Structural, Innovation and Efficiency Effects of Environmental Regulation: Evidence from China’s Carbon Emissions Trading Pilot

Abstract

Conventional wisdom argues that environmental regulation can trigger both structural adjustments and enhanced innovation. We test this conjecture by using a difference-in-differences approach to analyze the impacts of China’s carbon emission trading (CET) pilot policy on energy consumption. We find that compliance with the CET regulation has triggered statistically significant adjustments in energy structure, industrial structure, and technological innovation. Adjustments in industrial structure also contribute to enhanced total factor energy efficiency, whereas increased technological innovation has mixed effects on energy efficiency. We show that in the short run, government-led innovation does not immediately contribute to improvement in energy efficiency, whereas enterprise-led innovation has a negative impact. It indicates that CET regulation can affect energy efficiency through industrial structure and technological innovation. Overall, our results provide new evidence for the strong version of the Porter hypothesis. Our results also provide strong scientific support for China’s recent transition towards market-based carbon mitigation strategies.

1 Introduction

As the world’s largest carbon emitter and energy consumer since 2007 (IEA 2009), China has faced mounting domestic and international pressure to reduce energy consumption and mitigate greenhouse gas (GHG) emissions. In response, China has taken increasingly proactive measures to address energy conservation and GHG emissions. The government has also recently experimented with various market-based policy instruments, justified by the commonly held belief that such instruments are more flexible and potentially more cost-effective than conventional command and control policies. The grand experiment was the carbon emission trading (CET) pilot launched in 2011 (NDRC 2011). With China’s long-awaited national emissions trading scheme officially launched in late 2017 (NDRC 2017), the government seems to have firmly committed to the market-based approach. Yet, evidence on the effects of the CET pilot policy and particularly its conservation mechanisms has been very limited.

In principle, carbon mitigation compliance in the CET pilot can be achieved through two main mechanisms. On the one hand, the CET pilot policy could have facilitated a faster energy structure transition towards cleaner energy sources (Nayyar et al. 2014). However, in the short run, the mitigation potential through such a mechanism may be very limited given China’s heavily coal-dependent energy structure. On the other hand, improved energy efficiency can also help to reduce GHG emissions. Regulatory pressure may affect the layout and adjustment of industrial structure such that the overall energy efficiency is improved (David and Sinclair-Desgagné 2005). Regulation can also promote enhanced green technology innovation to increase energy efficiency in production processes (Cagno and Trianni 2013; Zhang et al. 2017; Huang et al. 2018).

However, the empirical evidence on the efficiency or productivity effect of environmental regulation has been mixed. The traditional wisdom among economists concerning environmental regulation is that it comes at an additional cost imposed on firms. Environmental regulation, such as an emission trading scheme, forces firms to allocate resources to pollution reduction, which is not beneficial for productivity or efficiency improvement (Gray 1987). Recent empirical studies supporting this argument include Rogge et al. (2011), Greenstone et al. (2012), and Bergquist et al. (2013). Rogge et al. (2011) is of particular relevance to our present study; it questioned whether the EU carbon trading mechanism triggered technology innovation. The study found that the EU carbon trading system could not provide sufficient incentives for innovation activities in the German power industry.

In contrast, the Porter hypothesis (PH) argues that environmental regulations do not inevitably hinder productivity improvement (Porter 1991; Porter and van der Linde 1995). Strict but well-designed regulations (often referring to market-based instruments) can trigger innovation that may partially or even fully offset the compliance costs. Innovation is broadly defined (Porter and van der Linde 1995), which may include genuine technological innovation, but can also include a change of input mix, a change of product portfolio, or a reduction of wasteful resource use. There is also a vast body of literature supporting the PH (e.g., Rubashkina et al. 2015; Calel and Dechezlepretre 2016; Zhao and Sun 2016; Albrizio et al. 2017). In contrast to Rogge et al. (2011) who found no evidence for the enhanced innovation impact of the EU carbon trading system, Calel and Dechezlepretre (2016) found that the EU carbon trading system increased regulated enterprises’ low-carbon technology by 10%.

Our study contributes to the debate in three ways. First, there has been a small but growing literature testing the PH in the Chinese context (e.g., Tanaka et al. 2014; Wang and Shen 2016; Stavropoulos et al. 2018; Xie et al. 2017a, b). However, most previous studies have examined the impact of environmental regulation from the perspective of innovation activities, factor productivity, or competitiveness (Ambec et al. 2013). We offer a different perspective by investigating the energy conserving effect of environmental regulation. A carbon trading scheme is by design targeted at carbon emission reduction and thus its induced innovation impact, if any, is most likely to be linked to energy consumption. In this study, we explore whether China’s CET pilot has changed its energy consumption profile and the mechanisms through which changes are induced.

Second, these existing Chinese studies used waste mitigation expenses, pollution levy, total discharge of various wastes or a composite of these discharges, and environmental protection investments as indicators of regulatory stringency. However, these indicators are notorious for measurement error due to pollution multidimensionality. These measures also present a simultaneity issue if environmentally adjusted productivity is used for analysis. Recent works measuring regulatory stringency have increasingly used quasi-natural experiments as a solution to simultaneity, based on the assumption that exogenous shocks determine regulatory stringency (Greenstone et al. 2012; Hering and Poncet 2014; Brunel and Levinson 2016; Cai et al. 2016; Chen et al. 2018). Zhang et al. (2017) used a DID method by treating the CET pilot policy as innovation and tested whether the CET pilot policy had an impact on emission reduction. Following these studies, we treat the implementation of the CET pilot as a quasi-natural experiment and employ a DID approach to the analysis of the energy conservation effect of regulatory changes (i.e., the CET pilot policy).

