4.1 Baseline Regression Results
First our study tests direct effect of policy interaction item on full factor energy efficiency only with time fixed effects and region fixed effects, also, without control variables. Its results are shown in column (1) and coefficient of our core explanatory variable is positive but insignificant. Columns (2) to (6) gradually add control variables one by one. It could be seen that the core coefficient has become gradually bigger and more significant. When it comes to column (6) this core coefficient is significantly positive at statistical level of 1%, illustrating that compared with regions without carbon trading policy, those pilot regions do benefit from this policy. To sum up roughly, there is a positive impact of carbon emission trading on improving energy efficiency.
Turning to the control variables, the negative coefficient of avegdp may be because greater avegdp means promoted regional economic level and this may bring about policy dilution, which we will give a detailed expression in following heterogeneity part. Twp represents industrial structure, and its positive coefficient implies that industrial structure optimization probably helps adjust overall energy structure and make it fitted to real needs. Spd without doubt has a significant coefficient as its growth equals to greater degree of pollutant emissions and it is surely detrimental to clean and efficient energy structure, thus being harmful for raising energy efficiency. The significant negative coefficient of pat may be a result of the time lag of innovation capacity. In other words, though its enhancement indicates greater regional revolution potential, it still cannot have an immediate impact on regional willingness to act for green innovation right after the policy’s implementation. As for negative coefficient of dens, we could see that higher density of population causes more energy consumption especially coal consumption in absolute value and it increases difficulty for reforms around energy structure.
Table 3
Baseline regression results
Variables | EFF (1) | EFF (2) | EFF (3) | EFF (4) | EFF (5) | EFF (6) |
did | 0.0085 (0.0524) | 0.0268 (0.0549) | 0.0214 (0.0544) | 0.0452 (0.0540) | 0. 0962* (0.0533) | 0.1998*** (0.0573) |
avegdp | | 0.0002 (0.0002) | 0.0002 (0.0002) | 0.0001 (0.0002) | -0.0001 (0.0002) | -0.0003* (0.0002) |
twp | | | 1.3036** (0.5184) | 1.5574*** (0.5153) | 1.6537*** (0.4979) | 1.8204*** (0.4842) |
spd | | | | -0.0021*** (0.0006) | -0.0027*** (0.0006) | -0.0027*** (0.0006) |
pat | | | | | -0.0037*** (0.0008) | -0.0027*** (0.0008) |
dens | | | | | | -33.2251*** (7.9384) |
Cons | 1.0775*** (0.0304) | 1.0605*** (0.0341) | 0.4594* (0.2414) | 0.5029** (0.2375) | 0.5509** (0.2295) | 1.9875*** (0.4090) |
Time Fe | Yes | Yes | Yes | Yes | Yes | Yes |
Region Fe | Yes | Yes | Yes | Yes | Yes | Yes |
R² | 0.7765 | 0.7775 | 0.7828 | 0.7913 | 0.8063 | 0.8187 |
Obs | 300 | 300 | 300 | 300 | 300 | 300 |
Note: In the parenthesis are standard deviations (Std). Meanwhile, *, ** and *** indicate significance levels at 10%, 5%and 1%. Contents in the parentheses of the following tables have the same implications in like manner.
4.2 Robustness Tests
(1) Replace the Explained Variable
In the baseline regression, energy efficiency measured in full factor is used as the explained variable. Since this paper's measurement for energy efficiency is centered around output, the replacement of the explained variables draws on the experience from research of Bu et al. (2022) who have expressed in their research the definition of energy efficiency and the means of measurement: energy productivity is used to represent the energy efficiency of enterprises, which enables this indicator(efficiency) to be comparable with other indicators of factor measurement. According to their literature, the measurement formula used single factor measurement: Efficiency = regional GDP/Energy Input. Then a robustness test is conducted. The following results indicates that after replacing the previous explained variable, the implementation of carbon emission trading still has a strongly significant positive impact on energy efficiency.
