STIRPAT model, one of the most representative methods in environmental impact assessment, is widely used to study CO₂ emissions and the influencing factors (Li et al., 2016; Yi et al., 2019; Yue et al., 2022; Ziyuan et al., 2022). STIRPAT model overcomes the limitations of single elasticity in the IPAT model (Dietz and Rosa, 1994; Huo et al., 2021). Compared with the LMDI method, the STIRPAT model allows estimating the impact of each factor as a parameter, which makes up for the deficiency that the LMDI model cannot measure elasticity (Cai et al., 2017; Chen et al., 2018). The STIRPAT model is as follows: I i t = α P i t b × A i t c × T i t d × e i t (1) Where I represents environmental impact; P denotes population; A represents affluence level; T represents technical level; a , b , c , and d are the parameters. e is the random disturbance. The logarithmic form of the model is: ln ⁡ I i t = α + b ln ( P i t ) + c ln ( A i t ) + d ln ( T i t ) + e i t (2)

The fixed-effect panel threshold model is proposed by Hansen, (1999), which is used to describe the jumping character or structural break in the relationship between variables (Hansen, 1999; Hansen, 2000). The setting of the single-threshold model is as follows: y i t = μ i + β 1 x i t I ( q i t ≤ γ ) + β 2 x i t I ( q i t > γ ) + e i t (3) Where i and t represents cities and years. y i t is the dependent variable; x i t is the explanatory variable; q i t is the threshold variable; μ i is the individual effect; e i t ∼ i i d ( 0 , δ 2 ) is the random disturbance. I ( · ) is the indicator function. The γ is a specific threshold value. β 1 and β 2 are the parameters to be estimated, respectively.

In this study, the extended STIRPAT model, including the urban population size, industrial structure, technical level, per capita GDP, and foreign direct investment, are used to investigate their effects on CO₂ emissions. By choosing the urban population size and industrial structure as the threshold variables, respectively, the threshold-STIPPAT models are as follows: ln ( c o 2 ) i t = μ i + λ Z i t + β 1 ⁡ ln ( s c a l e i t ) I ( s c a l e i t ≤ γ ) + β 2 ⁡ ln ( s c a l e i t ) I ( s c a l e i t > γ ) + e i t (4) ln ( c o 2 ) i t = μ i + λ Z i t + β 1 ⁡ ln ( s c a l e i t ) I ( ln ( i s ) i t ≤ θ ) + β 2 ⁡ ln ( s c a l e i t ) I ( ln ( i s ) i t > θ ) + e i t (5) ln ( c o 2 ) i t = μ i + λ Z i t + β 1 ⁡ ln ( i s ) i t I ( s c a l e i t ≤ ξ ) + β 2 ⁡ ln ( i s ) i t I ( s c a l e i t > ξ ) + e i t     (6) Where γ , θ , and ξ denote the threshold values. ln ( c o 2 ) i t represents the logarithm of CO₂ emissions; ln ( s c a l e i t ) is the logarithm of the urban population size; ln ( i s ) i t is the logarithm of industrial structure. Z i t is a set of control variables, including technical factor ln ( r d ) i t , per capita GDP ln ( g d p ) i t , and foreign direct investment ln ( f d i ) i t . Eq. 4 investigates the staged effect of urban population size on CO₂ emissions with urban population size as a threshold variable. Eq. 5 examines the phased effect of population size on CO₂ emissions with industrial structure as a threshold variable. Eq. 6 investigates the staged impact of industrial structure on CO₂ emissions with urban population size as the threshold variable.

To investigate whether the influence of urban population size on CO₂ emissions presented U shape obtained in previous studies, this paper adds the quadratic term of urban population size to construct the following two-way fixed effects panel model: ln ( c o 2 ) i t = ϕ 0 + ϕ 1 ⁡ ln ( s c a l e i t ) + ϕ 2 [ ln ( s c a l e i t ) ] 2 + ϕ 3 ⁡ ln ( i s ) i t + λ Z i t + μ i + η t + ε i t (7) Where, μ i is the individual effect; η t is the time effect; and ε i t the disturbance. The control variables are consistent with Eqs. 4–6. As reference, Eq. 7 without the quadratic term of urban population size is expressed by Model 1. Eq. 7 is denoted by Model2. In Model 2, the mean center processing was conducted on the urban population size, then the residual-centralization program (Zhang and Rajagopalan, 2010; Geng et al., 2012; Yu and Tao, 2015) was used to eliminated the multicollinearity of ln ( s c a l e i t ) and [ ln ( s c a l e i t ) ] 2 .

