The effect of automation on firms’ carbon dioxide emissions of China

Abstract

This paper empirically analyzes the impact of automation upon firms’ carbon dioxide emissions (CO₂ emissions) of China by using data for the period 1998–2009. Our research yields a few findings. First, we find that automation as measured by the robot density can reduce firms’ CO₂ emissions intensity. Specifically, 1% increase in the robot density leads to a 0.018% decrease in CO₂ emissions intensity. Second, we find that automation reduces firms’ CO₂ emissions intensity by promoting firms’ technological innovation and improving management efficiency. Finally, we find that automation exerts a greater impact on reducing CO₂ emissions intensity for firms in industries with high CO₂ emissions intensity rather than low CO₂ emissions intensity, and for firms in capital-intensive industries rather than non-capital-intensive industries, as well as firms in industries with high servitization of manufacturing rather than low servitization of manufacturing. Moreover, the mitigating effects of automation have been given greater play on firms’ CO₂ emissions intensity after the global financial crisis.

1 Introduction

With the development of economy, the greenhouse gases with carbon dioxide (CO₂) as the most part aggravate the deterioration of the environment (Kuang and Lin 2021) and cause global warming (Du et al. 2018; Qin et al. 2021). It is predicted that without appropriate measures taken to stop the growth of CO₂ emissions, the earth’s temperature will rise by 3–4 degrees centigrade compared with that in the pre-industrial era (Chen and Lin 2021). Therefore, many countries, including China, have formulated plans to achieve carbon neutrality.¹ Studies have proved that the digital technology plays an important role in energy utilization and environmental improvement (Ciocoiu 2011; Bakhriddinovna 2021; Li et al. 2022a, b). With the continuous development of automation driven by computer numerical control machinery, industrial robots and artificial intelligence (AI), automation is becoming the core technology of CO₂ emissions reduction (Alarfaj et al. 2020). For example, the GoBe Robots can help to reduce CO₂ emissions by thousands of tons.² However, unlike the vigorous development at the practical activity, whether the automation can promote CO₂ emissions reduction, especially for developing countries like China which has become the largest CO₂ emitter in the world (Yu and Yu 2019), is still an open question in the existing literature.

Previous studies have researched CO₂ emissions reduction from the perspectives of trade (Guo et al. 2010; Richter and Schiersch 2017; Cole et al. 2013; Forslid et al. 2018), environmental regulation (Wen and Liu 2022), energy consumption (Ahmad et al. 2016), economic growth (Li 2010; Meng et al. 2011; Zhang et al. 2012), urbanization (Liu et al. 2011; Zhang and Lin 2012; Zhang et al. 2014) and technology development (Ang 2009; Moyer and Hughes 2012; Li et al. 2022a, b). In these studies, the measurement of CO₂ emissions is mostly at national and regional levels (Guo et al. 2010; Su and Ang 2015; Wang et al. 2020; Li et al. 2022a, b). For example, based on China’s input-output table, Wang et al. (2020) study China’s overall CO₂ emissions intensity by using multi-region multiplication structure decomposition. While there is a small part of the literature focusing on the CO₂ emissions intensity at the firm level using the fuel-specific emissions factors (Richter and Schiersch 2017; Forslid et al. 2018), which is highly related to our research. This strand of the literature mainly focuses on the impact of exports on firms’ CO₂ emissions intensity. Specifically, based on the data of German, Richter and Schiersch (2017) derive empirical strategies from the firm production function and find that the firms’ export intensity is negatively related to the CO₂ emissions intensity. This negative correlation is also confirmed with Swedish firms data (Forslid et al. 2018) and Japanese firms data (Cole et al. 2013). In addition, studies from the perspective of technological development in the above literature are also highly relevant to our research. Many studies confirm the key role of technological development including technology transfer (Ang 2009), digital technology (Moyer and Hughes 2012; Nguyen et al. 2020; Li et al. 2022a, b; Zhang et al. 2022) in CO₂ emissions reduction at the country, province and city levels. Among this literature, Li et al. (2022a, b) use country-level data to examine the impact of industrial robot applications on CO₂ emissions intensity, and find that the application of industrial robots can reduce CO₂ emissions intensity, which is similar to our study. In addition, some researchers believe that the increase of patent filings (Ma et al. 2021), especially the environmental patent filings (Hashmi and Alam 2019; Wang et al. 2012), also plays an important role in reducing CO₂ emissions of the country and the province. However, technological progress can also increase CO₂ emissions due to its effects on output growth (Van den Bergh 2011; Li and Jiang 2016; Zhang et al. 2017). In this regard, some literature finds the dual effects of technological development upon CO₂ emissions (Yang and Li 2017). The existence of rebound effect makes the impact of technological development on CO₂ emissions reduction more complicated. As a specific manifestation of digital technology progress, automation is an important part and a major innovation of Industry 4.0 (Li et al. 2022a, b), while few studies focus on the impact of automation on firm CO₂ emissions.

