Data and sample
This paper selects Chinese listed companies from 2010-2018 as the initial sample. The above listed companies are also screened according to the following criteria: (1) listed companies with special treatment (ST/PT) and financial listed companies are eliminated; (2) insolvent listed companies, i.e. those whose debt ratio is already higher than 100%, are eliminated; (3) listed companies that are merged and restructured are excluded; (4) those companies listed in the current year are excluded; (5) listed companies with incomplete data on certain indicators are excluded. Finally, 14,110 observations were obtained, involving 2,282 unique listed companies. To mitigate the influence of potential outliers, all continuous variables are winsored at the level of 1%. The basic data of the dependent variables (TFP) are from CSMAR database and calculated by LP method and OP method. The independent variables (ER, ER2 and ER3) are constructed according to the frequency of environmental-related vocabulary and the proportion of words in environmental-related paragraphs in annual government work reports at the city level. In addition to regulatory capture from the China Research Data Service Platform (CNRDS), other firm-level data comes from the China Stock Market and Accounting Research (CSMAR) Database. The region-level data such as per capita GDP, industrial structure, population density, foreign direct investment and government science expenditure scale come from the Economy Prediction System (EPS) Database.
Total factor productivity (TFP). TFP has been a core measure of input-output efficiency, but there has been a lack of academic consensus on how to measure TFP (Syverson 2011). Since non-parametric methods such as data envelopment analysis may not avoid simultaneity and selection bias problems (Gatto et al. 2011), parametric methods such as stochastic frontier analysis rely heavily on the assumption of TFP distribution (Xiao et al. 2021). In light of these deficiencies, we employ two semi-parametric methods proposed by Levinsohn and Petrin (2003) and Olley and Pakes (1996), called LP and OP methods, respectively, are the most widely used to measure enterprises' TFP (Ai et al. 2020; Peng et al. 2021). The LP method is an improvement of the OP method due to its unbiased consistent estimation results and its superiority in mitigating both the endogenous problems and sample loss (Cai and Ye 2020). Therefore, in the following analysis, only the enterprises' TFP estimated by LP method is analyzed, and the enterprises' TFP obtained by the OP method will be used as the benchmark regression robustness test.
Environmental regulation (ER). At present, there is no unified standard on the measurement of environmental regulation intensity in academia. In order to further discuss the impact of environmental regulation on enterprise's TFP, this paper needs to construct appropriate environmental regulation variable. The existing research measures environmental regulation are usually based on pollution intensity (Ren et al. 2018), pollution control cost (Wang and Shen 2016), operating cost of pollution control facilities (Becker et al. 2013), the number of environmental protection personnel (Zhou et al. 2017), and the synthetic index of environmental regulation (Ai et al. 2020), these indicators often focus on one aspect of government environmental governance, while China's environmental regulation means include both economic means and legal and administrative means. Therefore, the above indicators are difficult to measure the overall appearance of government's environmental governance, and there is obvious endogeneity between these indicators and economic development (Cai et al. 2016). As such, they may suffer from severe measurement errors and simultaneity bias (Yang and Song 2019; Peng et al. 2021). The government work report is an outline for the administration and implementation of decisions and resolutions of organs of power in accordance with the law, it is a programmatic document to guide the government work, so it can reflect the government's environmental governance policy comprehensively. The proportion of environment-related vocabulary in a city's government annual work report is commonly used by the public to measure the amount of actual effort that local government has exerted in environmental governance (Chen et al. 2018). In addition, since the government work report is often prepared at the beginning of the year, it is not affected by the economic development of that year, which can effectively alleviate the endogenous problems (Yang and Song 2019).
For this reason, referring to Chen et al. (2018), this paper selects the frequency of environment-related vocabulary in each city's annual government work report as a proportion of the total frequency of vocabulary as a proxy variable for environmental regulation. But Chen et al. (2018) only selected five environment-related vocabulary, including “environment”, “energy consumption”, “pollution”, “emission reduction” and “environmental protection”, the vocabulary of environmental category is not specific. Compared with Chen et al. (2018), this paper chooses a richer vocabulary, which can more comprehensively reflects the strength of government's environmental governance.
However, the above method is not so perfect, because we can not rule out the influence of vocabulary such as “political ecology”, “vicious cycle” which are not related to environmental protection. Therefore, we employ the total number of words in the environment-related paragraphs as a proportion of the total number of words in that annual government work report (ER2) as a robustness test. Furthermore, the environmental-related vocabulary selected in this paper is subjective and random, in order to avoid estimation bias due to subjectivity and randomness, based on the original environmental-related vocabulary, this paper deleted four vocabulary with low frequency, such as “cycle”, “sustainable development”, “greening”, “particulate matter”, to construct the independent variable ER3 as the robustness test.
