Abstract

To explore the effectiveness of the carbon emission trading mechanisms in pilot regions, we first employ the Difference-in-Differences (DID) model. This approach involves identifying the impact of the policy by comparing performance differences between the treated group (pilot regions) and the control group (non-pilot regions), while controlling for other factors. DID model is a widely-used methodological framework for examining the impact of policy interventions. Considering the availability of carbon emission and energy data, our study will focus on analyzing data at the regional level in China¹⁾ to evaluate the impact of the emission trading pilot program. Specifically, the six pilot regions (Beijing, Shanghai, Tianjin, Chongqing, Guangdong, and Hubei,) will be treated as experimental group, while all other major regions that were not selected as pilots will serve as control units for comparison purposes. To specify the DID model, we will construct a regression equation as follows:

In the specified DID model, the carbon emission outcome in region i in year t is denoted by y_it. The variable treat_it is used to signify whether a province or region i implements a carbon trading pilot in year t. If a pilot is implemented, treat it is assigned a value of 1, otherwise, it is assigned a value of 0. The control variables Controls_it accounts for other factors that may affect carbon emissions. Ehrlich and Holdren (1971) suggested that population, affluence, and technological factors are three important aspects affecting environmental outputs and Kaya (1989) suggested that population, per capita GDP, energy intensity, and emission coefficient are the four main factors affecting carbon emissions. Following the previous studies (Luo et al., 2023), we incorporate eight variables into the model as Controls_it, population aging, population urbanization, the share of the working-age population, economic size, energy intensity, industrial structure, freedom of trade, and technological progress. Additionally, the time and individual fixed effects are represented by γ_t and μ_i, respectively. And the error term is represented by ε_it.

Nonetheless, utilizing the proportional growth rate in the Kaya equation enhances effectiveness. Additionally, the power-law formulation of the extended and improved STIRPAT model (Dietz and Rosa, 1997), which builds upon the IPAT equation (Ehrlich and Holdren, 1971), enables the linearization of nonlinear relationships via logarithmic transformation. This adaptation renders it compatible with linear regression models. Hence, our model will subject all variables to logarithmic transformation.

In observational studies, treatment and control groups may exhibit differences even before the intervention. These differences can affect the likelihood of individuals receiving treatment, thereby leading to biased outcomes. The PSM-DID model is a two-step approach that aims to increase the robustness of the results. In the first step, the sample is adjusted for possible systematic differences using the propensity score matching (PSM) model. PSM estimates the probability of the treated group assignment based on observable characteristics, allowing the treated and control groups to be balanced on observable covariates. In the second step, the net effect of the policy is assessed using the DID model, which compares the changes in outcomes between the treated and control groups before and after the policy intervention. The PSM-DID model can increase the internal validity of the study by reducing the bias due to unobservable confounding variables and selection bias.

The PSM model is used to find a comparable province j in the control group for the treated province i based on observable variables. Specifically, a province j in the non-pilot provinces (control group) is selected such that its observable characteristics are sufficiently similar to those of province i, ( x_i ≈ x_j ). Based on the assumption of ignorability, when individual characteristics of provinces have no influence on the decision to implement carbon emission trading, we assume that the probability of implementing the policy is similar for province i and province j. Therefore, the systematic differences between the treated and control groups are resolved. However, matching in high-dimensional space may result in data sparsity and difficulty finding suitable matches or may result in a small number of matched pairs that are insufficient for subsequent analysis. Rosenbaum and Rubin (1983) proposed the use of propensity scores to address these issues. The propensity score is the probability of an individual receiving a certain treatment given a set of covariates. Traditional matching methods that use multiple observable variables are difficult to implement in practice. PSM model replaces the multidimensional covariates with a one-dimensional variable - the propensity score P(x)0, which simplifies the matching process by allowing researchers to match based on a single variable.

Before applying the DID model, verifying the “Parallel Trends Assumption” is essential. This assumption suggests that, without the intervention, the trends in the outcome variable for both treatment and control groups would be similar over time. In the absence of treatment, the temporal difference between these groups would remain stable. Once the assumption is confirmed, we align the treatment and control groups using Propensity Score Matching. Subsequently, the DID model is utilized to evaluate the variations in carbon dioxide emissions in provinces affected by the carbon emissions trading system, before and after the policy implementation. And evaluate the average treatment effect on the treated (ATT) for the carbon emissions trading system policy participants.