For the effects of automation, the existing literature mainly focuses on the point of employment (Acemoglu 2002; Autor and Salomons 2018; Acemoglu and Restrepo 2019), income inequality (Aghion et al. 2019; Hémous and Olsen 2022), productivity (Graetz and Michaels 2018; Acemoglu and Restrepo 2020), economic growth (Acemoglu and Restrepo 2017; Graetz and Michaels 2018), as well as environmental pollution (Liu and De Giovanni 2019; Li et al. 2022a, b). And environmental pollution is highly relevant to our research. Existing research believes that robot technology can help firms to reduce resource waste and pollution emissions, thereby mitigating environmental pollution (Liu and De Giovanni 2019). From an empirical perspective, based on the Environmental Kuznets Curve (EKC) model, Li et al. (2022a, b) find that the application of industrial robots can reduce a country’s CO₂ emissions intensity through green total factor productivity and energy intensity. Besides that, there is also a strand of literature studying the impact of AI on environmental pollution (Ye et al. 2020; Ghane et al. 2016), which is also highly relevant to this study. Specificially, on the one hand, the adoption of AI can optimize firms’ production conditions and improve the efficiency of pollution control (Ye et al. 2020), and formulate reasonable and efficient emission control measures (Guozhen et al. 2016), thus reducing environmental pollution (Ghane et al. 2016). While on the other hand, with the popularity of machine learning systems, AI systems consume too much power and generate a large amount of CO₂ emissions (Henderson et al. 2020). Strubell et al. (2019) find that training a AI processing system can generate 1400 pounds of CO₂ emissions. Therefore, AI has rebound effects on CO₂ emissions. Through the above discussion, the existing research has not empirically explored the effects of automation on firms’ CO₂ emissions intensity of China and its mechanism.

To address this issue, we focus on the role of automation in manufacturing industries in reducing firms’ CO₂ emissions intensity of China. Our starting point is that automation in an industry can be identified from the robot density, which equals the number of robots per million hours worked (Aghion et al. 2019). To put this idea into practice, we use highly refined firm-level data on China’s manufacturing firms for the period 1998–2009, which comes from three sources-the China’s Environmental Statistics Database (CESD), the China’s Industrial Enterprise Database (CIED), as well as the International Federation of Robotics (IFR).

Through empirical analysis, our findings can be summarized as follows. We first find that there is a negative effect of automation on firms’ CO₂ emissions intensity. Specifically, 1% increase in the robot density (the proxy for automation) in the industry leads to a 0.018% decrease in firms’ CO₂ emissions intensity. This relationship is verified to hold over solving endogeneity and various robustness checks, including the use of alternative measures, changing the time interval of research sample, controlling industry-time trend, and so on. Then we find two potential mechanisms to explain how automation affects CO₂ emissions intensity at the firm level. One side is through promoting the firms’ technological innovation, while the other is through improving the firms’ management efficiency. Both channels cause the reduction of firms’ CO₂ emissions intensity. Finally, through heterogeneous effects analysis, we find that automation exerts a greater impact on reducing CO₂ emissions intensity for firms in industries with high CO₂ emissions intensity rather than low CO₂ emissions intensity. Besides, automation has significant effects on reducing CO₂ emissions intensity for firms in capital-intensive industries, but not for those in non-capital-intensive industries. Furthermore, compared to firms in industries with low servitization of manufacturing, automation has a greater impact on reducing CO₂ emissions intensity for firms in industries with high servitization of manufacturing. Moreover, the empirical results also show that the mitigating effects of automation have been given greater play on firms’ CO₂ emissions intensity after the global financial crisis.

Our study contributes to the existing literature in three ways: firstly, different from few existing literature focusing on developed countries’ firm CO₂ emissions intensity and export (Forslid et al. 2018; Richter and Schiersch 2017; Cole et al. 2013), we not only investigate the impact of automation in the manufacturing industry on firms’ CO₂ emissions intensity of China, which is the largest developing country and the main driver of the increase of global CO₂ emissions,³ but we also explore the mechanisms of the effects. Therefore our analysis provides empirical evidence for developing countries, not only China, to develop digital technology to reduce CO₂ emissions. Secondly, we not only integrate highly refined research data, but also extend the time span of the research data to 2016 at the industry level to explore the impact of automation on the industry’s CO₂ emissions intensity from 2006 to 2016. More specifically, we have obtained the data of the China’s Environmental Statistics Database (CESD), which covers major pollutant emissions and various energy consumption indicators at the firm level collected by the National Bureau of Statistics of China (NBSC). This dataset is monitored and inspected by China’s local environmental protection departments at any time and is considered as the most comprehensive and reliable pollution emissions dataset at the firm level in China. We combine the data of CESD with the data of the China’s Industrial Enterprise Database (CIED), and further match with the robot data provided by the International Federation of Robotics (IFR). With them we finally form the highly refined research data. Thirdly, our research is different from the existing research on automation and CO₂ emissions at the macro levels (Li et al. 2022a, b), which may ignore the firm specific characteristics. We have creatively constructed the CO₂ emissions intensity for China’s manufacturing firms by using the sectoral method (Liu et al. 2015; Shan et al. 2018). Based on this indicator, we can investigate the heterogeneity of CO₂ emissions at the firm level.

The rest of the paper proceeds as follows. Section 2 introduces the theoretical framework of our research. Section 3 lays out the research design, including the empirical model, data sources, and the description of variables. Section 4 shows our main empirical results. Section 5 discusses some extensions of the heterogeneous analysis for the main findings. And Section 6 concludes the conclusion.

