Driving Factors of NOx Emissions in China: Insights from Spatial Regression Analysis

Abstract

China’s rapid industrialization and urbanization have led to significant nitrogen oxide (NOx) emissions, contributing to severe atmospheric pollution. Understanding the driving factors behind these emissions is crucial for effective pollution control and environmental management. Therefore, this study is an attempt to provide insights into the influence of socioeconomic factors and explore spatial dependencies of NOx emissions in China in 2022 employing spatial regression models (SRMs). Among the SRMs considered, the spatial Durbin model (SDM) is identified as the most suitable for analyzing regional NOx emissions. The study highlights the importance of controlling electricity consumption and vehicle emissions for addressing air pollution in Chinese regions. Specifically, a one billion kilowatt-hour increase in electricity consumption leads to approximately 549.6 tons of NOx emissions, and an increase of 1000 vehicles in a region results in an average increase of 7113.4 tons of NOx emissions in the same region. Furthermore, per capita consumption expenditure (PCEXP) and research and development (R&D) expenditure exhibit negative direct and spillover impacts. Contrary to previous studies, this research finds that changes in urban population density do not have a significant direct or indirect effect on NOx emissions within the studied areas. Moreover, we conducted additional investigations to assess the effectiveness of government action plans in reducing NOx emissions. Specifically, we evaluated the impact of Phases 1 and 2 of the Clean Air Action Plan, launched in 2013 and 2018, respectively, on the socioeconomic drivers of NOx emissions. Therefore, the data were modeled for the years 2013 and 2017 and compared to the results obtained for 2022. The findings indicate that over the entire period (2013–2022), the emission controls mandated by the action plan resulted in significant reductions in the impact of many of the studied NOx drivers. In conclusion, based on the results, this study presents recommendations to mitigate NOx emissions.

1. Introduction

One important and interesting class of chemical compounds that adds to air pollution is nitrogen oxides (NOx). This family consists of seven compounds. Among them, nitrogen dioxide (NO₂), which mostly comes from human activities, is the most widespread kind of NOx in the Earth’s atmosphere. NO₂ is not a dangerous air pollutant by itself, but it reacts with various components in the atmosphere, forming tropospheric ozone (O₃) and acid rain. It is important to keep in mind that tropospheric ozone, which is found in the air we breathe, is the type of ozone of concern that we want to reduce; see [1]. Although China’s rapid industrialization and urbanization have greatly benefited the national economy over the past few decades, they also exposed it to serious environmental issues; see [2,3]. According to the Community Emissions Data System (CEDS) 2024, China emerged as the largest global emitter of pollutant emissions from 2000 to 2022, as illustrated in Figure 1, which highlights the top four countries in terms of six types of air pollutants. To combat air pollutants, China has primarily targeted air pollutants such as O₃ and NOx. Consequently, our study will specifically focus on exploring the driving factors of NOx emissions.

In 2013, the Chinese government launched the Clean Air Action Plan (State Council of China, 2013), in an effort to address the severe issue of air pollution, and as a result, the majority of air pollutant concentrations have rapidly decreased. The annual average of NOx emissions was significantly reduced by 21% during 2013–2017. However, O₃ concentrations were increased. Subsequently, the second phase of the Clean Air Action Plan was launched in 2018 (State Council of China, 2018), with new emission controls for O₃, as outlined in [4].

Recent research indicates that despite an increase in fossil fuel consumption, NOx emissions in China have consistently declined from 2020 to 2022. The slight reduction observed in 2020 can be attributed to decreased transportation activities resulting from COVID-19 lockdowns. Subsequently, with stricter air pollution regulations in place for both the transportation and industrial sectors, the following years, 2021 and 2022, showed substantial decreases in NOx emissions, accounting for almost 70% of the total NOx reduction, as shown in [5].

This is consistent with the information highlighted in Figure 1, illustrating a significant decrease in emissions of most pollutants in China since 2013. Overall, Figure 1 reinforces China’s efforts to improve air quality and demonstrates the progress made in reducing pollutant emissions. It serves as visual evidence supporting the statement that China’s initiatives have had a positive impact on addressing environmental concerns and reducing pollution levels.

While much previous research on NOx pollution primarily focused on natural science and technological aspects, limited attention has been given to exploring the influence of socioeconomic factors and spatial dependencies. Therefore, this study focuses on discussing the impacts of these factors by utilizing data from 31 regions in China in 2022 employing the SRMs, which is a valuable tool in analyzing the spatial dependencies among different regions. The research conducted by [6] reveals that China’s NO₂ emissions have a significant spatial dependence. This means that NO₂ levels in a specific location are influenced by both the characteristics of that location as well as the NO₂ levels of nearby locations. Consequently, neglecting the spatial effects and relying on traditional regression models like ordinary least squares (OLS) and generalized least squares (GLS) can lead to biased estimates. Therefore, it is essential to utilize a spatial analysis approach to adequately address this issue.

The literature review of the main determinations of NOx emissions will be separated into three subsections to serve our empirical study.

1.1. The Built Environment Characteristics

Indeed, a large number of researchers have looked into the relationship between different characteristics of the environment and air pollution. Land usage, traffic and roadways, and land development are a few of these characteristics. A summary of the literature exploring these relationships can be found here:

• Land Use Characteristics: Zones, commercial areas, or residential neighborhoods can significantly impact air pollutant levels. Research by [7] demonstrated that areas with heavy commercial and industrial activities tend to have higher NO₂ concentrations in Seoul, Korea. Additionally, studies by [6,8] highlighted the role of land use patterns, such as green infrastructure, in lowering the concentrations of NOx and NO₂.
• Traffic and Roadways Characteristics: Numerous studies have demonstrated that the road width and the % of roads in an area may be significant factors affecting the amount of air pollution; see [9,10,11]. Additionally, other road characteristics, such as road density, traffic congestion, and the number of bus stops per unit area, have been considered as determinants of air pollution in studies like [12,13]. These findings highlight the importance of considering traffic and road-related factors when examining the factors influencing air pollution levels.
• Land Development Characteristics: Factors such as population density, building density, building height, and building types have been linked to air pollutant levels. A study by [14] demonstrated that higher building height may contribute to increased air pollution concentration, whereas wider streets often result in lower pollutant concentrations.

