3.1. Model specification

In this section we provide the specification of the estimated model. In contrast to previous research, we proxy trade levels by not only the degree of openness and but also the proportion of exporters at the city level. We follow the theoretical work of Copeland and Taylor (Reference Copeland and Taylor1994, Reference Copeland and Taylor1995, Reference Copeland and Taylor2005) to study how economic activities, trade in particular, can impact environmental quality through three channels: scale, composition and technique effects. We adapt the model to the case of China as follows:

(1)\begin{equation} \hat{P}_{it}=\hat{S}_{it}+\hat{\varphi}_{it}+\hat{e}_{it}+\hat{\mathbf{Z}} _{it}, \end{equation}

where $\hat {P}_{it}$ represents the emissions in Chinese city $i$ in year $t$. $\hat {S}_{it}$ captures the scale effect. The expansion of economic activities that is induced by trade in China generates more consumption of natural resources, which can further increase the emissions in city $i$ in year $t$. The composition effect is captured by the second term, $\hat {\varphi }_{it}$, representing the output composition. If countries produce a high ratio of pollution-intensive goods, environmental degradation is triggered. However, an environmentally-friendly result can occur if countries produce less-polluting goods. It is not a priori clear what the effect of trade for the case of China is. The third term $\hat {e}_{it}$ in the model indicates the technique effect. Multinational firms run operations with innovative technologies and advanced management, and further on trade promotes the entry of these firms and their advantages. Higher level technologies reduce emissions and enable clean industrial production, potentially reducing pollution in China. Moreover, $\hat {\mathbf {Z}}_{it}$ represents other economic controls related to the trade effect on the environment for Chinese cities.

In our model specification, the annual concentration of particulate matter 2.5 (PM2.5) suspended in the air at the city level is the dependent variable used to identify environmental quality. We use this pollutant for two reasons: first, it has the most severe effect on public health in China (Cao et al., Reference Cao, Liang and Niu2017); and second, these data are available at the city level.Footnote ³

Including further control variables, our final specification is given by:

(2)\begin{align} EmissionDen_{it}=&\,\beta _{0}+\beta _{1}({industry/GDP})_{it}+\beta _{2}(K/L)_{it}\nonumber\\ &+\beta _{3}ln(FDI_{it})+\beta _{4}Trade_{it} \nonumber\\ &+\beta _{5}ln(Wage_{it})+\beta _{6}ln(EleComsumption_{it})+\varepsilon _{it}, \end{align}

where $ln(.)$ denotes the natural logarithm. The dependent variable ($EmissionDen_{it}$ in the regression) is the emissions of PM2.5 per square km in city $i$ in year $t$.

We include income ($Wage_{it}$ in the regression) as an explanary variable in our analysis, to capture the scale effect (Copeland and Taylor, Reference Copeland and Taylor2005; Cole et al., Reference Cole, Elliott and Zhang2011, Baghdadi et al., Reference Baghdadi, Martinez-Zarzoso and Zitouna2013; among others).

The composition effect is captured by the share of the industrial output to GDP ($(industry/GDP)_{it}$ in the regression) in our analysis as in previous studies (He and Richard, Reference He and Richard2010; Aller et al., Reference Aller, Ductor and Herrerias2015; among others). As China is a manufacturing country, emissions in the secondary sector are considered to be the main source of China's pollution.

We further control for the technique effect by including foreign direct investment ($FDI_{it}$ in the regression) in our specification. The role of investment in determining the pollution level has been identified by Cole et al. (Reference Cole, Elliott and Zhang2011). Since multinational firms have advanced technologies and efficient management in clean production, they bring these advantages with them when they invest/relocate in developing countries, with a beneficial impact on environmental quality. This finding is also confirmed by the study of Zhu et al. (Reference Zhu, Duan, Guo and Yu2016) who find that FDI helps to reduce carbon emissions, especially in low- and middle- income countries. A contrasting hypothesis is discussed by Antweiler et al. (Reference Antweiler, Copeland and Taylor2001); namely that FDI could harm the environment. Specifically, this theory is the PHH: multinational firms run their operations in countries with lax environmental policies to relocate pollution-intensive industries there, in order to have a lower abatement cost.

