Digital Transformation, Service-Oriented Manufacturing, and Total Factor Productivity: Evidence from A-Share Listed Companies in China

Abstract

Promoting corporate total factor productivity is important content for high-quality development in the manufacturing industry. Under the background of digitalization, probing whether digital transformation affects corporate total factor productivity is worth further study. We investigate the digitalization level of China’s listed manufacturing companies from a micro perspective using text analysis methods and empirically study whether and how corporate digital transformation affects its total factor productivity. We reveal that: First, moderate digital transformation improves the total factor productivity of manufacturing firms significantly, and there is a nonlinear inverted U-shaped effect between them, which is robust across different methods of measuring digital transformation and TFP and samples and passes the sensitivity analysis. Second, mechanism examination verifies that the promotion effects that digital transformation provides on corporate total factor productivity is mainly through promoting service-oriented manufacturing, technological innovation, and enterprise-scale effect. Third, that promotion effect is more intensified when the company is state-owned, or the company is non-key pollution monitored. Finally, with the deepening of digitalization, its positive promotion effect on corporates’ total factor productivity varies—showing a significant threshold feature, which is much stronger when the degree is lower than the critical value. This paper argues that digitalization has a marginal decreasing boosting effect on enterprise total factor productivity after reaching a certain degree. The study plays an important role in guiding the current promotion of enterprise digital transformation.

1. Introduction

The 20th Party Congress report proposed that “we should adhere to the theme of promoting high-quality development, integrate the strategy of expanding domestic demand with deepening supply-side structural reform, enhance the endogenous power and reliability of the domestic circulation, improve the quality and level of the international cycle, accelerate the construction of a modern economic system, and focus on improving total factor productivity”. Promoting total factor productivity is important content and performance for high-quality development and an inevitable requirement to achieve economic sustainability. Being a vital pillar of the real economy, China’s total manufacturing output value has achieved first for many consecutive years in the whole world. However, at present, the main resources of the Chinese manufacturing industry are still concentrated in the middle and even low-end fields, while the high-end field has been blank for a long time, so it still faces problems such as unreasonable structure and production efficiency to be improved. The fundamental reason is that the ability for independent innovation is not strong since some key resources, such as talents, basic materials, technology, etc., are not mastered. It is still facing the risk of a “stuck neck” in many high-end fields. In addition, there are still many gaps with developed countries, such as resource utilization efficiency, service-oriented manufacturing, product quality, and efficiency, so the Chinese manufacturing industry is “big but not strong, comprehensive but not excellent” [1]. Problems such as resource allocation and technological innovation have restricted the further progress of corporate total factor productivity in the Chinese manufacturing industry [2,3]. Young et al., (2003) calculated that during 1978–1998 China’s average annual growth rate of TFP (total factor productivity) was only 1.4% based on traditional statistical corrections, taking into account the effects of labor dividends and transfers and education improvements [4]. Studies have shown that digitization can improve information transparency inside and outside enterprises by reducing communication costs [5,6] and can influence the innovation activities of enterprises [7] to improve their productivity [6].

As digital technology is easy to acquire for extensive applications, the digital economy has become the most dynamic and innovative new economic form at present and a “new engine” to promote economic development and progress [8]. According to China’s 2022 Digital Economy Development Report, the digital economy accomplished CNY 45.5 trillion within the year 2021, accounting for 39.8 percent of China’s GDP, with a nominal increase of 16.2 percent over the previous year. Since the 18th Party Congress, the Chinese government has emphasized the vital role played by the real economy, particularly the manufacturing industry, and has put forward a bundle of policies to encourage them to accelerate implementing digital transformation. The eight strategies proposed in “Made in China 2025” include “promoting digital, networked, and intelligent manufacturing”. The “14th Five-Year Plan” of Intelligent Manufacturing Progress also proposes that by 2025, more than 70% of manufacturing enterprises above the scale should realize digital networking. The 20th Party Congress Report put forward “promote the high-end, intelligent and green development of the manufacturing industry”. The essence of intelligent manufacturing is to use the Internet, the Internet of Things, artificial intelligence, big data, and other technologies to carry out product design, production, and manufacturing based on scientific models. The basis is to have corresponding data and intelligent reasoning processes, and the reasoning process needs data in each link of the production process. Therefore, in this sense, digitization is the premise of intelligent manufacturing. In addition, the outbreak of the novel coronavirus in 2020 has had a huge impact on the production and lifestyle of the whole society, which accelerates companies’ application and dissemination of digital technology. A considerable quantity of firms carried out digital transformation and actively explored new growth points for their survival and development.

Digital transformation is a common sense and inevitable trend in the development of Chinese manufacturing enterprises. In the future, the digitalization of China’s manufacturing industry should not only be limited to simple business model innovation but also should carry out deeper penetration and upgrading in the production and operation, technology management, service strategy, and other aspects of the entire enterprise. We will continue to improve our ability to produce high-quality products and services that meet market demand, thus promoting the high-end and intelligent areas of the Chinese manufacturing industry. Objectively speaking, with data added in as a momentous production element, digital transformation has been likely to affect the manufacturing function to some extent [9,10,11], thus causing changes in production efficiency. So, does—and how does—digital transformation impact the TFP of manufacturing firms? This is an issue of both theoretical significance and practical value and is worth exploring and pondering.

Research on the digital economy is extensive, mainly from the basic connotation [12], the appraisal of the digital economy [13], the business value and governance of data sharing [14,15], the sharing economy based on digital platforms [16], and the value creation of corporate innovation activities [17]. With the accelerated fusion of the real economy and the digital economy, our traditional economic paradigm is changing [18,19,20], and more and more scholars are delving into the digital transformation of micro-enterprises. The process of enterprises applying digital technology to collect information, process data, and create a series of auxiliary decisions is enterprise digitalization [21].

The theoretical mechanisms and economic impact of enterprise digital transformation have been studied from different perspectives. Porter and Heppelmann (2014) discussed the enterprise management transformation triggered by digitalization from the aspects of user value and competition [22]. Chen et al., (2022) adopted a multivariate case study approach from the perspective of dynamic capabilities, exploring the data-driven mechanism of effective knowledge management and promoting enterprise innovation transformation [23]. In addition, Chen and Kamal (2016) investigated how digital transformation affects enterprise boundaries in the view of the professional division of labor [24]. In addition, digital transformation is considered to have some influence on enterprise performance [25], corporate innovation performance [26], corporate capital market performance [27], and corporate market power [28]. These studies provide a rich literature base and ideological inspiration for this paper. Existing studies on how digital transformation affects corporate total factor productivity are still general. Some studies focus on its prompting role, while few scholars discuss the moderate issue of enterprise digital transformation. This paper makes some beneficial attempts in the following aspects: (1) Considering that excessive or too rapid digitalization may also inhibit corporate total factor productivity, we put forward an inverted U-shape nonlinear hypothesis and conduct an empirical test. Further analysis finds that the positive promotion effect of digitalization varies with the deepening of digitalization, showing significant threshold characteristics. (2) Given the strategic transformation of the current manufacturing industry from product-oriented to consumer-centered service-oriented, this paper incorporates service-oriented manufacturing into the analysis framework, distinguishes it into embedded service-oriented manufacturing and mixed service-oriented manufacturing according to whether the service business and the core strategy of the enterprise are consistent. Then, this paper discusses its role in moderating the relationship between digitalization and corporate total factor productivity. At last, this paper empirically tests that. (3) Considering the network effect and synergy effect of digitalization, its positive impact on corporate technological innovation and also scale development are discussed, as well as their mechanism role in further enhancing corporate total factor productivity. Since the data from non-listed companies are hard to acquire, our study is based on only listed companies, so whether the conclusions apply to the whole manufacturing industry in China needs further in-depth discussion.

2. Theoretical Analysis and Research Hypotheses

Improving total factor productivity is one internal requirement for a sustainable and healthy economy, which is, in turn, an important embodiment of achieving advanced economic development. Scholars have explored the influencing factors of total factor productivity from multiple perspectives. From a macro perspective, researchers have examined its influencing factors from public infrastructure [29], industrial structure change [30,31], exchange rate changes [32,33], and import and export Trade [34,35]. Macro research mainly explores the influencing factors of TFP at the regional or national level. As the essential units for the macro, corporate behaviors are increasingly valued by researchers. From a micro perspective, studies have revealed the effects of open capital accounts [36], corporate cash holdings [37], and equity liquidity [38] on firm total factor productivity. The above studies, from macro to micro, provide good experience and inspiration for understanding the impact mechanism of total factor productivity.