Third, despite the fact that China’s national emissions trading scheme was officially launched in late 2017 (NDRC 2017), rigorous assessment of the impacts of the preceding pilot is lacking. Deng et al. (2018) conducted a comprehensive survey of the enterprises covered in the seven CET pilots and offered many insightful observations about their operations and responses to the CET pilots; however, the analysis is primarily qualitative. Both Zhu (2017) and Zhang et al. (2017) investigated the environmental abatement effects of the CET pilots. Zhang et al. (2017) is the closest to the present work in that it also used a DID method to study the impact of the CET pilot policy. However, their study focused on how innovations have contributed to emission reduction and whether the CET pilot policy has had an impact on emission reduction. Energy efficiency was used as an indicator of innovation performance but their study did not show whether the CET has influenced such innovation performance. Our study complements their research by further unraveling this latter link. Specifically, we investigate the structural and innovation mechanisms through which the CET pilot policy has changed the energy consumption profile (and hence the emissions).

The remainder of this paper is organized as follows. Section 2 introduces the background of the CET policy and analyzes possible emission mitigation mechanisms. Section 3 describes the data and methods including the derivation of energy efficiency and the identification strategy for the regulatory impacts of the CET pilot policy. Section 4 presents empirical results and discussions, as well as robustness checks. Section 5 concludes the paper with policy suggestions.

2 Policy Background

China has surpassed the United States to become the world’s largest carbon emitter since 2007 (IEA 2009). Under the dual pressure of emissions reduction and sustained economic growth, China has proactively experimented with various market based policy instruments aimed at cost-effective emissions reduction. The government has embraced emissions trading more than most developing countries primarily through the CET pilot in seven municipalities and provinces since 2011 and then a national trading market from late 2017 (NDRC 2011, 2017). In a substantial effort to lay the foundations for future cost-effective alternatives, China launched seven regional pilot markets (Beijing, Shanghai, Tianjin, Guangzhou, Shenzhen, Hubei, and Chongqing) for carbon trading to gain experience ahead of the planned nationwide emission trading scheme (ETS) (NDRC 2011).

The pilot regions are given considerable leeway to design their own schemes. There are common features but also substantial differences across pilot schemes in cap measurement, allocation, entry threshold as well as punitive measures for non-compliance (Zhang 2015). Table 1 presents the main characteristics of each of the CET pilot schemes. Although there are differences in the set of industries covered, most sectors included are from the second industries (i.e., manufacturing and construction). They are typically emission sources that have relatively good data and large reduction potential. Agricultural sectors are excluded at this stage.

Table 1 Characteristics of the Chinese CET pilot schemes.

As a cap and trade policy, pilot schemes first need to quantify baseline emissions which were measured using one or a mixture of three methods: historical emission levels, verified industry baseline emission levels, or historical emission intensity. Different measuring methods can also be used in different industries even in the same pilot region (e.g., Shanghai and Hubei). Once baseline emissions are measured, pilot schemes set annual caps that are either consistent or over-comply with the CO₂ intensity reduction targets specified in the regional 5-year development plans. The caps in Beijing, Shanghai, Tianjin and Shenzhen remained unchanged for the period of 2013–2015. Chongqing set a cap that declined 4.13% annually. To account for a substantial proportion of heavy industries and accelerated industrialization, Guangdong and Hubei set their caps annually according actual conditions. Overall, these caps represent an annual reduction of 3.43–15.47% reduction in CO₂ intensity for the period of 2013–2015. The caps are binding and recent research has shown that the compliance rates are close to 100% in most pilot regions (Zhang 2015). The majority of capped quotas are allocated freely to engage enterprises and encourage participation. To avoid dramatic price fluctuation as observed at the early stage of the EU ETS, all pilot schemes in China has incorporated some mechanism to address supply–demand imbalance and associated price uncertainty. In addition to the majority free allocation, pilot schemes typically also keep some quotas for adjustments and new entrants. To facilitate market engagement and price formation, some pilots also choose to auction off a smaller proportion of the quotas and reserve some quotas to maintain price stability. For comparison purposes, we plotted observed carbon prices in China’s pilot trading schemes and EU ETS, together with implied range of carbon prices for the US (Meng 2017) and the 2nd-stage carbon price in Australia for corresponding periods.¹ Except for the relatively large fluctuations at the early stage of pilot, carbon prices in China’s pilot carbon markets quickly converged and stabilized at a range below US$ 10 per metric ton (Fig. 1).

Fig. 1

Carbon prices in China, Europe, US and Australian markets

The CET pilot regions extend from the eastern coast to the central inland. Figure 2 presents the locations of 37 CET pilot cities and 246 non-pilot cities in our final sample. The 37 CET pilot cities contributed to roughly 26.7% of the gross domestic product (GDP) and 19% of the country’s population in 2014. Over 2000 enterprises were included in the pilot, accounting for 20% total national energy consumption and 16% total national carbon emissions in 2014 (Zhu 2017).

Fig. 2

Locations of China’s CET pilot and non-pilot cities. This is only a schematic map and does not indicate definite boundaries

3 Data and Methods

3.1 Data Description

The data used in this study were primarily sourced from the China City Statistical Yearbook (CCSY) (NBS, 2004–2016]. We also cross-check the data that are available from provincial statistical yearbooks. Removing observations with missing data, we collected data for 283 prefectural-and-above (PAA) level cities in China from 2003 to 2015, including 37 pilot cities and 246 non-pilot cities. All monetary variables were deflated to constant 2003 prices.

A substantial effort has been made to compile a consistent set of input–output data at the PAA-level, including GDP, labor (total number of employees), capital, energy, and CO₂ emissions. Nominal GDP is deflated to constant 2003 prices using province-specific GDP deflators. Capital is estimated from annual fixed asset investment data using the perpetual inventory method following Xie et al. (2017a, b). Nominal investment data were first deflated to 2003 constant prices using province-specific investment deflators. Total number of employees by PAA city, fixed asset investment, nominal GDP by PAA city, and GDP deflators by province were directly sourced from the CCSYs.

