The Spatial Correlation between Foreign Direct Investment and Air Quality in China and the Potential Channel

: Due to the attention given to air pollution, the impact of foreign direct investment (FDI) on air quality has led to many discussions on this topic; however, there is a lack of literature discussing the correlation between FDI and air quality from a spatial perspective. In China, the discontinuity of ground monitoring data further limits research in this area. Using a new air pollution dataset, this paper constructs a dynamic panel of 259 prefecture-level Chinese cities over the period 2013–2018 and reveals that FDI on average induces the pollution halo effect in host cities but shows direct correlation with air pollution in the outer conurbation areas. Further examination supports the main ﬁndings by showing that FDI presents the same correlation with coal consumption and thermal power generation of the local and the outer conurbation areas. The heterogeneity analysis ﬁnds that the industrialization stage, ecological construction, and technology development are important moderators for FDI’s pollution effect. The ﬁndings of this paper generate potential policy implications for regional green development regarding FDI.


Introduction
While foreign direct investment (FDI) plays a key role in the economic development of host countries [1], host countries with relatively weak environmental institutions may attract FDI to pollution-intensive sectors, which leads to environmental deterioration [2]. For this topic, the largest developing country, China, is of great study value. China, ranking as the largest FDI recipient among developing countries, has experienced both rapid economic development and observable environmental decay since establishing its world trade system. Comparing 2018 with 2011, when China first joined the World Trade Organization, annual FDI increased from 4688 million dollars to 121,257 million dollars, an annual average growth rate of 21.09% (data cited from China National Bureau of Statistics, website: http://www.stats.gov.cn/, 1 December 2020). Meanwhile, China has been experiencing environmental decay, and its air quality deterioration has caused global concern [3]. Although atmospheric protection has reached unprecedented levels, according to the 2018 China Environmental Status Bulletin, 217 cities, or 64.2% of the 338 cities in China, still exceeded the ambient air quality standard. While FDI promotes economic development, the role of FDI in changes in the air quality of host countries has become an important topic of discussion in the environmental and economic fields.
The existing literature generally supports two opposite hypotheses regarding the correlation between FDI and environmental indicators. One of these hypotheses, the pollution haven hypothesis (PHH), argues that FDI has a negative impact on the environment in host countries because FDI transfers pollution-intensive industries to countries with relatively weak environmental institutions and leads to environmental decay in these countries [4,5]. Another hypothesis, the so-called pollution halo hypothesis, suggests that FDI shows a positive effect on the environment in the host countries because FDI promotes consistent with the situation of the eastern region but not always reflective of the situation of the western or the central regions because cities that accept FDI are mostly located in the eastern area. [10,11,23]. In addition, due to the availability of data on FDI origins, several studies distinguish FDI as ethnically linked FDI (FDI from Hong Kong, Taiwan, and Macau) and non-ethnically linked FDI (FDI from other countries) and find that ethnically linked FDI often appears to have pollution halo effects, while nonethnic FDI tends to have opposite effects or insignificant effects [1,24]. Some articles distinguish FDI's effects based on the socio-economic characteristics of the host regions, such as the industrialization stage (IS), institutional development, and human capital [1,9,25]. These factors are often found to reduce FDI's pollution haven effect. The existing literature argues that FDI simultaneously affects the environment through scale effects, technique effects, and composition effects [26]. Empirical papers that follow this argument employ simultaneous equations to estimate FDI's effects and find that FDI often shows negative environmental externalities through scale effects and composition effects, while its positive environmental externalities are often achieved through technique effects [12,[27][28][29]. In this way, the overall effect of FDI that most researchers prefer to present depicts only the dominant effect. Further mechanism studies are seldom carried out based on a heterogeneity analysis and thus this issue needs to be further explored.
Here, a panel of 259 Chinese prefecture-level cities from 2013 to 2018 and the dynamic spatial Durbin model (SDM) are used to conduct a spatial analysis between FDI and air quality, and to examine the existence of PHH in China in the last few years. This paper is different from the existing literature and makes contributions in the following ways: (1) this paper considers FDI's pollution effect and decomposes the effect into long-term/short-term effects and direct/indirect effects to compare the dynamic and spatial differences; (2) it updates the observation period to 2018 and compares the effects of FDI on six pollutants to provide robust conclusions; and (3) this paper further examines the spatial correlation by employing energy variables and traffic variables (the potential channel) and exploring the potential moderating effect of city-specific variables.