Poncet et al. (Reference Poncet, Hering and Sousa2015) employ the ratio of capital to labor as a further control to investigate the impact of factor endowment on emissions. Other studies such as Managi et al. (Reference Managi, Hibiki and Tsurumi2009) and Cole and Elliott (Reference Cole and Elliott2003) also include this variable in their model to examine the trade–emission nexus. We follow their method to measure the factor endowment ($({K}/{L})_{it}$ in the regression) for Chinese cities.

Energy consumption is also considered as the main source of emissions, as proven in the previous literature (Cole et al., Reference Cole, Rayner and Bates1997; Poncet et al., Reference Poncet, Hering and Sousa2015; Zhu et al., Reference Zhu, Duan, Guo and Yu2016; among others). We proxy the energy consumption by using the consumption of electricity ($EleComsumption_{it}$ in the regression) in each Chinese city for each year.

We use two variables to proxy trade intensity and evaluate the impact of trade on pollution in this study. The first variable is the annual trade openness index at the city level, and the second is the proportion of exporters in each city for each year. The openness index, computed as the trade intensity (the ratio of the sum of exports and imports to GDP) in city $i$ in year $t$, is included as $Openness_{it}$ in our specification.

Furthermore, we use a second variable, $Exporter_{kit}$, in our analysis: the ratio of exporting firms to total firms in city $i$ in year $i$. We consider this rate in two stages. First, we include the proportion of exporters in order to understand whether the increasing exporters can affect the environmental quality of Chinese cities. A high ratio of exporters to total local firms shows the city's trade intensity, captured at the plant level. Holladay (Reference Holladay2016) claims that, for the same level of output, exporters pollute less than non-exporters. Thus, we hypothesize that, at the same GDP level, cities with more exporters may have lower emissions. Therefore, the exporters in different industries might be the source for this link. To study this aspect in more depth, we divide exporters by different industrial sectors (mining, pollution-intensive manufacturing, less-polluting manufacturing, and electricity sectors). These rates are computed as follows:

(3)\begin{equation} Exporter_{kit}=\frac{X_{kit}}{N_{kit}}, \end{equation}

where $Exporter_{kit}$ is the proportion of exporters in sector $k$ in city $i$ in year $t$; $X_{kit}$ denotes the number of exporting plants in sector $k$ in city $i$ in year $t$; and $N_{kit}$ is the total number of firms in sector $k$ in city $i$ in year $t$.

Finally, we split the industrial sectors into four different categories according to their emissions. The pollution information about major industrial emissions data for 41 industries in China is shown in figures A1 and A2 in the online appendix. The emissions related to industrial smoke dust for each manufacturing industry are reported in figure A1. The most polluting industries in this regard are: non-metallic manufacturing; electricity, heating production and supply; smelting and pressing ferrous metal manufacturing; chemical manufacturing; petroleum, coal and other fuel processing; and non-ferrous metal manufacturing. The emissions of industrial solids are reported in figure A2, indicating substantial emissions in some industries: electricity, heating production and supply; smelting and pressing ferrous metal manufacturing; coal and mining industry; ferrous metal mining industry; and non-ferrous metal mining industry. From the information above, we identify the following as pollution-intensive manufacturing industries: paper manufacturing; petroleum, coal and other fuel processing; chemistry manufacturing; non-metallic mining manufacturing; smelting and pressing ferrous metal manufacturing; and smelting and pressing non-ferrous metal manufacturing.Footnote ⁴ Combining the pollution information, we can further divide the industries that are included in the secondary sectors into four categories: the pollution-intensive manufacturing sector, the less-polluting manufacturing sector, the mining sector and the electricity sector. Last, we calculate the rate of exporters introduced above for each of these categories separately. These variables are $ExpoterMin_{it}$, $ExpoterMaPo_{it}$, $ExpoterMaNonpo_{it}$, and $ExpoterEle_{it}$ for the proportion of exporters in the mining, pollution-intensive manufacturing, less-polluting manufacturing and electricity sectors, respectively.