A consistent and comprehensive account of total energy consumption of all sectors at the PAA level is not publicly accessible. Recent studies have mostly proposed to estimate directly from energy sources that have publicly accessible data without making strong assumptions about the composition of the complete energy structure (Glaeser and Kahn 2010; Chen et al. 2011; Xie et al. 2017a, b; Wu and Ma 2018). Although the estimated measure is still not a comprehensive account of the total energy consumption, it captures the bulk of it. It provides a reliable and consistent basis for comparison across PAA cities, and allows easy estimation of CO₂ emissions. In this study, we follow Xie et al. (2017a, b) and calculate the total energy consumption and CO₂ emissions from four sources, including electricity, natural gas, liquefied petroleum gas, and transportation.

We use the innovation index (Kou and Liu 2017) to measure enterprise-led innovation activities. The innovation index provides an aggregated account of patents and entrepreneurship by industrial sector and PAA city from 2001 to 2016, respectively. To calculate PAA city innovation index, the patent data are obtained from the National Intellectual Property Administration and the enterprise registration capital data are obtained from the National Industry and Commerce Administration. According to the micro data and location of enterprises, they calculate the PAA city innovation index. We also collected from the CCSYs the ratio of fiscal expenditure on scientific and technological development to GDP as an indicator of government-led technology innovation. All other variables used in the regression analyses were also obtained from the CCSYs with some variables deflated and/or transformed. Industry structure is indicated by the ratio of GDP share of the secondary industry. The China City Statistical Yearbook divides Chinese industry into three parts, the primary industry, secondary industry, tertiary industry. The secondary industry includes Mining, Manufacturing, Electricity, Heat, Gas and Water production and supply, Construction. Table 2 presents the definition and summary statistics of all variables.

Table 2 Description of variables and data

3.2 Total Factor Energy Efficiency

Data envelopment analysis (DEA) is widely applied to measure efficiency when multiple inputs and outputs (e.g., GDP and CO₂ emissions) are involved. The main advantage of DEA over parametric approaches is that it does not need to specify the functional form of the production relationship. However, conventional DEA models based on radial efficiency measures ignore the role of slacks and may overestimate efficiency when there exist non zero slacks in the constraints (Fukuyama and Weber 2009). Tone (2001) proposed a slacks-based model (SBM), which was a non-radial model, and was widely applied to environmental performance assessment. The model is monotone, decreasing in each input and output, and allows the estimation of target input of each input. In this study, we construct an SBM to estimate the total factor energy efficiency. The conventional SBM model is constructed on the premise of constant returns to scale under the assumption of output changes proportional to input changes. However, according to microeconomics, production may not operate at the optimal scale in real production due to the restriction of factor inputs in an imperfectly competitive market. Therefore, we build the model on the premise of variable returns to scale, as decision management units (DMU) in such a model are allowed to operate at a non-optimal scale, and it satisfies real operation (Banker et al. 1984; Banker and Thrall 1992).

Suppose there are n DMUs [e.g., cities], each DMU has three factors including inputs (x), desirable outputs ($y^{g}$) and undesirable outputs ($y^{b}$). These factors can be represented by vectors $x \in R^{m}$, $y^{g} \in R^{{s_{1} }}$ and $y^{b} \in R^{{s_{2} }}$, respectively. In this study, we suppose that each DMU uses capital, labor, and energy as inputs to generate the gross domestic product (GDP), the desirable output, and CO₂ emissions, the undesirable output. The production technology can be defined as follows:

$$P(x) = \left\{ {(y^{g} ,y^{b} )\left| {x\;produce\;(y^{g} ,y^{b} ),\;x \ge X\lambda ,y^{g} \le Y^{g} \lambda ,y^{b} \ge Y^{b} \lambda ,\sum \lambda = 1,\lambda \ge 0} \right.} \right\} ,$$

(1)

where $\lambda$ is a non-negative multiplier vector and $\sum \lambda { = 1}$, $\lambda \ge 0$ corresponds to the variable returns to scale (VRS) technology.

Following Cooper et al. (2007), we have the SBM model under VRS:

$$\begin{aligned} \phi^{*} & = \hbox{min} \frac{{1 - \frac{1}{m}\sum\nolimits_{i = 1}^{m} {\frac{{s_{{i_{0} }}^{ - } }}{{x_{{i_{0} }} }}} }}{{1 + \frac{1}{{s_{1} + s_{2} }}\left( {\sum\nolimits_{{r_{1} = 1}}^{{s_{1} }} {\frac{{s_{{r_{10} }}^{g} }}{{y_{{r_{20} }}^{g} }}} + \sum\nolimits_{{r_{2} = 1}}^{{s_{2} }} {\frac{{s_{{r_{20} }}^{b} }}{{y_{{r_{20} }}^{b} }}} } \right)}} \\ & \quad s.t.\;\;\;\;x_{0} = X\lambda + s_{0}^{ - } \\ & \quad \qquad y_{0}^{g} = Y^{g} \lambda - s_{0}^{g} \\ & \quad \quad \quad y_{0}^{b} = Y^{b} \lambda + s_{0}^{b} \\ & \quad \qquad \sum \lambda = 1 \\ & \quad \qquad s_{0}^{ - } \ge 0,\;s_{0}^{g} \ge 0,\;s_{0}^{b} \ge 0,\;\lambda \ge 0 \\ \end{aligned} ,$$

(2)

where $i = 1,2, \ldots ,m$; m is inputs number. $r_{1} = 1,2, \ldots ,S_{1}$ are desirable outputs. $r_{ 2} = 1,2, \ldots ,S_{ 2}$ are undesirable outputs. $S_{ 1}$ is desirable number and $S_{ 2}$ is undesirable number. $s_{0}^{ - }$, $s_{0}^{g}$ and $s_{0}^{b}$ are slack variables of factor inputs (potential reduction), desirable output (potential increase) and undesirable output (potential reduction), respectively.