1.2. Economic Development

Generally, economic development can have both positive and negative effects on NOx emissions. In this context, [6] stated that certain researchers have embraced the notion of an inverted U-shaped association between economic growth and air pollution. Here are a few important things to think about in this relationship:

• Industrial Activities: Economic development often leads to increased industrialization and economic activities such as manufacturing, power generation, and transportation. These activities can contribute to higher NOx emissions, particularly from industrial processes and combustion sources like power plants and factories.
• Energy Consumption: Economic development is usually accompanied by increased energy consumption. If the energy is predominantly derived from fossil fuels like coal, oil, or natural gas, it can result in higher NOx emissions.
• Scientific Research: R&D investments play a crucial role in pollution reduction efforts. As countries develop scientifically, they may adopt cleaner technologies, such as more efficient power plants, advanced monitoring systems, and sustainable materials. These advancements can help mitigate NOx emissions.
• Environmental Regulations: Economic development is often accompanied by the implementation of environmental regulations and policies aimed at reducing pollution. These regulations can include emission standards, fuel quality requirements, and emission control measures. When effectively enforced, such regulations can lead to a reduction in NOx emissions.

One of the more intriguing studies that looked at the relationship between NO₂ pollution and economic development was Han et al.’s analysis [6], which employed the spatial lag model (SLM) to analyze data from 333 prefecture-level cities during the period from 2016 to 2018. The following variables were used in this study to quantify economic development: per capita GDP, natural gas consumption, residential natural gas consumption, industrialization percentage, technology investment percentage, percentage of green coverage, and number of vehicles.

1.3. Meteorological Factors

The impact of current and past meteorological factors on NOx emissions has been extensively studied. Among these studies is the one conducted by the study by [15], which used hourly data from 2015 to 2017 from the western Polish city of Wrocław to assess the impact of wind speed, air temperature, sunshine duration, air pressure, and relative humidity, in addition to traffic flow, on both NO₂ and NOx. By utilizing built random forest (RF) models, the study discovered that including lagged or independent variables can significantly increase the performance of the model and suggest unexpected relationships or dependencies. For example, the study concluded that wind speed increases the importance for NOx prediction with a two-hour delay.

As for the studies conducted in China, Ju et al. [16] examined the impact of meteorological factors like wind speed, temperature, and humidity on the NO₂ in the troposphere during the COVID-19 lockdown period in Wuhan, China, in 2020. They used an RF model to estimate potential NO₂ levels had the lockdown not occurred. The results of this study showed a significant reduction in NO₂ concentrations, ranging from 11% to 65% lower compared to normal periods. This study concluded that tailored emission-reduction policies accounting for varying meteorological conditions are necessary to improve air pollution control efforts.

In Shandong Province, China, changes in air pollutants, specifically NO₂, were examined by Zhao et al. [17] from 2013 to 2019. Up until 2017, NO₂ concentrations showed a steep reduction, which was followed by a more moderate decline. Short-term fluctuation was more influenced by seasonal meteorology, but inter-annual variations in meteorological circumstances accounted for just a small percentage (3.40–18.60%) of the long-term decline in NO₂ levels from 2015 to 2019. Ambient NO₂ concentrations were lowered by climatic circumstances that supported diffusion in the summer and winter and increased by those that promoted accumulation in the spring and fall. Winters had the greatest influence on pollutant dispersion due to excellent weather, which may reduce NO₂ concentrations by as much as 31.0% in comparison to years with unfavorable weather patterns.

Lin et al. [18] investigated the effects of meteorological factors on long-term changes in NO₂ concentrations across China during summer periods from 2013 to 2020. While initial ground measurements showed decreasing NO₂ levels in eastern, central, and southeastern regions, after removing impacts from meteorological conditions like reduced wind speeds, lower temperatures, and higher humidity that tend to limit NO₂ dispersion, the trends changed. In eastern and central China, although meteorology acted to significantly depress NO₂ concentrations, the underlying trend without those meteorological influences was an increase in NO₂ from 2013 to 2020, suggesting these areas were in a volatile organic compound (VOC)-limited regime for ozone formation. However, in southeastern China, NO₂ levels decreased even after accounting for meteorology, implying a shift towards a NOx-limited or mixed VOC/NOx-limited ozone formation regime in that region during this period. This study highlights that properly separating out meteorological effects is crucial for understanding the ozone sensitivity regimes across different regions based on the underlying NO₂ levels and trends driven by emissions changes.

The geographically and temporally weighted regression model was used by Yi et al. [19] to examine the effects of socioeconomic and meteorological variables on NO₂ concentrations in cities in mid-eastern China between 2015 and 2021. According to their data, most cities’ NO₂ concentrations have decreased by more than 10% since 2015—Bozhou, in particular, has seen a drop of 50.5%. On the other hand, NO₂ concentrations have increased in several parts of Jiangsu and Anhui. There is a notable degree of regional variability in the correlation between NO₂ concentrations and variables that influence them. Yi et al. [19] provided useful knowledge and direction for developing air emission-reduction programs in several cities in mid-eastern China.

The following sections are organized as follows: Section 2 discusses the data and study methodology. Section 3 and Section 4 highlight empirical findings. Section 5 contains the main concluding remarks.