Carbon trading offers two main pathways for emission reduction. Firstly, the government imposes emission caps aligned with policy goals and distributes initial emission rights to firms. If a company exceeds these limits, it must either reduce production or buy additional emission rights from others. Secondly, carbon trading inherently prompts companies to lessen emissions by enhancing energy efficiency. Companies that improve energy utilization efficiency are more apt to generate excess emission rights, benefiting from the sale of these certificates. Thus, efficiency-focused firms are motivated to continually refine their energy structures and increase efficiency, leading to more certificate sales. Conversely, less efficient companies, facing ongoing costs from buying emission rights and future emission constraints, are pushed towards emission reduction by improving energy utilization. Consequently, under carbon trading policies, boosting energy efficiency emerges as a primary strategy for emission reduction among most firms (Huang et al., 2018; Zhang et al., 2018). This study attempts to investigate how energy efficiency acts as a mediating factor in carbon emission reductions within the framework of carbon trading, examining to what extent improvements in energy utilization efficiency indirectly contribute to emission reduction.

To explore the underlying reduction mechanisms of the ETS, this study employs the mediating effect model based on the influential framework introduced by Baron and Kenny (1986), aiming to illuminate the intricate relationships among the variables. To be more precise, the model consists of three discernible stages. In the initial stage, the model adheres to a base regression framework, similar to Eq.(1), where the treatment effect is solely regressed against the dependent variable (carbon emissions). This stage seeks to establish the extent to which China’s pilot ETS effectively contributes to carbon emission mitigation. Advancing to the second stage, the treatment effect is regressed against the mediating factor (energy intensity), enabling a deeper understanding of the policy’s influence on these intermediary factors. Specifically, a positive value for ρ₁ in Eq.(3) would suggest a promoting effect, whereas a negative value would indicate a suppressive impact. In the final stage, the treatment effect, mediating factor, and dependent variable are integrated into a unified regression framework. If the coefficient of δ₁ in Eq.(4) loses its significance or diminishes in size compared to β₁ in Eq.(2), while the coefficient of δ₂ maintains statistical significance, the verification of the mediating role of the energy intensity and the underlying mechanism of ETS in reducing carbon emission is established. Furthermore, if the coefficient of δ₂ is greater than zero, it signifies that enhancements in the mediating variable (i.e. lower energy intensity level) lead to a decrease in carbon emissions. The mediating effect model explained above is formulated as follows.

Where the variable lnEI_it represents the mediating effect within province or region i at year t. Other symbols employed in these models hold identical interpretations as those presented in Eq.(1).

In order to evaluate the effectiveness of the carbon emission trading mechanisms and elucidate the underlying policy mechanisms, this study analyzes a representative sample comprising 30 provinces and regions across China. The study covers the period from 2005 to 2019. The dependent variable is the aggregate carbon emissions within the province or region i during the year t. The core independent variable in our DID model is whether each region initiated the carbon trading pilot (referred to as “treat”) in 2013 or not. The primary objective of this variable is to assess the impact of carbon trading implementation on the regional emissions level . Furthermore, as explained in subsection 3.3, the study incorporates the mediating variable - energy intensity, which represents the regional consumption of fossil energy per unit of GRDP. This variable is included to analyze the mechanism through which the implementation of the carbon trading market influences the dependent variables. Based on the existing literature, this study incorporates several control variables to mitigate potential bias that may arise from omitting other significant explanatory variables. In line with Luo et al. (2023), this study takes into account several pertinent control variables. These variables include population aging, population urbanization, the proportion of the working-age population, per capita GRDP, industrial structure, freedom of trade, investment amount completed in industrial pollution control, and technological progress. By integrating these demographic, social, and economic factors into the analysis, we aim to comprehensively examine and account for their potential influence on the relationship under investigation. Table 1 provides a concise description along with the sources of the major variables.

All the explanatory variables used in this paper are from the National Bureau of Statistics of China (NBS, 2023), while the data on total carbon emissions were obtained from the China Stock Market And Accounting Research (CSMAR; CSMAR Database, 2023) database. The detailed timeline of China’s carbon trading pilot program is sourced from the China Energy Network (CEN; China Energy Network, 2023). Table 2 provides the descriptive statistics of the selected variables for all provinces and municipalities between 2005 and 2019. The statistics include the mean, standard deviation, maximum, and minimum values after the logarithmic transformation.