Hu and Wang (2006) was the first to propose the total factor energy efficiency (TFEE) indicator as the ratio of target energy input suggested by DEA to actual energy input. Based on the definition, Lin and Tan (2016) used a meta-frontier slack-based DEA model to estimate TFEE in Chinese four sectors. Meng et al. (2018) also combined SBM model to estimate energy efficiency of Chinese 30 regions. According to the definition of TFEE and the superior of SBM model mentioned above, following Hu and Wang (2006) and Lin and Tan (2016), TFEE in region n can be estimated from the following equation:

$$TFEE_{n,t} = \frac{Target\;energy\;consumption\,(n,t)}{Actual\;energy\;consumption\,(n,t)} = \frac{{e_{n,t} - S_{n,t}^{e} }}{{e_{n,t} }} = 1 - \frac{{S_{n,t}^{e} }}{{e_{n,t} }} ,$$

(3)

where $e_{n,t}$ and $S_{n,t}^{e}$ are actual energy consumption and potential energy reduction (i.e., slacks in energy consumption) in region $n$ at year $t$. The target energy input of each region can be estimated from Eq. (2). The TFEE defined in Eq. (3) inherits the characteristics of our production technology [Eq. (1)] that considers not only energy inputs, but also undesirable output CO₂ emissions. According to Eq. (2), the target energy input is estimated under the constraints of minimized energy input and undesirable output. As such, the TFEE reflects not only the energy conservation effect but also the emission reduction effect, which is particularly relevant for the analysis of carbon mitigation regulations, such as the CET pilot.

3.3 Emission Mitigation Compliance Mechanisms

The cap and trade policy like the CET pilot is set up to meet mitigation targets in a cost-effective and green manner under a market-based mechanism. Compared with other types of environmental regulation, the main advantage of the CET pilot is to achieve cost-effective emission mitigation through trading. It creates a market for reducing carbon emissions by giving monetary value to the cost of carbon mitigation (Oh and Chua 2010). Enterprises can sell emission quotas to obtain profits if they can reduce emissions at low costs, while those with high emission reduction costs need to purchase quotas to maintain production. Apart from trading, conventional mitigation compliance mechanisms still apply.

First, emission mitigation regulations can effectively limit or reduce the consumption of emission intensive energy carriers, such as coal (Nayyar et al. 2014), and thus facilitate a transition towards energy carriers with high calorific value but low carbon emissions, such as natural gas (Johansson 2006). Johansson (2006) indicated that emission trading with free emission quota allocation has incentives for emission mitigation through fuel substitution. The climate policy increases the cost of emission-intensive fuels significantly, which may force industrial sectors to consider fuel switching where the switching cost is lower than the compliance cost otherwise. This may involve transition from coal to oil and natural gas in the short run and renewables in the long run. As energy demand is inevitably on the rise due to rapid urbanization and industrialization, it is only judicious to divert energy towards less emission-intensive fuels.

Second, emission mitigation regulations can also induce energy efficiency improvements by changing industrial structure and catalyzing efficiency-oriented technology progress. On the one hand, regulations can induce the adjustments of the production portfolio at the enterprise level and changes of the intra-industry production structure, inter-industrial structure, and inter-regional industry layout at the regional level. Energy-inefficient and hence emission-intensive production is more sensitive to mitigation regulations and is more likely to be scaled down (Cole et al. 2005), or relocated from regulated regions to unregulated regions (Copeland and Taylor 1994). As a market-based regulation tool, the CET pilot policy mainly selects high-energy consumption sectors, such as iron and steel, chemicals, and non-ferrous metals as pilot industries. The CET pilot policy may initially encourage these high-energy consumption companies to optimize their production methods and reduce costs within the industry. As time passes, inter-industrial structure and inter-regional industry adjustment may occur. Such structural adjustments will be reflected in energy efficiency improvements in regulated regions and help to achieve compliance with mitigation regulations (Adom et al. 2012; Jin 2012).

On the other hand, the PH argues that environmental regulations can induce enhanced green technology innovation (Porter 1991; Porter and van der Linde 1995). In the context of the CET pilot, such green innovation is most likely oriented towards energy efficiency improvements. The PH propose that regulatory pressure would have forced enterprises to make adjustment to existing production or innovative new ways that are less energy- and, hence, pollution-intensive (Cagno and Trianni 2013; Zhang et al. 2017). Although technology innovation may take a certain amount time to affect energy efficiency, in the long term, the cost changes caused by the CET pilot policy may force enterprises to make adjustments to the existing production methods and carry out innovation. Such innovation reduces not only pollution governance expenditure, but also energy demand and utility.

Simultaneously, technology progress may also have a rebound effect, the well-known phenomenon that improving energy efficiency may save less energy than expected due to substitution and income effects (i.e., the backfire hypothesis). The net effect of regulation on efficiency improvement is often subject to uncertainty and the existing research provides little support for the backfire hypothesis (Gillingham et al. 2016).

3.4 Identification Strategy

The empirical analysis in this study appeals to a panel data set from 283 PAA-level Chinese cities, including 37 CET pilot cities. The first section asks whether the CET pilot has had any impact on energy consumption structure, industrial structure, and innovation activities. It also provides evidence about the influencing mechanisms of CET pilot policy on energy efficiency. We employ a DID approach to isolate the regulation-related casual effects. Specifically, we estimate the following equations:

$$IND_{it} ,RD_{it} ,ES_{it} = \alpha_{0} + \alpha_{1} CET_{i} + \alpha_{2} POST_{t} * CET_{i} + \alpha_{3} x_{it}^{'} + \gamma_{t} + \varepsilon_{it} ,$$

(4)

where x_it^′ is a vector of control variables, which includes FDI, government intervention (GOV), urbanization level (URBAN) and economic level (per capita GDP). $\gamma_{t}$ and $\varepsilon_{it}$ correspond to year fixed effects and an idiosyncratic error term. $IND_{it}$ is industrial structure. $RD_{it}$ denotes technological innovation, we select two types of technological innovation including enterprise-led innovation and government-led innovation, which are represented by RD_PRIVATE and RD_GOV, respectively. $ES_{it}$ is energy structure. As $IND_{it}$, $RD_{it}$, and $ES_{it}$ are ratio value and zero value of them accounts for 16% of our final sample, we run regressions with $IND_{it}$, $RD_{it}$, and $ES_{it}$ in levels instead of logs, as it allows us to include zero values.²POST_t is a dummy for the enter time of the CET pilot policy. CET_i is a dummy for pilot cities. POST_t is 1 since 2011, and CET_i of the pilots is 1. $\alpha_{ 2}$ in Eq. (4) identifies the industrial structure, innovation, and energy structure effects of the CET pilot respectively. x_it^′ is a vector of control variables including PGDP, government intervention, urbanization and FDI.

