China’s Low-Carbon City Pilot Policy, Eco-Efficiency, and Energy Consumption: Study Based on Period-by-Period PSM-DID Model

Abstract

The sustainable development of Chinese cities is of long-term significance. Multiple environmental regulatory instruments aim to promote the parallel advancement of environmental conservation and economic growth. This study examines three batches of low-carbon city pilot (LCCP) programs, employing eco-efficiency as the outcome variable. Using conventional difference-in-differences (DID) models, time-varying DID models, and period-by-period propensity score matching DID (PSM-DID) models with city and time fixed effects, we investigate the comprehensive impact of pilot policies on both economic and environmental performance. Eco-efficiency, measured through the Data Envelopment Analysis (DEA) model, exhibits a strong correlation with energy consumption patterns, as carbon emissions and air pollutants predominantly originate from non-clean energy utilization. The analysis reveals that LCCP policies significantly enhance eco-efficiency. These findings demonstrate robustness across placebo tests, endogeneity treatments, and alternative outcome variable specifications. The first and third LCCP batches significantly improve eco-efficiency, whereas the second batch demonstrates no statistically significant effect. Significant impacts emerge in regions where cities hold pilot status while provinces do not; conversely, regions where both cities and provinces participate in pilot programs show no significant effects. Finally, from an energy consumption perspective, policy recommendations are proposed to further enhance eco-efficiency through regulatory instruments.

1. Introduction

China’s socio-economic development has ushered in a new era of high-quality growth defined by a green and low-carbon transformation, while the construction of its ecological civilization persists in a critical phase fraught with mounting pressures and a challenging journey ahead. Since the 1970s, the concept of sustainable development has gradually evolved. Theoretical frameworks including industrial ecology, eco-efficiency, eco-design, and the X-factor revolution constitute the conceptual foundation of sustainable development, with eco-efficiency representing one of the most quantifiable analytical approaches. In 1990, Schaltegger and Sturm first formally introduced the concept of eco-efficiency within academic discourse [1]. Subsequently promoted by the World Business Council for Sustainable Development (WBCSD), eco-efficiency has garnered considerable scholarly attention within sustainable development research, emerging as a prominent focus among researchers and corporate managers.

According to the latest Sustainable Development Goals assessment in the Energy Progress Report [2], economic growth remains intrinsically linked to increased resource utilization, particularly energy consumption. Enhancing energy efficiency and optimizing the energy mix represent critical pathways for improving eco-efficiency. Key drivers facilitating concurrent economic growth and greenhouse gas mitigation include the deployment of new clean energy power generation infrastructure and the expanded integration of renewable energy sources. Such initiatives are fundamental for sustaining economic advancement while simultaneously mitigating atmospheric pollutant emissions, thereby fostering sustainable economic development. Therefore, the structure of energy consumption is particularly important for enhancing eco-efficiency. Relevant studies have shown that utilizing photovoltaic technology in the construction industry can significantly reduce energy consumption [3]. Additionally, China Mobile has proposed a deep reinforcement learning algorithm called DeepEnergy, which aims to enhance the carbon efficiency of mobile networks and reduce carbon emissions [4]. Significant progress has been made in low-carbon, with innovations including customized carbon accounting methodologies for different wood types [5], carbon price equilibrium mechanisms [6], and the carbon storage enhancement from harvested timber products [7]. Forestry demonstrates particular advantages in achieving higher emission reduction targets at lower costs [8]. Furthermore, environmental governance, green finance, production-related emission reduction technologies, and energy production emission reduction technologies all play vital roles in advancing carbon neutrality [9].

In the course of economic development, governments and industry sectors globally have fulfilled their respective functions. Particularly following the proposition of low-carbon development objectives, a multitude of policy instruments have been promulgated and implemented across diverse governmental departments and dimensions to curtail global carbon emissions, safeguard resources and the environment, and realize sustainable development for human societies. With the evolution of regulatory economics, notably the emergence of social regulation theory, scholars have progressively adopted the concept of environmental regulation, which possesses distinct economic attributes. Environmental regulation was initially advanced by the American economist Marshall, who conducted a theoretical analysis of environmental issues and introduced the concepts of internal and external economies. It refers to governmental participation in market activities via regulatory mechanisms, aiming to compensate for market failures through macroeconomic intervention, thereby simultaneously achieving energy conservation, emission reduction, and economic growth. Subsequently, in 1991, Professor Porter of Harvard University, adopting a dynamic perspective, postulated that judicious environmental policies, while elevating corporate production costs, concurrently incentivize green innovation. This innovation ultimately counterbalances regulatory expenditures, enabling firms to attain a “win–win” scenario between economic development and environmental protection [10]. This theoretical proposition is designated the “Porter Hypothesis.”

The National Development and Reform Commission issued notices pertaining to the low-carbon city pilot (LCCP) policy in 2010, 2012, and 2017. The first batch encompassed 8 cities and 5 provinces, the second batch included 28 cities and 1 province, and the third batch comprised 45 cities. The LCCP policy exhibits dual regulatory attributes: manifest environmental regulation and latent environmental regulation [11]. From a policy instrument perspective, it functions as a binding agreement, constituting a form of mandatory environmental regulation. However, regarding the selection process for pilot areas, it demonstrates characteristics of voluntary environmental regulation. In the absence of mandatory directives, regions proactively apply for LCCP designation and autonomously formulate pilot implementation plans. These plans target low-carbon production and consumption patterns, aiming to foster urban development within a constructive and sustainable social ecosystem.

Nevertheless, a critical gap exists in understanding whether these LCCP initiatives successfully realize the low-carbon societal objectives envisioned by policymakers and whether they effectively enhance regional eco-efficiency. These constitute the core questions this study seeks to address. Consequently, this research aims to empirically investigate the impact of voluntary environmental regulation policies, specifically the LCCP, on city-level eco-efficiency. We seek to answer the following key questions: How does a city’s regional context influence the effectiveness of these policies? Does the temporal classification of a city’s pilot batch (i.e., first, second, and third) affect policy outcomes? Is the policy effectiveness altered if the city’s province is concurrently designated as a pilot area? Furthermore, given its significance as an environmental factor, does energy structure mediate the effects of the LCCP policy? Addressing these crucial research questions forms the central objective of this study.

Existing research on low-carbon city pilots (LCCP), while substantial, exhibits three primary limitations. First, most studies rely solely on samples from the initial or second pilot batches, examining the effects of a single cohort. Second, previous research predominantly concentrates on macro levels, such as provincial or industry analyses, with significantly fewer investigations conducted at the city level. Third, models assessing policy effectiveness often employ traditional difference-in-differences (DID) approaches, which suffer from numerous constraints in practical application, necessitating further verification of their empirical conclusions. This study directly addresses these gaps: it investigates policy effects critically at the city level and utilizes robust time-varying DID and period-by-period PSM-DID models for comprehensive testing and comparative analysis, strengthening the research findings; furthermore, it encompasses a broader timeframe with more comprehensive data, yielding empirical conclusions of greater credibility that rigorously satisfy robustness requirements.

The remainder of this paper is organized as follows: The Section 2 summarizes the relevancy of existing environmental policy instruments with eco-efficiency. The Section 3 introduces the eco-efficiency measurement model and the time-varying DID model used in this study. The Section 4 explains data sources and processing methods. The Section 5 presents the empirical analysis results with a brief interpretation. The Section 6 further conducts robustness tests using period-by-period PSM-DID models, among others, The Section 7 analyzes the heterogeneity of LCCP policies, The Section 8 conduct a pathway analysis, The Section 9 provides an empirical summary and offers recommendations.

2. Literature Review

Currently, there are mainly two methods for measuring the green and sustainable development level. One is the index method, in which a three-level indicator system is established and a green development index is constructed by integrating information from all levels of indicators [12]. This index primarily includes three categories: economic growth and greening degree, resource and environmental carrying capacity, and government policy support. However, this method uses expert weighting, which introduces some subjectivity and can lead to biased conclusions. Moreover, these three indicators overlook the impact of green development on economic growth, making the evaluation dimension overly singular. The other method is the efficiency approach, using data envelopment analysis (DEA) and similar techniques to measure input–output efficiency. It emphasizes the core characteristics of green development: achieving higher economic output with less resource input while reducing environmental pollution. Key concepts include eco-efficiency and environmental efficiency. The concept of eco-efficiency is proposed by Schaltegger and Sturm, defining it as the ratio of economic output to total environmental pollution [1]. In 1992, WBCSD formally proposed the concept of eco-efficiency. In 1998, based on input–output theory, OECD first extended the measurement of eco-efficiency to enterprises, industries, governments, and even entire societies, viewing it as a comprehensive measure of input–output relationships that include environmental impacts. The European Environment Agency believes that the key to eco-efficiency is to achieve more welfare with fewer resource inputs. It is found that changing the proportion of renewable energy can increase the technical eco-efficiency, with the proportion of fossil fuel consumption and eco-efficiency presenting an inverted U-shaped functional relationship [13]. The above definition and calculation of eco-efficiency take into account both ecological environment and economic development, so as to avoid the problem of a single dimension.