2 Theoretical framework

As previously mentioned, as an important part of Industry 4.0, the application of automation technology represented by industrial robots has provided many economic benefits, such as increasing production flexibility, reducing production costs, improving productivity, and mitigating resource waste in production (Gaggl and Wright 2017; Graetz and Michaels 2018; Aghion et al. 2019; Acemoglu and Restrepo 2020; Li et al. 2022b). Although the use of industrial robots does not necessarily give priority to reduce environmental pollution, the series of automation technologies driven by robot promote pollution treatment facilities upgrading thereby improving the efficiency of pollution control (Ye et al. 2020; Javaid et al. 2021; Liu and De Giovanni 2019), thus reducing CO₂ emissions (Zhang et al. 2022). What’s more, the application of industrial robots can help to increase production efficiency by replacing labor, which can expand firms’ output with the same input, decreasing resource and energy consumption, further reducing CO₂ emissions (Graetz and Michaels 2018). In addition, the application of industrial robots can promote firms’ intelligent transformation, which can also improve environment from the demand side and further promote social sustainability (St Louis and Cazier 2010). For example, in recent years, the widespread of shared bicycles has greatly reduced the frequency of private car travel, thus reducing energy consumption (Li et al. 2020). Therefore, based on the forgoing analysis, we propose the first hypothesis:

Hypothesis 1. Automation can reduce a firm’s CO₂ emissions intensity.

We further investigate that how automation reduces firms’ CO₂ emissions intensity. Automation contains digital algorithms and some other technologies. Its high-precision attribute characteristics provide firms with superior hardware infrastructure and digital algorithm technology, which supports R&D innovation (Trajtenberg 2018; Vannuccini and Prytkova 2021). What’s more, the use of industrial robots in production helps to promote firms’ technological innovation through knowledge spillover, learning ability improvement and R&D investment (Liu et al. 2020). And industrial robots also contribute to green innovation (Liu and De Giovanni 2019). Meanwhile, using robots to collect, integrate, classify and process user data is helpful to optimize firms’ business strategy and improve the process of R&D (Cockburn et al. 2018). This supports firms to invest more funds, personnels and some other factors in R&D innovation, thereby driving firms’ innovation activities.

To sum up, automation technology can promote firms’ technological innovation. And technological innovation is the main driver of firms’ environmental governance (Grossman and Krueger 1991; Wang et al. 2012; Hashmi and Alam 2019; Ma et al. 2021), which plays a key role in reducing firms’ CO₂ emission intensity (Du et al. 2019; Yang et al. 2020). Therefore, the above discussion leads to the second hypothesis:

Hypothesis 2. Automation can reduce a firm’s CO₂ emissions intensity by promoting technological innovation.

In addition, automation technology can help to establish an environmental information coordination center, integrating online monitoring of data information, real-time transmission of video images, and real-time reporting of environmental pollution (Guozhen et al. 2016). This optimizes the efficiency of the firm’s pollution management, guiding managers to accelerate the elimination of high energy consumption and pollution production links while making more efficient and cleaner improvement on the existing production, further reducing the firm’s CO₂ emissions intensity. In addition, the extensive use of industrial robots in production can help the optimization of firms’ production factors and promotion of the reconfiguration of production resources, thus optimizing the production process and the product management system, further improving energy efficiency and promoting firms’ green transformation (Ye et al. 2020; Wang et al. 2022). The reason is that, on the one hand, compared with labor, automation equipment can more accurately simulate the physical parameters and performance differences of products, thus reducing the rework and defective rate, which reduces the corresponding energy consumption and pollutant generation (Cai et al. 2013). On the other hand, automation system with a large amount of data collection, integration and analysis can help firm to adopt the lean management principle in the activities and processes of the organization (Khuntia et al. 2018). This improvement accurately identifies the customer needs and masters the production and transaction dynamics, thus promoting the effective coordination between the human resources inside firm and the customer needs outside.

Therefore, automation provide a large amount of market information for firm production and maximize firm management efficiency, thus reducing ineffective inventory, the corresponding energy consumption and CO₂ emissions. In this regard, the above discussion leads to the third hypothesis:

Hypothesis 3. Automation can reduce a firm’s CO₂ emissions intensity by improving management efficiency.

3 Research design

In this section, we present the research design. We begin at the baseline model, and then introduce the variables and data with which we estimate the model.

3.1 Model

The goal of our empirical strategy is to evaluate the impact of automation on firms’ CO₂ emissions intensity. Consistent with Cherniwchan (2017), we set the following empirical formula as our baseline model.

$${lnCO}_{{2}_{it}}=\alpha +{\beta }_{1}{lnau}_{jt}+{\beta }_{2}\overrightarrow{X}+{\gamma }_{i}+{\delta }_{t}+{\varepsilon }_{it}$$

(1)

where i, j, and t represent the firm, two-digit CIC (the China Industrial Classification) industry, and year respectively. ${\mathrm{lnCO}}_{{2}_{\mathrm{it}}}$ is CO₂ emissions intensity at the firm level, which is calculated as the logarithm of the ratio of a firm’s CO₂ emissions to its real output. ${\mathrm{lnau}}_{\mathrm{jt}}$ represents the automation of the two-digit CIC industry and is measured by robot density. In addition, $\overrightarrow{X}$ represents a vector containing control variables which may potentially affect ${\mathrm{lnCO}}_{{2}_{\mathrm{it}}}$. ${\upgamma }_{\mathrm{i}}$ and ${\updelta }_{\mathrm{t}}$ are the firm, and year fixed effects, respectively. ${\upvarepsilon }_{\mathrm{it}}$ is the error term.