An endogeneity problem with the income variable is also mentioned in the previous literature. For instance, potential endogeneity issues are raised by income–environmental policy and the income–trade link (Frankel and Rose, Reference Frankel and Rose2005; Baghdadi et al., Reference Baghdadi, Martinez-Zarzoso and Zitouna2013, among others). To avoid this issue, we adopt an IV strategy by employing a set of variables to mitigate the endogeneity by following the measure of Baghdadi et al. (Reference Baghdadi, Martinez-Zarzoso and Zitouna2013).Footnote ⁵ Finally, the instruments that we adopt in this paper are the one-year lagged of income, population and the enrollment rates of primary school and university students.

3.2. Data

The data used cover the period from 1998 to 2007. Three levels of data are employed: pollution data, city-level administration data and firm-level data. They are drawn from different datasets. Descriptive statistics are presented in table 1. The correlation information for the data can also be found in table A1 (online appendix).

Table 1.

Descriptive statistics

Notes: The data for each region is also reported in the table. West, Central, and East refers to the data for the cities located in the western, central, and eastern regions.

Our data on the pollutant PM2.5 are taken from the Socioeconomic Data and Applications Center (SEDAC), which is managed by the NASA Earth Science Data and Information System project.Footnote ⁶ The emission measurement used in this paper is the annual concentration of the PM2.5 value in 287 Chinese cities. It consists of annual concentrations (micrograms per cubic meter) of ground-level fine particulate matter (PM2.5), with dust and sea salt removed.

The prefecture-level data on 287 Chinese cities are drawn from China's City Statistic Yearbook. The variables available at this level include GDP, population, city area, investment, education and wage. City name and location information (western, central and eastern regions) are shown in table A2 (online appendix). Moreover, the data on the total volume of trade (the sum of the export and import values) are drawn from the Local Statistical Yearbook in each city. We compute the ratio of the sum of the export and import values to GDP by combining these two datasets.

Precise industry data are not available at the city level. We aggregate the firm-level data in order to proxy the city-level industry data. These firm data are drawn from the annual surveys from the Chinese National Bureau of Statistics. This dataset includes the firms in China that have total sales of more than 5 million Chinese Yuan, which is the so-called ‘above-scale’. These data include a four-digit Chinese Industry Classification (CIC) code, as well as location, export value, founding year and value-added, for each of these firms. Data on the number of different firms in each city in each year are drawn and aggregated from this level.

4.1. Openness

Table 2 reports the results for the impact of openness on emissions. We first employ fixed effects by controlling for the city- and time-invariant impact. These results are shown in the first two columns. Column (1) reports the results for the full sample of Chinese cities. The positive sign of the openness variable suggests that the high intensity of trade activities has a detrimental effect on the environment by increasing emissions of PM2.5. The magnitude of the coefficient indicates that a one-unit increase in openness, ceteris paribus, would lead to an average increase of ${31.17}~\mu {\rm g/m}^3$ of PM2.5 in every square km for the cities in China. The region-specific openness effects are reported in column (2). From these results, we find that the detrimental impact of trade mainly happens in the cities of the western region, since only those cities show that trade has a statistically significant effect on environmental quality. The coefficient indicates that a one-unit increase in openness, all else being equal, would lead an increase of $39.32~\mu {\rm g/m}^{3}$ of PM2.5 in every square km for the cities located in the western region on average. As a comparison, we then perform the random effects by including year dummies within our model, and the results are shown in columns (3) and (4).Footnote ⁷ Column (3) reports the results for all Chinese cities on average. The coefficient for openness shows a negative sign, nevertheless, it is insignificant. We observe that, when we count for the region-specific openness effect, as the results show in column (4), the variables for the three regions all have negative signs. However, only the coefficient for the cities in the eastern region is statistically significant. This result indicates that, all else being equal, a unit increase in openness would reduce PM2.5 per square km by $337~\mu {\rm g/m}^{3}$ on average for cities in the eastern region of China. Columns (5) and (6) show the result of the estimation that controls for the income variable by using the IV strategy. Although all the tests show the validity of the instrument, the results for the full sample show that the coefficient for openness in column (5) is not statistically significant. The results for the openness effect in different regions are reported in column (6). The coefficients for the western and the central regions are not statistically significant. Openness has a beneficial impact on the environment for the cities located in the eastern region. The magnitude of coefficient indicates that, all else being equal, a unit increase in openness in the eastern region would reduce an average of $258.3~\mu {\rm g/m}^{3}$ of PM2.5 per square km. All in all, we find that trade openness has a detrimental impact on environmental quality, which is different from the conclusion of Poncet et al. (Reference Poncet, Hering and Sousa2015). However, a recent study by Afesorgbor and Demena (Reference Afesorgbor and Demena2019) finds that most of the papers reach a conclusion that trade increases pollution, and this impact can vary depending on different pollutants.