Existing studies generally agree that, as a general-purpose technology, information technology has a wide range of vertical and horizontal external penetration [39], as well as substitution of other elements such as labor and capital [40] and synergy [41] and plays a rapid substitution effect and synergy effect in corporate production and operation activities, thus significantly promoting their value creation activities. Information is essential to the operation of enterprises, not only facilitating managers to make various effective decisions but also improving the productivity of enterprises [6]. Since Industry 4.0, information from production, R&D (research and development), and sales can be digitized and smartened through the Internet, the Internet of Things, artificial intelligence, and other key technologies to improve corporate production efficiency and quality. Accompanied by the development and application of digital technology, data as a resource in the operation of manufacturing enterprises has become increasingly prominent. Data added as a new element is likely to affect the traditional manufacturing function of micro-enterprises, and so does their total factor productivity [9,42]. The digital transformation of manufacturing enterprises is an important stage for the high fusion of the real and digital economy, which is a combination of manufacturing enterprises and digital technologies such as intelligent manufacturing, industrial Internet, and digital twin to comprehensively reconstruct business processes and business models, thereby improving the networking and digitalization level of their R&D, operation, and services. This innovation and transformation should be reflected in their total factor productivity. Researchers argue that the big data application has stimulated enterprises to invest in data solutions and other aspects, and further discussed its commercial value in the view of data asset use, found that enterprises’ BDA (big data application) assets have an obvious positive effect on the improvement of corporate practice [43]. From the perspective of the industrial ecosystem, different enterprises form modular, symbiotic cooperation and highly autonomous organism by providing complementary products, services, or innovations [44]. Data as the new production factor can optimize the allocation of traditional factors by improving the interaction efficiency between enterprises and further drive the improvement of industrial efficiency, therefore undoubtedly contributing to improving corporate total factor productivity. In addition, digitalization can also enhance enterprise productivity by reducing internal and external transaction costs and promoting specialization [24]. It can be said that corporate digital transformation does effectively promote its total factor productivity to some extent. Considering that digital information overload may occur redundancies, make effective information annihilated and reduce decision-making efficiency, coupled with the huge cost of digital transformation investment, the cases of transformation failure such as GE, P&G, and other companies show that excessive or too fast digital transformation may also be detrimental to the development of enterprises [45]. Accordingly, this paper proposes:

Hypothesis 1.

Moderate digital transformation effectively improves enterprise total factor productivity, while digitalization beyond a certain degree may inhibit the increase of that, i.e., enterprise digital transformation and its total factor productivity show a nonlinear inverted U-shape correlation.

The reason why digital transformation could improve corporate total factor productivity is not only through reducing costs and increasing efficiency across intelligent manufacturing processes but also through the mechanisms of service-oriented manufacturing, technological innovation, and scale effect. Among them, service-oriented manufacturing mainly refers to the embedded service activities closely related to its core business carried out by manufacturing enterprises after they have reached a certain level of development, which is conducive to expanding multi-dimensional value-added and thus enhancing the status of their industrial value chain [46]. Technological innovation is mainly manifested as technological progress achieved by manufacturing enterprises relying on 4D printing, the Internet of things, and other technologies to activate innovation activities and improve innovation efficiency. The scale effect is manifested as the improvement of production efficiency achieved by digital manufacturing firms due to the gradual relaxation of their production scale restrictions.

In terms of digital transformation promoting service-oriented manufacturing, the Internet, Internet of Things, and artificial intelligence can quickly identify and respond to changes in user demand, enhance the interaction between enterprises and users, effectively alleviate information asymmetry, and reduce the uncertainty faced by enterprises in operation to promote corporate supply quality and efficiency. 4D printing and other digital technologies can promote the modular and flexible production of enterprises, enable enterprises to flexibly provide a series of product clusters, and continuously push back customer order separation points in the large-scale manufacturing process, thereby breaking the paradox of service-oriented manufacturing and manufacturing efficiency such as high cost of personalized, customized solutions, and drive the transition of traditional manufacturing enterprises to user value-oriented and service-oriented manufacturing. With the increasing demand for personalized and customized services, more and more manufacturing companies are carrying out product service business. The White Paper of China’s Digital Enterprise points out that improving the ability to serve customers is one of the core goals of enterprise digital transformation and that service-oriented manufacturing shifted from the past product-centric to customer-centric, such as collecting and processing user data through digital technology to better guide R&D activities, optimize the configuration of business processes, carry out precision marketing and provide personalized services in responding to user needs, thus further enhance customer satisfaction and loyalty, and finally improve the value creation capabilities and total factor productivity. With the concept of service relevance, the service business supplied by manufacturing enterprises can be subdivided into embedded services or mixed services. When an enterprise carries out embedded services, which are closely related to its core business, it will significantly promote the enhancement of enterprise value [47]. By integrating product and service resources, embedded services can also extend to the upstream of the value chain around its core products, thus improving customer satisfaction and product-added value and also enhancing the competitive advantage of manufacturing enterprises [46]. The embedded services carried out by manufacturing enterprises mainly include product installation and maintenance testing, solutions, software information solutions, technology development, and services, etc. Mixed services are services that are less relevant to its core business strategy but for entrance to a new industry or more profits, mainly including traditional logistics, retail, and other non-core businesses, as well as hotel catering, real estate development, property management, investment finance, and other service businesses unrelated to the main business. In short, by enhancing service-oriented manufacturing, especially embedded service-oriented manufacturing, digital transformation could promote corporate total factor productivity in the manufacturing industry.

In terms of digital transformation promoting corporate technological innovation, the widespread application of digital technology makes data a new production input, which is a technological advance in itself. Technological upgrading has changed the way of collaboration inside and outside enterprises, highlighting the role of data-driven innovation and promoting the improvement of enterprise productivity. First, enterprise digital transformation provides an open and shared information exchange platform, which can accelerate the flow of knowledge and information elements, lower the endeavor of information searching and matching, form an open also networked innovation ecosystem, promote the quality and efficiency of enterprises in absorbing and transforming knowledge into innovation results [48], and also increase product innovation by promoting the scope economy within the enterprise [7]. Second, under the current consumer-led trend, digital transformation is changing corporate innovation patterns. Traditional manufacturing enterprises are product-centric and lack timely responses to user needs. Digital technologies empower enterprises to establish extensive value connections and real-time interactions between enterprises and users and among enterprises so that enterprises can grasp immediate changes in user demand and market supply, clarify the direction of product innovation, reduce trial and error costs, and further improve innovation efficiency. Third, digitalization can also reconstruct the organizational structure and management innovation of enterprises, improve the organizational and management efficiency of R&D activities, and further promote enterprise innovation. With the application and penetration of digital technology, the communication and collaborative activities between various functional departments in the enterprise are increasing, and the corresponding organizational structure also needs to be adapted to the trend of faster communication, higher efficiency, responsive networking, and flattening, which is conducive to weakening the boundary barriers between different departments of the enterprise, promoting the blend of more innovation resources, then improving corporate innovation capability and productivity. Therefore, digital transformation could enhance corporate total factor productivity of the manufacturing industry with the aid of technological innovation.

In terms of the corporate scale effect brought by digital transformation, manufacturing firms in traditional industrial economies are often constrained by factors such as operation and management capabilities, asset stock, and transaction costs and cannot quickly and flexibly adjust the production scale to the optimal scale, namely, the lowest long-term average cost, to obtain maximum production profits. At present, although the fixed cost investment is high in the early stage of digital transformation, the marginal cost of applying digital technology is very low or even negligible due to its non-competitive, reprogrammable, and mobile characteristics, which can continuously reduce the marginal production cost and average production cost of manufacturing enterprises’ products or services, thus relaxing the limitations on the scale expansion of manufacturing enterprises [49]. Therefore, digital transformation is beneficial for manufacturing companies to acquire economies of scale to further expand the boundaries of their production possibilities. Digital technology also accelerates communication and cooperation between industries and enterprises. In these cases, the transformation of corporate organization, production, and management modes have changed the traditional scale return curve, to a certain extent showing the phenomenon of increasing return to scale, that is, the productivity of enterprises is improved with the expansion of company scale. In addition, that scale effect brought by corporate digital transformation is also reflected in the scope economy in some way. On the one hand, the application of digital technology accelerates information exchange, division, and cooperation of labor within the enterprise, promoting the migration of knowledge, technology, and other resources among different products or services. On the other hand, extensive digital connectivity enables manufacturing enterprises to quickly identify and actively respond to market information and then apply the feedback to their production behavior, thus increasing product diversity and improving production efficiency. Therefore, digital transformation can promote corporate total factor productivity by expanding the scale of enterprises. So, this paper puts forward:

Hypothesis 2.