3.2 Measurements

3.2.1 CO₂ emissions intensity

CO₂ emissions intensity at the firm level is measured by the logged value of the ratio of a firm’s CO₂ emissions to its real output (deflated by the ex-factory price index of industrial producers). As a robustness check, the real industrial value added (deflated by the ex-factory price index of industrial producers) is also used as an alternative variable of a firm’s output. Although firms do not directly report CO₂ emissions, they report information about energy use in detail. Therefore, following Liu et al. (2015) and Shan et al. (2018), we use sectoral method to calculate CO₂ emissions at the firm level. The formula for calculation is as follows:

$${CO}_{{2}_{ik}}={AD}_{ik}\times {NCV}_{k}\times {CC}_{k}\times {O}_{ik}$$

(2)

$${CO}_{{2}_{i}}=\sum\nolimits_{k=1}^{3}{CO}_{{2}_{ik}}$$

(3)

where k = 1,2,3, represents three types of energy, including coal, fuel oil and natural gas, respectively. ${\mathrm{CO}}_{{2}_{\mathrm{ik}}}$ is CO₂ emissions generated by the consumption of energy k by firm i. ${\mathrm{AD}}_{\mathrm{ik}}$ is the consumption of energy k by firm i. ${\mathrm{NCV}}_{\mathrm{k}}$ is the net heat value, which is generated by per unit energy k. ${\mathrm{CC}}_{\mathrm{k}}$ is CO₂ emissions generated by the unit net heat value of energy k. ${\mathrm{O}}_{\mathrm{ik}}$ is the oxidation rate, which is the oxidation conversion rate of energy k combustion in firm i.

Based on the formulas (2) and (3), we use various energy consumption indicators offered by CESD to calculate CO₂ emissions for China’s manufacturing firms from 1998 to 2009⁴ and further depict its dynamic trend in different regions and detailed industries. As Fig. 1 shows, from 1998 to 2009, CO₂ emissions in China increased from 5.399 kt to 75.463 kt.

Fig. 1

The dynamic trend of CO₂ emissions: 1998–2009

Figure 2 shows that CO₂ emissions in China’s manufacturing mainly come from firms in central and western regions, which has experienced a great growth. About CO₂ emissions in China’s different manufacturing industries, Fig. 3 shows that CO₂ emissions in the following industries are high, including petroleum processing, coking and nuclear fuel processing industry (25), ferrous metal smelting and calendering processing industry (32). While CO₂ emissions in other industries are relatively low, especially the textile and garment, shoe and hat manufacturing (18), leather, fur, feather (down) and its products (19) and cultural, educational and sports goods manufacturing (24). The main reason is the great differences of energy consumption among various industries.

Fig. 2

The dynamic trend of the CO₂ emissions of manufacturing firms in China’s different regions: 1998–2009

Fig. 3

The CO₂ emissions in China’s different manufacturing. Note: 13-agricultural and sideline food processing industry; 14-food manufacturing; 15-beverage manufacturing; 16-tobacco products industry; 17-textile industry; 18-textile and clothing, shoes and hats manufacturing; 19-leather, fur, feather (down) and its products; 20-wood processing and wood, bamboo, rattan, palm and grass products; 21-furniture manufacturing; 22-papermaking and paper products industry; 23-reproduction of printing and recording media; 24-cultural, educational and sports goods manufacturing; 25-petroleum processing, coking and nuclear fuel processing industries; 26-chemical raw materials and chemical products manufacturing; 27-pharmaceutical manufacturing; 28-chemical fiber manufacturing; 29-rubber products industry; 30-plastic products industry; 31-non metallic mineral products industry; 32-ferrous metal smelting and rolling processing industry; 33-non ferrous metal smelting and rolling processing industry; 34-metal products industry; 35-general equipment manufacturing; 36-special equipment manufacturing; 37-transportation equipment manufacturing; 39-electrical machinery and equipment manufacturing; 40-manufacturing of communication equipment, computers and other electronic equipment; 41-manufacturing of instruments and meters, culture and office machinery; 42-handicrafts and other manufacturing; 43-waste resources and waste material recycling and processing industry

3.2.2 Automation

Following Graetz and Michaels (2018), we use the robot density (the logged value of the number of robots per million hours worked) from IFR as the proxy of automation. With this index, we depict the dynamic trend of automation at the industry level in Fig. 4, and further investigate the dynamic trend in different regions and specific industries, which are shown respectively in Figs. 5, and 6. Overall, Fig. 4 shows that prior to 2006, the robot density of China’s manufacturing remained at 0. After 2005, except for the large decrease in 2008 due to the global financial crisis, the robot density is increasing in the other years, which is in line with the actual situation in China.

Fig. 4

The dynamic trend of the robot density: 1998–2009

Fig. 5

The dynamic trend of the robot density in China’s different regions: 1998–2009

Fig. 6

The robot density in China’s different manufacturing. Note: The industry introduction is consistent with that in Fig. 3

Figure 5 shows that the total dynamic trend of the robot density in different regions are consistent with that in Fig. 4, while the robot density in the eastern region of China is significantly higher than that in the central and western regions. This is because compared to the central and western regions, the eastern region has a higher level of economic development, investing more in research and development, and has more obvious advantages in talent accumulation as well as more complete infrastructure. We then further explore the robot density in the detailed industries, which is shown in Fig. 6. It indicates that the robot density of the plastic products industry (30) and the rubber products industry (29) are higher than that in the other industries.

