar to the amounts in 2017 and 2018, and it accounts for 21.75% total human source CO emissions (Wang et al. 2005). In 2003, coal-fire sources contributed to 58% CO emissions in Tianjin (Zhao and Ma 2008), which indicates that mobile sources accounted for at most 42% of CO emissions. The vehicle share of CO is different across cities and time periods, but we can use the statistics in the literature to do inference. Wang et al. (2005)’s statistics should be close to the average value in China, while Zhao and Ma (2008)’s statistics can be used as an upper bound. Thus, mobile sources account for about 20% to 40% of CO emissions in China. The shares of vehicle-emitted CO to the total mobile-source CO from 2015 to 2018 are 86.9%, 87.7%, 87.8%, and 92.6% (MEE 2019). The average is 88.75%. With a simple multiplication, vehicle emissions contribute about 17.75% to 35.50% CO.

The mobile-source contribution to $NO_x$ was 31.64% on average in mainland China, 2015 (MEE 2016a). This percentage varies across different cities, ranging from 10 to 50% (MEE 2019). $NO_x$ mainly consists of NO and $NO_2$. The $NO_2 / NO_x$ ratio has increased during these years. The chemical reaction between NO and $O_3$ forms $NO_2$, so the $NO_2 / NO_x$ ratio varies with the concentration of $O_3$. The $NO_2 / NO_x$ ratio from vehicle emissions has a large range from 2% to 70% (Carslaw 2005, Itano et al. 2014, Wild et al. 2017, Alvarez et al. 2008). In addition, not all $NO_2$ is directly emitted from vehicles, some is formed by $NO_x$. The shares of vehicle-emitted $NO_2$ to the total mobile-source $NO_2$ from 2015 to 2018 were 92.2%, 92.5%, 92.8% and 92.7% (MEE 2019). The average is 92.55%. Using the mobile-source $NO_x$ multiply the $NO_2 / NO_x$ ratio and the vehicle-to-mobile share $NO_2$, I calculate a vehicle share $NO_2$ roughly ranges from 0.19% to 32.40%.

In the 2018 China Annual Environmental Report, $SO_2$ from vehicle emissions accounts for less than 0.3%. The 2010 MIX Asian Emission Inventory shows that 0.82% $SO_2$ in mainland China comes from the Transportation sector. Figure 11 in the Appendix shows increasingly stringent automobile diesel and gasoline standards in sulfur. The percentages of vehicle ownership following different standards in 2015 are 7.5% for Standard I and below, 8% for Standard II, 51.6% for Standard III, 30.5% for Standard IV, and 1.4% for Standard V and above. These percentages change to 3%, 4.5%, 19.1%, 42.5%, 30.9%, respectively, in 2018 (MEE 2016b, 2019). With cleaner fuel and cleaner vehicles, the same amount of vehicles on roads leads to lower sulfur emissions. In a word, vehicle emissions contribute little to the total $SO_2$ in mainland China.

$O_3$ is not directly emitted by vehicles but formed through the chemical reaction of $NO_x$ and VOC with heat and sunlight. The two regimes of ozone formation are the $NO_x$-limited regime and the $NO_x$-saturated regime (McCubbin and Delucchi 1999). Salvo and Wang (2017) concludes that the $NO_x$-limited regime usually applies to the rural areas, the amount of $O_3$ increases with $NO_x$, and it is insensitive to VOC; the $NO_x$-saturated regime usually applies to urban areas, and $O_3$ increases with VOC and decreases with $NO_x$. Jin et al. 2017 studies New York, London, and Seoul and finds increasingly longer NOx-limited $O_3$ chemistry in the warm season. The changing formation regimes of $O_3$ across spaces and time make it difficult to infer $O_3$’s vehicle share.

The share of the mobile-source PM also varies widely across cities. Mobile source contribution to PM2.5 ranges from 45% in Beijing to 13.5% in Hengshui in 2017 (MEE 2017). Zhang et al. (2017) finds that 1.8% PM10 emissions in the US come from on-road mobile sources, 24% PM10 emissions in the UK comes from transportation sources. Therefore, it is also difficult to determine a range of PM’s vehicle shares.