Table 4
Robustness test: replace the explained variable
Variables | EFF1(single factor) |
did | 0.2707*** (0.0640) |
avegdp | -0.0003 (0.0002) |
twp | 1.8805*** (0.5404) |
spd | 0.0021*** (0.0007) |
pat | 0.0054*** (0.0009) |
dens | 13.2464 (8.8598) |
Cons | -0.4810 (0.4565) |
Time FE | Yes |
Region FE | Yes |
R² | 0.8188 |
Obs | 300 |
(2) Parallel Trend Test |
The parallel trend hypothesis is key prerequisite for unbiased estimation of DID model. Only if there exists no systematic difference between control regions and treated regions before policy implementation can we rely on DID model to test policy impact. Figure 2 gives a rough description of the two-way trend where both control and treated regions stay relatively paralleled to each other before the policy impact in 2014. After 2014 meanEFF of treated regions as a whole stay above that of control regions and it gradually grows up. Such is our verified perquisite, and it proves our model reliable. Figure 3 gives a dynamic test where we find that for the years before policy impact (pre3 dropped considering collinearity) the coefficients of cross-term pass the significance test because their confidence intervals have crossover point with zero line. This illustrates that there is no significant change in energy efficiency differences between control regions and treat regions, emphasizing the validity of parallel trend hypothesis. And after policy implementation the estimated coefficients are all significantly positive, which means the trading policy does have positive effect and there’s not so much time lag.
(3) PSM-DID Test
For DID validity we also have to make sure that policy implementation is not random and there exists no selection bias. To testify this, we have used PSM method (propensity score matching) to select individuals from control regions that are similar to those in treated regions and subsequently made corresponding regression. Based on logit model, our paper has taken k-nearest neighbors within caliper range and conducted one-to-one matching, with control variables selected as covariables.
Kernel density distributing situation is shown in Fig. 4—treated regions and control regions have become better in fitting after matching. Displayed in Table 5 are regression results after matching. We could see from them that core coefficient remains strongly significant and positive, indicating the policy’s great positive effect on energy efficiency, verifying the robustness of our baseline empirical results.
Table 5
Variables | EFF |
did | 0.2186*** (0.0770) |
avegdp | -7.09e-06 (0.0004) |
twp | 6.1255*** (1.9871) |
spd | -0.0041** (0.0017) |
pat | -0.0026** (0.0011) |
dens | -87.6257 *** (19.8279) |
Cons | 2.1095** (1.0282) |
Time FE | Yes |
Region FE | Yes |
R² | 0.9390 |
Obs | 71 |
(4) Placebo Test |
Besides tests above, we must exclude influence from other factors as well. Hence, we conducted a placebo test by randomly selecting and generating treated groups for totally 500 times. Figure 5 has reported our results and shown P-value and kernel density distribution. Seen from the figure, most of these coefficients assemble near 0 and their mean value is -0.00613, obviously different from baseline coefficient which is 0.199842. Besides, most of the P-value appears to be above baseline P-value (0.001) and are insignificant. All these results have verified that the trading policy’s impact on energy efficiency has nothing to do with other unobserved factors’ influence, in other words, the impact comes from the policy itself.
4.3 Mechanism Test
Results of our mechanism test is shown as Table 6 as follows.
Columns (1) and (2) in Table 6 empirically illustrates whether green innovation can be used as a mechanism variable to expound the transmission from carbon trading to energy efficiency. As shown in column (1), did and green innovation show an insignificant positive relationship, while in column (2), coefficients of both did and gi appear to be significantly positive. That is, from the test, it cannot be concluded that implementation of carbon trading has promoted green innovation, even though green innovation does improve energy efficiency, and consequently hypothesis 1 is rejected.
In response to the rejection of hypothesis 1, our paper considers that the reasons might be as follows. Firstly, the "weak Porter hypothesis" is not valid in some regions, i.e., the impact of environmental regulations on innovation is likely to be ineffective at preliminary stage. From our findings, joint by research of Zhang et al. (2022) and study of Xie and Thompson (2022), some of the relevant literature indicates that there exists a "U" shape between environmental regulation and technological innovation, and the weaker environmental regulation is detrimental to technological innovation——only when the intensity reaches the inflection point will it have a positive impact. There are sectors like LC (low value-added & cleaner) industry which aren’t significantly green innovation-driven, and environmental regulations’ influence is complex due to industrial heterogeneity.