Figure 1 and Figure 2 show the scatter plots of urban population size and CO₂ emissions, industrial structure, and CO₂ emissions, respectively.

As shown in Figure 1, CO₂ emission increased markedly with the growth of urban population size. As shown in Figure 2, in the process of industrial structure adjustment, the proportion of the secondary industry with high energy consumption decreased, and the proportion of the tertiary industry dominated by services, emerging industry, and low-carbon industry increased, the CO₂ emissions gradually decreased.

However, the scatter plot and linear regression were difficult to show the staged heterogeneity of the influence of urban population size and industrial structure on CO₂ emissions. Therefore, it was necessary to use the threshold-STIRPAT model to study the phased effects of urban population size and industrial structure on CO₂ emissions.

For the sake of analyzing the impact of urban population size on CO₂ emissions, the xtscc command in the Stata software was used to estimate the two-way fixed effects panel model. The results are listed in Table 1.

As shown in Table 1, Model 1 showed that the impact of urban population size on CO₂ emissions was positive. The variance inflation factor (VIF) was 2.26 in Model 2, indicating no severe multicollinearity. In Model 2, the coefficient of the quadratic term of the urban population size was negative, but it was not statistically significant. The coefficient of urban population size was significantly positive, indicating the relationship between urban population size and CO₂ emissions was linear.

At the same time, using the urban population size as the threshold variable, Eq. 4 examined whether the impact of urban population size on CO₂ emissions differed with the change of urban population size. In the process of threshold estimation, two tests are needed (Huang et al., 2018). The first is whether there is a threshold, and the second is whether the estimated threshold value is equal to the actual threshold value. The existence of the urban population size threshold was tested by using the Bootstrap method. The test’s F statistic and p value are shown in Table 2.

As shown in Table 2, the p value of the single threshold was 0.1110, failing to pass the significant test. Thus, the null hypothesis H₀ that there was no single threshold effect was accepted. In other words, urban population size had a significantly positive effect on CO₂ emissions, and there was no staged heterogeneity in such effect.

Model 1 indicated that a 1% rise in urban population size could increase the CO₂ emissions by 0.1479%. The impact of industrial structure on CO₂ emissions was significantly negative. Increasing the proportion of the tertiary industry reduced CO₂ emissions significantly, which was consistent with the previous conclusion (Shan et al., 2017). Overall, the high-carbon development mode can be avoided by guiding the adjustment in the industrial structure to the tertiary industry with high added value and low energy consumption.

The influence of technical factors on CO₂ emissions was positive, inconsistent with previous literature (Sheng et al., 2019). Compared with the developed countries, technological innovation in most developing countries increased CO₂ emissions (Kumar and Managi, 2009). As the world’s largest developing country, China’s research and development spending focused on technology that facilitated production rather than reduced energy consumption and CO₂ emissions (Yu and Du, 2019; Lin and Wang, 2021). As such, China’s technical factor tended to increase CO₂ emissions, which was confirmed in this study.

A 1% increase in per capita GDP could bring the rise of CO₂ emissions by 0.1476%. A 1% increase in foreign direct investment led to a 0.0038% reduction in CO₂ emissions. The pollution paradise was not confirmed in this study, which was consistent with the viewpoints of other studies (Yi et al., 2019; Yu and Du, 2019).

Using Eq. 5, the phased effects of urban population size on CO₂ emissions at different stages of industrial structure were investigated. The results of the industrial structure threshold test are shown in Table 3.

As shown in Table 3, the number of thresholds was determined. From the p value in Table 3, the single threshold test of industrial structure rejected the original hypothesis of linear model ( β 1 = β 2 ) at the significance level of 1%. And the double threshold test rejected the null hypothesis (H₀) of a single threshold at the significance level of 5%. However, the p value of the third threshold was 0.458, indicating it accepted the null hypothesis of no three-threshold effect.

As shown in Table 4, the industrial structures in different cities were divided into three intervals by two threshold values: interval 1 (lnis ≤ −1.3774), interval 2 (−1.3774 < lnis ≤ 0.1992), and interval 3 (lnis > 0.1992).

Table 4 indicated that the impact of urban population size on CO₂ emissions had nonlinear characteristics at different stages of industrial structure.

Then, the likelihood ratio (LR) statistic was used to test whether the estimated threshold value was equal to the actual value.

As shown in Figure 3, the LR statistics of the two threshold values were both significantly less than the critical value of 7.35 (marked with red dotted line in the figure). In other words, the threshold estimator for the industrial structure was true and effective.