3.2.3 Other variables

We further include the following variables in the analysis: (1) Firm age (age). We first calculate the differences of the current year and the firm’s year of establishment, and then take the logarithm of the sum of this value and 1. (2) Capital intensity (lnkl). It is the natural logarithm of the ratio of the firm’s real fixed assets (deflated by the fixed asset investment price index) to the number of its employees. A bigger value of lnkl indicates the firm is more capital intensive. (3) SOE (if_soe). If the firm is state-owned, the value is 1, otherwise it is 0. (4) FOE (if_fdi). If the firm is foreign-owned, the value is 1, otherwise it equals 0. (5) Financing capability (fc). It is measured by the ratio of firm’s interest expenditure to its real fixed assets (deflated by the fixed asset investment price index). The larger this variable’s value, the stronger the firm’s financing capability, which means it has less significant financing constraints. (6) Market concentration (hhi). We measure this variable by Herfindahl–Hirschman index at the two-digit CIC industry level, which is equal to the square sum of the market shares of the firms in the two-digit CIC industry. A bigger value of hhi indicates a more concentrated market. The above variables are all calculated based on the information of CIED. (7) Environmental regulation (envi). Environmental regulation is measured by the logged value of the number of environmental administrative punishment cases in the region according to China Statistical Yearbook on environment. A bigger value of envi indicates more serious environmental regulation. (8) Upstream AU index (UAU). It is measured by the the robot density of various manufacturing from IFR and the input-output relationship among industries from the 2007 National Input and Output Table which published by the NBSC. The detailed measurement is shown in the formula (5) of Section 5.2. (9) Alternative measurement of firm’s CO₂ emissions intensity (lnCO₂_1). We use the logged value of the ratio of firm’s CO₂ emissions to its real industrial value added (deflated by the ex-factory price index of industrial producers) with CESD and CIED. (10) Robot stock (lnaustock). It equals to the logged value of the robot stock from IFR. (11) Patent quality (patent_quality). Following Boh et al. (2014), we use the knowledge width method to measure it. (12) Patent quantity (patent_quantity). Patent quantity is the logged value of the total number of firm’s patents application in year t. The data used to measure patent quality and patent quantity are from China’s State Intellectual Property Office (SIPO), which contains detailed information of all patent filings since 1985, and patents are dated by the year of application. In particular, it includes the filing date, the applicant’s official name and address, the patent’s name and type, and the IPC classifications. (13) Management efficiency (m_efficiency). Management efficiency is equal to the logged value of the ratio of firm’s real sales (deflated by the ex-factory price index of industrial producers) to management cost, both of which can be got from CIED.

3.3 Data sources

To study the relationship between CO₂ emissions intensity at the firm level and robot density at the industry level, we rely on three datasets from 1998 to 2009-the China’s Environmental Statistics Database (CESD), the China’s Industrial Enterprise Database (CIED), as well as the International Federation of Robotics (IFR). These three databases provide the necessary data information for our research. More specifically, CIED is from the National Bureau of Statistics of China (NBSC) and includes rich information on all state-owned and above-scale non-state-owned firms’ characteristics, such as name, location, industry, opening year, balance sheet, employees, and so on. It has the advantages of large sample, comprehensive indicators and long time interval. CESD comes from NBSC, considered to be the most comprehensive and reliable environmental industrial firm dataset in China. It covers firms whose emissions of major pollutants accounts for more than 85% of the total emissions of various regions and reports the detailed information, such as firms’ fundamental information (name, legal person code and so on), major pollutant emissions (industrial wastewater, SO₂ and soot), as well as various energy consumption indicators (coal, fuel oil and natural gas and so on). IFR, a unique international association, provides the stock and installation of industrial robots that meet ISO standards in 75 countries or regions in the world from 1993 to 2019, and is classified by industry and use.

With regard to the match of these databases, firstly, we merge CIED and CESD, and the details are as follows: (1) Following Brandt et al. (2012) and Wang et al. (2018), we merge these two databases using firm code, firm name, and the firm’s administrative division code for the same year in turn. (2) We then use the firm’s abbreviated name and the province code to merge these two databases. Secondly, through the conversion of different kinds of industry code, we integrate the robot data in IFR with the firms’ data obtained in the first step. We make the following processing: (1) Based on the original robot data from IFR, we extract the robot data of all China’s industries. (2) We merge the robot data of industries in IFR with the two-digit CIC codes industry of the National Economic Industry Classification (see Table 8 in Appendix for the industry conversion details), realizing the integrating of robot data with China’s firms’ data. Finally, based on Feenstra et al. (2014) and Yu (2015), we also carry out a series of cleaning and treatment on the research sample. In detail, we exclude the observation that do not conform to one of the following accounting standards: (1) if the number of a firm’s employees are less than 8; (2) if one of the firm’s following indicators, such as age, real output, real sales, real industrial value added, and the net value of fixed assets are less than or equal to 0. In addition, due to lacking energy consumption data and poor data quality in 2010 and beyond, CO₂ emissions cannot be measured. Therefore, the final sample interval is from 1998 to 2009. Table1 summarizes the descriptive statistics of the main variables in this research.

Table 1 Descriptive statistics

4 Empirical results

In this section, we first present the estimation results of the model as well as the instrumental variable estimation. We then report the mechanism results about firm technological innovation promotion and management efficiency improvement. Finally, we make a set of robustness checks with the baseline model.