Columns (3) and (4) empirically test whether clean energy transition can improve efficiency by improving energy structure. In column (3), did has a significant negative impact on eins, indicating that, significantly, the implementation of carbon trading reduces the coal consumption intensity and improves the overall energy structure. While in column (4), did still shows a significant positive impact on EFF, and the coefficient of eins is also significantly negative. All these above indicates that hypothesis 2 is strongly valid.
This conclusion can also be corroborated by the study of Ren and Sun (2022). This trading has forced enterprises to change their pollutant emission decisions and has reduced CO2 emissions by reducing the regional coal consumption and adjusting the energy structure. This means that in pilot regions clean energy transition has been realized and corresponding achievements have been made.
In summary, the transmission mechanism of carbon emission trading to improve energy efficiency in the present situation is mainly through reducing coal consumption density, promoting energy structure, then realizing transition into clean energy.
Table 6
Variables | gi (1) | EFF (2) | eins (3) | EFF (4) |
did | 0.0622 (0.0511) | 0.1896*** (0.0570) | -0.0474*** (0.0121) | 0.1763*** (0.0588) |
gi | | 0.1647** (0.0696) | | |
eins | | | | -0.4976* (0.2946) |
avegdp | -0.0009*** (0.0001) | -0.0001 (0.0002) | -0.00002 (0.00003) | -0.0003* (0.0002) |
twp | 0.2684 (0.4319) | 1.7762*** (0.4802) | -0.1474 (0.1026) | 1.7470*** (0.4844) |
spd | 0.0006 (0.0006) | -0.0028*** (0.0006) | 0.0006*** (0.0001) | -0.0024*** (0.0006) |
pat | 0.0023*** (0.0008) | -0.0031*** (0.0009) | -0.00003 (0.0002) | -0.0027*** (0.0008) |
dens | 12.4003* (7.0817) | -35.2675*** (7.9149) | 2.4074 (1.6816) | -32.0273*** (7.9414) |
Cons | -0.5515 (0.3649) | 2.0783*** (0.4072) | 0.3760*** (0.0866) | 2.1746*** (0.4223) |
Time FE | Yes | Yes | Yes | Yes |
Region FE | Yes | Yes | Yes | Yes |
R² | 0.5440 | 0.8226 | 0.6994 | 0.8207 |
Obs | 300 | 300 | 300 | 300 |
4.4 Heterogeneity Test
(1) Heterogeneity at the Regional Economic Level
Although economic development level in all the pilot regions where carbon finance has already gone into effect is high, there still appears to be some differences in economic level and development degree among these regions. Due to the relatively small number of pilot provinces and cities, this paper uses weighted coefficients of per capita GDP and CPI to obtain the coefficients of comprehensive economic development level of each region and divides the sample into two groups according to these coefficients: regions with higher level in comprehensive economic development and regions with relatively lower level in comprehensive economic development. Then our paper analyzes the heterogeneity rendered by regional economic level. The results of regression are shown as follows.
Contrary to our common conjecture, the following table shows that carbon emission trading has a more significant positive effect on those regions with relatively lower level of economic development. In other words, we cannot conclude from the current panel data whether the implementation of carbon trading has any negative or positive impact on the energy efficiency in these economically more-developed regions.
The possible explanations are as follows. Firstly, the positive impact on these economically less-developed regions undoubtedly reflects the effectiveness of carbon trading implementation. The prevalent existence of many energy-intensive industrial enterprises in economically less developed regions still means that for these regions carbon finance mechanism is timelier and more demand-fitted, and its corresponding policies are less diluted.