During the process of economic development, governments and industrial sectors across various countries have played their due roles. Particularly since the proposal of low-carbon development goals, multiple policies have been promulgated and implemented across various sectors and dimensions, aiming to control global carbon emissions, protect resources and the environment, and achieve sustainable development for human society. With the advancement of regulatory economics, especially the emergence of social regulation theory, scholars have gradually adopted the concept of environmental regulation, which carries stronger economic attributes. Environmental regulation was first proposed by the American economist Marshall, who analyzed environmental issues at a theoretical level and introduced the concepts of internal economies and external economies. Environmental regulation refers to the government’s intervention in market activities through regulatory means, with the objective of addressing market failures through macro-control, thereby enabling energy conservation, emission reduction, and economic development to proceed in parallel. In 1991, Professor Porter from Harvard University, based on a dynamic perspective, proposed that reasonable environmental policies may increase production costs for enterprises but can also incentivize green innovation, ultimately offsetting regulatory costs and enabling enterprises to achieve a “win–win” scenario of economic development and environmental protection. This theory is known as the “Porter Hypothesis.” For instance, China’s smart city pilot policy reduced carbon emissions from industrial enterprises by 23% through measures such as strengthening environmental regulation and encouraging green technology innovation [14], which aligns with the research conclusions of the Porter Hypothesis.

The research of many scholars has confirmed the significant promoting effect of China’s policy instruments on eco-efficiency. In provincial-level studies, Li used the DEA method to measure the eco-efficiency of 30 provinces in China 1997–2010 from the perspective of local government competition. The study found that after 2003, the role of local policy instruments in regional eco-efficiency shifted from “restraint” to “promotion,” or the “restraint” effect weakened [15]. A super-efficiency SBM model including non-desirable outputs was adopted to measure the eco-efficiency values of Chinese provinces, finding that different types of policy instruments have varying impacts on regional eco-efficiency [16]. There is a “U”-shaped relationship between “governance input-type” environmental regulations and eco-efficiency, while “economic incentive-type” environmental regulations had little influence on eco-efficiency at both the national and regional levels. Wang and Sun used a system GMM model to verify the impact of environmental regulations on China’s regional eco-efficiency, concluding that the influence of environmental regulations on green economic efficiency varies across different characteristics in both the national sample and sub-samples of eastern, central, and western regions [17]. At the micro level, Fang empirically analyzed low-carbon technology innovation enterprises and found that environmental policies have significant positive effects on low-carbon technology innovation, with different types of policy instruments showing varying degrees of significance [18]. The research by You and Ouyang shows that fiscal decentralization significantly inhibits the green innovation efficiency of industrial enterprises, with different policy instruments exhibiting significant differences [19]. The interaction effect between fiscal decentralization and environmental regulation is particularly pronounced in “administrative order-type” environmental regulations. Based on these research conclusions, this paper uses LCCP as an environmental regulatory tool as an explanatory variable and eco-efficiency as the dependent variable, to set the rational model to reveal the relationship between them.

3. Methods

3.1. Eco-Efficiency Model

Eco-efficiency is essentially a measure of the degree of sustainable development. The Data Envelopment Analysis (DEA) model uses quantitative data on economic and environmental factors from input–output perspectives to model without human intervention, yielding more objective measurement results. The non-parametric technical efficiency analysis method was first proposed by Charnes, Cooper, and Rhodes in 1978 [20]. For panel data in this study that includes observations at multiple time points, it is necessary to analyze productivity changes, i.e., using the Malmquist Total Factor Productivity (TFP) index to measure eco-efficiency. The Malmquist TFP Index (MI) was originally derived from Malmquist (1953) [21], and R Fare, Grosskopf, Lindgren, and Roos used the DEA method to calculate the Malmquist index first [22]. Therefore, this study evaluates eco-efficiency using a direction distance function model of Malmquist, which is generally known as the Malmquist–Luenberger index. The ML index is calculated using a fixed reference period in 2001. The fixed reference model (fixed reference Malmquist) is an ML calculation method proposed by Berg, Forsund, and Jansen [23]. It uses the single-period frontier of a certain fixed period as the reference frontier for calculating MI(t − 1, t) in each period. The specific calculation formula is shown in (1).

$$\mathrm{M}\mathrm{I}(\mathrm{t}-1,\mathrm{t})={\displaystyle \frac{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e} \mathrm{f}({\mathrm{x}}_{\mathrm{t}},{\mathrm{y}}_{\mathrm{t}})}{\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e} \mathrm{f}({\mathrm{x}}_{\mathrm{t}-1},{\mathrm{y}}_{\mathrm{t}-1})}}$$

(1)

$\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e} \mathrm{f}({\mathrm{x}}_{\mathrm{t}},{\mathrm{y}}_{\mathrm{t}})$ represents the DEA efficiency value in year t obtained from the reference-fixed frontier, and $\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e} \mathrm{f}({\mathrm{x}}_{\mathrm{t}-1},{\mathrm{y}}_{\mathrm{t}-1})$ represents the DEA efficiency value in year t − 1 obtained from the reference-fixed frontier.

The input and output indicators selected for the Malmquist–Luenberger index measurement are shown in

Table 1. Input indicators and output indicators.

Categories	Indicators	Code	Specific Indicators Constitute
Input factors	Capital	CI	Investment in fixed assets (CNY 100 million)
	Labor force	LI	Total number of employees (10 thousand)
	Science and technology	SI	Expenditure on science and technology (CNY 10 thousand)
	The sources of energy	EI	Annual electricity consumption (10,000 KWH)
Expected outputs	Economic output	EC	Per capita GDP (CNY 10 thousand)
	Green output	GR	Park green area (square meters)
Unexpected outputs	Pollution output	AQI	Air Quality Index (AQI)

. The input indicators include capital investment, labor input, technology investment, and energy input. The expected output indicators are economic output and green output. The non-expected output indicator is pollution output. Specifically, capital investment is represented by annual fixed asset investment (billion yuan), labor input by total number of employees (ten thousand people) annually, technology investment by expenditure on science and technology (ten thousand yuan), and energy input by annual electricity consumption (million kilowatt-hours). Economic value is indicated by per capita GDP (ten thousand yuan), green output by per capita park green space area (square meters), and pollutant output by Air Quality Index (AQI). AQI is a dimensionless index that quantitatively describes air quality conditions. The higher the AQI value, the more severe air pollution, making it even more harmful to human health. AQI includes six pollutants: SO₂, NO₂, CO, O₃, PM10, and PM2.5. The regional GDP is calculated at actual prices using 2000 as the base period.

3.2. Classical DID Model

The classical difference-in-differences (DID) model, is a commonly used non-experimental method for assessing the impact of policies or events. The DID model infers causal effects by comparing differences between an intervention group and a control group before and after the implementation of the policy. The core idea is to eliminate the influence of time trends and inter-group differences through two differences, thereby accurately estimating the causal effect of the policy. The DID model is widely applied in policy evaluation, economic research, and other fields. Here, it is necessary to assess the impact of LCCP on eco-efficiency. Therefore, cities affected by the policy are treated as the intervention group, while those unaffected are treated as the control group. By comparing the changes in eco-efficiency before and after the policy’s implementation, the effect of the policy can be estimated. In the model, the dependent variable is eco-efficiency, and the key independent variable is a dummy variable for LCCP policy. Based on the previous literature, economic development level, urbanization level, fiscal decentralization degree, science and technology expenditure, environmental regulation degree, and employment structure are selected as control variables. The codes and meanings of each variable are shown in

Table 2. Variable definitions.