4.1 Baseline results

Table 2 reports the results of the effect of automation expressed by the robot density of the manufacturing industry on the CO₂ emissions intensity at the firm level based on model (1). Column (1) do not control any fixed effects and column (2) controls year fixed effects, while column (3) presents the results including year and firm fixed effects. These results indicate that automation decreases CO₂ emissions intensity of firms. The coefficient in column (3) indicates that a 1% increase in the robot density of the manufacturing industry is connected with a 0.018% decrease in firms’ CO₂ emissions intensity.

Table 2 Baseline results

4.2 2SLS estimates⁵

Endogenity caused by missing variables may affect the consistency of estimation results. Therefore, we build an upstream AU index (UAU) as an instrument. This instrumental variable is based on the the robot density of various manufacturing industries and the input-output relationship among industries, which is measured as follows: we first calculate the input-output weight with the complete consumption coefficient.⁶ Below is the formula derive the measure:

$${\text{weight}}_{lj} = a_{lj} + \sum\limits_{k = 1}^{n} {a_{lk} a_{kj} + \sum\limits_{s = 1}^{n} {\sum\limits_{k = 1}^{n} {a_{ls} a_{sk} a_{kj} + ...} } }$$

(4)

where weight_lj is the complete consumption coefficient of the manufacturing industry j to the upstream industry $l$, which is the sum of the direct consumption and indirect consumption of the goods or services of industry $l$ when industry $j$ provides a unit product for final consumption. The first term on the right side of Eq. (4) represents the direct consumption of the industry $j$ to the industry $l$, the second term represents the first round of indirect consumption, and so on. Similarly, the N + 1 term is the n-th round of indirect consumption.

Then, after obtaining the consumption weight of a manufacturing industry to other upstream industries, we multiply the weight by the robot density of each upstream manufacturing industry, through which we can quantify the absorption of automation factors of the manufacturing industry from all upstream industries. Finally, we aggregate this index based on the same industry to get the UAU of the manufacturing industry of interest. Specifically, the formula is as follows:

$$UAU_{jt} = \sum\limits_{l} {au_{lt} \times \omega_{ljt} }$$

(5)

where, au_lt is the the robot density of the upstream industry $l$ in year t. $\omega_{ljt}$ is the consumption weight of the downstream industry on the inputs of industry $l$ in year t, and UAU_jt is the upstream automation of manufacturing industry $j$ in year t.

Since the industry can significantly increase the use of robot with the help of high automation intermediate inputs, which means the automation in a industry is closely related to that in the upstream industries. Therefore, UAU meets the correlation requirements of instrumental variables. At the same time, the firms’ pollution in the industry can hardly affect the automation in upstream industries. Therefore, the instrumental variables are not related to the residual term, which can meet the exogenous requirements. In general, satisfying the assumptions of correlation and exclusionary restrictions, the instrumental variable UAU is reliable and effective. We will verify this in subsequent empirical analysis.

With UAU as the instrumental variable, we employ two-stage least squares to estimate baseline model (1). The estimation results are shown in columns (1)-(2) of Table 3. The weak identification test and the under identification test shows that the instrumental variables can meet requirements. The results of the first stage show that UAU is significantly positively related to lnau. The second stage results show that after solving endogenous problems, the coefficient of lnau on firms’ CO₂ emissions intensity is still significantly negative. Furthermore, we also overcome endogenity by using a 1-year lag of lnau as an alternative instrumental variable (L.lnau). The estimation results are shown in columns (3) and (4) of Table 3. The relevant tests also supports the correlation assumptions and exclusionary restrictions of instrumental variables, and the regression results are still consistent with that in the baseline model.

Table 3 2SLS estimates

4.3 Mechanisms

Through the theoretical analysis in the previous part, we propose that automation can reduce a firm’s CO₂ emissions intensity by promoting technological innovation and improving management efficiency. As stated above, technological innovation is expressed by a firm’s patent quantity and patent quality. Specifically, patent quantity (patent_quantity) is measured by the logged value of the total number of a firm’s patents applied in one year. And patent quality (patent_quality) is measured by the knowledge width method. Management efficiency (m_efficiency) is the logged value of the ratio of a firm’s real sales (deflated by the ex-factory price index of industrial producers) to management cost.

Based on these variables, we conduct a detailed analysis and investigation of the above-mentioned channels from the perspective of empirical analysis. The results are displayed in Table 4. Columns (1)-(4) shows that automation can significantly improve a firm’s patent quality and increase the number of a firm’s patent filings. Meanwhile, the coefficients of patent_quality and patent_quantity on firm’s CO₂ emissions intensity are all significantly negaitive. This supports the hypothesis 2 that the reduction effect of automation on a firm’s CO₂ emissions intensity is through promoting a firm’s technological innovation, measured by the firm’s patent quantity and patent quality. Column (5)-(6) demonstrates that the coefficient of lnau on a firm’s management efficiency is significantly positive and the coefficient of m_efficiency on a firm’s CO₂ emissions intensity is significantly negaitive, implying that automation can effectively improve a firm’s management efficiency. It supports the hypothesis 3 that automation can reduce a firm’s CO₂ emissions intensity by improving its management efficiency.

Table 4 Mechanisms

4.4 Robustness checks

To test the stability of our baseline results, in this part we conduct a series of robustness checks in Table 5, including alternative measurement of CO₂ emissions intensity and automation, further dealing with the potential reverse causality problem, deleting outliers, controlling industry-time trend, changing time period of the research sample, controlling trade liberalization effect and constructing the industrial robot density at the firm level.