Economically more-developed regions have gradually made many high-energy-consuming enterprises withdrawn from the market before the prevalence of carbon finance, which means that the energy consumption intensity of their industrial enterprises is currently quite low, hence the carbon trading policy has a lower marginal effect than that of other regions. Besides, as research of Berkhout et al. (2000) corroborates, there exists energy rebound effect and what we consider is that for those more developed regions their already-low energy consumption intensity makes the policy’s effect diluted in the long run. The economically more-developed regions show an overall surplus of emission allowances and a dysfunctional price mechanism because carbon emission allowances allocated to them do not match their actual situation. Moreover, many key enterprises still lack the ability to convert extra allowances into productivity after receiving them.
Table 7
Heterogeneity at the regional economic level
Variables | (1) Regions with Relatively Higher Level in Comprehensive Economic Development | (2) Regions with Relatively Lower Level in Comprehensive Economic Development |
EFF | EFF |
did | -0.0382 (0.0928) | 0.2694*** (0.0634) |
avegdp | -0.0003** (0.0002) | -0.0004** (0.0002) |
twp | 1.7623*** (0.4958) | 1.7679*** (0.4787) |
spd | -0.0023*** (0.0006) | -0.0026*** (0.0006) |
pat | -0.0025*** (0.0009) | -0.0025*** (0.0008) |
dens | -19.2311** (8.8305) | -23.0696*** (7.0942) |
Cons | 1.3563*** (0.4425) | 1.5535*** (0.3750) |
Time FE | Yes | Yes |
Region FE | Yes | Yes |
R² | 0.8102 | 0.8226 |
Obs | 300 | 300 |
(2) Heterogeneity at the Regional Marketization Level |
Now, a new way of classification is introduced to rank the pilot regions with the help of marketization index. All the samples are hereby divided into two groups: regions with higher level of marketization and regions with lower level of marketization.
The regression results are displayed in Table 8. In the group of lower marketization level, carbon emission trading shows an insignificant negative impact on energy efficiency, while in the group of higher marketization level the core coefficient is significantly positive at 1% level. Therefore, it can be concluded that carbon trading policy is more significantly effective in regions of higher marketization level. Possible explanations are given as follows.
Our findings together with the research of Hong et al. (2022) consider that in regions of lower marketization level, the group structure of investors and traders in the carbon market appears single, with liquidity of investment and financing being weaker compared to other regions. Also, the relevant information is often incomplete and unsymmetrical. As a result, there is a lack of both willingness to trade and a sound market order, making the overall carbon financial trading process not able to perfectly operate. Moreover, we consider that lower level of marketization implies tighter government control over the market, greater influence of macro intervention, severer market segmentation, and more room for rent-seeking. The state-owned enterprises in those regions thus wield industrial property rights and monopoly power, squeezing out many private enterprises which have the motivation to innovate and the ability to improve their production capacity, meanwhile controlling most of the capital inflows generated by carbon emission trading and obtaining the greatest degree of autonomy in carbon resource allocation. This behavior, however, brings about the disproportionately uneven distribution of carbon resources throughout the market, resulting in market failure, deterioration of industrial structure and lower allocation efficiency.
Contrary to low marketization’s disadvantages, as for high-marketization regions the information fluxes freely, and trading appear to be more flexible, with less room for central monopoly and rent seeking. Trading markets there may have been more fitful for enterprises’ need and it can rely on proper market readjustment to achieve dynamic balance.
Table 8
Heterogeneity at the regional marketization level
Variables | (1) Regions with Lower Level of Marketization | (2) Regions with Higher Level of Marketization |
EFF | EFF |
did | -0.1127 (0.0800) | 0.3377*** (0.0634) |
avegdp | -0.0003** (0.0002) | -0.0003* (0.0002) |
twp | 1.6371*** (0.5027) | 1.4487*** (0.4739) |
spd | -0.0022*** (0.0006) | -0.0026*** (0.0006) |
pat | -0.0024*** (0.0009) | -0.0025*** (0.0008) |
dens | -15.7379* (8.2858) | -24.9166*** (6.9850) |
Cons | 1.2493*** (0.4098) | 1.7761*** (0.3725) |
Time FE | Yes | Yes |
Region FE | Yes | Yes |
R² | 0.8115 | 0.8291 |
Obs | 300 | 300 |