Table 2.

The impact of openness

Notes: Robust standard errors in parentheses.

***, **, *, + denote significance at the 0.1, 1, 5 and 10% level, respectively.

Dependent variable : the concentration of PM2.5 per square kilometer in city $i$ in year $t.$

The underidentification test is based on an LM version of the Anderson (Reference Anderson1951) canonical correlation test and its p-value (Chi-sq(4)) is less than 0.1, indicating that the test rejects the null hypothesis that the equation is underidentified. The weak identification test is based on the Cragg-Donald Wald F statistic. The F-statistic is above 10, which indicates the validity of the instrument. The Sargan-Hansen test is the overidentification test of all instruments. The p-value (Chi-sq(3)) is greater than 0.1, meaning that the instruments are valid; otherwise, the instruments would not be valid.

The other controls have the expected signs and are statistically significant overall. For the variable that captures industry composition, the positive signs show that the higher the proportion of industrial output to total domestic output, the more polluted is the environment. The factor endowment, which is proxied by the ratio of capital to labor, is also positive. This result indicates that those cities that are relatively endowed with capital have higher emissions. Turning to the results for FDI, they are positive and statistically significant when fixed and random effects are employed. The positive signs for FDI might prove the existence of the PHE for Chinese cities. The low cost of production regarding the environmental policy or technology standards in China attracts foreign investment; thus, these multinationals do not have the positive impact of improving the local environment. This result is in line with the research of Cole et al. (Reference Cole, Elliott and Zhang2011). Furthermore, when we control for fixed effects, the coefficient of electricity consumption is negative, which could be explained by the fact that the consumption of electricity is not the main source of PM2.5 emissions. Finally, the coefficients for the wage are positive, which means that the environment deteriorates as the local income level increases.

4.2. The impact of exporters

In the previous subsection, we show that openness can have a detrimental impact on the environment for Chinese cities, on average, and this effect is heterogeneous across regions. We suppose that a city's specialization in a sector, such as manufacturing, might have a more detrimental impact on the environment than the other sectors. Hence, it is interesting to analyze how exporters could affect environmental quality.

Exporters can have a different impact on the environment compared to local firms (Holladay, Reference Holladay2016). In this subsection, the proportion of exporters is used to evaluate the impact of trade on emissions from the perspective of firms. We first compute the ratio of number of exporters to total firms for each Chinese city using the industrial firm-level data (in the secondary sector) and then split exporters into different sectors $k$ in city $i$ in year $t$.

Table 3 reports the impact of the proportion of exporters on environmental quality at the city level. Columns (1), (3), and (5) report the results from employing respectively the fixed-effects, random-effects and IV methodology for the full sample of Chinese cities. However, the coefficients for $Exporter_{it}$ in these columns show no statistical significance. When we focus on the region-specific exporters effects, we find the heterogeneous effects among regions. These results are shown in columns (2), (4) and (6). The second column reports the impact of the proportion of exporters on the environment by controlling the time and city fixed effects. The coefficient of exporters shows a positive sign for those cities located in the western region, which indicates that, when the proportion of exporters in the western region is higher, so are emissions in the region. Meanwhile, the coefficient of the proportion of exporters of the central region has a negative sign, meaning that more exporters in this region can reduce the local emissions of PM2.5. The effect in the eastern region does not show any statistical significance. The results across regions by applying random effects and including the year dummies are reported in column (4). The coefficient for the cities in the western region shows a positive sign, which indicates that more exporters in this region can increase emissions. Then, the sign of coefficient for exporters in the central region is statistically insignificant. The effect of the eastern region shows a beneficial impact on the environment. Finally, the result of an estimate that instruments the income variable is reported in column (6). In this case, the coefficients on the exporters for the three regions are not statistically significant. All in all, the high proportion of exporters in cities located in the western region has a harmful impact on the environment. Moreover, having more exporting firms in the central and eastern regions can reduce the emissions in these areas.

Table 3.

The impact of exporters

Notes: Robust standard errors in parentheses.