Spatial Autocorrelation
Before the econometric model is developed, this paper employs global Moran's I (GMI) to examine the spatial autocorrelation of six measures of air pollution and the key explanatory variable, FDI. GMI is calculated with STATA 16, based on the following formula [9,13]: where y stands for the examined variable, and y and Var(y) are its mean and variance, respectively. w ij is the spatial weight between city i and city j. N is the total number of cities. Z score is employed to identify the significance of GMI and is calculated as follows: Moran's I is a standardized index with a range of [-1,1]. A negative value indicates a negative spatial correlation, a positive value indicates a positive spatial autocorrelation, and the values of zero indicate there is no spatial autocorrelation [13,27].
In addition to global Moran's I, a local index of spatial agglomeration (LISA) map is employed to present the regional spatial autocorrelation of the examined variable. The LISA map is drawn with ArcMap 10.7, based on local Moran's I (LMI). The formulas to calculate LMI are as follows: where w ij is the same as in Formula (1), and Zi(j) is the standardized observed value of city i(j), calculated as follows: when LMI is significantly positive, there is either an H-H cluster if high value is surrounded by similarly high values or an L-L cluster if low value is surrounded by similarly low values. When LMI is significantly negative, there is either an H-L outlier if the high value is surrounded by dissimilarly low values or an L-H outlier if the low value is surrounded by dissimilarly high values [13]. The LISA map, with color, shows H-H and L-L clusters, H-L and L-H outliers, and areas with insignificant LMI.

Model Specifications
To deal with the spatial autocorrelation of variables, the spatial lag model (SLM), spatial error model, and spatial Durbin model (SDM) are often used. Compared with ordinary econometric models, the SLM considers the spatial autocorrelation of the explained variables. The SEM considers the spatial autocorrelation of missing variables, that is, the spatial autocorrelation of the error term. The SDM considers not only the spatial autocorrelation of the explained variable but also that of the explanatory variables. The specifications are as follows, respectively: SEM : SDM : In the SDM, the causal effect of the explanatory variables on the explained variable is decomposed into direct effects (the local causal effect) and indirect effects (the spatial spillover effect). When there is no significant spatial autocorrelation of the explanatory variables, the SDM degenerates into the SLM, and the direct effect reported by the SDM equals the total effect reported by the SLM [20]. The SDM does not consider the spatial correlation of the error term [13]. Lesage and Pace (2009) proposed that the cost of ignoring the spatial autocorrelation of the explanatory variables is higher than that of ignoring omitted variables. Ignoring the spatial autocorrelation of the error term reduces only the efficiency of the model; however, a maximum likelihood estimation (MLE) can compensate for the lack of efficiency of OLS.
Many studies ignore the time lag effect of pollution, although the pollution in an earlier period is proven to have a significant impact on the current period [9]. This article takes the time lag effect of the explained variable into consideration, that is, a dynamic panel is constructed to avoid potential estimation bias. A dynamic specification is available only for the SDM and SLM; therefore, this paper employs the following dynamic SDM: where AP stands for air pollution and Control includes a series of control variables that are both correlated with AP and FDI. The spatial weight matrix w ij is a geographical inverse distance matrix calculated with the coordinates of city centers that come with CN-reanalysis observations through the econometric software STATA 16.0. γ i controls the city fixed effect and η t controls the year fixed effects.
To further validate the main regression and explore the moderating effect of cityspecific variables, this paper carries out a heterogeneity analysis by employing the interaction terms of FDI with the interested urban variables and the spatial factors of these terms. The model is specified as follows: where AP represents PM 2.5 , PM 10 , SO 2 , O 3 , and CO, sequentially (the regression of NO 2 is not presented because its dynamic term is omitted). FDI's direct pollution effect (the effect of FDI i,t on AP i,t ) is expressed by β 1 + θ 1 Factor i,t , and the spatial spillover effect of FDI (the effect of FDI i,t on AP j,t ) is expressed by β 2 w ij + θ 2 w ij Factor i,t .
This paper further explores the potential channels through which FDI may affect air quality. The model specification is as follows: where  10 , and CO, are sequentially used as proxies of air pollution. This paper exploits the CN-reanalysis dataset to eliminate the endogeneity of the locations of ground monitoring stations. This dataset not only includes these six main pollutant concentrations but also resamples the observations into grids of 0.25 • × 0.25 • through atmospheric transportation models. This paper first extracts observations of each city and then averages these values to obtain the daily average of pollutant concentration of the city. Finally, the daily values of the city average are again averaged into annual values. Averaging over space and long periods is efficient for eliminating data uncertainty.