Indicator	Variable	Definitions
MI	Eco-efficiency	The Malmquist productivity index with non-expected output
LnGdp	Level of economic development	Logarithm of GDP per capita
Urban	Urbanization level	Urban resident population/total resident population
FisDec	Degree of fiscal decentralization	General revenue of government finance/general expenditure of government finance
SEd	Spending on science and technology	Science and technology expenditure/general government expenditure
EnRe	Degree of environmental regulation	Comprehensive utilization rate of general industrial solid waste
Emp	Employment structure	Number of employees in the tertiary industry/total number of employees

. The model equation is expressed as Equation (2).

$${\mathrm{y}}_{\mathrm{i}\mathrm{t}}=\mathsf{\alpha }+\mathsf{\gamma }{\mathrm{D}}_{\mathrm{t}}+\mathsf{\beta }{\mathrm{T}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}_{\mathrm{i}}+\mathsf{\delta }{\mathrm{D}}_{\mathrm{t}}\times {\mathrm{T}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}_{\mathrm{i}}+\sum {\mathrm{X}}_{\mathrm{i}\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}} (\mathrm{i}=1,\dots \mathrm{n}; \mathrm{t}=1,\dots \mathrm{T})$$

(2)

Among them, D_t is the experimental period dummy variable, taking values of 0 or 1, where 0 indicates before the policy implementation and 1 indicates after the policy implementation. Treat_i is the policy dummy variable, also taking values of 0 or 1, where 0 indicates the control group without the policy, and 1 indicates the intervention group with the policy. The coefficient of the interaction term $\mathsf{\delta }$ represents the policy effect. ${\mathrm{X}}_{\mathrm{i}\mathrm{t}}$ denotes the control variable, ${\epsilon }_{it}$ is the residual term, n represents the number of individuals, and t represents the period.

The classical DID controls the individual fixed effect and time fixed effect on the traditional DID model, removes the independent variables, and retains the interaction term of policy dummy variable and time dummy variable. The model expression is shown in Equation (3).

$${\mathrm{y}}_{\mathrm{i}\mathrm{t}}=\mathsf{\alpha }+\mathsf{\delta }{\mathrm{D}}_{\mathrm{t}}\times {\mathrm{T}\mathrm{r}\mathrm{e}\mathrm{a}\mathrm{t}}_{\mathrm{i}}+{{\mathsf{\phi }}_{\mathrm{i}}+{\mathsf{\tau }}_{\mathrm{t}}+\mathsf{\epsilon }}_{\mathrm{i}\mathrm{t}} (\mathrm{i}=1,\dots \mathrm{n}; \mathrm{t}=1,\dots \mathrm{T})$$

(3)

Among them, ${\mathsf{\phi }}_{\mathrm{i}}$ is the individual fixed effect and ${\mathsf{\tau }}_{\mathrm{t}}$ is the time fixed effect.

3.3. Time-Varying DID Model

The LCCP program has released three batches of pilot lists since 2010. Therefore, the time-varying DID analysis model is used for our benchmark regression analysis. Cities selected in the three batches of LCCP lists serve as the intervention group samples, with Treat_i set to 1. Cities not included in any batch lists serve as the control group samples, with Treat_i set to 0. The dummy variable D_t for policy implementation time is set to 1 in the year following the list’s rollout, and it is set to 0 in the year of rollout and all previous years. The interaction term D_it represents the product of the policy implementation dummy variable Treat_i and the implementation time dummy variable D_t. Thus, except for the intervention group samples, where D_it is 1 in the year following selection into the LCCP program and onwards, D_it remains 0 in all other years for these cities, while the D_it for the control group samples is always 0.

4. Data Sources

The data used in this study come from the China Urban Statistical Yearbook from 2001 to 2021. Samples with incomplete data, especially those lacking key variable data, were excluded. The data underwent truncation at the 1% and 99% levels to avoid the impact of extreme outliers on regression results. Since a few cities are also designated as LCCP by their respective provinces, to maintain consistency in the research dimensions, LCPP was omitted. Cities were classified as intervention group samples based on whether they were selected as LCCP. The total sample includes 278 cities, with 67 in the intervention group and 211 in the control group; the total dataset size is 5282, with 1273 in the intervention group and 4009 in the control group. Detailed data is presented in

Table 3. Descriptive statistics.

Variable	Obs	Mean	Std. Dev.	Min	Max
MI	5282	2.7220	2.0410	0.3130	10.7490
LnGdp	5282	10.2710	0.8900	8.2760	12.4340
Urban	5282	0.5030	0.1720	0.1580	0.9460
FisDec	5282	0.4710	0.2210	0.1090	1.0240
SEd	5282	0.0130	0.0140	0.0010	0.0750
EnRe	5282	79.9130	22.4620	10.3600	114.8500
Emp	5282	0.5480	0.1340	0.2250	0.8640

.

5. Benchmark Analysis Results

5.1. Parallel Trend Test

The time trend chart can roughly assess the trend of the intervention group and the control group before and after the project was established, which is a prerequisite for DID research. As can be observed in Figure 1a, the eco-efficiency of the intervention group and the control group are basically synchronized and moving in the same direction. Figure 1b shows the time trend for the first batch of LCCP, Figure 1c shows the second batch trend, and Figure 1d shows the third batch trend. It can be observed that after the establishment of the second and third batches of LCCP, the gap in eco-efficiency values between the intervention group and the control group has widened, thus preliminarily indicating that the policies of the second and third batches of LCCP have had more significant effects.

Similarly, using a fixed effects model to eliminate the impact of individual effects and control variables on the outcome variable yields residuals. Plotting these residuals against time for both groups results in a cleaner time trend chart. As shown in the Figure 2, the trends of the intervention group and control group converge, with the residual of the control group being relatively higher. Therefore, this graph indirectly suggests that there is some possibility that the two groups meet the assumption of parallel trends.

Further, the event analysis method is used to test the parallel trend. The general model construction of an event analysis is shown in Equation (4).

$${\mathrm{Y}}_{\mathrm{i},\mathrm{t}}={\mathsf{\beta }}_0+{\sum }_{-\mathrm{k}}^{\mathrm{s}}{\mathsf{\beta }}_{\mathrm{m}}{\mathrm{D}}_{\mathrm{i},\mathrm{t}}^{\mathrm{m}}+\mathsf{\gamma }{\mathrm{X}}_{\mathrm{i},\mathrm{t}}+{\mathsf{\phi }}_{\mathrm{i}}+{\mathsf{\tau }}_{\mathrm{t}}+{\mathsf{\epsilon }}_{\mathrm{i},\mathrm{t}}$$

(4)

The time dummy variable ${\mathrm{D}}_{\mathrm{i},\mathrm{t}}^{\mathrm{m}}$ represents the observations of the first k years, the current year, and the following s years before becoming a pilot, while $D_{i,t}^m$ of non-pilot cities is always set to 0. K = 6 and s = 8 are taken in this study.

As shown in Figure 3, the coefficients before the project were not significant and relatively low. It is obvious that there was no significant disparity in eco-efficiency in the two groups, thus satisfying the common trend assumption. Further, from a dynamic perspective, after the project was established, the pilot program began to improve eco-efficiency (d₀ being the year following the announcement of pilot city). The impact coefficient of LCCP remained significantly positive and steadily increased over the next three years. Although it started to decline slightly in the fourth year, it still maintained a positive effect, demonstrating that the LCCP policy can generate policy effects on improving eco-efficiency.

5.2. Time-Varying DID

The results of the time-varying DID (

Table 4. Time-varying DID analysis results.

	Classical DID		Time-Varying DID
Dit	2.349 *** (0.095)	0.374 *** (0.067)	2.435 *** (0.097)	0.385 *** (0.067)	0.321 *** (0.072)	0.131 ** (0.065)
LnGdp		1.828 *** (0.039)		1.820 *** (0.039)		2.232 *** (0.065)
Urban		1.339 *** (0.219)		1.882 *** (0.225)		0.989 *** (0.220)
FisDec		−1.234 *** (0.170)		−0.510 *** (0.193)		−0.219 (0.192)
SEd		−13.978 *** (1.807)		−13.613 *** (1.808)		−6.419 *** (1.781)
EnRe		0.001 (0.001)		0.002 * (0.001)		0.005 *** (0.001)
Emp		3.595 *** (0.192)		3.636 *** (0.197)		2.452 *** (0.201)
Individual-fixed	No	No	Yes	Yes	Yes	Yes
Time-fixed	No	No	No	No	Yes	Yes
R-squared	0.059	0.347	0.112	0.647	0.583	0.682
Observations	5282	5282	5282	5282	5282	5282

Notes: Robust standard errors in parentheses (*** p < 0.01, ** p < 0.05, and * p < 0.1).

) illustrate that in a two-way fixed effects model with five-year plans as time effects and cities as individual fixed effects, the effect of LCCP is evidently positive. The parameter estimates for the classical DID model, individual fixed effects model, and two-way fixed effects model are all positive. Taking the two-way fixed effects model as an example, the policy coefficient is 0.131, indicating that cities implementing the policy have an average eco-efficiency 0.131 higher than those not implementing the policy.

6. Robustness Test

6.1. Placebo Test

6.1.1. Change the Timing of Policies

Changing the timing of policy implementation includes two scenarios: advancing or delaying the policy timing, as well as more generalized methods that randomize the timing of policies (often used for randomly advancing policy timing). If the regression coefficient is not statistically significant when the policy timing is advanced, it indicates that the policy has taken effect, thus validating the robustness of the conclusion in reverse. When the policy timing is delayed, the coefficient remains significant; generally, an increase in absolute value suggests a more pronounced and sustained policy impact.