The second section then focuses on the energy efficiency effect of the CET pilot and asks whether the implementation of the CET pilot led to different variations of energy efficiency in the CET pilot cities as compared to non-CET pilot cities.

$$\ln (TFEE_{it} ) = \beta_{0} + \beta_{1} CET_{i} + \beta_{2} POST_{t} * CET_{i} + \beta_{3} x_{it}^{'} + \gamma_{t} + \varepsilon_{it} ,$$

(5)

where $\ln (TFEE_{it} )$ is the estimated total factor energy efficiency [Eq. (3)] of city i at year t in logs. x_it^′ is a vector of control variables that have been shown to affect energy efficiency. First, the Chinese goal of energy conservation and CO₂ reduction could be achieved through industry structure optimization (Jin 2012). Wang et al. (2013) indicated that industrial structure is one of the driving factors of energy intensity. We use the ratio of GDP share of the secondary industry to reflect it. Second, we include technological progress in energy-saving and emission mitigation, as it is undoubtedly a significant method to improve energy efficiency (Wang et al. 2014). Wang et al. (2014) analyzed Chinese energy efficiency from static and dynamic perspectives, and found that technical progress is the major motivation for energy efficiency improvement. Third, we use FDI to reflect openness level. Some scholars have analyzed the impact of FDI on energy intensity (Hübler and Keller 2010; Elliott et al. 2013). Fourth, we use population density to estimate urbanization, as Zhang et al. (2014) found that urbanization leads to carbon emission intensity growth. Fifth, we use the ratio of fiscal revenue to GDP to estimate government intervention. Han et al. (2018a, b) indicated that excessive government intervention would lead to distortions in resource allocation, which was not conducive to improving energy efficiency.$\beta_{2}$ in Eq. (5) only captures the average regulatory impact on energy efficiency. It is also of interest to know whether the CET impact is lagged, prolonged, or decays. We employ the following difference-in-difference-in-differences (DDD) specification to identify the dynamic effects of the CET pilot policy on energy efficiency:

$$\ln (TFEE_{it} ) = \beta_{0} + \beta_{1} CET_{i} + \beta_{ 2} x_{it}^{'} + \sum\limits_{j = 201 1}^{2015} {\beta_{j} } POST_{t} * CET_{i} * year^{j} + \gamma_{t} + \varepsilon_{it} ,$$

(6)

where $year^{j}$ are year dummies, and $\beta_{j}$ corresponds to the marginal effect of CET on energy efficiency in each of the post-treatment year j relative to the pre-treatment period.

In the last section, we conduct a series of robustness checks and sensitivity analyses. The validity of our estimation hinges crucially on the assumption that there is no differential pre-treatment trend between the CET pilot and non-CET pilot cities. We first verify that there is no significant deviation from this assumption (Kudamatsu 2012). We also check whether our results are robust to the choice of the treatment year. The CET pilot was intensively discussed and officially announced in 2011 (NDRC 2011). Although preparatory work and actual start time of regulatory changes slightly varied across different pilot cities, the market could well take the discussion and announcement as strong signal of commitment to regulatory changes and respond with immediate actions. We thus conduct sensitivity analyses taking 2012, 2013, 2014, or 2015 as the alternative starting year of the CET pilot. We address the concern that our results may be biased due to omitted variables at the city level by conducting a placebo test in which we randomly assign the CET pilot cities (see, e.g., Chetty et al. 2009; La et al. 2012; Cai et al. 2016, for similar practices). As a further check that there may be an omitted selection, we conduct the propensity score matching (PSM) and rerun the regressions on the matched sample.

4 Empirical Results

4.1 Energy Structure, Industrial Structure, and Innovation

Strict but flexible emissions mitigation regulations, such as emission trading schemes, provide incentives for a great variety of compliance responses, such as fuel substitution, reduced production of carbon intensive products, and energy efficiency improvements through enhanced innovation activities. Table 3 presents the results from the estimation of Eq. (4). Column 2 shows that the CET pilot had a positive effect on energy structure by increasing the proportion of natural gas consumption and the effect is statistically significant at the 1% level. It shows that the implementation of the CET pilot policy can enhance the proportion of clean energy consumption (e.g., natural gas). The estimated effect (0.035) represents a substantial average treatment effect on the treated (ATT) compared to the sample mean of 0.06 for the treatment group reported in Table 2. The magnitude of the effect may be biased upwards as the available data only present a partial account of total energy consumption. Our results, nevertheless, suggest that there is a significant energy structure effect of the CET pilot.

Table 3 The structure and innovation impacts of the CET^†

Even though most sectors covered in the CET pilot regions are from the secondary industry, regulatory pressure has not shifted production out of the industry (Column 2). Contrarily, the impact of the CET pilot on the GDP share of the secondary industry appears to be positive. This result suggests a strong decoupling effect of the CET pilot. However, the effect (0.006) is not statistically significant at the 10% level and does not have much practical significance compared to the average GDP share of the secondary industry (0.49). This result is perhaps very sensible given that we have only a few years of trading pilot. And one would not expect dramatic changes in industry shares within such a short period of implementation of the policy. However, the CET pilot does allow mitigation compliance through trading among covered sectors, which are primarily from the secondary industry. The scale effect and technology spillover effect of the secondary industry can be significant. Brännlund et al. (2014) also found that pollution reduction policy could increase the output value of Swedish manufacturing and reduce pollution emissions at the same time. The CET pilot may encourage structural shifts within the secondary industry, which optimize their production methods and promote green innovation. We unfortunately cannot identify such intra-industry shifts given the available data.