***, **, *, + denote significance at the 0.1, 1, 5 and 10% level, respectively.

Dependent variable : the concentration of PM2.5 per square kilometer in city $i$ in year $t.$

The underidentification test is based on an LM version of the Anderson (Reference Anderson1951) canonical correlation test and its p-value (Chi-sq(4)) is less than 0.1, indicating that the test rejects the null hypothesis that the equation is underidentified. The weak identification test is based on the Cragg-Donald Wald F statistic. The F-statistic is above 10, which indicates the validity of the instrument. The Sargan-Hansen test is the overidentification test of all instruments. The p-value (Chi-sq(3)) is greater than 0.1, meaning that the instruments are valid; otherwise, the instruments would not be valid.

Exporters have an uncertain impact on the environmental quality, on average, for the full sample of Chinese cities. Moreover, this effect varies across the regions. To examine the impact of exporters in more depth, we split the industries into four sectoral categories by using firm-level data. Doing so provides us with further information on exporters’ environmental performances across industries. We assume that if there are more exporters in ‘dirty’ sectors, the region has a comparative advantage in producing ‘dirty’ goods. Eskeland and Harrison (Reference Eskeland and Harrison2003) discuss how foreign investors prefer to invest in those countries with a low cost of the factors they need, which means that foreign investment/exporting firms might have preference for pollution-intensive or less-polluting sectors. Hence, it could be of interest to investigate the impact of exporters in different sectors. These results are reported in table 4. The variables in this estimation – $MinExporter_{it}$, $PoExporter_{it}$, $NonpExporter$ and $EleExporter$ – capture the ratio of exporters to total firms in the mining, pollution-intensive manufacturing, less-polluting manufacturing and electricity sectors, respectively. The coefficients are not statistically significant for the full sample of Chinese cities, and they do not show any statistical significance for cities in the central region, although they do show such significance for cities located in the western and eastern regions. For cities in the western region, the results show that a higher proportion of exporters in the less-polluting manufacturing sector increases the local emissions of PM2.5. Regarding the results of the eastern region, the variables show that more exporters in the mining sector could mitigate the emissions. Furthermore, exporters in the less-polluting manufacturing sector can have a larger positive impact on environmental quality than the other sectors, as the magnitude of the coefficient shows.

Table 4.

The impact of the exporters’ rate in the four sectors

Notes: Robust standard errors in parentheses.

***, **, *, + denote significance at the 0.1, 1, 5 and 10% level, respectively.

Dependent variable : the concentration of PM2.5 per square kilometer in city $i$ in year $t.$

5.1. The environmental Kuznets curve

As the first robustness check, we include income squared in the regressions to examine the nonlinear relationship between income and pollution. In the previous section, we highlighted the importance of the income effect, which is also linked to the EKC. Again, this relation is that, at the early stage of development, pollution increases as income increases. Once the local income reaches a certain amount (which varies across countries), called a turning point, emissions decrease when income increases due to an awareness in society of the importance of environmental protection. Nevertheless, the EKC effect does not necessarily exist in the case of all pollutants. For example, Carson (Reference Carson2010) finds that there is no EKC effects for the emissions of ${\rm CO}_{2}$; Bradford et al. (Reference Bradford, Fender, Shore and Wagner2005) confirm this finding by examining several emissions.

To understand whether there is an EKC effect on PM2.5 for Chinese cities, we include an additional variable, wage squared. Our model is as follows:

(4)\begin{align} EmissionDen_{it}^{\prime }=&\,\beta _{0}+\beta _{1}^{\prime }({industry/GDP} )+\beta _{2}^{\prime }(K/L)_{it}+\beta _{3}^{\prime }ln(FDI_{it})+\beta _{4}^{\prime }Trade_{it} \nonumber\\ &+\beta _{5}^{\prime }ln(Wage_{it})+\beta _{6}^{\prime }(lnWage_{it-1})^{2}+\beta _{7}^{\prime }ln(EleComsumption_{it})+\varepsilon _{it}^{\prime } \end{align}