The policy implementation times for the three batches of pilot cities were advanced by three years and postponed by two years, respectively, to observe whether the policy effects are significant. Since the third batch of pilot cities was determined in 2017, delaying it by three years would result in insufficient post-policy data, so only a two-year delay is applied. Specifically, the first batch was advanced to 2007, 2008, and 2009 and postponed to 2011 and 2012; the second batch was advanced to 2009, 2010, and 2011 and postponed to 2013 and 2014; and the third batch was advanced to 2014, 2015, and 2016 and postponed to 2018 and 2019. The results from the DID model using bidirectional fixed effects are shown in

Table 5. Estimates of policy timing changes.

	Pre3	Pre2	Pre1	Lat1	Lat2
Dit	0.073 (0.062)	0.068 (0.062)	0.057 (0.064)	0.144 ** (0.067)	0.144 ** (0.071)
Individual-fixed	Yes	Yes	Yes	Yes	Yes
Time-fixed	Yes	Yes	Yes	Yes	Yes
R-squared	0.682	0.682	0.682	0.682	0.682
Observations	5282	5282	5282	5282	5282

Notes: Robust standard errors in parentheses (** p < 0.05).

. The results indicate that the policy effects were not significant when advanced by three years but were significant when postponed by two years, with the coefficient of 0.144 being greater than the actual policy year’s coefficient of 0.131, consistent with the logic that the effect strengthens one year after implementation. This confirms the stability of the conclusion that LCCP policies notably optimize eco-efficiency.

Furthermore, the following method is used to randomly advance the policy timing: assuming that the selected LCCP remain unchanged. If city i is designated as LCCP in year t, then any one year within the time range [2003, t − 1] is randomly selected as the year when city i is designated as LCCP. Using this new sample, the baseline can be re-estimated to obtain the estimated coefficient of variable DID. This process is repeated 1000 times, resulting in the probability distribution shown in Figure 4.

As can be seen from Figure 4, the coefficients basically satisfy the normal distribution, and the average value is less than the real coefficient estimate. That is, the random advance of the project time of LCCP will lead to a significant decline in eco-efficiency of pilot cities by LCCP. This also confirms from the counterfactual perspective that LCCP improves the eco-efficiency of pilot cities.

6.1.2. Fictional Intervention Group

The cities identified as low-carbon pilot projects in the intervention group should be regarded as a new control group and cities with the same sample size should be selected as the intervention group in the control group to form a new intervention group, ensuring that the policy implementation time and batch of cities in the new intervention group are consistent with those in the original intervention group. That is, if n cities were designated as low-carbon pilots in year t, then n cities should implement policies in the new control group in year t. Based on this, the new sample should be used to re-estimate the benchmark model, and the above process should be repeated 1000 times to complete one placebo test.

It is estimated that the mean value of the coefficient of the variable DID is −0.0273, which is much smaller than the regression result of 0.131, as shown in Figure 5. This indicates that the policy effect of LCCP shows obvious regional characteristics, and the significance of the improvement effect of eco-efficiency of LCCP in the original sample is not accidental.

6.1.3. Randomize Policy Time and Sample of the Intervention Group at the Same Time

From all samples, a random selection of the same number as the original intervention group was made, and the implementation time of policies was randomly generated to construct a new intervention group with both pilot cities and policy times being randomly assigned. Based on this, the baseline regression model was re-estimated, and the experiment was randomly repeated 500 times. As shown in the Figure 6 below, after randomization, the coefficients of the DID term are concentrated around 0, with most p-values greater than 0.1, and the random coefficients are generally located to the left of the true value of 0.131. This indicates that after double randomization, the policy effects have significantly weakened in terms of both significance and intensity, indirectly confirming the robustness of the original conclusion.

6.2. Endogeneity Problem: Period-by-Period PSM-DID

Due to the fact that low-carbon pilot city policies are, to some extent, “determined by regional willingness,” meaning that whether each region implements these policies is an autonomous choice, there may be a “self-selection bias.” Self-selection bias refers to the decision-making process where each sample decides whether to participate in the experiment being endogenous. In reality, whether a city implements a policy is more based on considerations such as geographical location, economic development, and industrial structure. This aligns with the actual situation, where policies are generally first implemented in demonstration zones, which tend to have higher levels of economic development. Generally speaking, factors influencing whether a place implements a policy also affect its eco-efficiency. These factors can be divided into observable and unobservable factors. First, observable factors are measurable; these variables are related to eco-efficiency and can therefore be incorporated into our model, even if they may have collinearity with the policy implementation dummy variable. However, a certain degree of collinearity is not a problem. Since it is impossible to identify all observable variables or because observable variables influence the outcome variable in a non-linear manner, these unobservable factors (or their non-linear forms) are left out of the model and placed in the disturbance term, leading to correlation between the disturbance term and the policy implementation dummy variable, thus causing endogeneity and ultimately resulting in biased estimates of the interaction terms between policy and time. Second, unobservable factors are unmeasurable; therefore, these unobservable factors must be included in the disturbance term, which also poses an endogeneity issue, leading to estimation bias. Therefore, in order to make the test results more convincing, the period-by-period PSM-DID model is used for further accurate regression.

6.2.1. Preference Score Matching

The covariates used to match the sample of the intervention group with that of the control group include six control variables from the baseline model plus the logarithm of foreign investment, LnFdi. This is because the previous literature has already provided sufficient evidence that these seven variables have a significant impact on eco-efficiency, which will not be elaborated here due to space constraints. As shown in Figure 7, both kernel density curves exhibit considerable deviation before and after matching, but the mean line distance shortens after matching, making the two curves more closely aligned. This suggests, to some extent, that annual matching is effective. After matching, the gap in eco-efficiency between the intervention group and the control group narrows, indicating that the characteristics of the control group are more similar to those of the intervention group after matching. Therefore, it can be inferred that the new control group better meets the common trend test compared to the original control group and effectively eliminates endogeneity.

6.2.2. Balance Test

A balance test examines whether there are significant differences in the values of matched covariates between two groups. If the differences are not obvious, it indicates that the matching effect is good, making the DID regression using such a matched sample more robust. Since propensity score matching for panel data needs to be conducted annually, examining systematic biases among covariates across different years can only be performed within the same year. Matching samples from different years are not comparable, and technically, conducting a balance test after merging into a panel dataset is also impractical. Here, we refer to the method of Xie et al. to compare the logit regression results before and after matching across different years [24]. As can be seen from

Table 6. Standardized deviation variation table.

	Unmatched	Mean		%Reduct		t-Test		V(T)/V(C)
Variable	Matched	Treated	Control	%bias	bias	t	p > t
Lngdp	U	10.683	10.141	61.5		19.59	0	1.21 *
	M	10.582	10.586	−0.4	99.4	−0.09	0.93	0.89 *
Urban	U	0.59055	0.47505	66.5		21.74	0	1.45 *
	M	0.56682	0.56929	−1.4	97.9	−0.33	0.745	0.93
Fisdec	U	0.61282	0.42596	87.1		28.2	0	1.36 *
	M	0.58791	0.58541	1.2	98.7	0.27	0.79	1
Sed	U	0.02036	0.01084	62.7		22.28	0	2.64 *
	M	0.01802	0.01733	4.5	92.8	1	0.318	0.89
Enre	U	83.873	78.656	24.8		7.25	0	0.60 *
	M	83.504	83.001	2.4	90.4	0.62	0.535	0.84 *
Emp	U	0.54781	0.54793	−0.1		−0.03	0.979	1.03
	M	0.53999	0.53761	1.8	−2006.4	0.42	0.675	0.92
Lnfdi	U	10.693	9.3412	68		21.6	0	1.17 *
	M	10.492	10.433	3	95.7	0.78	0.438	1.54 *

* If variance ratio outside [0.90; 1.12] for U and [0.89; 1.12] for M.

, the standardized bias (%bias) of most variables after matching is relatively small, and the t-values do not reject the null hypothesis of no systematic bias between the treatment group and the control group. Figure 8 also shows that the standard deviations of all variables are reduced after matching.