The results in the last two columns of Table 3 also indicate positive impacts of the CET pilot on innovation activities. It effectively stimulated enterprise-led innovation activities in the form of patent accumulation and entrepreneurship in the pilot regions. The result provides empirical support for the weak version of the PH (Ambec et al. 2013). This is well in line with the literature testing the PH in the Chinese context (Tanaka et al. 2014; Wang and Shen 2016; Stavropoulos et al. 2018; Xie et al. 2017a, b), although most existing studies have approached the PH from the stronger perspective. Apart from the enterprise-led innovation activities and possibly new venture capital flowing into the pilot regions, these areas are more likely to benefit from special funds allocated by the central and local governments as part of the pilot scheme. The estimated effects on government-led and enterprise-led innovations (0.001 and 0.089) both represent substantial ATT given the sample means of 0.002 and 0.22 for the pilot cities.

The results presented in Table 3 are largely consistent with what has been found in firm-level surveys. For instance, in a recent survey of the enterprises covered in the CET pilots, Deng et al. (2018) found that most enterprises in all pilot areas have responded to the new regulation by upgrading technology. In addition, the enterprises in pilot areas other than Guangdong and Chongqing also responded by switching fuels, whereas those in Guangdong chose to respond more by controlling output mix.

4.2 Average and Dynamic CET Impacts on Energy Efficiency

Table 4 presents the average and dynamic impacts of the CET pilot on total factor energy efficiency (TFEE) at the PAA level. Our dependent variable TFEE takes on limited values and is left-censored. We thus employed a random-effects interval-data regression model. Panel models with fixed individual effects in this case are widely understood to be biased and inconsistent (Greene 2004a, b). Because TFEE was estimated from input–output data, we also excluded energy structure (ES) and GDP from the set of control variables due to endogeneity concerns.

Table 4 Average and dynamic CET impact on total factor energy efficiency^†

We explored a range of different specifications. Column (1) is the baseline specification including only variables of our primary interest. Columns (2) and (3) include additional control variables sequentially. The results in columns (2) and (3) show that on average, regions with a higher GDP share of secondary industry also present higher TFEE. The TFEE is negatively associated with enterprise-led innovations (through patent accumulation and entrepreneurship), but not with government-led innovations (through fiscal scientific and technological expenditures). This is probably because a substantial proportion of government-led innovation expenditures flow to research institutes and universities to support basic scientific and technological R&D. The impact on regional energy efficiency may take a long time to be visible, depending on the commercialization competence in the regional economy. Due to the bulky nature of the upfront innovation expenditures, a negative association between enterprise-led innovation activities and efficiency performance in the relatively short term is not uncommon (Berman and Bui 2001; Greenstone et al. 2012). Enterprise-led innovation efforts are often directly oriented towards immediate application. The capital costs associated with innovations are often high and such investment will have both short-term and long-term effects on resource allocation and efficiency performance. In the context of CET, enterprise-led innovations are likely to be biased towards pollution mitigation, which may crowd out efficiency-oriented innovations and put future efficiency at risk (Acemoglu, 2002; Lanoie et al. 2008). In the short term, technology innovation activities may themselves be energy-intensive and exert an immediate and negative impact on energy efficiency.

The CET impact on energy efficiency can be channeled through an industrial structure effect and innovation effects, which we explored in Table 3 and controlled for in Table 4. Efficiency improvement can also be achieved through a mix of changing energy structure, intra-industry adjustments, or unpatented efficiency innovations. We unfortunately cannot identify these effects separately. Nevertheless, after controlling for the industrial structure, government-led innovations, and enterprise-led patented innovations, there remains a positive and statistically significant CET impact on TFEE regardless of the model specifications. The estimated coefficient of the DID term CET*POST is reasonably robust and falls roughly in the range of 0.10–0.14. With logged TFEE as the dependent variable, this result translates to an average CET impact of roughly a 10% increase in TFEE. This result seems to provide empirical support for the strong version of the PH (Ambec et al. 2013). Our result is consistent with Schleich et al. (2009) who also found that the EU emission trading scheme stimulated the improvement of energy efficiency.³

Column (4) in Table 4 presents the results for the dynamic specification [Eq. (6)]. While the average CET impact on TFEE is roughly a 10% increase for the period under study, the impact evolves with the gradual implementation of the CET over time. Market seems to have responded immediately with the discussion and announcement of the CET policy in 2011. However, the impact appeared the strongest in the following year. The TFEE was significantly lifted by about 20% as compared to the pre-treatment period. The effect remains positive and statistically significant within 4 years of the CET announcement.

4.3 Robustness Checks

4.3.1 Pre-treatment Trends

The validity of our estimation hinges crucially on the assumption that there is no differential pre-treatment trend between the CET pilot and non-CET pilot cities. A concern with the results in Table 4 is that the cities covered in the CET pilot might already have been on a higher or lower efficiency trajectory. The cities might even have been selected precisely on an efficiency criterion, that is, the assignment of the pilot cities may not be random. The fact that the pilot cities are mostly located in relatively more developed regions further aggravates the concern. A significant deviation from the assumption of parallel pre-treatment trends will lead to a biased estimate of the CET treatment effect. We address this concern by conducting the following parallel trend test (see, e.g., Kudamatsu 2012; Alder et al. 2016, for similar practices):

$$TFEE_{it} = \sum_{n = - 8}^{ 4} (\rho_{t} \times I_{t}^{t - post} \times CET_{i} ) + \lambda x_{it}^{'} + \gamma_{t} + \varepsilon_{it} ,$$

(7)

where $I_{t}^{t - post}$ takes the value of 1 when $t - post = n$, and 0 otherwise. $post$ indicates the year of the implementation of the CET pilot (i.e., 2011). Given the period of our sample, $n$ can take the values of − 8, − 7, − 6, …, 0, …, and 4. By comparing the coefficient estimates $\rho_{t}$ before and after the treatment, one can tell whether or not the parallel pre-treatment trend assumption holds.