Table 5 Robustness check results I

In the baseline model, CO₂ emissions intensity at the firm level equals to the logarithm of the ratio of a firm’s CO₂ emissions to its real output, and in column (1) of Table 5, we replace the firm’s output to the real industrial value added, thus constructing a new indicator lnCO₂_1. In our baseline results, automation is measured by the robot density. In column (2) of Table 5, this variable is replaced by the robot stock, which is the logged value of the robot stock (lnaustock). To address the endogeneity of the potential reverse causality problem, in column (3) of Table 5, we use a 1-year lag of lnau (l.lnau) to replace lnau and in column (4) of Table 5, we use a 1-year lag of other control variables to replace the original control variables. The results of above robustness checks, which are presented in columns (1)-(4) of Table 5, are consistent with the baseline results, which preliminarily proves the robustness of the results in this paper.

In column (1) of Table 6, we run the baseline model after deleting the top and bottom 1% values of CO₂ emissions intensity to avoid the influence of data thick tailed distribution on the coefficient. Considering that the unobservable industry-time trend factors may affect the baseline results, based on the baseline model (1), we further control the industry-time trend in column (2) of Table 6. Figure 4 shows that prior to 2006, the robot density of China’s manufacturing industry remained at zero, and began to increase rapidly after that. In order to avoid the influence of this trend on the estimation results, in column (3) of Table 6, we limit the research sample from 2006 to 2009 and conduct regression again. Considering that the rapid development of automation after 2005 may be caused by trade liberalization, following the previous literature (Lu et al. 2017), in column (4) of Table 6, we control the time trend through introducing the interaction term between the average tariff of the industry and a dummy variable of the year that equals 1 if it is the statistical year and 0 otherwise. In addition, following the idea of standard shift-share or “Bartik” (Bartik 1991), we use the ratio of firm’s employees to the industry’s employees to construct the industrial robot density at the firm level. We then run the baseline model using this indicator as the robustness test. The estimated results are displayed in column (5) of Table 6.

Table 6 Robustness check results II

The results of above robustness checks are shown in columns (1)-(5) of Table 6, indicating that they do not alter the interpretation of our main findings regarding that automation can significantly reduce CO₂ emissions intensity at the firm level, which further indicates that our finding remains robust after taking into consideration the different kinds of robustness checks.

Finally, due to the availability of data, the time interval of data used for firm level analysis above is 1998–2009. Considering the potential impact of the differences between 1998–2009 and the current period may influence the research conclusion, we further use the data at the industry level from 2006 to 2016⁷ for further robustness test. Specifically, the regression model at the industry level analysis is as follows:

$${\mathrm{lnCO}}_{{2}_{\mathrm{jt}}}=\mathrm{\alpha }+{\upbeta }_{1}{\mathrm{lna}u}_{\mathrm{jt}}+{\upbeta }_{2}\overrightarrow{\mathrm{X}}+{\upgamma }_{\mathrm{j}}+{\updelta }_{\mathrm{t}}+{\upvarepsilon }_{\mathrm{jt}}$$

(6)

where j and t represent the two-digit CIC industry and year respectively. ${\mathrm{lnCO}}_{{2}_{\mathrm{jt}}}$ is CO₂ emissions intensity at the industry level. ${\mathrm{lnau}}_{\mathrm{jt}}$ represents the automation of the two-digit CIC industry and is measured by robot density. $\overrightarrow{\mathrm{X}}$ represents a vector containing control variables which may potentially affect ${\mathrm{lnCO}}_{{2}_{\mathrm{jt}}}$, including real industry scale (the logged value of real industrial sales), industry structure (the ratio of real industrial sales to total real industrial sales), R&D intensity (the ratio of scientific R&D expenditure to real industrial sales), energy consumption (the logged value of coal consumption), energy intensity (energy consumption per unit of real industrial sales) and export intensity (the ratio of export to real industrial sales). ${\upgamma }_{\mathrm{j}}$ and ${\updelta }_{\mathrm{t}}$ are the industry, and year fixed effects, respectively. ${\upvarepsilon }_{\mathrm{jt}}$ is the error term. The data used in this model mainly come from four sources: China Industrial Economic Statistics Yearbook, China Industrial Statistics Yearbook, China Science and Technology Statistics Yearbook and China Energy Statistics Yearbook.

The estimation results of model (6) are shown in column (6) of Table 6. It shows that the main variable of interest, automation (lnau), is still significantly negative, which indicates that automation can help to reduce CO₂ emissions intensity, consistent with the baseline result.

5 Analysis of heterogeneous effects

In this section, we investigate the heterogeneity performance of impact of automation on firms’ CO₂ emissions intensity in terms of four dimensions: CO₂ emissions intensity, capital factor intensity, the manufacturing servitization and financial crisis.

5.1 High CO₂ emissions intensity versus low CO₂ emissions intensity

Considering that the impact of automation on CO₂ emissions intensity at the firm level may be different due to the extent of CO₂ emission intensity in industries. Therefore, we divide the sample into high and low CO₂ emissions intensity industries by the average CO₂ emissions intensity in all industries (if the industry belongs to the high CO₂ emissions intensity industries, emi equals 1, otherwise it equals 0), and we add the interaction term of lnau and emi (lnau × emi) in the model (1). The results are displayed in column (1) of Table 7. It shows that automation has a more significant role in CO₂ emissions intensity reduction for firms in industries with high CO₂ emissions intensity, but not for firms in industries with low CO₂ emissions intensity. This means that in order to make green and low-carbon development come true, automation plays a key role in China’s high CO₂ emissions industry.