Other Variables
Following the existing literature, this study selects the population, economic development, average wage, IS, ecological construction, and technology level as control variables [3,9,30,31].

1.
Population: Since the annual average population better represents the number of actual residents than the household population, the population is measured as the annual average population (Pop) provided in the China City Statistical Yearbook.

2.
Economic Development: The GDP is a commonly used indicator that characterizes the total economic output of a region, and thus here is used as the variable for the economic development of each city. The GDP deflator is employed to convert all data into a comparable amount based on 2013 values. GDP per capita (GDP/Pop) is employed here. To address the potential nonlinear effects of economic development on air pollution, the quadratic term of GDP per capita is also controlled for in the regressions [13].

3.
Attraction to Talent: The average wage of laborers (AW) is used to measure the level to which a city attracts talent since a higher wage can attract more advanced laborers. Additionally, in contrast to the average years of education, the wage level directly represents the quality of human resources.

4.
Ecological Construction: Green areas are used as proxies for the ecological construction level of a city. The green areas per capita (EC/Pop) are employed in the model.

5.
Industrialization Stage: The industrialization stage is an important mediator of FDI's pollution effect and is generally discussed in studies herein [9]. The industrialization stage is measured as the proportion of tertiary industry to secondary industry (IS). 6.
Technology: Expenditures on science and technology per capita (Tech/pop) are used as proxies for the technology level of a city. 7.
Energy Production: Emissions from coal-based electricity generation are important sources of local air pollution. Out of environmental consideration, thermal power generation per capita (Thermal/Pop) is employed. 8.
Energy Consumption: Emissions from energy consumption are the main sources of air pollution among which coal consumption is mainly from the industrial sector and gasoline consumption is an ideal agent of emissions from the transportation sector. Therefore, this paper employs total energy consumption per capita (Total Energy/Pop) to control emissions from energy consumption. In addition, coal consumption per capita (Coal/Pop) and gasoline consumption per capita (Gasoline/Pop) are employed to study the potential channel through which FDI affects the air quality.  Table 1.

Spatial Autocorrelation Tests
Cross-sectional data on the six pollutants and FDI from 2013 to 2018 were used to calculate the GMI, and the results are shown in Table 2. These seven variables all have significant positive GMI, which means that high-high clusters and low-low clusters exist. PM 10 has the strongest spatial autocorrelation, while FDI ranks the lowest. For the air pollutants, the value of GMI is close to that reported in Huang   The LISA map of the six pollutants and FDI was then calculated and displayed on maps with ArcGIS. From 2013-2018, the spatial agglomeration did not change much, except for O 3 (the change is presented in Appendix B Figure A1). Due to space considerations, Figure 1 displays only the LISA map for 2018 as an example to show the spatial distribution of air pollution and FDI. For air pollution, high-high clusters are mainly concentrated in the middle and lower reaches of the Yellow River and the Yangtze River, which indicates that these areas have high pollution levels and high spatial correlation values for air pollution. Low-low clusters are mainly concentrated in the southeast and the northeast, which indicates a low pollution level and weak spatial correlations. The FDI shows a much weaker spatial correlation than air pollution, and its low-low clusters are mainly concentrated in the southwest and the northeast, while the high-high clusters are mainly concentrated in the Yangtze Delta. The LISA map again illustrates the need to incorporate the spatial lag term of both the explained variable and explanatory variables into the spatial econometric model.