B: Regression Discontinuity in Time Approach at City Level

I conduct city-level RDiT analysis and regress the city-level RDiT estimation results on the sets of control variables same as in Sect. 5.2. The density distributions of the RDiT estimation results are presented in Fig. 15. The density distributions look more multimodal than Fig. 5 because the number of estimation results is smaller. But we can still observe the negative shifts of the estimation results from air pollution alert triggered policies to traffic-triggered and seasonal and regular policies. The second-stage regression of the RDiT estimation results on the two most influential factors, pollution concentration and policy details, are presented in Tables 9 and 10 with robust standard errors.

The city-level RDiT results are consistent with the monitor-level RDiT results in Sects. 5.1 and 5.2. In addition, they further verify the conclusions on some pollutants’ policy responses and are more consistent with the pooled panel fixed-effect results in Sect. 5.3. For example, in the implementation criteria section of Table 10, the traffic-triggered policies and seasonal and regular policies do not correlate to more negative estimation results for $SO_2$ than the air pollution alert triggered policies; only the events-triggered policies correlate to more negative estimation results for $SO_2$. These results match with the chemical property of $SO_2$ that it almost does not respond to driving restriction policies but respond to industrial pollution reduction policies applied at the same time of the events-triggered policies. All the coefficients under the OE policies have smaller magnitudes and larger standard errors than in Table 5, verifying further that the OE policies do not have significantly different estimation results than the ODW policies.

C: Block Bootstrap of the Regression Discontinuity in Time Approach

I block bootstrap the full RDiT analysis by resampling the 54 treated cities 50 times. The simulated distribution of second stage coefficients for pollutant concentration and policy implementation criteria are presented in Figs. 16 and 17.²⁹ The results in Sect. 5.2 are still robust.

D: Balance Table of the Pooled Panel Fixed-Effect Approach

Crump et al. (2009) characterizes the optimal sub-samples with propensity scores lie between [0.1, 0.9]. There are 41 cities in the treatment group and 90 cities in the control group after trimming the sample. The distributions of the propensity scores for the treatment and control cities are drawn in Fig. 12, and it shows that the common support is big enough for propensity score matching. Table 8 shows that the mean city characteristics in the treatment group and control group look different from one another in some variables, but the t-statistics for equal means t-test in column 4 show that we do not reject the null hypothesis that the treatment group and the control group have equal means in these variables.³⁰ To balance these observables further, I reweight the control group by the estimated propensity scores ($\left( {{\text{weight}}\;{\text{ = }}\;\frac{{p\left( {\left. D \right|X} \right)}}{{1-p\left( {\left. D \right|X} \right)}}} \right)$). The reweighted means for the control group are reported in column 3, and the corresponding t-statistics are reported in column 5. Among the 11 variables listed, 9 of them have smaller absolute values of t statistics in column 5 than column 4, showing an improvement in the comparability between the treatment and control groups.

E: Figures

See Figs. 8, 9, 10, 11, 12, 13, 14, 15, 16 and 17.

Fig. 8

The number of unhealthy days calculated by PM2.5

Fig. 9

Population density in 2015

Fig. 10

GDP in 2015

Fig. 11

Fuel standards upgrading timeline (Note The numbers associated with each fuel standard are the maximum sulfur content (ppm).)

Fig. 12

Common support of propensity score matching

Fig. 13

Daily maximum pollutant concentration

Fig. 14

Distribution of the RDiT estimation results without some weather variables in regression

Fig. 15

Distribution of city-specific RDiT estimation results by policy implementation criteria

Fig. 16

Coefficients of pollution concentration

Fig. 17

Distribution of the estimated coefficients by policy implementation criteria

F: Tables

See Tables 8, 9, 10, 11, 12, 13 and 14.

Table 8 Mean city characteristics, 2013

Table 9 Influence of pollution concentration on the city-specific RDiT estimation results

Table 10 Influence of policy details on the city-specific RDiT estimation results

Table 11 Descriptive statistics: 2013 city characteristics

Table 12 Influence of weather on the RDiT estimation results

Table 13 Influence of 2013 city characteristics on the RDiT estimation results

Table 14 AIC and BIC for different orders of time polynomials in RDiT estimation for CO