Table 7 Heterogeneous analysis

5.2 Capital-intensive firms versus non-capital-intensive firms

Considering that different capital intensive firms may have different degrees of robot usage, which leads to heterogeneous effects of automation upon firms’ CO₂ emissions intensity. We divide the sample into capital-intensive firms (cap == 1) and non-capital-intensive firms (cap == 0), and add the interaction term of lnau and cap (lnau × cap) in the model (1). The regression results are shown in columns (2) of Table 7, which indicates that the coefficients of the interaction term are negative and statistically significant. This indicates that automation has a stronger negative impact on the CO₂ emissions intensity of capital-intensive firms. The reason is that compared with non-capital-intensive firms, the resources like capital, technology and other resource elements are more abundant in capital-intensive firms, which can give better play to the promotion effects of automation in technological innovation. According to mechanism analysis above, automation can reduce firms’ CO₂ emissions intensity by promoting technological innovation, which means that automation has a stronger negative impact on CO₂ emissions intensity of capital-intensive firms.

5.3 High servitization of manufacturing versus Low servitization of manufacturing

Promoting the service-oriented development of manufacturing industry is not only the main trend of the world’s industrialization development, but also an important strategic measure for the upgrading of China’s manufacturing industries. So how firms respond to the combined effect of automation development and the servitization of manufacturing by adjusting their CO₂ emissions intensity is worthy of investigation. We use the World Input-Output Database (WIOD) to construct the service-oriented indicators at the industry level (See Table 9 in Appendix for industry conversion details). Based on this indicator, we divide the sample into high servitization of manufacturing (ser == 1) and low servitization of manufacturing (ser == 0) by the mean of all the manufacturing industries. We then add the interaction term of lnau and ser (lnau × ser) in the model (1). The regression results are shown in column (3) of Table7. It shows that the coefficient of interaction term is significantly negative, indicating the reduction effects of automation upon CO₂ emissions intensity mainly reflects in firms with high servitization of manufacturing. The main reason is that, on the one hand, many service products, including business training, technology extension and product design, are knowledge intensive and technology intensive, which can play a synergistic role with intelligent robots, bringing a higher level of CO₂ emissions reduction while improving production and management efficiency. On the other hand, the in-depth application of services such as financial leasing and logistics can bring auxiliary support for the intelligent development and improvement of firms’ business efficiency, which also helps firms achieve CO₂ emissions reduction goals by virtue of automation.

5.4 Pre-global financial crisis versus post-global financial crisis

The global financial crisis in 2008 has brought different degrees of shocks to countries around the world. Therefore, after the crisis, in order to improve the ability to cope with a new round of potential crises and risks, countries around the world started to upgrade the industrial structures. In this process, as an important tool to improve productivity and upgrade the industrial structure, automation is widely applied into production. In order to examine whether the impact of automation on the CO₂ emissions intensity of firms is different from that after the crisis, we add the interaction term between the financial crisis (cris, it equals 1 if the statistical year is 2008 and otherwise it equals 0) and automation (lnau) in the model (1). The results are shown in Column (4) of Table 7. It indicates that the coefficient of the interaction term lnau × cris is significantly negative, which means that automation plays a greater role in mitigating firms’ CO₂ emissions intensity after the financial crisis.

6 Conclusion

This paper analyzes the impact of automation on firms’ CO₂ emissions intensity of China between 1998 and 2009 using data from the China’s Environmental Statistics Database, China’s Industrial Enterprise Database, and the International Federation of Robotics. After carefully empirical analysis with focus on China’s manufacturing firms, we yield a series of conclusions.

First, we find that automation can reduce firms’ CO₂ emissions intensity. This relationship is verified to hold over solving endogeneity and various robustness checks, including the use of alternative measures, resetting the time span of research sample, controlling industry-time trend, using industry level data from 2006 to 2016 to expand the sample time to 2016 and so on. Second, the reduction effects of automation on firm CO₂ emissions intensity are mainly through promoting technological innovation expressed by patent quantity and patent quality and improving management efficiency. Lastly, through heterogeneous analysis, we find that automation has a greater impact on reducing CO₂ emissions intensity of firms in industries with high CO₂ emissions intensity. Besides, compared with non-capital-intensive firms, automation can significantly reduce CO₂ emissions intensity of capital-intensive firms. Furthermore, the empirical results also show that the reduction effects of automation on firms’ CO₂ emissions intensity are mainly reflected in firms of high servitization of manufacturing. Moreover, the mitigating effects of automation have been given greater play on firms’ CO₂ emissions intensity after the financial crisis.

With the development of digital technology, automation has gradually become the core technology of the new round of industrial revolution. Our findings have the implications for the sustainable development of developing countries like China who creates great CO₂ emissions, indicating the importance of promoting the deep integration of automation technology with manufacturing industries and strengthening the research and development of automation technology, especially in the field of CO₂ emissions reduction research.

Appendix

Tables 8 and 9

Table 8 Industry conversion table from IFR to CIC

Table 9 Industry conversion table from WIOD to CIC