2. Materials and Methods

2.1. Data and Study Area

To model regional NOx emissions in China, we took into account 5 explanatory variables and 31 geographical regions sourced from the China Statistical Yearbook 2023 (because there are no data for Taiwan, Hong Kong, and Macao), compiled by the National Bureau of Statistics of China. Their descriptive statistics are shown in

Table 1. Definitions and the annual statistics of all relevant variables in 2022 ($n$ = 31).

Dimension	Variable Name	Definition	Measuring Unit	Mean	Std.	Min.	Max.
Dependent	NOx	Nitrogen Oxides Emissions in Waste Gas	1000 metric tons	290.5	187.7	34.9	769.6
Economic Development	Electricity	Electricity Consumption	billion k.w.h	278.6	202.9	11.9	787.0
	PCEXP	Per Capita Consumption Expenditure of Household	1000 CNY	24.0	7.7	15.9	46.1
	R&D	Expenditure on Research and Development	billion CNY	62.5	79.9	0.2	321.8
The Built Environment Characteristics	Vehicles	Number of Vehicles in Operation	1000 unit	22.7	16.7	0.9	66.6
	Pop.	Population Density of Urban Area	1000 Pearson/sq.km	3.2	1.1	1.5	5.4

.

According to Table 1, in 2022, the annual average of NOx emissions in China was recorded at 290.5 thousand tons, with a standard deviation of 187.7. The lowest recorded value was in the “Hainan” region at 34.9 thousand tons, while the highest was in the “Shandong” region at 769.6 thousand tons. This highlights the considerable regional disparities in NOx emissions. One possible explanation for this variation is the diverse stages of economic growth in different regions of China, and this is what we will investigate throughout our study, especially since the statistics of the explanatory variables differ greatly among these regions.

2.2. NOx Emissions Spatial Visualization across Chinese Regions

The study attempts to explore changes in spatial patterns of NOx emissions across 31 Chinese provinces (because there are no data for Taiwan, Hong Kong, and Macao) using the Multi-resolution Emission Inventory model for Climate and Air Pollution Research (MEIC) database in the years 2000, 2010, and 2020, which can be accessed through the website http://meicmodel.org.cn, as depicted in Figure 2.

The colors on the map represent different levels of emissions. Red and blue shades indicate lower and higher intensities, respectively. This visual representation tells us that the number of provinces with high levels of emissions increased during the period from 2000 to 2010. After 2010, the enforcement of emission-reduction policies resulted in a notable decrease and mitigated the expansion trend, aligning with the observed changes in NOx emissions reported by [20].

In 2020, the geographical distribution analysis of NOx emissions revealed that the “hot” spots or areas with high-emission accumulation are mostly located in the eastern and northern regions, specifically in Shandong, Hebei, and Jiangsu. Conversely, areas with low-emission accumulation, which is indicated by the “cold” spots, are located in the southwestern, southern, and northwestern regions, particularly in Tibet, Hainan, and Qinghai. In general, these patterns suggest the potential presence of spatial dependence among different regions in terms of NOx emissions. It is crucial to prioritize pollution control measures in these high-emission agglomeration areas and develop effective strategies to prevent the emergence of new zones.

Figure 2. Spatial distribution of provincial NOx emissions (a) in 2000, (b) in 2010, and (c) in 2020. Source: http://meicmodel.org.cn (accessed on 24 June 2024) [21,22].

Figure 2. Spatial distribution of provincial NOx emissions (a) in 2000, (b) in 2010, and (c) in 2020. Source: http://meicmodel.org.cn (accessed on 24 June 2024) [21,22].

2.3. Sectoral Contributions to the NOx Emissions

Using the MEIC database, we plotted Figure 3 and Figure 4, which highlight the reduction trends in NOx emissions in 5 main sectors and 22 subsectors during the period from 2000 to 2022.

The analysis revealed that with the execution of the Action Plan for the Prevention and Control of Air Pollutants in 2013 and the Three-Year Action Plan to Win the Battle against the Blue Sky, NOx emissions began to decline in almost all sectors; see [23]. Figure 3 highlights that the primary contributors to NOx emissions were the industry, power plants, and transportation sectors. At the sub-sector level, Figure 4 shows that power generation, diesel vehicles, and off-road mobile sources were the main sources of NOx emissions.

Figure 3. China NOx emissions by main sector. Source: http://meicmodel.org.cn (accessed on 24 June 2024) [24,25,26,27,28].

Figure 3. China NOx emissions by main sector. Source: http://meicmodel.org.cn (accessed on 24 June 2024) [24,25,26,27,28].

Figure 4. China NOx emissions by sub-sector. Source: http://meicmodel.org.cn (accessed on 24 June 2024). [24,25,26,27,28].

Figure 4. China NOx emissions by sub-sector. Source: http://meicmodel.org.cn (accessed on 24 June 2024). [24,25,26,27,28].

2.4. Spatial Regression Models

As we concluded above, NOx air pollutants in China typically demonstrate spatial characteristics as they disperse through the air. Consequently, due to the likely presence of spatial dependence among adjacent regions, it is advisable to utilize an SRM capable of controlling interaction effects among various spatial units.

In modeling terms of spatial econometrics, three distinct types of interaction effects can be identified to explain the reliance of observations in a certain region on observations from neighboring regions. Therefore, it is essential for the researcher to first determine whether spatial effects are present in the data and, if so, ascertain what type of spatial effects need to be considered. These include (i) a spatially lagged dependent variable [Equation (1)], (ii) a spatially autocorrelated error term [Equation (2)], (iii) a spatially lagged independent variables [Equation (3)], or (iv) a combination of them, as mentioned in

Table 2. All possible spatial interaction combinations in the SRMs.