Results of the Main Regression
This section shows the estimation results of Model (8). Based on the results of the Hausman tests, the following regressions were conducted with fixed effects. Supplementary Appendix C Table A2 shows the dynamic term (namely, τ in Model (8)) and other information regarding the regressions, which validates the inclusion of the dynamic factor and the spatial factor of the explained variables.
In contrast to the existing literature that controls for the spillover effect to eliminate estimation bias, this article addresses the spatial spillover effect of FDI and the local impact of FDI. Empirical studies relying on simple spatial estimators may lead to biased conclusions [3]. To identify the impacts arising from different sources, Lesage and Pace (2009) suggest reporting the summary indicators of the direct, indirect, and total effects. The calculation methods and significance of these three estimators are detailed in multiple studies [32][33][34].

Short-Term Effects
Employing a dynamic panel enables us to distinguish the impact of the explanatory variables on the explained variables in the long term and short term. Table 3 reports the short-term direct, indirect, and total effects of the explanatory variables on air pollutants.      Robust standard errors are in parentheses, *** p < 0.01, ** p < 0.05, * p < 0.1.
The dynamic term of NO 2 is omitted due to collinearity; therefore, results regarding the short-term effect of NO 2 are not available. In Table 3, Panel A, the coefficients of FDI/GDP on PM 10 , PM 2.5 , and SO 2 are significant and negative and those on O 3 and CO are negative but insignificant. The coefficients in Panel A indicate the local-to-local effect of FDI on air pollution. Therefore, the statistical results support a reverse correlation between FDI inflow and local air pollution, which supports the pollution halo hypothesis from the local-to-local perspective. The results are robust across pollutant types.
Panel B presents the coefficients indicating indirect effects (namely the local-tosurroundings effect). The coefficients of FDI/GDP on PM 10

Long-Term Effects
The coefficients of the direct and indirect effects of FDI and other variables on air pollution are consistent with the short-term case, among which the short-term and longterm effects of FDI/GDP on PM 10 and SO 2 are robust while the long-term effects on PM 2.5 ( Table 4 Panel A and B) are insignificant. Furthermore, the effects of FDI on O 3 and CO are not robust/significant, which indicate the correlations between FDI and these pollutant types are weak. Since the temporal lag terms in the regression of NO 2 are omitted, the insignificant coefficients may result from the setting of the models instead of a weak correlation between FDI and NO 2 .  Robust standard errors are in parentheses, and *** p < 0.01, ** p < 0.05, * p < 0.1.
In brief, although FDI, to some extent, shows heterogeneous effects across the pollutants, it generally shows reverse correlation with the air pollution of host cities but direct correlation with air pollution in the outer conurbation areas.