Digital transformation improves corporate total factor productivity mainly through technological innovation, service-oriented manufacturing—especially the embedded one closely related to its core strategy—and scale effects.

Furthermore, this paper argues digital transformation exerts an asymmetric impact on corporate total factor productivity. Companies carrying out digital transformation expend high fixed-cost investment at the beginning, and they cannot quickly reap the returns from that [45]. With the deep integration of digitalization and manufacturing entities, on the one hand, the increase in collaboration among various departments of the company improves the operational efficiency of the enterprise; on the other hand, the constant application of digital technology may gradually bring about a series of transformation effects such as product and service upgrading, technological innovation capability improvement, and increasing returns on production scale through network effects and Metcalfe effects. If there is an inverted U-shape correlation between them, digital transformation within a certain range tends to exert a positive promotion effect, first rising and then decreasing to zero; that is, it first reaches the inverted U-shape apex, then turns into a negative inhibitory effect. When the degree of digital transformation rises to a particular inflection point, that is, the threshold, its positive effect is the strongest, and beyond that point, it still shows the promoting effect but is obviously weakened. In other words, the influence on corporate total factor productivity brought by digital transformation at different stages may differ to some extent, among which the positive effect shows the law of diminishing marginal effect. Therefore, this paper argues:

Hypothesis 3.

The promotional influence on corporate total factor productivity brought by digital transformation at different stages is asymmetric.

3. Study Design and Data Sources

3.1. Model Settings

Considering the current measurement problems of digital transformation, as well as the data availability from micro-enterprises, this paper uses text analysis to construct indicator variables of enterprise digital transformation. To test Hypothesis 1 and explore whether there is an inverted U-shape nonlinear correlation between digital transformation and corporate total factor productivity, we set the following specific regression model:

$$TF{P}_{ijt}={\alpha }_0+{\alpha }_{1}digita{l}_{ijt}+{\alpha }_{2}digs{q}_{ijt}+{\alpha }_{3}control{s}_{ijt}+{\epsilon }_{ijt}$$

(1)

Among them, ${\alpha }_0,$ ${\alpha }_1,{\alpha }_2{,\alpha }_3$ are the regression coefficients to be estimated; The subscripts i, j, and t indicate the corresponding company, industry, and year; ${TFP}_{ijt}$ indicates corporate total factor productivity; ${digital}_{ijt}$ is the degree of digital transformation while ${digsq}_{ijt}$ is that square term, which is the core explanatory variable of our study; and ${controls}_{ijt}$ are the relevant control variables. If the regression coefficient ${\alpha }_1$ is positive and ${\alpha }_2$ is negative significantly, it indicates that digital transformation exerts an enhancing effect on corporate total factor productivity, and with the deepening of that transition, they have an inverted U-shape relationship.

Service-oriented manufacturing, technological innovation, and scale effects are not only strongly relevant to company performance but also directly affect its total factor productivity. Moreover, to test Hypothesis 2 and examine the indirect impact mechanism between digital transformation and firms’ total factor productivity, we further draw on the research of Liu and Qiu (2016) [50], examine the relationship between possible mechanism variables which are the explained variables and previous core explanatory variable digital, and construct the following model:

$$me{d}_{ijt}={\alpha }_0+{\alpha }_{1}digita{l}_{ijt}+{\alpha }_{2}control{s}_{ijt}+{\epsilon }_{ijt}$$

(2)

where ${med}_{ijt}$ represent the service-oriented manufacturing, scale, and technology innovation of enterprise i in industry j of period t, respectively.

To test Hypothesis 3 and explore whether digital transformation exerts a significant asymmetric nonlinear effect on corporate total factor productivity, this paper uses the panel threshold model of Hansen (1999) [51] and constructs the following model:

$$TF{P}_{ijt}={\alpha }_0+{\alpha }_{1}digita{l}_{ijt}I(digita{l}_{ijt}\le \omega )+{\alpha }_{2}digita{l}_{ijt}I(digita{l}_{ijt}>\omega )+{\alpha }_{3}control{s}_{ijt}+{\epsilon }_{ijt}$$

(3)

Among them, ${digital}_{ijt}$ is both the core explanatory and the threshold variable; ω is the particular threshold to be estimated, which can divide the digitization degree of the study samples into two intervals, and the regression coefficients estimated by the model across different intervals vary significantly; I(.) is an indicator function, when the condition in parentheses is satisfied, the value is 1, and 0 otherwise.

3.2. Variable Selection

3.2.1. Explained Variables

Total factor productivity (TFP). Since both the LP (Levinsohn and Petrin)and OP (Olley and Pakes) methods could solve the challenge of simultaneity of its estimation well, and the LP method has the advantage of using the more accessible intermediate inputs as proxy variables to estimate productivity, this paper follows the current popular LP method to calculate the total factor productivity of enterprises for benchmark regression analysis [52], while adopts the OP and other methods estimated total factor productivity for subsequent robustness tests [53].

3.2.2. Explanatory Variables

Digital transformation (digital). Measurement of corporate digital transformation varies in the current study, and most of that is qualitative analysis or survey research. For example, some scholars designed scales according to the business scope of the company involved in digital technology and constructed enterprise digital index through the data obtained from the survey research [54]. With the rapid development and application of big data technology, scholars have begun to build indicator variables through text analysis methods with the help of Python machine learning [55]. Since the MD&A (management discussion and analysis) part of the company’s annual report contains important strategic operation information of the enterprise, this paper builds digital transformation indicators with text analytics on annual reports. The specific steps are: (1) Crawling the 2013–2020 annual reports of A-share manufacturing companies listed on Giant Tide Information Network through Python, converting them into text files, and extracting the text of their MD&A part, respectively. (2) Drawing on the practice of existing research, by retrieving important policy documents released on government websites in recent years, combined with classic literature on the digital economy, extract keywords related to digitalization. After Python word segmentation processing and manual recognition, the digital dictionary of this paper is finally determined, including the “underlying digital technology” and “digital technology application” pertaining to big data, cloud computing, artificial intelligence, and blockchain. (3) Using Python’s Jieba library to segment the previously sorted and extracted text, count the digitized related word frequencies of the sample companies in sample years according to the determined digital dictionary and then sum them up to obtain digitized word frequencies. This paper defines the digital transformation metric (digital) as the logarithmic value of the total digital keywords’ frequencies appearing in the MD&A section of the sample companies’ annual reports after adding one.

3.2.3. Control Variables

To mitigate the effect of possible confounding variables so as to improve the accuracy of model estimation, the following control variables are selected regarding Zhong et al., (2021) [56]: (1) Return on assets (Roa), reflecting the ability of enterprises in creating value by using existing assets and resources, and the effective use of assets and resources is beneficial to improve corporate total factor productivity, measured by after-tax profit to the annual average balance of total assets. (2) Asset–liability ratio (Lev), calculated by the ratio of total corporate liabilities to its year-end total assets. (3) Age of the company (Age), defined as the logarithm of the particular year minus the firm’s established year plus 1. (4) Cash holding level (Cash), defined as the sum of cash and cash equivalents at the end of the year to corporate total assets. Sufficient cash holding is essential to satisfy the vast expenditure and ensure the smooth implementation of corporate digital transformation. (5) Equity concentration (Share), measured by a company’s holdings owned by its top ten shareholders. (6) Intangible assets ratio of enterprises (Intanr) reflects the non-monetary assets of the company, defined as the rate of intangible assets to year-end total assets. (7) Market power (Marketp), the intensity of external market competition in which the enterprise is included, may affect its business strategy and then affect its operating efficiency. This paper measures it by the Lerner index, specifically measured by the ratio of the gap between a corporate main business income and that cost to its main business income. (8) Enterprise nature (soe), whose value is 1 when an enterprise is state-owned and 0 otherwise.