Model Name		Spatial Interactions
		Term	Count
SLM	Spatial Lag Model	${\mathrm{W}}_n{\mathrm{y}}_n$	1
SEM	Spatial Error Model	${\mathrm{W}}_n{\epsilon }_n$	1
SLX	Spatial Lag of X Model	${\mathrm{W}}_n{\mathrm{X}}_n$	K
SAC	Spatial Autoregressive Combined Model	${\mathrm{W}}_n{\mathrm{y}}_n{\mathrm{and}\mathrm{W}}_n{\epsilon }_n$	2
SDM	Spatial Durbin Model	${\mathrm{W}}_n{\mathrm{y}}_n$ and ${\mathrm{W}}_n{\mathrm{X}}_n$	K + 1
SDEM	Spatial Durbin Error Model	${\mathrm{W}}_n{\mathrm{X}}_n$ and ${\mathrm{W}}_n{\epsilon }_n$	K + 1
GNS	General Nesting Spatial Model	${\mathrm{W}}_n{\mathrm{y}}_n$ and ${\mathrm{W}}_n{\mathrm{X}}_n$ and ${\mathrm{W}}_n{\epsilon }_n$	K + 2

. For further details, refer to [29,30]. Figure 5 shows the three types of interaction effects between two locations.

To address the spatial interaction effects observed in a dependent variable, we can employ the SLM when the spatial autoregressive coefficient, λ, exhibits statistical significance—this suggests the importance of considering the spatial autocorrelation in the dependent variable. The SLM equation can be expressed as follows:

$${\mathrm{y}}_n={\mathsf{\lambda }}_0{\mathrm{W}}_n{\mathrm{y}}_n+{\mathrm{X}}_n{\beta }_0+{\epsilon }_n,$$

(1)

where ${\mathrm{y}}_n$ is an $\left(n\times 1\right)$ vector of the dependent variable, n represents the number of observations, ${\mathsf{\lambda }}_0$ denotes a spatial autoregressive coefficient, ${\mathrm{W}}_n$ is an $\left(n\times n\right)$ spatial weight matrix of non-negative known constants providing information about the spatial relationships and connectivity between different units or locations in the dataset, ${\mathrm{W}}_n\mathrm{y}$ is a spatially lagged dependent variable, ${\mathrm{X}}_n$ is an $\left(n\times \mathrm{K}\right)$ matrix of exogenous explanatory variables, K represents the number of explanatory variables, ${\beta }_0$ is a $\left(\mathrm{K}\times 1\right)$ vector of parameters, and ${\epsilon }_n$ is an $\left(n\times 1\right)$ vector of independently and identically distributed $\left(iid\right)$ disturbances such that ${\mathsf{\epsilon }}_n~\left(0,{\sigma }^2{\mathrm{I}}_n\right)$. Unlike the SLM, the spatial error model (SEM) is utilized when there is a need to consider the interdependence among error terms that exhibit dependence on each other. In SEM, the statistical significance of the $\rho$-value indicates the need to control for spatial autocorrelation within the error term. The SEM equation is stated as follows:

$$\begin{array}{c}{\mathrm{y}}_n={\mathrm{X}}_n{\beta }_0+{\epsilon }_n;\\ {\epsilon }_n={\rho }_0{\mathrm{W}}_n{\epsilon }_n+v_n,\end{array}$$

(2)

where ${\rho }_0$ is a spatial autocorrelation parameter and $v_n$ is a spatially uncorrelated error term.

Instead of focusing on endogenous interaction effects or spatial dependency in error terms, spatial econometrics uses the spatial lag of X model (SLX) to study the spatial interaction effects of independent variables. As an extra explanatory variable, the spatial lag of regressors, denoted as ${\mathrm{W}}_n{\mathrm{X}}_n$, is added. The equation for the SLX can be expressed as follows:

$${\mathrm{y}}_n={\mathrm{X}}_n{\beta }_0+{\mathrm{W}}_n{\mathrm{X}}_n{\gamma }_0+{\epsilon }_n,$$

(3)

where ${\gamma }_0$ is just like ${\beta }_0$ of order $\left(\mathrm{K}\times 1\right)$. The spatial spillover effects of this model coincide with the parameter estimates ${\gamma }_0$ of the ${\mathrm{W}}_n{\mathrm{X}}_n$ variables that capture the average value of the neighboring observations of X. The direct effects coincide with the parameter estimates ${\beta }_0$ of the ${\mathrm{X}}_n$ variables.

2.5. Data of Spatial Weights Matrix

To conduct our analysis, we created the spatial weights matrix using two different structures to assess the matrix density impact on empirical inference.

1.

The K-nearest Neighborhood Matrix (KNN): the KNN matrices define the spatial connectivity between neighboring regions based on their proximity. In this structure of spatial weights, each spatial unit is connected to a given number of other units (K units)—here, we use K = 6, where for spatial unit $i$, a particular element of the weight’s matrix, ${\mathrm{w}}_{\mathrm{ij}}$ takes a value of one when the centroid of unit $\mathrm{j}$ is one of the K-nearest centroids and takes zero otherwise; refer to Equation (4).

$${\mathrm{w}}_{\mathrm{ij}}=\left\{\begin{array}{cc}1& {\mathrm{if}\mathrm{d}}_{\mathrm{ij}}\le {\stackrel{⃛}{\mathrm{d}}}_{i\left(n\right)}\forall i\ne \mathrm{j}\mathrm{and}i,\mathrm{j}=1,\dots ,n;\\ 0& \mathrm{otherwise},\end{array}\right.$$

(4)

where ${\mathrm{w}}_{\mathrm{ij}}$ is an element in the matrix ${\mathrm{W}}_n$ that is predefined to represent the strength of the interaction between region $i$ and region $\mathrm{j}$ within a set of geographical units, ${\mathrm{d}}_{\mathrm{ij}}$ denotes the distance between the two regions $\left(i\right)$ and $\left(\mathrm{j}\right),$ which is an indicator of their spatial proximity or remoteness from each other, and $\stackrel{⃛}{\mathrm{d}}$ refers to a critical distance beyond which there is no spatial influence between spatial units.