To mitigate potential bias arising from individual differences between the experimental and control groups, this paper employes the PSM-DID model. This approach aims to ensure the accuracy of the policy treatment effects by accounting for confounding factors. The test results demonstrate that the coefficients associated with the effect of the carbon trading policy on total carbon emissions remain statistically significant across multiple matching techniques, including first-neighbor matching, second-neighbor matching, fourth-neighbor matching, and kernel density matching. These findings suggest that the observed reduction in total carbon emissions within the pilot regions cannot be solely attributed to individual differences within those regions. Consequently, these results further validate the robustness of the base regression test findings, supporting the study’s conclusions. In this study, we present the results obtained through fourth-neighbor matching techniques. The corresponding findings are summarized in Table 3.

Table 3 presents the estimation results for Eq.(1). In column 1, the PSM-DID model result of a simple regression is presented where no control variables are included. It implies that China’s pilot ETS has negative overall impact on carbon emissions. When control variables are incorporated in our main model (column 2), the carbon trading policy exhibit a significant negative effect on carbon emissions. The implementation of the policy results in a substantial decrease of 19.5% in total carbon emissions within the pilot area during the study period. These results indicate that when we account for other related factors, the effect of carbon trading policy becomes even greater. Also, the fixed effects model produces similar results. Carbon trading policy’s significant impacts on carbon emissions reduction are consistently found across various model specifications and estimations. These findings suggest that carbon trading policy plays an important role as a driver of carbon emissions reduction in China.

It is important to acknowledge that the estimations mentioned above are contingent upon certain underlying assumptions, such as the parallel trend assumption of the DID model. Additionally, the empirical results can be susceptible to interference arising from the selection of measurement methods, particularly in the calculation of carbon emissions. Environmental science, especially in the context of calculating carbon emissions, employs diverse measurement methodologies. These include direct methods, like using sensors to monitor emission sources, and indirect methods, which rely on statistical data related to energy consumption or economic activities. The choice of measurement approach can yield different outcomes. Direct measurement is often more precise in reflecting actual emissions, whereas indirect measurement hinges on various assumptions and conversion factors, introducing a higher degree of uncertainty. Considering these sources of uncertainty, this paper undertakes additional measures to reinforce the robustness of the regression results. Specifically, preexisting trend tests and placebo tests are conducted. These tests aim to further scrutinize the validity and reliability of the regression findings, taking into account potential confounding factors and addressing concerns regarding the impact of measurement methods.

4.2.1. Parallel-trends test

Ensuring similar variation tendencies in the treatment group and control group is a critical precondition for the Difference-in-Differences strategy. To verify the parallel trend assumption, this study conducts two tests based on the methodology proposed by Roberts and Whited (2013). Firstly, the annual carbon emissions of the experimental group and control group are plotted before and after applying a logarithmic transformation. As illustrated in Fig. 1 (a) and (b), a parallel trend is observed prior to the implementation of the Emission Trading policy (before 2013). Following the policy’s implementation, the experimental group experiences a significant decline in carbon emissions starting from 2013, with a steady decrease in subsequent years. In contrast, the control group continues to exhibit an upward trend in carbon emissions. Secondly, a parallel-trends test is conducted using Stata 17 software to evaluate whether the control and treatment groups followed a parallel path before the policy implementation. The hypothesis of interest is H₀ : linear trends are parallel, and F-test indicates that there is insufficient evidence to reject the null hypothesis of parallel trends. The test result, combined with the graphical analysis, supports the parallel-trends assumption.

Fig. 1. 
Parallel trends of carbon emissions

4.2.2. Placebo test

The DID placebo test revolves around the concept of manipulating the treatment group or policy time for test purpose. By introducing a “pseudo-policy group,” researchers can assess the significance of its coefficient and identify potential biases in the original estimation. If the coefficient of the “pseudo-policy group” remains statistically significant, it suggests that the observed changes in the explained variable may be influenced by factors other than the intended policy intervention or random variations. This indicates a potential bias in the original estimation results. To conduct the placebo test, researchers typically re-estimate the regressions using data from years preceding the actual policy implementation. By examining whether the coefficients of the policy dummy variables remain significant in these placebo years, researchers can assess the robustness of the original estimation and determine if the observed effects are indeed driven by the policy intervention.