The results of testing for an EKC effect for Chinese cities are shown in tables A3 and A4 (in the online appendix). Table A3 shows the impact of openness on environmental quality, which allows us to test whether the EKC effect is present in China. In columns (1), (2) and (3), we include fixed and random effects, and use an IV strategy, respectively. Focusing on the variables of wage and its square term, we find an EKC effect for Chinese cities, since the coefficients of wage report positive signs and its squared terms are negative for three cases. These results indicate an inverted-U curve for the income-emissions relationship. The EKC effect, shown through the impact of exporters on emissions, is shown in table A4. Using the same method as in table A3, columns (1), (2) and (3) of table A4 report the results of the fixed effects, random effects, and IV estimations of our regression, respectively. The results also show that the coefficients for the wage term are positive, and those for its squared term show a negative sign. From these results, we cannot reject the EKC hypothesis for Chinese cities concerning PM2.5 emissions. Finally, the impact of trade on environmental quality is in line with the baseline results.

5.2. Additional variables

Some additional variables can be interesting to include to control for scale and manufacturing size. We use a set of additional variables in our second robustness check. As argued by Cherniwchan et al. (Reference Cherniwchan, Copeland and Taylor2017), the value-added produced by firms can be translated into the scale of industry output. We aggregate the value-added data at the firm level for the four industrial sectors classified in this paper. The results are shown in tables A5 and A6 of the online appendix.

Table A5 reports the results of our specification by adding the value-added produced in different industrial sectors, through the impact of openness on the environment. However, the openness variable shows a significance only in column (1), when we include fixed effects. This result confirms the detrimental impact of openness on environmental quality for Chinese cities as in our baseline results. The results employing random effects and IV methodology are presented in columns (2) and (3); they are not statistically significant. Furthermore, the additional variables are $Avmining_{it}$, $AvMapo_{it}$, $AvManonpo_{it}$ and $AvElec_{it}$, representing the value added in the mining, pollution-intensive manufacturing, less-polluting manufacturing and electricity sectors, respectively. The variables of value-added produced in the mining and pollution-intensive manufacturing sectors do not show statistical significance. Meanwhile, the coefficients of $AvManonpo_{it}$ and $AvElec_{it}$ report negative signs. This result indicates that the higher value-added produced in the less-polluting manufacturing and electricity sectors can improve environmental quality in terms of PM2.5 emissions for Chinese cities.

The results of adding the value-added variables to investigate the impact of exporters on PM2.5 emission are presented in table A6. We study this issue by including the fixed and random effects, as well as implementing an IV strategy; the results are reported respectively in columns (1), (2), and (3). As in our baseline results, the impact of exporters remains inconclusive, as they do not show any statistical significance. Regarding the additional variables, the coefficients of $AvManonpo_{it}$ and $AvElec_{it}$ report negative signs, which have the same results as in table A5.

5.3. Endogeneity issue

Although we have controlled the endogeneity of the income variable, some argue that the variable of trade should be instrumented as well. As Frankel and Rose (Reference Frankel and Rose2005), Managi et al. (Reference Managi, Hibiki and Tsurumi2009) and Baghdadi et al. (Reference Baghdadi, Martinez-Zarzoso and Zitouna2013) argue, another endogeneity issue arises between trade and income since trade increases the income levels; conversely, citizens with higher incomes will demand more exports and imports. The general strategy is to control the income variable by using the growth model and the trade variable with the gravity model. Since the interregional data for the Chinese case are lacking, we follow the studies of Chintrakarn and Millimet (Reference Chintrakarn and Millimet2006) and Managi et al. (Reference Managi, Hibiki and Tsurumi2009) to employ a GMM estimation by employing the method of Arellano and Bond (Reference Arellano and Bond1991). We instrument both the income and trade variables by using their lags. The results are reported in table A7. The first column shows the results of the impact of trade openness on the environment. The coefficient for openness is negative but it is not statistically significant. The impact of the exporters is reported in column (2). The coefficient for exporters shows also a negative sign and insignificance. Precise information regarding the sector is provided in column (3). The only significant value for the less-polluting manufacturing sector shows that a larger proportion of exporters in this sector increases the PM2.5 emissions for Chinese cities. For the other controls, the results are not always statistically significant. The corresponding tests indicate that first order autocorrelation is present in the data, whereas second order autocorrelation is not, as expected in this setting. Regarding the Hansen tests, they cannot reject the null hypothesis that the instruments are valid according to the results in columns (2) and (3) in table A7.