One of the primary advantages of this approach is its simplicity in interpreting spatial relations. However, its main drawback is the lack of consideration for the intensity of the distance. For instance, if we set the distance of influence at one kilometer, this approach assumes that a spatial unit located 10 m away from another unit holds the same weight as a unit positioned one kilometer away from it. This limitation becomes apparent when comparing the weights assigned to spatial units at different distances, as highlighted by [31].

2.

The Gaussian Transformation Matrix: the main advantage of the Gaussian function is that it assigns higher weights to spatial units that are closer together and lower weights to those that are farther apart, with the weights decreasing as the distance increases, refer to Equation (5).

$${\mathrm{w}}_{\mathrm{ij}}=\left\{\begin{array}{cc}{\left[1-{\left({\mathrm{d}}_{\mathrm{ij}}/\stackrel{⃛}{\mathrm{d}}\right)}^2\right]}^2& {\mathrm{if}\mathrm{d}}_{\mathrm{ij}}\le \stackrel{⃛}{\mathrm{d}}\forall \mathrm{i}\ne \mathrm{j}\mathrm{and}\mathrm{i},\mathrm{j}=1,\dots ,n;\\ 0& {\mathrm{if}\mathrm{d}}_{\mathrm{ij}}>\stackrel{⃛}{\mathrm{d}}\mathrm{or}\mathrm{i}=\mathrm{j}.\end{array}\right.$$

(5)

In this particular case, one of the main challenges is determining the optimal value of the critical distance $\stackrel{⃛}{\mathrm{d}}$. For this analysis, denote $\stackrel{⃛}{\mathrm{d}}=Max\left(\min \left({\mathrm{d}}_{\mathrm{ij}}\right)\right),$ which is calculated as 1199. The distances between the centroids of Chinese regions are obtained from geographical data available at https://www.mapdevelopers.com/distance_from_to.php (accessed on 10 February 2024); see

Table 3. Summary statistics of the distances (in kilometers) between centroids in Chinese regions.

Min. Distance	Average Distances		Max. Distance		Std. of Distances
104	1450		3739		761
Links		W1: K-Nearest Neighborhood		W2: Gaussian Transformation
Total number of nonzero links		186		396
Percentage nonzero weights		19.35		41.2
The average number of links		6		12.8

. To enhance the interpretability of the matrix, it is row-standardized following the method described by [32].

2.6. Our Empirical Framework

In this section, we will outline the empirical framework used in our study. The framework shown in Figure 6 is designed to choose the most suitable model in a spatial analysis context.

3. Results

3.1. Multicollinearity Diagnostics

To improve the reliability of our results, we performed multicollinearity diagnostics before conducting the SRMs. Multicollinearity is defined as a very high degree of linear correlation between the regressors in the model. This issue may disrupt the integrity of data, which could result in unreliable statistical inferences. This leads to inaccurate estimates, inflated standard errors of those estimates, incorrect non-significant p-values, a decline in the significance of partial t-tests, and undermining the model’s predictive ability.

The variance inflation factor (VIF) is utilized to determine if there is multicollinearity among regressors. As a general guideline, VIFs surpassing 10 indicate a significant presence of multicollinearity that necessitates corrective actions; see [33,34]. The VIF results are (7.08) for electricity consumption, (1.61) for PCEXP, (7.17) for R&D, (8.30) for number of vehicles, and (1.23) for population density. This demonstrates that there are no high multicollinearity issues in our dataset. Consequently, we can perform our analysis using the five explanatory variables listed in Table 1.

3.2. Exploratory Spatial Dependence

To understand the spatial structure of NOx emissions, we first employ the widely used classical Lagrange Multiplier (LM) tests and their robust counterparts, as developed by [35,36,37]. Based on the findings presented in

Table 4. LM test results for various spatial weight matrices.

Test	W1: K-Nearest Neighborhood		W2: Gaussian Transformation
	Lag	Error	Lag	Error
LM	5.0579 *	0.1963	5.3316 *	0.0007
RLM	9.2684 **	3.4068	7.9755 **	2.6447

The superscripts ** and * refer to the significance levels at 0.01 and 0.05 respectively.

, it is evident that there is no significant spatial error present in the data for both weights’ matrices. Therefore, we can exclude the SE term from our analysis.

As a next step, the likelihood ratio (LR) test is employed to verify the hypothesis that the SDM can be simplified to the SLM. The LR test is used to compare the goodness of fit of two hierarchically nested models, where one of them is a restricted version of the other. In our case, the SDM includes SLM. Based on the results presented in

Table 5. Results of the LR tests.

Spatial Weights Matrix	Assumptions	LRchi2(5)	p-Value
W1	SLM nested in SDM	9.747	0.0827
W2	SLM nested in SDM	11.449	0.0432

, we find that the LR test statistic is statistically significant at 0.10 and 0.05 for both the KNN and Gaussian transformation matrices, respectively. This suggests that the SDM provides a significantly better fit than the SLM. Consequently, we reject the hypothesis of simplifying the SDM to the SLM in all scenarios.

3.3. Model Selection

The estimation results of the SEM, SLM, and SDM are summarized in

Table 6. W₁-based estimated SRMs results.