3.2.4. Mechanism Variables

We follow the practice by Josephson et al., (2016) [46], using the ratio of service income to the main business income in one year as an indicator of the degree of enterprise service-oriented manufacturing. Considering the definition of service relevance in existing studies [47], we further distinguish the service business carried out by manufacturing enterprises into embedded and mixed according to the closeness of the service products or business provided to their main business. Specifically, firstly, according to the national economic industry classification standard [57], analyze the corporate main composition (product) and main composition (industry) of manufacturing enterprises in the Wind database comprehensively to determine whether each enterprise is involved in the service industry or service business in its operation. If so, the annual service revenue is obtained by summing the data of the main business composition (product) and main business composition (industry), respectively, and the total service manufacturing (service) variable is calculated. Otherwise, the company is considered to have not implemented service-oriented manufacturing. Second, according to whether there is strategic alignment between the service industry or service business involved in the enterprise and its core business, if it matches or complements, it is classified as embedded service manufacturing (ser_1), and vice versa, mixed service manufacturing (ser_2). Embedded services include supply chain management, technical services, technology development, environmental protection services, system integration, equipment installation and maintenance testing, end-to-end solutions, main related supporting services, and other services closely related to the core business of the enterprise. Mixed-in services include real estate and property services, hotel services, finance, logistics services, trading business, leasing business, retail circulation, advertising business, tourism and catering, engineering construction, medical services, education and training, warehousing, investment management, and other services unrelated to corporate core business. In the process of manual sorting, for items that are difficult to distinguish which kind they are or if the calculation results based on the two types of data differ greatly, the final index calculation is determined after reviewing the detailed disclosure information such as the company’s business, main business composition, and company strategy in its corresponding annual report. As for the measurement of technological innovation, this paper constructs the index tec from the input–output perspective and defined as the ratio of R&D expenditure investment to the current year number of patents applied for. This paper uses size, the logarithmic value of the annual total assets of listed manufacturing firms, as an indicator of enterprise size.

3.3. Data Sources and Descriptive Statistics

We take all of the 2013 to 2020 years of China’s A-share manufacturing listed firms as the initial samples, and the reason for selecting this period is as follows: Before 2013, China’s digital economy had achieved certain development, but digital technology was still less penetrated and integrated into traditional manufacturing enterprises. Since 2013, as digital technology has been extensively applied, accompanied by the launch of the “digitalization” strategy at the national level, more and more manufacturing enterprises are carrying out digital transformation. The following samples are removed successively from the initial samples: (1) ST, *ST, and PT companies; (2) companies delisted in 2013 to 2020; (3) insolvent companies; and (4) companies with missing main variables. This paper finally obtained a total of 13,195 samples of 2239 listed manufacturing companies. The construction of variables related to digital transformation comes from the text analysis of the MD&A part in the corporate annual report, the financial data comes from the CSMAR (China Stock Market & Accounting Research) database, and other micro-enterprise data comes from the Wind database.

To reduce the probable interference of extreme values, this paper shortens the continuous variables by 1%. The main variables are described in

Table 1. Descriptive statistical analysis.

Variable	Obs	Mean	Std. Dev.	Median	Min	Max
TFP	13,195	8.972	1.015	8.880	5.664	12.549
digital	13,195	1.242	1.256	1.099	0	6.14
Age	13,195	2.901	0.302	2.944	1.609	4.143
Share	13,195	58.161	14.396	58.79	8.78	101.16
Cash	13,195	0.148	0.112	0.117	0.001	0.922
Intanr	13,195	0.046	0.04	0.0382	0	0.677
Lev	13,195	0.394	0.192	0.382	0.008	1.957
Marketp	13,195	0.291	0.174	0.259	−0.862	0.999
Roa	13,195	0.038	0.082	0.382	−3.994	0.786
soe	13,195	0.281	0.449	0	0	1

. The mean value of TFP is 8.972, while its median value, maximum value, minimum value, and standard deviation are 8.880, 12.549, 5.664, and 1.015, showing that there are some differences in TFP among different enterprises, although the overall fluctuation range is not large. The mean value of digital is 1.242, while its median value, maximum value, minimum value, and standard deviation are 1.099, 6.14, 0, and 1.256. That implies the overall degree of digital transformation in the samples is not high, while the extent of digital transformation varies greatly among different companies. As far as control variables, different enterprises also have significant differences in corporate age (Age), asset–liability ratio (Lev), return on assets (Roa), cash holding level (Cash), equity concentration (Share), the proportion of intangible assets (Intanr), market power (Marketp), and other aspects.

4. Empirical Analysis

4.1. Baseline Regression Analysis

The baseline regression analysis is presented in

Table 2. Baseline regression analysis.

	(1)	(2)	(3)	(4)
Variable	TFP	TFP	TFP	TFP
digital	0.2330 ***	0.1878 ***	0.3995 ***	0.3412 ***
	(0.0338)	(0.0291)	(0.0681)	(0.0611)
Age		0.3147 ***		0.3193 ***
		(0.0503)		(0.0504)
Share		0.0739 ***		0.0744 ***
		(0.0151)		(0.0151)
soe		−0.0050		−0.0042
		(0.0079)		(0.0078)
Cash		−0.0043		−0.0047
		(0.0078)		(0.0078)
Intanr		−0.0624 *		−0.0629 *
		(0.0365)		(0.0365)
Lev		0.2492 ***		0.2482 ***
		(0.0272)		(0.0272)
Roa		0.8539 ***		0.8505 ***
		(0.2289)		(0.2288)
Marketp		0.0349		0.0350
		(0.0460)		(0.0459)
digsq			−0.5978 ***	−0.5512 ***
			(0.1734)	(0.1531)
_cons	0.4359 ***	−0.5211 ***	0.4352 ***	−0.5211 ***
	(0.0017)	(0.1762)	(0.0017)	(0.1763)
N	13,195	13,195	13,195	13,195
adj. R2	0.298	0.390	0.300	0.391

Figures in parentheses are t-values, ***, and * represent regression results significant at the 1%, and 10% degrees, respectively.

. Among them, models (1)–(4) control the fixed effects of enterprises, industries, and years. The models’ ${\mathrm{R}}^2$ is between 0.298–0.391, indicating that the models’ explanatory power is well. Model (1) adds only the primary term, while model (2) also adds a bunch of control variables. The results show that, compared to model (1), the regression coefficient of the primary term digital in model (2) decreases slightly after adding the control variables but remains positive at the 1% significance level, implying that corporate total factor productivity does be impacted by the current digital transformation vigorously promoted by manufacturing enterprises, which increases with the deepening of digital transformation. This study argues that only considering the primary linear impact from digital transformation to corporate total factor productivity is incomplete, so the quadratic term digsq is added to models (3) and (4). The results in Table 2 show that regardless of whether control variables are added, the coefficient of the first term digital is positive at a 1% significant level, while that quadratic term digsq is oppositely negative at a 1% significant level, manifesting that the digital paradox does exist and moderate digital transformation can indeed play a positive role on corporate total factor productivity. Moreover, there is a certain nonlinear inverted U-shape trend between them with the deepening of digital transformation. Therefore, our previous theoretical analysis is reasonable, and Hypothesis 1 is verified. After substituting the coefficient estimated in model (4) into the model for derivation, the estimation of digital transformation corresponding to that apex at the inverted U-shape curve is about 5.130, indicating that when corporate digital transformation is about 5.130 in our research samples. In other words, when the measured frequency of digitalization-related words is about 168, the corresponding corporate total factor productivity reaches its maximum. Below this level, digital transformation exerts a positive influence on promoting corporate total factor productivity. However, after exceeding that level, the positive influence turns into a negative. According to the previous descriptive statistics, the average value of digital transformation in the samples is 1.242, and the median value is 1.099, which is far less than 5.130, coupled with the maximum value of the samples is 6.14, showing that the digital transformation extent of listed manufacturing firms is relatively low, most of them are still in the rising stage of the inverted U-shape curve, and vigorously impelling the digital transformation of manufacturing firms in China at the current stage is conducive to elevate their total factor productivity.

From models (2) and (4), the coefficients of the control variables Age, Share, Lev, and Roa are positive, which all go through the test at a 1% significant level, manifesting that the establishment of an enterprise, the shareholding ratio of the top ten shareholders, the asset–liability situation, and the assets income all have a certain positive impact on the total factor productivity, while Intanr’s coefficient is negative at 10% level, showing that higher proportion of intangible assets in enterprises is unfavorable to corporate total factor productivity. Other variables such as enterprise nature (soe), cash holding level (Cash), and market power (Marketp) are not significant, implying their unobvious impact on the total factor productivity of enterprises.