In this paper, the methodology is based on the study conducted by Shi et al. (2018). It involves employing a random selection process to create a “pseudo-policy dummy group” and conducting 500 iterations to assess the significance of the coefficients associated with this group. The initial regression analysis comprises a treatment group consisting of six regions and a control group consisting of 24 regions. For the placebo test, six regions are randomly selected from the total pool of 30 regions to form the “pseudo-treatment group,” while the remaining regions serve as the control group. A “pseudo-policy dummy variable” (interaction term) is then generated for regression analysis, with 2011 designated as the policy dummy time. The regression analysis is repeated 500 times, resulting in 500 sets of regression results that include the estimated coefficients of the “pseudo-policy dummy variables,” along with their corresponding standard errors and p-values. Finally, the distribution of the estimated coefficients and their p-values from the 500 “pseudo-policy dummy variables” is plotted, providing a visual representation of the placebo test results. Fig. 2 displays the results obtained from the placebo test.

Fig. 2. 
placebo test of carbon trading pilot

Fig. 2 displays the distribution of estimated coefficients for the 500 “pseudo-policy dummy variables” and their corresponding p-values. The x-axis represents the magnitude of the estimated coefficients, while the y-axis represents the density values and p-values. The curve represents the kernel density distribution of the estimated coefficients. The blue dots represent the p-values associated with the estimated coefficients. The vertical dashed line corresponds to the true estimate of the DID model (-0.195), and the horizontal dashed line represents the 10% significance level. Upon examination of the figure, it is observed that the estimated coefficients are primarily concentrated around zero. Additionally, a significant majority of the estimated values have p-values greater than 0.1, indicating insignificance at the 10% level. This suggests that our estimates are unlikely to have been obtained by chance nor driven by the influence of other policies or random factors.

According to our analysis so far, it is evident that the carbon trading policy significantly contributes to the mitigation of carbon emissions. We further investigate the underlying mechanisms that drive this impact and explore the specific paths through which the abatement effect is achieved. As outlined in the theoretical framework and hypothesis presented in the third part of this paper, we hypothesize that the carbon trading policy, as a policy intervention, leads to carbon emission reduction by influencing the energy intensity (mediating variable). To shed light on the influencing mechanism, Table 4 presents the analysis based on Eq.(2) to (4).

The estimation results presented in Table 4 overall suggest the mediating effect between the carbon trading policy and carbon mitigation. The mediating effect in this context is manifested through a decrease in energy intensity, which signifies an increase in the efficiency of comprehensive energy utilization. This finding suggests that the carbon trading policy induces carbon emissions reduction by promoting improvements in energy efficiency. Specifically, the mediating effect is tested in columns 1 to 5 of Table 4. The relationship between the carbon trading policy and energy intensity is examined first. As shown in columns 1-2, the estimated coefficient of carbon trading policy on energy intensity are negative and statistically significant. It indicates that the carbon trading policy has a significant influence on energy intensity and that there is a negative relationship between the two variables. As a result, it can be inferred that the mediating effect exists. Moving on to the columns 3-5, the mediating effect of energy intensity is further demonstrated by observing a decrease in the coefficient of the carbon trading policy and the positive coefficient of energy intensity variable. These results all together suggest that energy intensity plays a mediating role in the relationship between the carbon trading policy and carbon mitigation. Additionally, when control variables are introduced in the analysis (columns 2, 4, and 5), the significance of the coefficients and the overall effect remain consistent, providing consistent supporting evidence for the mediating effect of energy intensity.

Empirical results in Table 4 indicate that the total carbon emissions in the six pilot regions of China’s carbon trading policy experiment decreased by 27.2%, with a direct reduction of 9.1% due to the implementation of carbon trading pilot. Furthermore, the result indicate that the market mechanism significantly stimulate a reduction in energy consumption per unit of economic output in these regions, effectively reducing CO₂ emissions. Overall, it is estimated that improved energy efficiency in the ETS pilot region leads to a further reduction of approximately 18.1% (=0.190*0.955) in carbon emissions, significantly contributing to carbon neutrality and climate change mitigation.