4.1. Spatial Dependence Tests
We used Global Moran’s
I values to disclose the degree of autocorrelation for variables throughout the study region [
33]. Global Moran’s
I values fall between −1 and 1; when
I is greater than 0, positive spatial dependence is present, with higher values of Moran’s
I corresponding to a stronger positive spatial effect, and vice versa.
Table 3 presents our test results of Moran’s
I for global spatial dependence of the amount of pollutant emissions under the three different spatial matrices (i.e.,
Wpro,
Weco, and
Wpe). The Moran’s
I values are all above 0 and statistically significant at a 90% significance level or higher, thus implying the existence of positive spatial autocorrelations in the amounts of both the industrial SO
2 and CO
2 emissions in cities belonging to the same province (weighted by
Wpro), or with similar economic levels (weighted by
Weco). The results demonstrate that both the industrial SO
2 and CO
2 emission levels for all cities are not completely random but instead are in a positively related state of spatial dependence. This relationship will be biased if the estimation models are constructed without spatial effects [
34,
35].
Moreover, for the industrial SO2 emissions, the level of spatial dependence between cities with similar economic development is relatively low (the values of Moran’s I range from 0.033 to 0.154). At the same time, when the province factor is incorporated and hereafter the nested matrix Wpe is obtained, the value of Moran’s I for each year turns out to be greater than those under the Wpro or Weco. The results therefore reveal that the spatial dependence of SO2 emissions is much more apparent among cities that belong to the same province and are at similar economic levels, thus implying a much stronger IGC among them. The values of Moran’s I for CO2 emissions are also higher under the nested matrix Wpe than those under the other two matrices. We can therefore conclude that both of the spatial autocorrelations separately weighted by Wpro and Weco should be considered in order to capture the spatial dependence of pollutant emissions. Accordingly, the nested matrix Wpe was applied in the subsequent spatial estimations.
4.2. Spatial Econometric Estimation Results
For comparison purposes, the panel data models both without and with the spatial dependence of city-level SO
2 emissions (Case 1) and CO
2 emissions (Case 2) are estimated, and the results are presented in
Table 4. We first report the estimation results when spatial effects are not accounted for. The LM test suggests that either the fixed-effects (FE) or random-effects model is more appropriate than the classical model (not reported) in both cases. Furthermore, the Hausman test suggests that the FE model is better than the random-effects model. Finally, the LR test concludes that the two-way FE model is better than the one-way FE model. As the table shows, in both cases, the coefficient of ln
PGDP is consistently significantly positive and the coefficient of ln
PGDP2 is significantly negative, thus proffering robust evidence that an inverted U-shaped relationship between economic growth and pollutant emissions does exist. On the basis of the results reported in first column, the logarithm of per capita GDP corresponding to the peak of SO
2 emissions is 10.4464, indicating that SO
2 emissions in China have crossed the threshold and proceeded to the downward stage as a whole, which supports the findings of Jiang et al. [
11], Zhang et al. [
31], and Zhao et al. [
12]. However, the ln
PGDP at the peak of CO
2 emissions is 11.8304, indicating that carbon emissions in most of cities are still increasing rapidly. Similar evidence can be found from Kang et al. [
19], Wang and Ye [
15], and Yin et al. [
3].
According to Anselin [
34] and Pinkse and Slade [
35], given the condition that the spatial lagged and error correlations do exist, there will be serious consequences from ignoring these spatial correlations. If spatial lag dependence is ignored, ordinary least squares (OLS) estimators will be biased and inconsistent. If spatial error dependence is ignored, OLS estimators will be unbiased but inefficient, and the standard errors of the estimators will be biased. We report the results from verifying spatial dependence in
Table 5. According to Elhorst [
36], the significant results of the (robust) LM test response testify to the existence of spatial correlations, so an empirical analysis that does not consider the spatial effects would be subject to bias and hence not reliable [
2,
31]. The results of the Wald and LR tests are presented in the last two columns, and they reject the hypotheses of
θ = 0 and
θ = –
βρ, respectively, thus indicating that the SDM model is more suitable than either the SLM model or SEM model for this analysis [
30].
Estimation results from spatial Durbin models are reported in the last two columns of
Table 4. The Hausman and LR tests suggest that the two-way FE SDM is the most appropriate model in both cases. As can be seen in the first row, the estimated coefficients of lagged dependent variables (
Rho) are all positive at the 1% significant level in both cases, thus revealing that both the industrial SO
2 and CO
2 emissions of a certain city are likely to be influenced by other competing cities. The estimated coefficients of ln
PGDP and its quadratic term ln
PGDP2 are significantly positive and negative, respectively, verifying the validity of the EKC hypothesis. The product terms of the explanatory variables and spatial weights matrix
Wpe in the SDM reflect how these explanatory variables in other correlated cities affect the pollutants emissions of a certain city. In case 1, the coefficients of
Wpe*ln
PGDP and
Wpe*ln
PGDP2 are significantly positive and negative, respectively, indicating that a city’s industrial SO
2 emissions are sensitively correlated to its competing city’s economic levels in an inverted U-shaped way. In contrast, as the results in case 2 show, there appear to be no significant spatial correlations in CO
2 emissions. We can conclude, therefore, that the impacts of IGC exist only for the industrial SO
2 emissions, and they cannot be identified in the city-level CO
2 emissions––which is consistent with the factual background that only the industrial SO
2 emissions, and not the CO
2 emissions, have been set as the indicator for assessing the environmental performance of local officials.