	SEM	SLM	SDM
Electricity	0.8265 ***	0.8914 ***	0.7781 ***
PCEXP	−1.4382	−5.3303 *	−7.4567 ***
R&D	−2.1139 ***	−1.4474 **	−1.1030 **
Vehicles	9.4270 **	5.7152 *	5.7390 *
Pop.	3.1522	−5.5726	−4.4971
W1 × Electricity	----	----	0.2196
W1 × PCEXP	----	----	−0.6543
W1 × R&D	----	----	−2.6314 *
W1 × Vehicles	----	----	17.8451 *
W1 × Pop.	----	----	3.4417
$\widehat{\mathsf{\lambda }}$	----	0.4601 **	0.3894 **
$\widehat{\rho }$	−0.3916	----	----
Wald chi2 Test	878.01 ***	625.35 ***	854.15 ***
Wald Test of Spatial Terms	0.53	7.35 **	20.98 **
Pseudo R2	0.7957	0.8495	0.8707
Log Likelihood	−181.02	−178.23	−173.36
Sigma	82.44	74.76	64.39
AIC	376.04	370.46	370.71
BIC	386.08	380.50	387.92

The superscripts ***, **, and * refer to the significance levels at 0.001, 0.01, and 0.05, respectively.

and

Table 7. W₂-based estimated SRM results.

	SEM	SLM	SDM
Electricity	0.8894 ***	0.8592 ***	0.5877 ***
PCEXP	−1.3091	−5.2238 *	−8.2481 ***
R&D	−1.9841 ***	−1.5705 ***	−1.1965 **
Vehicles	7.9857 *	6.6370 **	7.6210 ***
Pop.	3.6789	−1.3910	−10.2712
W2 × Electricity	----	----	0.9291
W2 × PCEXP	----	----	16.0233*
W2 × R&D	----	----	−4.2146**
W2× Vehicles	----	----	12.2623
W2 × Pop.	----	----	−77.5137
$\widehat{\mathsf{\lambda }}$	----	0.4166 **	0.5056 ***
$\widehat{\mathsf{\rho }}$	−0.0179	----	----
Wald chi2 Test	418.05 ***	622.77 ***	918.32 ***
Wald Test of Spatial Terms	0.01	7.18 **	24.34 ***
Pseudo R2	0.7969	0.8429	0.8643
Log Likelihood	−181.28	−178.22	−172.4955
Sigma	83.83	74.92	62.16
AIC	376.56	370.44	368.99
BIC	386.60	380.48	386.20

The superscripts ***, **, and * refer to the significance levels at 0.001, 0.01, and 0.05, respectively.

for both weight matrices. The estimated models are compared based on several criteria, including Pseudo R², the log likelihood function, sigma, the Akaike Information Criterion (AIC), and the Bayesian Information Criterion (BIC). Specifically, the SDM estimated using the Gaussian transformation matrix (W₂-based SDM) exhibits higher values for Pseudo R² and the likelihood function, and lower values of sigma, AIC, and BIC compared to its counterpart estimated using the KNN matrix. This result suggests that incorporating the Gaussian transformation matrix in the SDM improves the model’s ability to explain and capture the variation in the dependent variable.

Furthermore, this model reveals that among the five explanatory variables, four variables demonstrate a significant effect on NOx emissions. This suggests that these variables are important factors in interpreting NOx emissions.

3.4. Heterogeneity Diagnoses

The Breusch–Pagan test is applied to examine if heteroscedasticity exists in the error term of the W₂-based SDM, see [38]. Based on the p-value of the Breusch–Pagan test in

Table 8. Results of Breusch–Pagan test of SDM based on W₂.

The Null Hypothesis	Test Statistic	p-Value
Homoscedasticity	8.2688	0.5073

, which is greater than the conventional significance level of 0.05, we fail to reject the null hypothesis of homoscedasticity. This indicates that the variability of the error term remains relatively consistent across different levels of the independent variables in the W₂-based SDM.

4. Discussion

In the context of SDM, it is crucial to recognize that the conventional interpretation of regression coefficients, which measure the impact of an explanatory variable on a dependent variable, does not apply to regression models incorporating spatial interaction effects. The inclusion of spatial interactions in SDM introduces additional complexity. This complexity arises from the presence of the dependent variable, ${\mathrm{y}}_n$, on both sides of the equation. On the left-hand side, we have the dependent variable, and on the right-hand side, we find the terms ${\mathsf{\lambda }}_0{\mathrm{W}}_n{\mathrm{y}}_n$ and ${\mathrm{W}}_n{\mathrm{X}}_n{\gamma }_0$, which captures the spatial interaction effects. To comprehensively understand and interpret the effects of changes in these models, it is essential to calculate both the direct and indirect effects of each explanatory variable. For more detailed guidance on this topic, refer to the works of [39,40,41].

In

Table 9. Direct and indirect effects of W₂-based SDM.

Variable	Direct Effects	Spillover Effects
Electricity	0.5496 **	0.4579
PCEXP	−9.4669 ***	14.6310 *
R&D	−0.9787 *	−2.6153 ***
Vehicles	7.1134 ***	6.0925
Pop.	−5.9065	−52.3982

The superscripts ***, **, and * refer to the significance levels at 0.001, 0.01, and 0.05, respectively.

, the first column presents the direct effects, which measure how NOx emissions vary within a specific region as a response to changes in a particular explanatory variable within the same region. The indirect impacts of changes in a particular region’s explanatory variable on changes in NOx emissions in surrounding regions are reported in the second column, which is titled “Spillover Effects.” Based on the information provided in Table 9, we draw the following conclusions:

Electricity consumption exhibits a significant and positive direct effect on NOx emissions, as indicated by a coefficient of 0.5496. This suggests that a one billion kilowatt-hour (k.w.h) increase in electricity consumption within the study areas leads to an approximate aggravation of 549.6 tons of NOx emissions in the same region. However, this increase does not significantly impact the surrounding regions. These findings align with the European Environment Agency (EEA), which has identified public electricity and heat production as the second-most-significant contributor to NOx emissions. In 2008, the EEA published an initial assessment indicating that implementing the best available techniques to improve the environmental performance of existing large combustion plants (LCPs) could potentially reduce NOx emissions by up to 59%; see [42]. In the context of addressing NOx emissions from electricity generation, China implemented an electricity price subsidy (EPS) policy in November 2011 to incentivize coal-fired power plants to install denitrification units.