Heterogeneity of FDI's Pollution Effect
The existing literature generally recognizes the heterogeneity of FDI's pollution effect; however, there is a lack of comprehensive heterogeneity analyses on urban variables under the spatial view which can provide valuable information for interregional cooperation in green development regarding FDI.
The results of these regressions are shown in Table 5 panels A-C. The Main coefficient indicates the moderating function of the studied variables on FDI's local-to-local effect. For example, if the main coefficient is significantly positive, the higher the value of the variable of the host city, and the FDI inflow is correlated with higher air pollution of the host city. The Wx coefficient indicates the moderating function of the studied variables on FDI's local-to-surroundings (or surroundings-to-local) effect. Specifically, if this coefficient is significantly positive: the larger the variable of the host city, the more FDI is correlated with air pollution in the outer conurbation areas.  Observations  1290  1290  1290  1290  1290  1290  1290  1290  1290  1290  City ID  258  258  258  258  258  258  258  258  258  258 The Main coefficient of the interaction term of FDI and IS (Table 5 Panel A) is significantly positive for PM 10 (p < 0.01), PM25 (p < 0.01), and SO 2 (p < 0.05), and insignificant for O 3 and CO. The spatially weighted coefficient of the interaction term, Wx, is significantly positive at the 1% level and larger in magnitude for all pollutants. This result implies that the IS (ratio of tertiary to secondary industry) can intensify FDI's pollution effect. In other words, FDI inflows to cities with higher tertiary industry ratios have a larger air pollution effect on the host city and even larger spillovers to neighboring cities. Areas with a higher IS are special economic zones: Beijing, Shanghai, Shenzhen, and tourist cities, and tourist cities account for the majority. These cities are characterized by low background concentrations of pollutants and underdeveloped industries (heavy industries in special economic zones are being reallocated to the outer conurbation areas). The relationship between pollution emissions and air quality usually has a marginal diminishing relationship. Therefore, the inflow of unit FDI into areas with better air quality has a greater marginal effect on pollution. Since the industrial sector is still an important flow of FDI, this suggests that special economic zones dominated by financial and service industries and tourist destination cities with better ecological conditions should pay more attention to introducing clean FDI from the service industry.
The Main coefficient and Wx coefficient of the interaction terms in Panel C and Panel D (FDI × EC and FDI × Tech) are all significantly negative at the 1% level for most of the pollutants (PM 10 , PM 2.5 , and SO 2 ), which indicate that the ecological construction level and technological development play a significant role in helping FDI achieve the pollution halo effect, to both the local and outer conurbation areas. The findings regarding FDI heterogeneity across ecological construction and technological development are consistent with expectations and findings in the existing literature [12].

Potential Channels
Section 3.3 explores the moderating effect of four urban variables which are not necessary to correlate with FDI inflows. However, FDI can affect air quality through several channels which can be explored by analyzing the potential mediators. Different from the moderating effect, the potential mediators must have a significant correlation with FDI. Therefore, this section studies the potential channels of FDI's pollution effects via Model (10). The control variables are the same as Table 3 unless extra clarification is required. Table 6 reports the results.
This section first examines FDI's contribution to the technology development of a city. As shown in Table 6, the coefficient indicating the direct effect of FDI on technology development is significantly positive (p < 0.05) while the coefficient indicating indirect effect is insignificant. The results support that one potential channel through which FDI benefits local air quality is by promoting local technology development.
The effect of FDI on IS is examined, but the coefficients indicating direct and indirect effects are insignificant. The results illustrate that although FDI does not have a significant effect on the ratio of tertiary-to-secondary industry, there is an overall tendency that FDI flows more into the secondary industry since the coefficients are negative. Therefore, IS acts as a moderator of FDI's pollution effect.  The effects of FDI on coal consumption are examined. The control variables eliminate total energy consumption and thermal power generation to avoid collinearity. The results show a significant (p < 0.01) negative relationship between FDI and local coal consumption but a positive relationship between FDI and coal consumption in the outer conurbation areas. The results support the main regression results and reveal one potential channel through which FDI's pollution effect is correlated with coal consumption. Coal is mainly consumed in the secondary industry and thermal power generation consumes the biggest amount of coal. Therefore, this article further examines the correlation between FDI and the thermal power generation amount. The direction and significance level of the coefficients are the same as those for coal consumption. The results are consistent with those on coal consumption and air quality, which suggest FDI's air pollution effect is significantly correlated with coal consumption and thermal power generation. Based on the fact that most big cities in China do not have thermal plants and buy electricity from the outer conurbation areas (because for environmental purposes, big cities that have higher economic development level and bigger attraction to FDI tend to reallocate the energyintensive industry to the outer conurbation areas which seek industrial development), the empirical results warn of the potential enhancing effect of FDI on the spatial energy-supply structure, which is cause for further concern by the city managers.
The effects of FDI on the transportation sector are examined. First, this paper estimates FDI's contribution to gasoline consumption. Tra1/Pop and Tra2/Pop are eliminated from control variables to avoid collinearity. The coefficient indicating direct effect is negative but at a lower significance level (p < 0.1). The coefficient indicating indirect effect is insignificant. Since gasoline is the main fuel of vehicles and most of the gasoline is consumed in the transportation sector, the results indicate that FDI's effect in the transportation sector is not as big or significant as that in the industrial sector. This article further estimates FDI's effect on highway passenger volume and highway freight volume, and the results support the statement.
This article further examines the correlation between FDI and average wage, and between FDI and ecological construction. The coefficients are insignificant, or significant at a low level (0.05 < p < 0.1). Therefore, EC acts more like a moderator instead of a channel variable.