4.2. Robustness Test

This paper examines the robustness from the following perspectives to verify the reliability of previous benchmark results.

• Change the measurement of TFP.

As an important measure of enterprise productivity, the current estimation methods of total factor productivity mainly pertain to LP, OP, OLS (ordinary least squares), FE (fixed effects), and GMM (generalized method of moment). To mitigate the interference from different measurement methods on the test results, this study uses the total factor productivity estimated by OP, OLS, FE, and GMM methods as the replacement index of TFP, respectively, expressed as tfp_op, tfp_gmm, tfp_ols, and tfp_fe, and regression analysis including the primary term, squared term, and control variables is conducted again. From

Table 3. Robustness test: change the measurement of TFP.

	(1)	(2)	(3)	(4)
Variable	tfp_op	tfp_gmm	tfp_ols	tfp_fe
digital	0.2569 ***	0.1684 ***	0.3881 ***	0.4036 ***
	(0.0547)	(0.0574)	(0.0664)	(0.0685)
digsq	−0.4801 ***	−0.2830 *	−0.6420 ***	−0.6662 ***
	(0.1419)	(0.1539)	(0.1706)	(0.1768)
Age	0.2282 ***	0.0940 **	0.3861 ***	0.4063 ***
	(0.0452)	(0.0445)	(0.0567)	(0.0589)
Share	0.0634 ***	0.0331 **	0.0884 ***	0.0921 ***
	(0.0140)	(0.0133)	(0.0174)	(0.0181)
soe	−0.0069	−0.0065	−0.0041	−0.0039
	(0.0075)	(0.0060)	(0.0090)	(0.0094)
Cash	−0.0046	0.0384 ***	−0.0227 ***	−0.0266 ***
	(0.0071)	(0.0083)	(0.0085)	(0.0088)
Intanr	−0.0636 *	−0.1147 ***	−0.0415	−0.0367
	(0.0367)	(0.0344)	(0.0424)	(0.0441)
Lev	0.1983 ***	0.0800 ***	0.3035 ***	0.3187 ***
	(0.0251)	(0.0244)	(0.0295)	(0.0303)
Roa	0.7591 ***	0.7033 ***	0.8844 ***	0.8976 ***
	(0.2243)	(0.2261)	(0.2298)	(0.2304)
Marketp	0.0162	0.0426	0.0262	0.0254
	(0.0447)	(0.0452)	(0.0477)	(0.0484)
_cons	−0.5130 ***	−1.0433 ***	−0.3002 *	−0.2526
	(0.1722)	(0.1735)	(0.1776)	(0.1784)
N	13,195	13,195	13,195	13,195
adj. R2	0.404	0.194	0.436	0.440

***, **, and * represent regression results significant at the 1%, 5%, and 10% degrees, respectively.

, we can see that the coefficients of digital are positive and go through the 1% statistical significance test, while the coefficients of digsq are negative by at least 10% significance, which supports the benchmark regression results. In a word, after changing the measurement method of total factor productivity, corporate digital transformation still tends to significantly promote their total factor productivity, and with the deepening of digital transformation, the two still show an inverted U-shape correlation, which supports the robustness of the benchmark results.

As for control variables, the coefficients of soe and Marketp are still insignificant, while that of Age, Share, Lev, and Roa are still positive, at least at 5% significance. However, unlike the benchmark analysis, in models (3) and (4), the Intanr’s coefficient is still negative but no longer significant. The coefficient for Cash in model (1) is not significant, while positive in model (2) and negative in models (3) and (4) significantly. Taken together, the results are generally robust after replacing the explained variables in the benchmark analysis with total factor productivity estimated by different methods.

2.

Change the measurement of digital transformation.

First of all, considering that the text words of the MD&A part in the annual reports of each enterprise vary to some extent, the ratio of total digital-related words number to its whole words of MD&A part in the corresponding annual report is added by 1 of the sample enterprise and then take that logarithm to measure the digitalization degree of the enterprise, which is expressed by digR and the square term is expressed by digRsq. Secondly, this paper forms a digital dictionary from five dimensions: big data, cloud computing, artificial intelligence, and blockchain, namely “underlying digital technology” and “digital technology application”, in which there are certain differences in the number of relevant words in each sub-dimension and may affect the measurement error of digital dictionary for different enterprises in different dimensions of digitalization. Therefore, this paper first counts the words frequency of digitization in each dimension to form a digital dimension indicator, then performs annual dimensionless processing on the five-dimensional indicators, and finally adds 1 to take the logarithm to construct the enterprise digital transformation index after adding together the standardized dimensional indicators, which is recorded as dig_n, and the square term is dig_nsq.

Table 4. Robustness test: change the measurement of digital transformation.

	(1)	(2)
Variable	TFP	TFP
digR	0.2235 ***
	(0.0558)
digRsq	−0.3936 ***
	(0.1220)
dig_n		0.0754 ***
		(0.0168)
dig_nsq		−0.0549 ***
		(0.0180)
Age	0.3312 ***	0.3270 ***
	(0.0510)	(0.0507)
Share	0.0731 ***	0.0723 ***
	(0.0151)	(0.0151)
soe	−0.0043	−0.0045
	(0.0081)	(0.0078)
Cash	−0.0051	−0.0061
	(0.0080)	(0.0079)
Intanr	−0.0645 *	−0.0654 *
	(0.0365)	(0.0364)
Lev	0.2526 ***	0.2518 ***
	(0.0272)	(0.0271)
Roa	0.8516 ***	0.8514 ***
	(0.2321)	(0.2289)
Marketp	0.0362	0.0325
	(0.0466)	(0.0460)
_cons	−0.5272 ***	−0.5234 ***
	(0.1786)	(0.1764)
N	13,145	13,195
adj. R2	0.388	0.388

*** and * represent regression results significant at the 1% and 10% degrees, respectively.

reports the empirical results. The regression coefficients of digR and dig_n, the two replacement indicators of enterprise digital transformation, are both positive at a 1% significance, and that of the corresponding square term digRsq and dig_nsq are negative at a 1% significance, which are consistent with Hypothesis 1 in our previous theoretical analysis and support the robustness of the previous baseline regression conclusions.

As for control variables, the coefficients of Age, Share, Lev, and Roa are still positive at a 1% significance level, while that of Intanr is still negative at a 10% level. The coefficients of other variables (soe), cash holding level (Cash), and market power (Marketp) are not significant. These things considered, the coefficient values of each control variable are close to those in the benchmark analysis, indicating that Hypothesis 1 still holds after changing the measurement method of digital transformation.

3.

Endogenous treatment.

Although the previous study takes into account different measurements of the core variables, the research may still have endogenous problems. On the one hand, promoting digital transformation could improve the corporate business environment and promote enterprise reform and innovation, thus enhancing their total factor productivity. On the other hand, companies with higher total factor productivity may also have more strength and willingness to carry out high-level digitalization to further improve their market position and profitability. Since the digital influence transformation exerted on corporate total factor productivity may objectively have a certain lag, we adopt the lag period of digital for regression analysis to mitigate the potential impact of this reverse causal relationship on the research conclusions.

Table 5. Baseline regression results using a one-period lagged digital.

	(1)	(2)
Variable	TFP	TFP
L.digital	0.1556 ***	0.2392 ***
	(0.0282)	(0.0615)
L.digsq		−0.3029 *
		(0.1726)
Age	0.3001 ***	0.3038 ***
	(0.0596)	(0.0598)
Share	0.0840 ***	0.0845 ***
	(0.0175)	(0.0175)
soe	−0.0082	−0.0078
	(0.0074)	(0.0073)
Cash	−0.0030	−0.0032
	(0.0089)	(0.0089)
Intanr	−0.0517	−0.0519
	(0.0389)	(0.0389)
Lev	0.2321 ***	0.2316 ***
	(0.0316)	(0.0316)
Roa	0.7876 ***	0.7869 ***
	(0.2320)	(0.2322)
Marketp	0.0590	0.0597
	(0.0494)	(0.0494)
_cons	−0.4713 ***	−0.4734 ***
	(0.1787)	(0.1789)
N	10,859	10,859
adj. R2	0.374	0.375

*** and * represent regression results significant at the 1% and 10% degrees, respectively.

reports the regression results that contain only the primary term L.digital, both primary term L.digital and quadratic term L.digsq, respectively. It can be found that model (1) performs re-regression merely on the lagged core explanatory variable L.digital, which coefficient is still positive at the 1% level of statistical significance while slightly smaller than in previous baseline regression. Except for Intanr is no longer significant, other control variables remain the same as in previous results. In model (2), the coefficient of L.digital is significantly positive, while that of L.digsq remains negative, with the significance level dropping to 10%. In addition, except for Intanr becomes negative and no longer significant, other control variables are in line with previous benchmark analysis. In summary, after analyzing the hysteresis core explanatory variables, the obtained results still support the robustness of our baseline analysis and verify hypothesis 1 convincingly.