Table 4 also reports the estimation coefficients of the control variables. One city’s population density (ln
PD) is not only negatively correlated with its SO
2 emissions, but it is also negatively impacted by its competing cities’ sulfur emissions, further supporting the spatial interaction effects on SO
2 emissions among local governments. In contrast, the CO
2 emissions have been significantly increasing with rising population density, but present no spillover effects across cities. Although the level of regional innovation and entrepreneurship (ln
RIE) is found to exert a positive but insignificant impact on pollutant emissions, a significant spillover effect of technology aggregation is found in the reduction of CO
2 emissions but not in the reduction of SO
2 emissions. We can infer then that the IGC regarding environmental indicators (e.g., the SO
2 emissions) would have handicapped the distribution of technologies designed for cleaner production and emissions reduction among local governments, which is consistent with the arguments of Eaton and Kostka [
23].
Table 6 reports the estimation results from the SDM models conducted based on the regional divisions (i.e., the eastern, central, and western regions). Because the diagnostic results suggest that the two-way FE SDM is the best fit, we used it to limit the interpretation of coefficient estimates. The results presented in first row show that the indicator
Rho is positive at the 1% significant level in both cases across all the regions, indicating a pervasive spatial dependence of pollutants emissions across cities that present high degrees of IGC. The estimated coefficients of ln
PGDP and ln
PGDP2 suggest that the EKC hypothesis is valid only in the eastern and central regions, whereas it does not operate in the western region. The results in the final column even reveal a significant positive correlation between economic growth and CO
2 emissions. In terms of the spillover effects, the estimated coefficient of
Wpe*ln
PGDP demonstrates that a western city’s CO
2 emissions are negatively correlated with its competing city’s economic development. The results jointly reveal that carbon emissions have been increasing precipitously with economic growth in western Chinese regions, and that siphoning effects of economic factors in urban development are also reflected in carbon emissions.
It is essential to use factor analysis to estimate the value change of the economic variable at the EKC turning point, so as to evaluate the effects of IGC on the income–pollution correlations. To this end, using the results of SDMs, we further calculate the direct, indirect, and total marginal effects of different influencing factors on SO
2 and CO
2 emissions for each regional division. The coefficients of ln
PGDP and ln
PGDP2 are reported in
Table 7, and the resulting economy–emissions nexuses are plotted in
Figure 1. Estimation results without spatial dependence are derived from two-way FE models, whereas estimation results identifying spatial effects are derived from the total effects in two-way FE SDMs. In
Figure 1, T1 marks the per capita GDP corresponding to the turning point of EKC with accounting for the spatial effects, and T2 marks the GDP at the turning point without accounting for the spatial effects.
As is shown in the first and fifth columns of
Table 7 and panels A and B of
Figure 1, when taking all of the cities as the sample, the EKC hypothesis is consistently valid in both cases. The level of per capita GDP corresponding to the peak of CO
2 emissions is higher than that of SO
2 emissions, whether or not the spatial effects are considered. Those results demonstrate that compared with the SO
2 emissions, which have already advanced into the downward stage, much greater economic costs will have to be paid to reach the turning point of the CO
2 emissions. Furthermore, in Panel A of
Figure 1, T1 is lower than T2, revealing that the per capita GDP corresponding to the peak of SO
2 emissions is estimated to be lower when the spatial effects are fully accounted for than when they are not. In contrast, in Panel B, T1 is greater than T2, which demonstrates that the spatial interaction has a negative effect on the reduction of carbon dioxide emissions. These results indicate that the IGC has significantly reduced the SO
2 emissions but has promoted the CO
2 emissions, which could be due to the fact that during the research period the reduction of SO
2 emissions had been set as a veto assessment indicator in the official promotion tournament, while the reduction of CO
2 emissions had not.
We also report the estimated results for different regional divisions in
Table 7, and we illustrate them in panels C–H in
Figure 1. Because the SO
2 emissions reduction was set as assessment indicator in 2007, the municipal governments in all three regional divisions gradually formulated positive IGCs that target the mutual performance of SO
2 emissions reduction and lead to a lower per capita GDP corresponding to the threshold of EKC than that when the spatial effects are not accounted for. Those results indicate that by assessing official performance with environmental indicators, for SO
2 emissions, we can achieve the turning point of EKC at a relatively low economic cost––a cost that at least is not as high as that in the eastern cities. In contrast, although the EKC hypothesis holds for CO
2 emissions in the eastern and central regions, in the western region there is an obvious upward trend that seems to be much more pronounced when spatial dependence is considered. This means that the western cities will experience a persistent and harsh increase of carbon dioxide emissions if there are no specific intervention policies in place. Kang et al. [
19], He and Lin [
4], Yin et al. [
3], and many others have provided similar evidence.