Several studies, including [43], investigate how China’s NOx emissions and NOx removal—that is, the amount of NOx that has been treated and is not released into the atmosphere—are affected by the EPS policy. A panel dataset encompassing 113 prefectural-level cities between 2008 and 2015 was employed in the study. According to their findings, the EPS policy reduced NOx emissions by 1.1% and increased NOx removal by 2.8% for every additional power plant in cities.

The PCEXP exhibits a significant negative direct effect on NOx emissions. This suggests that there will probably be a drop in NOx emissions in areas with higher PCEXP levels. An intriguing finding is that the PCEXP in a specific region yields a significant positive effect on NOx emissions in the surrounding regions. This implies that regions with higher PCEXP can indirectly influence and contribute to the rise of NOx emissions in neighboring regions. These findings align with the research conducted by [44], which highlights that economic development may initially hinder the reduction of NOx emissions until a critical threshold is reached, beyond which emissions begin to decline. The explanation for this phenomenon lies in the fact that regions with higher PCEXP and significant economic development tend to implement effective measures or adopt technologies that successfully mitigate NOx emissions.

As anticipated, there are negative direct and indirect effects of expenditure on R&D. This means that a CNY one million increase in R&D expenditure in a particular region results in a direct reduction of 9.5 tons of NOx emissions in that region, as well as an indirect reduction of 14.6 tons in surrounding regions. These results are in line with the research by [45], which shows that R&D investment significantly improves the quality of the environment.

In general, R&D expenditure is often related to government practices that focus on developing cleaner technologies, improving industrial processes, or implementing pollution control measures, which ultimately result in reduced emissions in both the specific region and its surroundings. By investing in R&D, governments can foster innovation and the creation of more sustainable solutions.

On the other hand, PCEXP can be related to individual practices that caused negative direct effects, or institutional practices that caused positive indirect effects. This negative direct effect may be for several reasons. For example, households with higher consumption expenditures may have access to more energy-efficient appliances and vehicles, leading to reduced emissions. Additionally, higher PCEXP may indicate a higher standard of living, which could be associated with greater environmental awareness and the adoption of eco-friendly practices. The positive indirect effect could include factors such as most manufacturing activities being on the outskirts of the regions, increased production, and the transportation of goods among regions to meet the demand generated by higher consumption expenditure, resulting in emissions in other regions.

It is important to note that both government R&D expenditure and individual/institutional practices can influence environmental outcomes. While government practices often have a broader impact and can drive systemic changes, individual and institutional practices play a significant role in shaping environmental sustainability as well. A combination of efforts from various stakeholders, including governments, businesses, and individuals, is necessary to address environmental challenges effectively and achieve sustainable development.

The number of vehicles in China has a significant positive direct impact on NOx emissions. Specifically, for every increase of 1000 vehicles in a particular region, the NOx emissions in that region increase by an average of 7113.4 tons. However, this increase does not significantly impact the surrounding regions. According to statistics from the China Environment Statistical Yearbook highlighted by [20], vehicle gas is now the nation’s second-largest contributor to NOx emissions. In the last few decades, the number of vehicles has increased significantly and quickly, especially in urban areas, and this upward trend is expected to keep going in the future. As a consequence, one of China’s biggest environmental problems is air pollution from automobile emissions. It is expected that reducing automobile emissions will be the most important step in tackling the nation’s air pollution problem.

However, some earlier studies have highlighted population density as an important factor contributing to the escalation of NOx emissions; see [46]. Our findings reveal that, in the studied area, population density may not be a key driver of NOx. The relationship between population density and NOx emissions can be influenced by various factors, including local conditions, urban planning, transportation infrastructure, and other contextual variables. These factors may vary across different regions and can impact the emissions differently. In this context, the study of [5] highlighted that some Chinese regions with high levels of industrialization and population density experienced a substantial reduction in NOx emissions in 2022. This reduction was attributed to densely populated regions, which were typically more influenced by the COVID-19 Omicron wave in 2022, imposed intensified lockdown measures. These lockdowns likely resulted in a decrease in industrial activities and transportation, leading to a significant reduction in NOx emissions.

The assessment of the government action plans is an essential part of the decision-making process to review their efficacy and to develop new policies. In this context, we conducted further investigation to assess how the first and second stages of the Clean Air Action Plan, which were launched in 2013 and 2018 sequentially, impact the socio-economic drivers of NOx emissions. Therefore, the W2-based SDM is employed for the years 2013 and 2017, and the results are compared to 2022—refer to

Table A1. W₂-based Estimated SDM Results in 2013, 2017, 2022.

	2013	2017	2022
Electricity	4.0101 ***	1.7316 ***	0.5877 ***
PCEXP	−27.8218 ***	−13.6778 ***	−8.2481 ***
R&D	−6.5645 ***	−0.3587	−1.1965 **
Vehicles	2.7478	−0.0310	7.6210 ***
Pop.	40.0340 *	14.1701	−10.2712
W2 × Electricity	0.1957	2.0129	0.9291
W2 × PCEXP	63.2048 ***	19.4349 *	16.0233 *
W2 × R&D	−3.1633	−7.1701 *	−4.2146 **
W2 × Vehicles	−13.9671 **	1.7215	12.2623
W2 × Pop.	−196.8724 **	−99.2287 *	−77.5137

The superscripts ***, **, and * refer to the significance levels at 0.001, 0.01, and 0.05, respectively.

and

Table A2. Direct and Indirect Effects of W₂-based SDM in 2013, 2017, 2022.