4.

Different sampling methods.

The digital transformation variables depicted through text analysis in this paper can comprehensively reflect the specific corporate digital transformation operation, but as an important carrier of public disclosure information of listed companies, the specific disclosure content of annual reports is affected by the public willingness of enterprises to a certain extent, and there may be certain differences among different enterprises. For example, some enterprises may exaggerate the implementation of digital transformation, while others may be more cautious in describing the specific transformation operation. To decrease the sample selection bias on test results, we first use simple random sampling to select 50% of the total samples in this paper and re-test. Second, considering that samples with too high or too low total factor productivity may have large sample differences from the overall, the samples of the top 20% and the bottom 20% of total factor productivity are excluded from testing, respectively.

Table 6. Different sampling methods.

	(1) 50% Random Sampling	(2) Drop Top 20% of the tfp	(3) Drop End 20% of the tfp
Variable	TFP	TFP	TFP
digital	0.3301 ***	0.2650 ***	0.3162 ***
	(0.0673)	(0.0624)	(0.0538)
digsq	−0.4921 ***	−0.3852 **	−0.4506 ***
	(0.1678)	(0.1501)	(0.1639)
Age	0.2115 ***	0.2732 ***	0.2497 ***
	(0.0564)	(0.0528)	(0.0494)
Share	0.0502 ***	0.0386 ***	0.0480 ***
	(0.0148)	(0.0144)	(0.0138)
soe	0.0088	0.0082 *	−0.0027
	(0.0065)	(0.0050)	(0.0052)
Cash	0.0079	−0.0083	0.0090
	(0.0099)	(0.0080)	(0.0086)
Intanr	−0.0845 ***	−0.0434	−0.0381
	(0.0297)	(0.0308)	(0.0271)
Lev	0.2738 ***	0.2057 ***	0.1870 ***
	(0.0241)	(0.0243)	(0.0236)
Roa	1.0757 ***	0.6594 ***	0.7332 ***
	(0.1321)	(0.2026)	(0.2124)
Marketp	0.0079	0.1148 ***	−0.0334
	(0.0433)	(0.0401)	(0.0495)
_cons	−0.6888 ***	−0.1547	−0.3044 *
	(0.1150)	(0.1581)	(0.1598)
N	6598	10,556	10,556
adj. R2	0.454	0.374	0.420

***, **, and * represent regression results significant at the 1%, 5%, and 10% degrees, respectively.

reports the test results for different sampling methods. In models (1)–(3), the coefficients of core explanatory variables are significantly positive for the primary term digital and significantly negative for the quadratic term digsq, and the coefficient values are not much different from the benchmark regression results. For control variables, coefficients are in line with benchmark results, except for intangible assets (Intanr), firm nature (soe), and market power (Marketp), whose coefficients are slightly different from the benchmark results. It shows that digital transformation could indeed promote corporate total factor productivity. Furthermore, there is an inverted U-shape relationship between them. Therefore, different sampling methods do not affect our basic conclusions.

4.3. Heterogeneity Testing

This study probes the effect that the digital transformation of manufacturing firms exerted on their total factor productivity, using our whole manufacturing samples in the previous test. The previous research verifies that digital transformation could indeed promote corporate total factor productivity; moreover, with the deepening of transformation, the two have a nonlinear inverted U-shape relationship, which passes several robustness tests. It is worth noting that enterprises with different property rights have great differences in financial conditions, policy support, and risk bearing, so the effect of digital transformation in different firms may vary, which requires considering the heterogeneity of the digital impact of enterprises with different property rights in the manufacturing industry. Combined with the urgent requirements of the current green and low-carbon development, considering that some enterprises have been subject to key pollution monitoring by relevant government departments, exploring the impact of digital transformation on total factor productivity by distinguishing whether they are key pollution monitoring enterprises is of immense practical guiding significance. These things considered, since we propose that digital transformation could improve corporate total factor productivity through that scale effect, we should specify the company size to examine whether larger firms show a better transformation effect. Our classification is based on the Statistical Standards for the Division of Large, Small, and Micro-Enterprises published by the National Bureau of Statistics. To be specific, firms with more than 1000 employees are divided into large firm groups, otherwise into small firm groups. So, this paper divides the whole samples by property rights attribute, pollution attributes, and company size and conducts group tests.

Table 7. Heterogeneity test based on the firm type.

	(1) soe = 1	(2) soe = 0	(3) Key pmu	(4) Non-Key pmu	(5) Large Firms	(6) Small Firms
Variable	TFP	TFP	TFP	TFP	TFP	TFP
digital	0.3489 ***	0.2527 ***	0.1884 *	0.2399 ***	0.2834 ***	0.1147
	(0.0825)	(0.0423)	(0.0965)	(0.0450)	(0.0536)	(0.1531)
digsq	−0.4968 **	−0.3533 ***	−0.0756	−0.2387 *	−0.4480 ***	0.0225
	(0.2501)	(0.1353)	(0.3407)	(0.1408)	(0.1637)	(0.4669)
Age	0.1798 ***	0.2792 ***	0.3223 ***	0.3157 ***	0.2640 ***	0.4175 ***
	(0.0575)	(0.0301)	(0.0738)	(0.0341)	(0.0450)	(0.1348)
Share	0.1152 ***	0.0423 ***	0.0897 ***	0.0682 ***	0.0344 ***	0.0730 **
	(0.0108)	(0.0079)	(0.0123)	(0.0081)	(0.0115)	(0.0313)
soe			−0.0057	0.0022	−0.0107 *	−0.0203
			(0.0061)	(0.0051)	(0.0060)	(0.0184)
Cash	−0.0020	0.0002	0.0168	−0.0052	0.0052	−0.0316 *
	(0.0111)	(0.0058)	(0.0113)	(0.0063)	(0.0110)	(0.0176)
Intanr	0.0312	−0.0427 ***	0.0329	−0.0756 ***	−0.0150	−0.1425 ***
	(0.0278)	(0.0136)	(0.0243)	(0.0148)	(0.0272)	(0.0488)
Lev	0.2818 ***	0.2391 ***	0.0996 ***	0.2701 ***	0.1885 ***	0.2364 ***
	(0.0184)	(0.0110)	(0.0196)	(0.0119)	(0.0181)	(0.0391)
Roa	1.5241 ***	0.7217 ***	0.5915 ***	0.7302 ***	1.1255 ***	0.7922 ***
	(0.0781)	(0.0323)	(0.0679)	(0.0365)	(0.0698)	(0.1131)
Marketp	−0.0478 *	0.0728 ***	0.1418 ***	−0.0323 *	−0.2291 ***	0.1395 ***
	(0.0247)	(0.0158)	(0.0274)	(0.0176)	(0.0262)	(0.0426)
_cons	−0.8400 ***	−0.4999 ***	−0.3029 ***	−0.2976 ***	−0.5443 ***	−0.6433 ***
	(0.0764)	(0.0587)	(0.0641)	(0.0504)	(0.0679)	(0.1103)
N	3709	9486	2888	8987	3313	902
adj. R2	0.308	0.300	0.120	0.224	0.401	0.162

***, **, and * represent regression results significant at the 1%, 5%, and 10% degrees, respectively.

describes the results. In models (1)–(2), the coefficients of digital are 0.3489 and 0.2527, both positive at a 1% level significantly, while the quadratic coefficients are −0.4968 and −0.3533 on at least a 5% significant level. However, the absolute values of the coefficients in the state-owned enterprise group are bigger than that in the opposite group. After calculation, the degree of digitization at the inverted U-shape vertex of the two models is 5.278 and 5.299, which are similar to each other and slightly higher than the benchmark analysis value of 5.130, indicating that although digital transformation exerts an obvious promoting effect on corporate total factor productivity in the manufacturing industry with different ownerships, state-owned enterprises seem to own much more strength and determination in promoting digital transformation under their natural reputation, policy, capital, and other advantages, and have achieved relatively better transformation results. Of course, while the extent of digital transformation is too high, the two in both groups show a similar inverted U-shape nonlinear relationship. Among control variables, the coefficient of Intanr in (1) is insignificant while significantly negative in (2), indicating that state-owned enterprises have relatively sufficient funds and stronger strength in implementing digital transformation, and their total factor productivity is not greatly affected by the proportion of intangible assets. In contrast, many non-state-owned enterprises often face more financial problems. Perhaps the excessive proportion of intangible assets in them inhibits their total factor productivity from increasing. The coefficient of market power (Marketp) is significantly negative in (1) while significantly positive in (2), indicating that due to the natural advantages, state-owned enterprises with too large market power may lack innovation motivation, which restrains their total factor productivity improvement, while non-state-owned enterprises often have more sense of distress after obtaining certain market power, and will continue to increase innovation investment to maintain market power they achieved, thereby their market power exerts active impact on their total factor productivity.