Conclusions and Policy Implications
Between 2013 and 2018, FDI showed a reverse correlation with local air pollution, which supports the existence of the pollution halo hypothesis. This finding is consistent with multiple articles that study the existence of PHH in China and other countries [11,12]. However, there is a direct correlation between FDI and air pollution in the outer conurbation areas. Some literature does report the spatial pollution spillover of FDI while reporting the pollution haven effect of FDI [9]. However, the existing literature has not reported the finding, by these authors, that FDI shows direct correlation with air pollution in the outer conurbation areas while it has a positive effect on local air quality. Therefore, this finding has become the main interest of this article.
In the heterogeneity analysis, this article finds that the industrialization stage, ecological construction level, and technology development level have a significant moderating effect on FDI's pollution effect. FDI tends to show a higher marginal effect on air pollution in cities with a higher tertiary-to-secondary industry ratio. According to the statistical data, cities with higher IS are mostly tourist destination cities. The common characteristic of these cities is that the background air pollution concentration is small. Therefore, the moderating effect of IS is consistent with the marginal diminishing relationship between emissions and air pollution. In addition, this article finds that FDI tends to show a lower marginal effect on air pollution in cities with higher technology development levels and ecological construction levels.
In the analysis of potential channels through which FDI affects air quality, the coefficients of FDI on coal consumption and the thermal power generation amount are consistent with the main findings; meanwhile, the coefficients of FDI on variables from transportation sectors and urban-specific characteristics are insignificant or significant at a low level (0.05 < p<0.01). These results further validate the main findings by providing a potential channel where FDI shows direct correlation with coal consumption and the thermal generation amount of the outer conurbation areas. One explanation is that FDI increases the electricity consumption of host cities, which are usually big cities that buy electricity from the nearby cities. In addition, the direct correlation between FDI and local technological development supports that FDI may improve the air quality by bringing advanced technology, which is the main mechanism of the pollution halo hypothesis stated in the existing literature.
The findings generally support the existence of the pollution halo effect of FDI in China in the past few years and provide primary evidence of FDI's direct correlation with air pollution concentration of the outer conurbation areas. The policy implications are mainly on the aspect of regional cooperation in green development regarding FDI. While FDI brings advanced technology into host cities, the urban and environmental managers should pay attention to the direct correlation between FDI and air pollution of the outer conurbation areas, especially in the aspect of the spatial pattern of the power supply. Regional heterogeneity: One of the striking findings of the paper shows that FDI has no significant impact on air quality in the central and western cities. The reason is that low level of FDI inflows to cities located in the center and west.  Appendix C Note: both the dynamic factor and spatial factor are significant at the 1% level, which supports the inclusion of the spatial and dynamic terms. However, the dynamic term is automatically omitted by STATA in the regression of lnNO 2 due to collinearity. The larger values of R 2 of regressions regarding lnPM 2.5 and lnSO 2 indicate the dynamic SDM has a better explanation on these two pollutants. Robust standard errors are in parentheses. *** p < 0.01, ** p < 0.05.