	2013		2017		2022
Variable	Direct	Spillover	Direct	Spillover	Direct	Spillover
Electricity	4.2677 ***	4.8962	1.7452 ***	2.1834 *	0.5496 **	0.4579
PCEXP	−22.5789 ***	99.6739 **	−13.5564 ***	19.5964 *	−9.4669 ***	14.6310 *
R&D	−7.2957 ***	−13.9000	−0.4051	−7.4935 *	−0.9787 *	−2.6153 ***
Vehicles	1.3889	−25.8343	−0.01987	1.7934	7.1134 ***	6.0925
Pop.	20.9563	−362.6874 **	13.5338	−102.7706 *	−5.9065	−52.3982

The superscripts ***, **, and * refer to the significance levels at 0.001, 0.01, and 0.05, respectively.

in the Appendix A for detailed results.

Over the whole period (2013–2022), the emission controls required by the first and second stages of the action plan led to significant reductions in the impact of many of the studied NOx drivers from 2013 to 2017 to 2022.

Our results indicate that the action plan has been highly effective in reducing the impact of negative NOx drivers in China. For instance, the direct impact of electricity consumption decreased from 4.2677 in 2013 to 1.7452 in 2017, further declining to 0.5496 in 2022. This means that an increase of one billion k.w.h in electricity consumption within the study areas resulted an approximate aggravation of NOx emissions by 4267.7 tons in 2013, 1745.2 tons in 2017, and 549.6 tons in 2022 in the same region. Additionally, both the direct and indirect impact of population density ceased to be significant by 2022.

In general, it can be said that this action plan provides a successful example for developing air quality policies in other developing countries. However, there is room for improvement. Two potential areas for enhancement include (1) implementing stricter measures to control the growth rate of vehicles, as the impact of the number of vehicles on NOx emissions increased between 2013 and 2022, and (2) launching initiatives to raise awareness among citizens and institutions about environmentally friendly practices. Notably, the plan also did not improve the impact of positive drivers, such as PCEXP and R&D expenditure on NOx emissions.

5. Conclusions

Environmental pollution issues are considered an unavoidable challenge that accompanies the process of economic development in any country. It is anticipated that China will require a great deal of time and effort to effectively manage and improve the situation. Thus, there is an urgent need to conduct comprehensive studies that delve into the socio-economic factors driving NOx emissions in China.

Based on the geographical dataset encompassing 31 regions in China in 2022, this paper used SRMs to investigate the socioeconomic determinants of NOx emissions, highlighting the importance of incorporating spatial dependencies and utilizing appropriate SRM for accurate results. The findings of our research confirm the following key points:

• The NOx emissions are spatially lagged-dependent and lagged-independent correlated, where the W₂-based SDM is deemed the most suitable model among the models considered based on the results of LM and LR tests and the goodness-of-fit criteria. The findings showed that the spatial dependence parameter is significant and has a moderate positive value.
• There is a slight difference between the two employed structures of spatial weights matrix, namely the KNN and Gaussian transformation matrices, in terms of the goodness-of-fit criteria mentioned in Table 6 and Table 7. For example, the W₂-based SDM yields a higher improvement.
• There are positive significant direct effects of electricity consumption and the number of vehicles on NOx emissions in Chinese regions. However, this increase does not significantly impact the surrounding regions.
• On the other hand, the R&D expenditure exhibits a negative significant direct and indirect effect, while the per capita consumption expenditure of households has a negative direct and positive indirect effect on NOx emissions.
• Interestingly, the urban population density does not show a significant impact on NOx emissions in Chinese regions.

To alleviate the challenge of environmental pollution, particularly NOx emissions, policymakers in China should prioritize the following pillars, which we recommend in the context of the Clean Air Action Plan assessment based on the W2-based SDMs for 2013 and 2017 and compared with the 2022 baseline study model:

• Strengthening spatial planning and coordination: Recognizing the spatial dependence of NOx emissions, policymakers should focus on spatial planning and coordination between regions. This includes implementing policies that ensure neighboring regions work together to reduce emissions collectively, considering the cross-border impact of pollutants.
• Encourage energy efficiency measures to reduce electricity consumption. This includes (1) raising awareness and incentivizing individuals and businesses to reduce their electricity usage through implementing energy-efficient technologies, improving insulation in buildings, and promoting energy storage systems; (2) implementing dynamic pricing schemes that encourage load balancing and reduce peak demand; and (3) retrofitting power plants with emission control technologies.
• Strengthen R&D efforts: Allocate resources for research and development in clean technologies and pollution control measures. This could involve funding and supporting research institutions, encouraging collaborations between academia and industry, and providing incentives for the development and implementation of innovative solutions to reduce NOx emissions.
• Control vehicle growth: Restricting the supply of license plate numbers can be an effective strategy to manage vehicle growth and reduce emissions in the short term. However, controlling the long-term growth in the number of private cars can be challenging. Governments should indeed take a comprehensive approach that includes encouraging the development of clean energy and promoting the use of low-emission vehicles, often referred to as “Green” or “Electronic” vehicles. This includes (1) providing financial incentives for green vehicle adoption such as tax credits, rebates, and subsidies, to individuals who purchase or lease green vehicles, and (2) investing in the development of charging infrastructure for electric vehicles to address range anxiety and promote their widespread adoption. Expanding the network of charging stations, implementing fast-changing technologies, and providing incentives for private and public entities to install charging infrastructure can support the growth of EVs.
• Enhance monitoring and enforcement: Strengthen monitoring systems for NOx emissions and enforce compliance with emission standards and regulations. This could involve investing in advanced monitoring technologies, increasing the frequency of inspections, and imposing strict penalties for non-compliance.