Figure 3 presents the estimated coefficients of $\rho_{t}$ with corresponding 95% confidence intervals. The assumption of parallel pre-treatment trends generally holds.⁴ Estimated differences between the treatment group and the control group are consistent with those reported in Table 4. The treatment effect remained statistically significant for the first 4 years after treatment. However, most effects estimated for the pre-treatment years are insignificant. More importantly we do not observe a trend that the difference between the two groups is consistently increasing or decreasing such that estimated post-treatment effects are biased upwards or downwards.

Fig. 3

Test of parallel trends

As an additional test for the assumption of pre-treatment parallel trends, we performed baseline regressions using each of the year during 2004–2010 as the false treatment year on the sub-sample for 2003–2010. If the parallel trend assumption holds, we would expect the coefficient of the false DID term (CET*POST) is neither statistically nor economically significant. The regression results are now reported in Table 5 of the revised manuscript. Taken together, we have no evidence against the parallel trend assumption for the pre-treatment period 2003–2010.

Table 5 Parallel trend tests using the sub-sample for 2003–2010^†

4.3.2 Propensity Score Matching (PSM)

To correct for the potential bias caused by nonrandom assignment of the pilot cities, we employed the treatment effect model using propensity score matching (PSM). Specifically, by conditioning on the set of variables $x_{it}^{'}$, we constructed a proper counterfactual sample of non-pilot cities that are observably similar to the pilot cities. Rather than linearly matching on the set of covariates $x_{it}^{'}$, we used PSM (to the nearest neighbor), which denotes the predicted probability that a particular city is assigned to the CET pilot. After the matching process, we then estimated regressions identical to the specifications of Eqs. (5) and (6) on the matched sample. Table 6 presents the results. The average and dynamic impacts of the CET pilot are reasonably robust to the matched sample. We still identify a positive and statistically significant impact of the CET pilot on TFEE, which remains significant within 4 years of the CET announcement.

Table 6 Results of PSM-DID^†

4.3.3 Temporal and Spatial Placebo Tests

The PSM can only mitigate selection bias due to observables, and our results may still be biased due to possible omitted variables at the year or city level. We address this concern by conducting two placebo tests with the variation in the treatment time and the random assignment of the CET pilot cities, respectively.

We first checked whether our results are robust to the assignment of treatment year. As we noted earlier, the CET pilot was officially announced in 2011; however, preparatory work and actual start time of trading slightly varied across different pilot cities. The range of sectors covered in each city may also expand sequentially. For instance, initial trading sectors in Shenzhen include electricity, water supply, and manufacturing, but construction and public transportation were later included in 2014. We thus conducted the sensitivity analysis taking 2012, 2013, 2014, or 2015 as the false treatment year of the CET pilot. For brevity, we only conducted the analysis for the average effect of the CET pilot. Note that we have identified a positive and statistically significant impact of the CET pilot on TFEE when 2011 is taken as the treatment year (reproduced in the first column of Table 7). The effect remain significant using 2012 as the treatment year; however, the positive coefficient on the DID term CET*POST loses its statistical significance and even becomes negative when we assign the treatment year 2013, 2014, or 2015 as shown in Table 7. This may be a result of more “true” treatment years being falsely included in the pre-treatment period as the false treatment year is pushed further to later years. Despite the varying start time of trading across pilot cities and the changing scope of pilot sectors over time, the signal of regulatory changes introduced by the CET pilot seem to have caused an immediate impact with the official announcement. This is again consistent with firm-level evidence presented in Deng et al. (2018). It was found that at the early stages of the ETS pilot, enterprises were concerned about the shortage in free allowances in the future and thus took very proactive mitigation strategies.

Table 7 Placebo tests of the CET treatment year (T.Y.) ^†

Following Chetty et al. (2009), La et al. (2012) and Cai et al. (2016) who select pilot samples randomly to perform robustness testing, we conducted a spatial placebo test by randomly assigning the pilot status to cities. Specifically, we randomly selected 37 cities from the total sample and constructed false treatment variable $CET^{false}$ based on the random sample. We then re-estimated the specification using Eq. (5). The randomization process ensures that the newly constructed DID term $CET^{false} * POST$ should have no impact on TFEE. In other words, a statistically significant coefficient on the DID term would indicate the misspecification of Eq. (5). We repeated the random sampling 1000 times and plotted the estimated coefficients on the false DID term in ascending order with corresponding 95% confidence intervals in Fig. 4. These 1000 coefficient estimates center around zero (mean = − 1.36e−04) and most estimates (79%) are not statistically significant at the 5% level. Our true estimate (0.137, marked by the solid line) is above the 99.3 percentile of the distribution of the 1000 coefficient estimates. Overall, these results do not suggest severe bias due to omitted variables in the estimate of the CET impact on TFEE.

Fig. 4

Estimated coefficients in ascending order with 95% confidence intervals

4.3.4 Stability of Unit Treatment Value Assumption

Our final robustness check is concerned with possible bias caused by shift in production across the treatment group and the control group. An important assumption for unbiased estimates of treatment effect is that the outcomes of the pilot cities are independent of the status of other cities, which is known as the stable unit treatment value assumption (SUTVA). If the CET policy causes energy-intensive production to transfer to cities exempt from the pilots, our estimates of the treatment effect of the CET policy would be biased. In a way, the inclusion of GDP share of the secondary industry as a control variable helps to mitigate this concern; however, it does not help to address the possibility of structural shift within the secondary industry. Following Fowlie et al. (2012), we perform a robust test on three subsets of the non-pilot cities to explore possible biases.