As shown in

Table 7. Logit regression results before matching.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
	2003b	2004b	2005b	2006b	2007b	2008b	2009b	2010b	2011b	2012b
LnGdp	0.4342	0.2764	0.2703	−0.3248	−0.4567	−0.2328	−0.1824	−0.0672	−0.1060	−0.3630
	(0.9743)	(0.6023)	(0.6343)	(−0.7216)	(−0.8859)	(−0.4675)	(−0.3588)	(−0.1369)	(−0.2111)	(−0.7252)
Urban	0.3887	0.7319	2.0257 *	1.9125	1.8362	1.6898	2.9644 **	4.4656 ***	3.5278 **	4.1952 ***
	(0.3490)	(0.6782)	(1.7335)	(1.3972)	(1.3009)	(1.0823)	(1.9966)	(2.9930)	(2.5077)	(2.6047)
FisDec	4.1379 ***	4.1090 ***	3.5619 **	5.1433 ***	2.6408 **	0.8521	0.9585	0.1302	2.0043	2.9139 *
	(3.1505)	(3.0738)	(2.4723)	(3.3534)	(2.0128)	(0.7731)	(0.6128)	(0.0866)	(1.3152)	(1.8540)
SEd	−63.9245	−76.4230	−98.4929	−16.4236	56.1680 **	74.2951 ***	66.0572 *	49.3896 *	32.1914 *	18.8283
	(−0.9483)	(−1.3628)	(−1.4370)	(−0.1729)	(2.1501)	(2.9410)	(1.8103)	(1.8786)	(1.8151)	(1.1183)
EnRe	0.0020	0.0026	0.0048	0.0009	0.0005	−0.0005	0.0033	0.0045	0.0064	0.0017
	(0.2909)	(0.3831)	(0.6749)	(0.1190)	(0.0664)	(−0.0738)	(0.4193)	(0.5805)	(0.7936)	(0.2350)
Emp	3.6022 **	3.0268 **	4.3901 ***	4.2069 ***	3.5034 **	3.6692 **	3.6561 **	4.1012 ***	4.2836 ***	3.7583 **
	(2.3996)	(2.1174)	(3.0546)	(2.8544)	(2.2632)	(2.2574)	(2.3911)	(2.7869)	(2.7115)	(2.5206)
Lnfdi	0.0731	0.1136	0.0900	0.0761	0.1131	0.1824	0.0405	0.0735	0.0619	0.0494
	(0.5605)	(0.9032)	(0.7408)	(0.5441)	(0.7711)	(1.2304)	(0.2527)	(0.4398)	(0.3966)	(0.3335)
N	278	278	278	278	278	278	278	278	278	278
Pseudo R2	0.1619	0.1622	0.1746	0.1775	0.1748	0.1760	0.1807	0.1788	0.1793	0.1709
	(11)	(12)	(13)	(14)	(15)	(16)	(17)	(18)	(19)
	2013b	2014b	2015b	2016b	2017b	2018b	2019b	2020b	2021b
LnGdp	−0.1403	−0.1236	−0.6166	−0.3863	−0.2820	0.3894	0.6336	0.3605	1.0948 *
	(−0.2694)	(−0.2302)	(−1.0597)	(−0.6638)	(−0.4611)	(0.6356)	(0.9176)	(0.5818)	(1.9218)
Urban	4.2759 **	4.2518 **	5.4612 ***	5.6291 ***	5.8614 ***	6.0809 ***	5.4680 **	4.8506 ***	3.0866 **
	(2.2750)	(2.3290)	(2.5932)	(2.6563)	(2.6860)	(2.6382)	(2.4023)	(2.5887)	(1.9743)
FisDec	1.4783	2.0088	4.2192 ***	3.5938 **	2.2585	0.1652	−0.2841	1.2378	−0.7433
	(0.8823)	(1.1957)	(2.5817)	(2.0653)	(1.3032)	(0.0915)	(−0.1452)	(0.6527)	(−0.4297)
SEd	36.1144 *	23.0323 *	18.8759	13.2753	15.7110	16.9796	18.0523	22.9261 *	16.9138
	(1.7787)	(1.6644)	(1.2368)	(1.0405)	(1.1249)	(1.2282)	(1.3441)	(1.7251)	(1.4978)
EnRe	0.0057	0.0079	0.0027	0.0038	0.0097	0.0134	0.0078	0.0110	0.0095
	(0.6586)	(0.9138)	(0.3505)	(0.4890)	(1.1520)	(1.5923)	(0.9412)	(1.2525)	(1.1552)
Emp	6.0189 ***	5.8113 ***	5.9136 ***	5.9587 ***	5.4924 ***	5.2800 ***	6.0372 ***	5.9990 ***	4.4083 ***
	(3.4929)	(3.6523)	(3.5347)	(3.6565)	(3.8272)	(3.7749)	(3.9170)	(4.0298)	(3.1742)
Lnfdi	0.1566	0.1777	0.0339	0.0880	0.2009 *	0.1297	0.1906 *	0.0843	0.1414
	(1.0699)	(1.3153)	(0.2424)	(0.6853)	(1.6600)	(1.3004)	(1.9150)	(0.6595)	(1.1070)
Observations	278	278	278	278	278	278	278	278	278
Pseudo R2	0.2039	0.2063	0.2313	0.2301	0.2364	0.2309	0.2399	0.2436	0.2160

Notes: Robust standard errors in parentheses (*** p < 0.01, ** p < 0.05, and * p < 0.1).

and

Table 8. Logit regression results after matching.

	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)
	2003a	2004a	2005a	2006a	2007a	2008a	2009a	2010a	2011a	2012a
LnGdp	−0.5644	0.3158	0.1415	−0.3440	−0.2104	−0.3924	0.2443	0.4435	−0.0491	0.4313
	(−1.0706)	(0.5727)	(0.2525)	(−0.5563)	(−0.3460)	(−0.6547)	(0.3711)	(0.7471)	(−0.0833)	(0.6917)
Urban	2.9719 *	0.4499	0.4826	1.2663	0.0695	0.5744	−1.3892	−0.5900	0.5963	0.0711
	(1.8151)	(0.3025)	(0.3114)	(0.7559)	(0.0437)	(0.3523)	(−0.7733)	(−0.2853)	(0.3528)	(0.0375)
FisDec	1.8020	0.8046	1.0440	1.7973	0.0306	1.4796	−0.8514	−0.7574	0.6342	−0.9004
	(1.0965)	(0.5155)	(0.6671)	(1.0351)	(0.0213)	(0.9520)	(−0.4492)	(−0.4681)	(0.3519)	(−0.4504)
SEd	−1.1 × 102	67.3873	−1.6 × 102	−62.6711	36.4138	24.3065	63.2775 *	16.3786	16.4340	14.1040
	(−1.0326)	(0.5536)	(−1.3843)	(−0.5698)	(1.0480)	(0.7342)	(1.9279)	(0.5081)	(0.8631)	(0.6690)
EnRe	−0.0057	0.0022	0.0004	−0.0013	−0.0059	−0.0100	0.0028	−0.0012	−0.0005	−0.0042
	(−0.5630)	(0.2377)	(0.0423)	(−0.1381)	(−0.6445)	(−0.9608)	(0.2893)	(−0.1302)	(−0.0440)	(−0.3969)
Emp	0.5059	1.0754	0.5182	0.4560	−0.0601	0.3310	−0.0781	0.2364	2.0201	−0.3492
	(0.2700)	(0.5293)	(0.2675)	(0.2482)	(−0.0327)	(0.1606)	(−0.0417)	(0.1261)	(1.2213)	(−0.1803)
Lnfdi	0.1127	−0.0109	−0.0365	−0.0164	0.0523	0.0239	−0.1226	0.0708	0.0138	0.0842
	(0.7967)	(−0.0813)	(−0.2809)	(−0.1221)	(0.3550)	(0.1581)	(−0.7439)	(0.4544)	(0.0797)	(0.5122)
N	126	124	128	132	126	125	121	122	125	122
Pseudo R2	0.0422	0.0223	0.0167	0.0139	0.0151	0.0272	0.0384	0.0165	0.0192	0.0288
	(11)	(12)	(13)	(14)	(15)	(16)	(17)	(18)	(19)
	2013a	2014a	2015a	2016a	2017a	2018a	2019a	2020a	2021a
LnGdp	0.4430	0.1624	−0.0744	0.3967	−0.1503	−0.0198	−0.4536	0.3098	−0.2524
	(0.6754)	(0.2284)	(−0.0995)	(0.5365)	(−0.1866)	(−0.0255)	(−0.5577)	(0.4259)	(−0.3508)
Urban	1.7801	−0.1778	−0.3741	−1.7313	2.7618	1.4108	2.9449	1.4836	0.7025
	(0.8201)	(−0.0739)	(−0.1308)	(−0.5935)	(0.8707)	(0.4786)	(1.0111)	(0.5682)	(0.3041)
FisDec	−1.0076	0.7001	1.5238	0.6489	−0.2337	−1.5095	1.9261	−0.7422	1.6424
	(−0.5086)	(0.3590)	(0.8354)	(0.3392)	(−0.1126)	(−0.8488)	(0.9870)	(−0.3263)	(0.8473)
SEd	−8.1101	15.7087	10.4282	−7.8896	15.1155	22.2366	4.5677	5.3367	15.6782
	(−0.3741)	(0.8765)	(0.6328)	(−0.5140)	(0.9115)	(1.3188)	(0.3267)	(0.3603)	(1.0938)
EnRe	−0.0032	−0.0046	0.0008	−0.0042	0.0012	0.0018	0.0046	0.0068	−0.0015
	(−0.3004)	(−0.4406)	(0.0865)	(−0.4239)	(0.1234)	(0.1932)	(0.4668)	(0.6951)	(−0.1510)
Emp	1.2598	1.5582	1.4378	0.1775	0.2813	1.7815	1.9490	2.1347	0.6847
	(0.6801)	(0.8756)	(0.7763)	(0.0812)	(0.1535)	(1.0215)	(1.0504)	(1.0993)	(0.4103)
Lnfdi	0.0821	−0.0964	−0.0634	0.1201	−0.1561	0.0817	−0.0554	0.0374	−0.0856
	(0.4994)	(−0.5277)	(−0.4022)	(0.8250)	(−0.9853)	(0.5894)	(−0.4026)	(0.2132)	(−0.4956)
Observations	124	120	115	114	114	119	129	115	120
Pseudo R2	0.0205	0.0152	0.0124	0.0201	0.0246	0.0280	0.0291	0.0215	0.0266

Notes: Robust standard errors in parentheses (* p < 0.1).