The specific construction steps of environmental regulation indicators in this research are as follows: First, manual collection of government work reports from prefecture-level and above cities in China for 2010-2018; Secondly, read each government work report in order to pick out the paragraphs devoted to the ecological environment; Finally, use Python 3.8 to calculate ER, ER2 and ER3 according to environmental-related text.
Enterprises' bargaining power (BP). There are few studies on enterprises' bargaining power, and most of them measure enterprises' bargaining power unilaterally by the amount of tax paid, the number of employees, and regulatory capture, and lack the organic integration of the three, thus failing to accurately characterize enterprises' bargaining power. Compared with other methods, the entropy evaluation method can eliminate the interference of human factors and make the evaluation results more scientific and reasonable. Therefore, we use the entropy evaluation method to compute a comprehensive score of enterprises' bargaining power, where corporate tax payments, number of employees, and regulatory capture all positively affect enterprises' bargaining power. Local government officials realize that political promotion mainly depends on economic performance (Zheng et al. 2015), so they often ignore the environmental violations of enterprises in exchange for local fiscal revenue, employment rate and economic growth (Jiang et al. 2014; Chen et al. 2018; Wang et al. 2018). Cai et al. (2011) found that travel and hospitality expenses under the administrative expense account in the financial statements of Chinese listed companies are often used by firms for bribes, seeking government support, etc., and that the item involves expenses for catering, entertainment, and other activities that are strongly associated with corrupt practices such as corporate bribery. And it has been shown that corrupt practices will reduce enterprises' TFP (Wu et al. 2017). Referring to Cai et al. (2011), this paper uses the sum of business entertainment fee and travel expenses to measure enterprises' regulatory capture behavior. Therefore, it is relatively reasonable and scientific to select the tax amount, the number of employees and regulatory capture to synthesize the bargaining power index.
Corporate ownership structure (Soe), Soe is 1 if it is a state-owned enterprise, 0 otherwise.
Executive compensation incentives (Eci), in this paper, the median of the executive compensation is used as the cut-off point, with the higher group assigns a value of 1 and the lower group assigns a value of 0.
Enterprise pollution level (Pol), Pol is 1 if it is a heavy-polluting enterprise, 0 otherwise. And the heavy-polluting enterprises are defined according to the Guidelines for Environmental Information Disclosure of Listed Companies issued by the Ministry of Environmental Protection in 2010 and the industry classification of China Securities Regulatory Commission in 2012.
Political constraints (PC), the sample cities are divided into provincial capital cities and non-provincial capital cities, and PC is 1 if it is a provincial capital city, and 0 otherwise.
To control for other important factors affecting enterprises' TFP, the control variables are selected mainly at the region- and firm-level in this paper. Referring to Hou et al. (2020), the region-level control variables mainly include GDP per capita (Pgdp), industrial structure (Indu), population density (Pden), foreign direct investment (FDI) and science expenditure scale (Sci) to control the effects of economic development level, industrial structure characteristics, human activity scale, foreign trade and R&D investment on enterprises' TFP. Referring to Ai et al. (2020) and Feng et al. (2020), this paper also incorporates firm size (Size), asset liability ratio (Lev), firm performance (Roa), capital labor density(CD), ratio of fixed assets(FR), and corporate governance, respectively. Among them, the variables of corporate governance level include the shareholding ratio of the largest shareholder (Shrcr), the duality of chairman and general manager (Duality), and the proportion of independent directors (Bodind).
Model specification and model approach
In order to examine the causal influence of environmental regulation on enterprises' TFP, we construct the following model:
Among them, where TFPit is the TFP of firm i in year t; ERmt is the environmental regulation intensity of city m in year t; Ximt is a set of control variables that include firm- and region-level; μi, νt andγj represent the firm, year and industry fixed effect, respectively; εit is the random error term.
To examine the moderating role of enterprises' bargaining power, the following model is designed in this paper:
BPit is the bargaining power of enterprise i in year t.
In order to explore the boundary when environmental regulation works, the following model is formalized:
Moderatorimt is the four moderating variables of heterogeneity analysis.
 Among them, the government work reports come from: http://www.drcnet.com.cn/www/int/.
 Specifically, the environmental terms selected in this paper include: pollution, emission reduction, ecology, PM2.5, pm2.5, haze, emission, air, pm10, PM10, green, environmental protection, energy conservation, dust, ammonia nitrogen, soot, atmosphere, sulfur dioxide, sewage, SO2, water conservation, nitrogen and oxygen, chemical oxygen demand, COD, energy consumption, forest coverage, low carbon, pollution control, wastewater, waste gas, carbon dioxide, consumption reduction, cycle, sustainable development, greening, particulate matter.
 Heavy-polluting industries specifically includes thermal power, iron and steel, cement, electrolytic aluminum, coal, metallurgy, chemical, petrochemical, building materials, paper, brewing, pharmaceutical, fermentation, textile, tannery and mining and other 16 types of industries.