If the introduction of the CET causes production to shift to non-pilot cities, and if such shift disproportionately affects pilot cities in proximity, we would expect to find larger treatment effects when the control group is restricted to nearby cities. We construct two subsets of the control group. In the first analysis, we remove the non-pilot city that is closest to each of the pilot city based on spatial distance (Han et al. 2018a, b). In the second analysis, we remove the non-pilot city that is farthest from each of non-pilot city. After accounting for overlaps, we end up removing 25 cities in the first subset and 45 in the second. Results based on the two subsets of the control group are reported in Column (2) and (3) of Table 8. To ease comparison, we also reproduce the baseline results based on the full sample. In either case, the results do not suggest substantial bias. In fact, the estimated treatment effect after removing the closest cities is almost identical to the baseline. The estimated treatment effect is slightly smaller rather than larger (as one would expect should upward bias arise).

Table 8 Robust tests for SUTVA^†

Bias could also arise if the CET policy has induced production to shift to regions with less stringent regulatory environment. If this occurred, we would also expect to find the treatment effect biased upward. Here we construct a third subset of the control group consisting only of non-pilot cities whose approved carbon intensity reduction targets are no higher than the pilot cities in the region with the lowest reduction target for the period of 2011–2015.⁵ Results based on this subset are reported in Column (4) in Table 8. We do not observe an upward bias. Instead, the estimated CET effect on energy efficiency is even smaller than the baseline results.

5 Conclusions and Policy Implications

China’s exploration and experiments with market-based environmental regulation—emission trading schemes in particular—has started since as early as the late 1980s. The most recent effort involves the seven region-wide and 6-year-long pilot policy of carbon emission trading, which culminated in the official launch of the national CET in late 2017 (NDRC 2017). Quantitative assessment of the conservation mechanisms and effects of the CET pilot will undoubtedly have important implications on the design and possible effects of the upcoming national scheme. Yet, despite the reasonably wide scope and long duration of the CET pilot, empirical evidence of its effects is lacking.

Our study complements Zhang et al. (2017) by unraveling the link between the CET pilot and environmental performance. Using PAA-level data and a DID approach, we show that the CET pilot has significantly improved the TFEE of the pilot cities as compared to the non-pilot cities. The market seems to have responded immediately in the first year of CET announcement. The CET impact on energy efficiency remains positive and statistically significant within 4 years of the CET implementation. Furthermore, the pilot also improves energy structure by increasing natural gas consumption and stimulates both government-led and enterprise-led innovation activities in the pilot regions although the long-term efficiency effect of innovation remains to be seen. The impact of the CET pilot on the GDP share of the secondary industry appears to be positive. Regulatory pressure has not shifted production out of the industry, while adjustments in industrial structure also contribute to enhance TFEE. Thus, the CET pilot has triggered statistically significant adjustments in energy structure, industrial structure, and technology innovation.

Our results have potentially significant implications for China’s policy-making in the next decade. Market-based regulatory tool such as the CET pilot should be promoted to accelerate emission reductions and achieve mitigation targets. Nurturing market environment and market-based regulatory capacity in mitigating carbon emissions will facilitate transition to low-carbon energy consumption and compliance with clean energy targets in the mid to long term. During emission quota allocation, industry and companies employing clean energy technology, such as wind and solar energy, should be given more policy bias. More detailed and nationwide carbon trading policy including emission quota allocation and pricing mechanism should be formulated in order to create and initiate a favorable environment for companies to transfer energy structure.

Rather than relying on technology mandates for innovation (as we did in past years), China’s central and local governments should further promote companies to make technical progress independently through the national ETS. The market-based approach can facilitate the achievement of mitigation targets through technology upgrading. More financial support policy and communication channels are required to boost the innovation application from research institutions to companies. Such efforts can transfer both government-led innovation and enterprise-led innovation into green productivity. As the carbon trading gives monetary value to the cost of carbon mitigation, a convenient transaction platform should also be established to maximize company profits through developing and deploying carbon mitigation technology innovation.

The recently launched national ETS takes the electricity sector as a breakthrough, which has the potential to ease the transition and adoption of cleaner generation technologies. It aims to optimize the energy consumption structure of the electricity sector and transform it into low-carbon development through market mechanisms. Besides scaling up the national ETS coverage, the government should include traditionally high-emission and output sectors in ETS, as CET does not shift production out of the industry and facilitates industrial structure adjustment. Since the proportion of high-pollution industries is much greater than that of high-tech industries in China, it is necessary to speed up the elimination of backward production capacity, improve the competitiveness of the industry, and enhance intra-industry development quality.

Due to the limitation of data availability, we only considered the energy structure of conventional energy carriers. Given China’s recent push for renewable energy development, investigating the impact of CET on renewables and interaction with other renewable energy development policies would be of great interest and an important area for further research.

Appendix

We select ten variables, including GDP per capita (PGDP), industrial structure (IND), government-led innovation (RD_GOV), enterprise-led innovation (RD_PRIVATE), FDI, government intervention (GOV), urbanization (URBAN), the ratio of capital to labor (K/L) and energy intensity (EI), to calculate the propensity. The Logit regression results are shown in Table 9. According to the matching process, we obtain a matched sample of 3519 observations. We find that most coefficients are statistically significant at the 1% level.

Table 9 The Logit regression results for the PSM^†

Table 10 compares variable means between pilots and non-pilots before and after matching. There are statistically significant differences between the two groups for most variables before matching; however such differences become insignificant after matching.

Table 10 Comparison between variables before and after employing matching

The comparison of logit regressions using unmatched and matched samples are shown in Table 11. Pseudo-R indicates the goodness-of-fit of the logit model based on the unmatched and matched samples. The LR statistic tests the joint insignificance of all variables in the model using unmatched and matched samples. B(%) denotes the standardized bias difference between the unmatched and matched samples. The different statistics all suggest that the systematic difference between the treatment group and control group is substantially reduced. Overall, there is insignificant difference between pilot cities and matched non-pilot cities after matching. The matched samples are now appropriate for our identification purposes.

Table 11 Results of test before and after matching