, the coefficients of most covariates have decreased in all years after matching, and most of these coefficients have become insignificant.The pseudo-R-squared values of all regressions have significantly decreased, indicating that there are no systematic biases in covariates between the two groups across different years, thus passing the balance test.

6.2.3. PSM-DID Analysis

The value of the new variable weight after matching indicates whether a sample is involved in the matching and its importance. As long as the weight of a sample is not empty, it means that the sample has participated in the matching process. These samples can then be used to estimate parameters in the DID regression model, which also helps alleviate selection bias issues present in benchmark regressions to some extent. Furthermore, they can be compared with the results of benchmark regressions to test their robustness. In fact, most of the literature follows this approach; however, some studies directly select samples that meet the common support hypothesis for inclusion in the DID regression model, such as Li and Xiao [25]. The common support hypothesis requires that the distribution of covariates between the intervention group and the control group have some overlap to ensure matching quality. This means that during the matching process, only covariates with similar distributions in both the experimental and control groups should be used to ensure that the matched cohorts are not only similar in treatment conditions but also in other important covariates. Meeting the common support hypothesis can improve the precision and stability of the match and reduce estimation bias, and due to the high matching requirements, the final sample size is relatively small.

Whether using samples with non-zero weights or those that meet the common support assumption for regression, a fundamental fact is often overlooked: matched control group samples may serve as matching objects for multiple intervention groups. Therefore, the importance of different-weighted control group samples in the overall control group sample varies. The higher the weight, the more times they are matched and thus should be given greater importance when participating in regression. A feasible approach is to replicate the matched samples from the control group based on their weights, which is known as frequency-weighted regression.

To facilitate the comparison of regression results across different models,

Table 9. Classical DID, time-varying DID, and PSM-DID results.

	(1)	(2)	(3)	(4)	(5)
	Classical DID	Time-Varying DID	PSM-DID (1)	PSM-DID (2)	PSM-DID (3)
Dit	0.4456 ***	0.1307 **	0.2694 ***	0.1849 ***	0.3165 ***
	(5.1248)	(2.0145)	(3.3393)	(2.7208)	(4.3162)
Individual-fixed	No	Yes	Yes	Yes	Yes
Time-fixed	No	Yes	Yes	Yes	Yes
Observations	5282	5282	2321	4531	3313
Adj. R2	0.3880	0.7871	0.7976	0.7946	0.8231

Notes: Robust standard errors in parentheses (*** p < 0.01,** p < 0.05).

presents five types of regression outcomes: two benchmark regressions, weights not being null, meeting the common support assumption, and frequency-weighted. The first column shows the OLS regression without fixed effects, where the policy variable Dit is significant but time and individual fixed effects are not included; the second column presents the asymptotic DID analysis under a dual fixed effects model, with a significant coefficient of 0.1307; the third column shows the PSM-DID regression using samples with non-null weights, where the coefficient is significant at 0.2694; the fourth column presents the PSM-DID regression using samples that meet the common support assumption, where the coefficient is significant at 0.1849; and the fifth column shows the frequency-weighted PSM-DID regression considering sample importance, where the coefficient is significant at 0.3165. Since the results of PSM-DID are all based on sample characteristics and matched year-by-year for control group samples and the control group samples were selected in three ways, theoretically, as long as the three results are consistent and significant, the effectiveness of low-carbon city pilots can be verified. For example, using samples that meet the common support assumption for PSM-DID regression (the fourth column), the policy coefficient is 0.1849. Compared to the coefficient of 0.1307 from the asymptotic DID regression (the second column), the policy effectiveness increases by 0.0542. This demonstrates that using an appropriate PSM-DID regression model can prevent underestimation of policy effectiveness.

6.2.4. Replace the Dependent Variable

The dependent variable is changed to GDP per unit of carbon dioxide emission. In order to avoid endogeneity, the GDP in the control variables is replaced by foreign direct investment and taken logarithmically. The results of the mixed OLS regression, double fixed effect regression, and PSM-DID analysis are shown in

Table 10. Regression results after changing the dependent variable.

	(1)	(2)	(3)	(4)	(5)
	Classical DID	Time-Varying DID	PSM-DID (1)	PSM-DID (2)	PSM-DID (3)
Dit	0.4016 ***	0.2702 ***	0.1335 ***	0.1826 ***	0.0818 ***
	(9.7883)	(10.1511)	(4.2065)	(7.3619)	(2.7464)
Observations	5282	5282	2321	4531	3313
Adj. R2	0.2044	0.7909	0.7805	0.7844	0.7974

Notes: Robust standard errors in parentheses (*** p < 0.01).

. The policy effect is also significant, which indicates that the regression results have certain robustness.

6.2.5. Modify the Desired and Undesired Outputs in the Eco-Efficiency Model

In the measurement of eco-efficiency, the desired green output per capita green space area was replaced with renewable energy consumption, while the undesired green outputs were changed to PM2.5 and CO₂ emissions. After obtaining the new MI eco-efficiency values, the PSM-DID test results remained consistent and statistically significant. The outcomes are presented in

Table 11. PSM-DID results after changing desirable and undesirable outputs.

PM2.5 as Undesirable Output
	(1)	(2)	(3)	(4)	(5)
	OLS	FE	Weight not vacancy.	On_Support	Weight_Reg
dit	0.3218 **	0.2974 ***	0.2678 *	0.3941 ***	0.3746 ***
	(2.4516)	(2.6238)	(1.6466)	(3.2737)	(2.7628)
lngdp	2.2797 ***	2.7285 ***	2.8478 ***	2.6937 ***	2.4925 ***
	(32.8319)	(24.0432)	(13.6406)	(21.7015)	(15.7856)
urban	−0.3482	0.4285	−0.7784	−0.0498	−0.8630
	(−1.0907)	(1.1154)	(−1.1057)	(−0.1193)	(−1.6177)
fisdec	−2.7827 ***	−0.3580	−0.7857	−0.4056	−0.4666
	(−13.8887)	(−1.0685)	(−1.3868)	(−1.1402)	(−1.1346)
sed	−37.0798 ***	−19.5161 ***	−24.3922 ***	−18.6950 ***	−16.9408 ***
	(−11.0205)	(−6.2707)	(−4.6189)	(−5.4160)	(−4.4417)
enre	−0.0016	0.0030	0.0028	0.0031	−0.0001
	(−1.0615)	(1.5730)	(0.6957)	(1.4006)	(−0.0459)
emp	3.8976 ***	3.2481 ***	2.8911 ***	3.7166 ***	3.5825 ***
	(13.8950)	(9.2556)	(4.4076)	(9.2357)	(7.5822)
N	5282	5282	2321	4531	3313
Adj. R2	0.3007	0.6742	0.6686	0.6820	0.7201

Notes: Robust standard errors in parentheses (*** p < 0.01,** p < 0.05 and * p < 0.10).

and

Table 12. PSM-DID results after altering desirable and undesirable outputs.