In models (3)–(4), the regression coefficients of digital in both groups are 0.1884 and 0.2399, all passing the statistical significance test of at least 10%, while the coefficients of the quadratic terms digsq are negative but not significant in (3), showing that promoting digital transformation does significantly improve the total factor productivity of both types of firms. The relationship between digital transformation and total factor productivity in key pollution monitoring unit (pmu) enterprises is more linear and positive, while in non-key pollution monitoring units, a certain inverted U-shape nonlinear relationship appears, indicating that key pollution monitoring units are limited to the influence of the current green development situation as they consume higher energy and emit more pollutants. The empowering role of digital transformation has not yet been fully exerted, and it is necessary to continue to explore and promote the digitalization process, especially for key polluting enterprises in the future. Among the control variables, Intanr is insignificant in (3) and significantly negative in (4), indicating that it has little impact on key pollution monitoring units but has a suppressive impact on general non-key pollution monitoring units. Marketp is positive in (3) and negative in (4) significantly, indicating that since key pollution monitoring enterprises are more supervised if they have certain market forces—that may help them improve the efficiency of pollution control and promote their total factor productivity, while the expansion of market power of general non-key pollution monitoring enterprises may not conducive to market competition and inhibit their total factor productivity.

As for the effect of company size, the coefficients of digital and digsq are 0.2834 and −0.4480 on at least a 1% significant level in model (5), which are similar to the baseline results, demonstrating that digital transformation in large firms exerts an obvious promoting effect on their total factor productivity in the manufacturing industry and the two show a similar inverted U-shape nonlinear relationship with the deepening of transformation. In contrast, the corresponding coefficients in model (6) are no longer significant, implying that digital transformation in small firms does not play the expected role in enhancing their total factor productivity. According to our previous analysis, firms implementing digital transformation cost high investment and endure plenty of risks, and perhaps small firms need more time and endeavor to benefit from that. Therefore, the effect digital transformation exerts on corporate total factor productivity has a significant difference between large and small firms; that is, company size is an important factor while implementing digital transformation. Among the control variables, Cash and Intanr are insignificant in (5) and significantly negative in (6), indicating that small firms have more fund shortage problems compared with large firms. Marketp is negative in (5) and positive in (6) significantly, indicating that since small firms often have various inadequate resources, if they have certain market forces, that may help them improve their reputation and capabilities to acquire more advantages to improve their performance and promote their total factor productivity, while the expansion of market power of large firms may not conducive to market competition and inhibit their total factor productivity.

4.4. Sensitivity Analysis

This study uses observational data to evaluate the effect on manufacturing total factor productivity exerted by its digital transformation. We conclude that moderate digitalization improves corporate total factor productivity, while digitalization beyond a certain level instead reduces it. To evaluate the influence of objective unmeasured potential confounders on the study results, sensitivity analysis is conducted in this paper. Referring to the regression results of model (2) in our previous baseline analysis, the point estimate between digital transformation and total factor productivity is 0.1878, the 95% confidence interval is [0.1307, 0.2449], and the sensitivity analysis results are described in

Table 8. Sensitivity analysis of confounding factors.

	(1) Zero Value Effect	(2) RR from 0.1879 to 0.2879	(3) RR from 0.1879 to 0.3879
Confidence Interval	[0.1307, 0.2449]	[0.1307, 0.2449]	[0.1307, 0.2449]
E-value (point estimate)	10.124	2.436	3.548
E-value (CI)	7.632	1.629	2.545

below. The point estimating E-value for the zero-value effect is 10.124, and the E-value for the upper CI (confidence interval) limit is 7.632 in column (1), indicating that the minimum correlation strength between unmeasured confounders and digitization and total factor productivity is 10.124 or 7.632 for the transfer of point estimates or CI to zero, while weaker confounders fail to do so. Column (2) shows that if the RR (risk ratio) estimate is converted to 0.2879, the E-value of the required point estimate is 2.436, and the E-value of the CI upper limit is 1.629. This indicates that an unmeasured confounder related to digitalization and total factor productivity could move RR from 0.18979 to 0.2879 if the minimum correlation strength reaches 2.436 times, while weaker confounders could not.

Similarly, to change the upper confidence limit from 0.2449 to 0.2879, unmeasured confounders with a strength of 1.629 associated with digitization and total factor productivity can do, while weaker confounders cannot. Column (3) shows if the estimate is converted to 0.3879, the E-value for the point estimate is 3.548, and the E-value for the CI upper limit is 2.545. This indicates that when a confounder associated with digitalization and total factor productivity shifts the observed risk ratio RR estimate from 0.1879 to 0.3879, the risk ratio RR of the confounder is at least 3.548. Similarly, an unmeasured confounder-related risk ratio RR of 2.545 for digitalization and total factor productivity can change the upper confidence limit from 0.2449 to 0.3879, while weaker confounders cannot. The results of comprehensive sensitivity analysis show that the benchmark regression in this paper is relatively robust.

5. Mechanism, Path, and Characteristics

5.1. Mechanism Path

The above describes the effect on manufacturing total factor productivity exerted by its digital transformation as a whole, which provides abundant empirical support to deeply understand the economic consequences of digital transformation. As the previous theoretical analysis described, the manufacturing digital transformation works through three channels: improving the level of service-oriented manufacturing, improving technological innovation capabilities, and expanding their own enterprises’ scale. According to the model (3) constructed above, this paper firstly uses service manufacturing (service), embedded service manufacturing (ser_1), and mixed service manufacturing (ser_2) of manufacturing enterprises as explained variables, and digital transformation (digital) as core explanatory variables for regression analysis. The results of this mechanism test are displayed in

Table 9. Test of the mechanism of action of service-oriented manufacturing.

	(1)	(2)	(3)	(4)	(5)	(6)
Variable	Service	Service	ser_1	ser_1	ser_2	ser_2
digital	0.3823 ***	0.3150 ***	0.6102 ***	0.5519 ***	0.0271	−0.0089
	(0.0969)	(0.0799)	(0.1424)	(0.1279)	(0.0617)	(0.0631)
Age		0.2711 *		0.5026 **		−0.0252
		(0.1484)		(0.2003)		(0.1170)
Share		−0.0243		0.0096		−0.0325
		(0.0287)		(0.0285)		(0.0256)
soe		0.0159		0.0217		0.0035
		(0.0145)		(0.0235)		(0.0117)
Cash		0.0160		−0.0124		0.0252
		(0.0215)		(0.0289)		(0.0187)
Intanr		−0.1690 ***		−0.1492 **		−0.0887
		(0.0653)		(0.0752)		(0.0583)
Lev		0.1625 ***		0.0616		0.1373 ***
		(0.0467)		(0.0476)		(0.0422)
Roa		0.4322 ***		0.0888		0.4129 ***
		(0.1506)		(0.1323)		(0.1460)
Marketp		−0.3894 ***		−0.1285 *		−0.3411 ***
		(0.1039)		(0.0743)		(0.0995)
_cons	0.2359 ***	0.0492	0.0218 ***	−0.1362	0.2423 ***	0.1400
	(0.0038)	(0.1534)	(0.0039)	(0.1580)	(0.0033)	(0.1462)
N	4215	4215	4215	4215	4215	4215
adj. R2	0.062	0.122	0.079	0.110	0.013	0.057

The reduction in the sample is due to the selection of companies that disclosed their service business in at least three consecutive years with complete data. ***, **, and * represent regression results significant at the 1%, 5%, and 10% degrees, respectively.