In the three stages of industrial structure, the driving effect of urban population size on CO₂ emissions first increased and then decreased. When the industrial structure was below the threshold of 0.25 (i.e., lnis ≤ −1.3774), the coefficient was 0.1332; When the industrial structure was between 0.25 and 1.22 (i.e., −1.3774 < lnis ≤ 0.1992), the coefficient was 0.1566; When the industrial structure was higher than the threshold value of 1.22 (i.e., lnis > 0.1992), the coefficient was 0.1495. Most of the cities in the sample were in the second stage (i.e., −1.3774 < lnis ≤ 0.1992), where the secondary industry led the economy. And most of the heavy industries in the secondary industry were energy-intensive industries (Hao et al., 2020; Li and Zhou, 2021). The expansion of urban population size in this stage provided human capital for industrial division, which would lead to further increase in energy consumption. In the third stage (i.e., lnis > 0.1992), the tertiary industry with low energy consumption and high added value led the economy. The service industries such as information, tourism, finance and logistics in the tertiary industry consumed less energy (Zhang et al., 2019b; Zheng et al., 2020; Yuan and Zhou, 2021). Taking Taiyuan as an example, its industrial structure exceeded 1.22 since 2012. From 2012 to 2017, the increase of Taiyuan’s population size did not lead to a significant increase in CO₂ emissions.

The actual situations in Anyang and Haikou were also consistent with the results. The industrial structure of Anyang was between 0.25 and 1.22, while the industrial structure of Haikou was higher than 1.22 over the period 2001–2017. Moreover, the increment difference of the two cities’ population size was relatively small (an increase of 1.02 million people in Anyang and an increase of 1.11 million people in Haikou). Over the period, Anyang’s CO₂ emissions increased by 24.25 million tons, while Haikou’s CO₂ emissions increased by just 9.57 million tons. The effects of urban population size change on CO₂ emissions were quite different in the two cities. The same increase of urban population size brought higher CO₂ emissions when the city’s industrial structure was in the second stage.

In Eq. 6, the urban population size was selected as the threshold variable to examine whether the impact of industrial structure on CO₂ emissions had a population size threshold effect. Table 5 displayed the results of the threshold effect test.

As shown in Table 5, there was a single threshold effect. The estimated results of Eq. 6 are listed in Table 6.

Table 6 showed that the threshold value was 1.3835 million people. Urban population sizes were divided into two stages and the influence coefficients of industrial structure on CO₂ emissions were 0.1567 and −0.0527, respectively.

From Figure 4, it could be seen that the threshold value of urban population size was true and effective.

The influence of industrial structure on CO₂ emissions was opposite when it exceeded the threshold, first promoting and then restraining. When the urban population size was below the threshold value of 1.3835 million people, an increase in the ratio of the tertiary industry to the secondary industry failed to achieve CO₂ emission reduction. The results also answered why industrial restructuring towards tertiary industry sometimes failed to achieve CO₂ emission reduction.

When the urban population size exceeded 1.3835 million people, the influence coefficient of industrial structure on CO₂ emissions was significantly negative. Such as Maanshan, a city dominated by secondary industry, exceeded the threshold value of 1.3835 million people in 2011. Moreover, its population size barely changed in 2011–2017. However, the ratio of the tertiary industry to the secondary industry increased rapidly (from 0.38 in 2011 to 0.83 in 2017). We found that its CO₂ emissions declined from 15.66 million tons in 2011 to 14.33 million tons in 2017. Similar for Daqing, with a population size more than 1.3835 million people in 2001–2017, its tertiary industry increased significantly from 2014, which was helpful to reduce CO₂ emissions. In cities with a population size larger than 1.3835 million people, adjusting the industrial structure to increase the tertiary industry and decrease the secondary industry was helpful to reduce CO₂ emissions.

This conclusion was consistent with the viewpoints of the previous studies from the perspectives of CO₂ emission intensity and economic performance. It was found that the inhibition effect of industrial structure on CO₂ emission intensity was enhanced when the urbanization level was improved (Yu et al., 2022). By introducing the corss term of urban population size and industrial agglomeration, Zhang et al. (2021a) found that the cities with a population more than 1.52 million was conducive to play the CO₂ emission reduction effect brought by industrial agglomeration (Zhang et al., 2021a). Cao (2017) deliberated the panel data of 277 cities and concluded that diversification in industry development patterns enhanced economic performance only when the population size exceeded 1.27 million people (Cao, 2017). From the perspective of CO₂ emission reduction, this paper concluded that the industrial structure should be adjusted according to the urban population size. The threshold value of 1.3835 million people was almost consistent with the previous literature.