CO2 as Undesirable Output
	(1)	(2)	(3)	(4)	(5)
	OLS	FE	Weight not vacancy.	On_Support	Weight_Reg
dit	0.5556 ***	0.4644 ***	0.4078 **	0.5321 ***	0.3746 ***
	(4.1510)	(4.1462)	(2.4970)	(4.4282)	(2.7628)
lngdp	2.1187 ***	2.5928 ***	2.7297 ***	2.6164 ***	2.4925 ***
	(29.9259)	(23.1229)	(13.0198)	(21.1201)	(15.7856)
urban	−0.4197	0.5267	−0.4477	0.2383	−0.8630
	(−1.2892)	(1.3875)	(−0.6332)	(0.5721)	(−1.6177)
fisdec	−2.4978 ***	0.5225	0.1159	0.4635	−0.4666
	(−12.2265)	(1.5779)	(0.2038)	(1.3055)	(−1.1346)
sed	−38.3605 ***	−17.7336 ***	−24.2723 ***	−18.1235 ***	−16.9408 ***
	(−11.1815)	(−5.7666)	(−4.5768)	(−5.2606)	(−4.4417)
enre	0.0040 ***	0.0005	−0.0039	−0.0016	−0.0001
	(2.5803)	(0.2892)	(−0.9785)	(−0.7062)	(−0.0459)
emp	2.8361 ***	1.8024 ***	2.5919 ***	2.4334 ***	3.5825 ***
	(9.9159)	(5.1980)	(3.9348)	(6.0588)	(7.5822)
N	5282	5282	2321	4531	3313
Adj. R2	0.2536	0.6734	0.6700	0.6769	0.7201

Notes: Robust standard errors in parentheses (*** p < 0.01, ** p < 0.05).

. The tables demonstrate that after altering the desired and undesired green outputs in the eco-efficiency model, the policy effects were significant and positive across the baseline regression analysis, fixed-effects models, and PSM-DID regression analysis.

7. Heterogeneity Analysis

7.1. Comparison of Effects in Different Regions

Based on the socio-economic development status of different regions, according to the “Several Opinions of the Central Committee of the Communist Party of China and the State Council on Promoting the Rise of the Central Region,” the “Implementation Opinions of the State Council on Several Policy Measures for the Large-Scale Development of the Western Region,” and the spirit of the report from the 16th National Congress of the Communist Party of China, China’s economic regions are typically divided into four major areas: Eastern, Central, Western, and Northeast. The Eastern region includes provinces and municipalities such as Beijing, Tianjin, Hebei, Shanghai, Jiangsu, Zhejiang, Fujian, Shandong, Guangdong, and Hainan; the Central region includes provinces such as Shanxi, Anhui, Jiangxi, Henan, Hubei, and Hunan; the Western region includes provinces and municipalities such as Inner Mongolia, Guangxi, Chongqing, Sichuan, Guizhou, Yunnan, Tibet, Shaanxi, Gansu, Qinghai, Ningxia, and Xinjiang; and the Northeast region includes provinces such as Liaoning, Jilin, and Heilongjiang.

After analyzing using the time-varying DID model and PSM-DID model, the results in

Table 13. PSM-DID analysis results of the four major economic regions.

	Eastern Region		Central Region		Western Region		Northeast Region
	Time-Varying DID	PSM-DID	Time-Varying DID	PSM-DID	Time-Varying DID	PSM-DID	Time-Varying DID	PSM-DID
Dit	−0.2905 ***	0.4708 ***	0.2344 **	0.3782 ***	0.5655 ***	0.3927 **	0.3696 *	0.7957 **
	(−2.9692)	(3.8665)	(2.1413)	(3.0587)	(3.7921)	(2.2147)	(1.9343)	(2.1383)
Individual-fixed	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Time-fixed	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Observations	1634	1092	1520	725	1482	926	646	118
Adj. R2	0.7847	0.7849	0.8186	0.8495	0.7956	0.8166	0.8112	0.8065

Notes: Robust standard errors in parentheses (*** p < 0.01, ** p < 0.05, and * p < 0.1).

show that the policy effects in all regions are significant. Notably, if the four economic regions are analyzed as independent sub-samples, the coefficients of the policy variables are higher, especially in Northeast China, where the coefficient of the policy variable is the highest. This indicates that the LCCP in the three northeastern provinces has the most noticeable improvement on eco-efficiency. However, due to the smaller sample size in the Northeast region, the results may not be robust enough.

To verify the robustness of the results in the Northeast region, the SCM-DID method was employed for further testing. The SCM method is suitable for scenarios where only one or a few individuals receive treatment, addressing the issue of failing to meet the parallel trend assumption by weighting and averaging the control group individuals. Arkhangelsky et al. (2021) combined DID and SCM, leveraging their respective advantages to estimate policy treatment effects, and proposed a novel method for estimating causal effects based on panel data—synthetic difference-in-differences (SDID) [26]. The SDID estimator assigns more weight to individuals more similar to the treatment group and to periods more similar to the treatment period. To compute the standard errors for the SDID method, clustered bootstrap sampling was used, treating the sample as the population and performing sampling with replacement. For this test, the number of sampling iterations was set to 1000 with each iteration requiring a complete SDID regression to obtain robust test results. The results are presented in

Table 14. Annual SCM-DID analysis results for Northeast China.

MI	ATT	Std. Err.	t	p > t	[95% Conf. Interval]
Dit	−0.21479	0.4102	−0.52	0.601	−1.01876	0.58919

. ATT is the average treatment effect, and the findings indicate that the results of the SCM-DID test were not significant (p-value far exceeding 0.05), suggesting that the significance of the Northeast region requires further investigation. However, it can be confirmed that the significant policy results in the Northeast region are not robust. Due to space constraints, further elaboration is omitted here.

7.2. Comparison of Effects of Different Batches

The overall model is analyzed using a sample of three pilot cities. Now, the three batches of pilots are divided into three intervention groups, with the control group remaining unchanged. This allows for further analysis of the differences in the effectiveness of each batch of pilots within the sub-samples. As shown in

Table 15. PSM-DID analysis results of each batch.

	The First Batch		The Second Batch		The Third Batch
	Time-Varying DID	PSM-DID	Time-Varying DID	PSM-DID	Time-Varying DID	PSM-DID
Dit	−0.2575	0.4606 **	0.0202	0.0400	0.1925 *	0.2662 **
	(−1.5630)	(2.1649)	(0.1954)	(0.3573)	(1.8092)	(2.4254)
Individual-fixed	Yes	Yes	Yes	Yes	Yes	Yes
Time-fixed	Yes	Yes	Yes	Yes	Yes	Yes
Observations	4180	956	4465	3492	4655	3814
Adj. R2	0.7675	0.8174	0.7675	0.7763	0.7708	0.7824

Notes: Robust standard errors in parentheses (** p < 0.05, * p < 0.1).

, only the third batch of policies has significant effects, with a policy coefficient of 0.1925 in the time-varying DID model, at the 10% significance level; the paired samples that meet the common support assumption are analyzed using the PSM-DID model, which shows that the first and third batches of policies have significant effects, with policy coefficients of 0.4606 and 0.2662, respectively, at the 5% significance level, while the second batch of policies does not have significant effects.

The initial cohort of LCCP centered on carbon emission intensity reduction via administrative directives. The second cohort sought to establish carbon emission peak targets while incorporating market-based mechanisms, notably carbon emissions trading. The third cohort probed pathways toward carbon neutrality, aligning with the “Dual Carbon” strategy through financial innovation, technology commercialization, and regional collaborative frameworks. A comparative analysis revealed that the administrative directive-driven approaches adopted by the first and third cohorts exhibited greater efficacy in enhancing ecological efficiency.

7.3. Comparison of Effects in Pilot Provinces

In the first batch of LCCP projects, five provinces—Guangdong, Liaoning, Hubei, Shaanxi, and Yunnan—were designated as LCPP. In the second batch, Hainan Province was also identified as LCPP, making it six low-carbon pilot provinces in total. Several cities in the third batch of pilot projects were located within these six provinces, which can be considered as secondary pilots, such as Shenyang, Shenzhen, Zhongshan, and Ankang. For this section, we will refer to these cities as secondary pilots, while the remaining cities in the intervention group will be called primary pilots, with the control group remaining unchanged. Therefore, the pilot cities can be divided into two categories to examine the policy overlay effect of secondary pilots: one category includes cities whose provincial-level pilot status has been established before; the other category includes cities whose provincial-level pilot status has not yet been established. This study aims to investigate whether the combined effect of LCPP and LCCP is more significant for improving eco-efficiency. After analyzing using both the period-by-period DID model and PSM-DID model, the results in

Table 16. PSM-DID analysis of pilot provinces and non-pilot provinces.

	Secondary Pilot Cities		Single Pilot Cities
	Time-Varying DID	PSM-DID	Time-Varying DID	PSM-DID
Dit	0.1968	0.1660	0.1095	0.2577 ***
	(1.3163)	(0.8980)	(1.5114)	(3.2784)
Individual-fixed	Yes	Yes	Yes	Yes
Time-fixed	Yes	Yes	Yes	Yes
Observations	1216	594	4066	3440
Adj. R2	0.7858	0.8277	0.7884	0.7993

Notes: Robust standard errors in parentheses (*** p < 0.01).

show that when the original intervention group is split into two subgroups—secondary pilot cities and single pilot cities—the policy effects change significantly. The policy effect of secondary pilots shows no statistical significance, possibly stemming from a smaller sample size; however, the PSM-DID model results for single pilots indicate a significant policy effect.

The poor performance of dual pilot cities may stem from overlapping regulations crowding out municipal measures. Since provincial administrative authorities hold higher responsibilities, rights, and interests than municipal ones, provincial measures could potentially override municipal initiatives. However, provincial measures cover broader areas and often lack rigorous supervision in urban implementation depth, which may lead to unsatisfactory outcomes in dual pilot cities.

8. Path Analysis

To further examine the pathways through which low-carbon pilot policies affect eco-efficiency, the energy consumption structure was incorporated as a mediating variable for path analysis.

Table 17. Test results of the mediating effect model.

	EnerStr	MI	MI
Dit	0.008 ** (3.071)	0.447 ** (4.891)	0.428 ** (4.690)
EnerStr			2.285 ** (4.595)
Sample size	4641	4641	4641
R2	0.152	0.364	0.367
Adjusted R2	0.151	0.363	0.366
F-value	F (7,4633) = 118.565, p = 0.000	F (7,4633) = 378.479, p = 0.000	F (8,4632) = 335.246, p = 0.000

Note: ** p < 0.05 (t-values in parentheses).

shows the relationship between the variables.

As shown in

Table 18. Summary of the effect analysis process.

	Process	Effect	SE	t	p	LLCI	ULCI
Direct effect	dit⇒MI	0.428	0.091	4.69	0	0.249	0.607
Indirect effect	dit⇒EnerStr	0.008	0.003	3.071	0.002	0.003	0.014
	EnerStr⇒MI	2.285	0.497	4.595	0	1.31	3.26
Total effect	dit⇒MI	0.447	0.091	4.891	0	0.268	0.626

, the mediation effect model involves three types of effects: a direct effect of 0.428, an indirect effect of 0.019, and a total effect of 0.447. When the mediating variable is the energy structure variable, the product of the estimated parameter of the low-carbon pilot city dummy variable on the energy consumption structure and the estimated parameter of the energy consumption structure on eco-efficiency yields the indirect effect value. The indirect effect value was subjected to Bootstrap sampling tests to ultimately verify the presence of a mediation effect. Using the Bootstrap sampling test method for the mediation effect analysis with 500 sampling iterations, the results showed that for the mediation path ‘dit⇒EnerStr⇒MI,’ the 95% confidence interval did not include the number 0 (95% CI: 0.000~0.005), indicating the existence of this mediation path. This demonstrates that the low-carbon pilot city policy indeed enhances eco-efficiency through the pathway of improving energy structure.

9. Conclusions

A rigorous empirical model analysis has shown that the LCCP has achieved meaningful effects on improving eco-efficiency. The main research conclusions of this paper are as follows: Under the bidirectional fixed effects of an urban and five-year plan, whether using the ordinary DID model, the time-varying DID model, or the period-by-period PSM-DID model, all analysis have shown the effectiveness of LCCP. Among these, the PSM-DID model can effectively eliminate the endogeneity issues of the ordinary DID and time-varying DID models. The results obtained after selecting sample ranges in three different ways are all significant. In terms of robustness testing, placebo tests were conducted using methods such as randomizing policy timing, randomizing intervention group samples, and simultaneously randomizing policy timing and intervention group samples. Robustness was also verified by changing the eco-efficiency model. Finally, the heterogeneity analysis of sub-samples shows that the LCCP has significantly improved eco-efficiency in the Eastern, Central, Western, and Northeastern regions, with the coefficients of policy variables being higher than those estimated for the overall sample. In the analysis of batch effects, the first and third batches of LCCP had a significant impact on improving eco-efficiency, but the degree of significance decreased compared to the initial assessment; the second batch of pilot policies did not pass the significance test, which may be related to insufficient institutional adaptation and resource allocation during the early stages of policy implementation. From the perspective of whether both provincial and municipal levels have initiated low-carbon city pilots, the policy effects in cities that were initially designated (i.e., both provincial and municipal levels included in the pilot) are not significant. In contrast, cities that were initially designated only at the municipal level show a notable improvement in eco-efficiency. This phenomenon may be due to issues such as dispersed policy resources and high coordination costs in initially designated cities, leading to insufficient policy synergy. Low-carbon pilot cities are committed to achieving urban low-carbon development, making significant efforts in areas such as enhancing the digital economy [27], improving energy structures, and industrial transformation.

Since the National Development and Reform Commission launched the program in 2010, many pilot cities have actively explored green and low-carbon development paths. In terms of energy structure adjustment, some cities have vigorously developed renewable energy, increased the proportion of clean energy, and reduced carbon emissions; in terms of industrial structure, they have accelerated the elimination of high-energy-consuming and highly polluting industries, fostering strategic emerging industries such as environmental protection, energy conservation, and new energy and promoting the low-carbon transformation of industries. Based on the research conclusions above, this paper puts forward suggestions on how to use environmental regulation tools more efficiently to improve eco-efficiency from the perspective of energy consumption structure.

(1)

Strengthen clean energy incentive policies: The government should increase subsidies for solar, wind, hydro, and biomass energy. For example, provide financial support to companies building solar power plants to reduce their initial construction costs and boost the share of clean energy in the energy consumption mix. Implement tax incentives for clean energy production, offering reduced corporate income tax for businesses engaged in clean energy production, encouraging them to expand their clean energy operations. Establish a green power certificate trading system to incentivize companies to produce more clean energy through market mechanisms, while also meeting the demand for green energy from high-energy-consuming enterprises, promoting a greener transformation of the energy consumption structure.

(2)

Enhance environmental standards for traditional energy: Strictly set emission standards for the use of coal, oil, and other traditional energy sources. Require companies to install advanced pollution control equipment to reduce pollutant emissions. Impose severe penalties such as hefty fines and production halts on non-compliant enterprises, compelling them to reduce reliance on high-pollution traditional energy. Raise the environmental entry threshold for traditional energy extraction, restricting large-scale mining activities that cause significant ecological damage, such as small coal mines and oil fields, to control the impact of traditional energy on the environment at its source. Promote technological upgrades in traditional energy companies to improve energy efficiency and lower pollutant emissions per unit of energy consumption.

(3)

Promote the diversification of energy consumption: Encourage businesses and residents to use multiple energy sources, reducing over-reliance on a single type. Promote the supply of natural gas, electricity, and other energy options in cities, offering users more choices. Support the development of distributed energy systems, such as constructing photovoltaic and wind power facilities in industrial parks and commercial complexes, to achieve local production and consumption of energy, enhancing the flexibility and reliability of energy utilization. Strengthen research and application of energy storage technologies, like battery storage, to address the intermittency issues of clean energy generation, promoting the stable and diversified development of energy consumption.

(4)

Guide enterprises towards green energy transition: The government can establish special funds to provide low-interest loans or financial support for companies transitioning to green energy, helping them purchase clean energy equipment and conduct technological research and development. Enterprises that actively transition should be given policy incentives, such as priority in energy supply and convenience in project approval. Organize industry associations and research institutions to offer technical consulting and training services for enterprises, assisting them in overcoming technical challenges during the transition. Through policy guidance, encourage enterprises to prioritize the use of clean energy in production processes, gradually reducing their reliance on traditional energy sources.

(5)

Strengthen the supervision of environmental regulation policies: Establish a dedicated supervisory body for the enforcement of environmental regulations to enhance daily oversight of corporate energy consumption. Utilize big data and Internet of Things technology to monitor real-time energy consumption and pollutant emissions from companies, ensuring strict compliance with environmental regulations. Hold accountable any departments or individuals who fail in their supervisory duties, enhancing the seriousness and effectiveness of policy implementation. Regularly publish information on corporate energy consumption and compliance with environmental regulations to foster public scrutiny, creating an atmosphere of societal participation that promotes the optimization of energy consumption structures and improves eco-efficiency.

(6)

The coordination of carbon pricing mechanisms and green financial instruments: Institutional design, market mechanisms, and financial innovation can form a synergistic closed loop of “policy guidance–market pricing–financial empowerment”. With the carbon pricing mechanism as the core, resource allocation is guided through market signals; green financial instruments are used as a lever to amplify the incentive effect of policies; and institutional innovation is taken as a guarantee to solve the problems of inter-departmental and cross-regional coordination. The carbon emission statistical accounting system of pilot cities needs to be aligned with the MRV standards of the national carbon market. For example, Huzhou City has established a “1 + 4” carbon emission control system, linking regional carbon budgets with quota allocation in the national market. Pilot cities should develop diversified financial instruments based on underlying assets such as carbon allowances, carbon sinks, and carbon footprints.