.

It can be seen that when taking service and ser_1 as explained variables, the coefficients of digital are 0.3823 and 0.3150, 0.6102 and 0.5519, significantly positive at a 1% level no matter whether we add control variables, while the regression coefficients corresponding to ser_2 are not significant. The estimated results in Table 9 approve of our previous theoretical analysis, namely, digital transformation can not only improve the service manufacturing level in manufacturing enterprises but also play an enhancing role in corporate total factor productivity through service manufacturing. After further distinguishing the service manufacturing into embedded and mixed, digital transformation mainly plays the mechanism role through the embedded service manufacturing business closely related to the core business of the enterprise, while the mixed service manufacturing business with low correlation with the core business of the enterprise does not play a significant mechanism impact. Accordingly, the mechanism of action of service-oriented manufacturing in Hypothesis 2 has been verified.

In

Table 10. Test of the mechanism of action of technological innovation and scale effect.

	(1)	(2)	(3)	(4)
Variable	tec	tec	Size	Size
digital	0.3215 **	0.3199 **	0.0262 **	0.0316 ***
	(0.1484)	(0.1470)	(0.0111)	(0.0113)
Age		0.0801		0.0059
		(0.0829)		(0.0199)
Share		0.0290		0.0194 ***
		(0.0178)		(0.0029)
soe		0.0013		−0.0014
		(0.0039)		(0.0009)
Cash		0.0106		0.0032 **
		(0.0114)		(0.0015)
Intanr		0.0356		0.0085 **
		(0.0435)		(0.0040)
Lev		0.0395		0.0127 ***
		(0.0249)		(0.0030)
Roa		−0.0818		0.0190 **
		(0.0893)		(0.0075)
Marketp		0.0835 *		0.0001
		(0.0463)		(0.0038)
_cons	0.0097 ***	−0.0539	0.0054 ***	−0.0329 ***
	(0.0026)	(0.0754)	(0.0005)	(0.0127)
N	2692	2692	13,195	13,195
adj. R2	0.037	0.037	0.086	0.111

The reduction in the sample is due to missing data that limit the sample of tec indicators that can be used to calculate input–output ratios. ***, **, and * represent regression results significant at the 1%, 5%, and 10% degrees, respectively.

, this paper turns to the mechanism identification tests for technological innovation (tec) and scale effects (size). In models (1) and (2), the regression coefficients of digital are 0.3215 and 0.3199, both significantly positive at a 5% level, indicating that manufacturing digital transformation has promoted R&D innovation capabilities and efficiency to some extent. On the one hand, enterprises need a lot of resource investment to promote digital transformation. In these cases, their R&D investments also increase. On the other hand, digital transformation contributes to forming a more efficient and transparent cooperation and communication ecosystem for enterprises’ R&D and innovation, thus promoting enterprise innovation input transformed into invention patents and other knowledge achievements. Further, digital transformation improves the efficiency and height of technological innovation, which in turn enhances their total factor productivity. The results in model (3) and model (4) also show that digital transformation is in favor of expanding the operating scale of manufacturing firms, which can generate the effects of scale cost and scope economy, then improves corporate total factor productivity. Accordingly, the mechanisms of technological innovation and scale effect in Hypothesis 2 are verified.

5.2. Asymmetric Features

To verify Hypothesis 3, this paper draws on the research of Hansen (1999) [51] to conduct empirical tests using the panel threshold model. Before our formal panel threshold regression estimation, the existence of the panel threshold must be judged, that is, whether there is a threshold interval between corporate digital transformation and total factor productivity, and if so, determine how many thresholds.

Table 11. Number of threshold tests.

Original Hypothesis	Threshold Value	Interval	Prob	Fstat
Single Threshold	4.094	[3.961, 4.081]	0.01 **	25.19
Double Threshold	3.135	[3.091, 3.178]	0.63	4.48

** represent regression results significant at the 5% degrees.

shows the tests of one or two threshold hypotheses for digital transformation as threshold variables after 300 times repeated sampling by the self-help method. Table 11 demonstrates that the threshold variable digital accepts a single threshold hypothesis at a 5% significance degree, in which F statistic and self-sampling p-value are 25.19 and 0.01, rejecting the hypothesis of no threshold and double threshold. The specific threshold variable value shows the 95% confidence interval of the single threshold estimate of 4.094 is [3.961, 4.081], indicating that when digital transformation reaches the threshold value of about 4.094 (the frequency of digitalization-related words reaches about 59 times), the improvement effect on total factor productivity exerted by digital transformation may have obvious interval difference.

This paper divides our sample into two groups by year according to this single threshold and shows the number of enterprises in each group from 2013 to 2020 in

Table 12. Number of enterprises in each digitalization interval 2013–2020 Units: pcs.

Group	2013	2014	2015	2016	2017	2018	2019	2020
digital ≤ 4.094	1252	1254	1332	1467	1595	1917	1973	2084
digital > 4.094	6	8	18	32	55	59	66	77
total	1258	1262	1350	1499	1650	1976	2039	2161

. It can be seen that, on the whole, the number of enterprises with different degrees of digital transformation in both groups tends to increase year by year, but the number of enterprises in the group with higher degrees is much smaller than that in the lower group. This is consistent with the average digital value of only 1.242 in the previous descriptive statistics, indicating that more and more manufacturing firms are launching digital transformation, which is a general trend; however, at present, the degree of digitalization in China’s manufacturing firms is still low, and vast of them have a lot of room for improvement yet.

To focus on whether digital transformation has an asymmetric enhancing effect on corporate total factor productivity, this paper makes an estimate based on the above single threshold regression model. As

Table 13. Threshold regression estimation: single threshold model.

Variable	Coefficient Estimates	OLS Standard Error	Robust Standard Error
Age	0.526 ***	0.0093	0.0188
Share	0.0912 ***	0.0072	0.0161
soe	−0.0076	0.0042	0.0094
Cash	−0.0129 *	0.0061	0.009
Intanr	−0.0557 ***	0.0145	0.0371
Lev	0.271 ***	0.0117	0.0257
Roa	1.133 ***	0.0448	0.1212
Marketp	−0.0477 **	0.0162	0.0448
digital.I (digital ≤ 4.094)	0.326 ***	0.0383	0.0623
digital.I (digital > 4.094)	0.165 ***	0.0328	0.029

***, **, and * represent regression results significant at the 1%, 5%, and 10% degrees, respectively.

shows, the coefficient of digital.I (digital ≤ 4.094) is 0.326, while that of digital.I (digital > 4.094) is 0.165, and both pass the 1% significant tests. Obviously, the positive promotion effect of digital transformation on total factor productivity in the range of digital ≤ 4.094 is nearly twice as much as that on the interval of digital > 4.094. This demonstrates a threshold effect convincingly. Therefore, the influence digital transformation on both sides of the threshold value of 4.094 exerted on enterprise total factor productivity is asymmetric, which verifies that Hypothesis 3 holds. This implies that when the digital transformation reaches the threshold point of 4.094, its positive influence on total factor productivity reaches its maximum. However, after beyond the threshold point, this positive impact gradually weakens, showing the law of marginal diminishing. Combined with the inverted U-shape vertex of 5.130 in the previous benchmark regression analysis, it is interesting that as corporate digital transformation moves forward constantly, its positive influence on total factor productivity is the greatest when it reaches the threshold point and then gradually decreases until the inverted U-shape vertex becomes 0, and finally turns into negative suppression.

6. Discussion and Conclusions

In the current society, China’s economy has entered a fresh, high-quality phase in which the manufacturing industry occupies a vital strategic position because of its characteristics of rapid technological progress, economies of scale, and significant technology diffusion effects. With the popular development of the digital economy, deeply discussing the influence manufacturing digital transformation exerts on its total factor productivity is conducive to fully understanding its influence on promoting the development of manufacturing firms so as to further guide enterprises’ investment decisions in digital capabilities and form important policy insights to guide their high-quality development. On the basis of existing literature, this paper theoretically analyses and empirically examines the influence that digital transformation exerts on corporate total factor productivity and internal mechanisms from the perspective of micro-enterprises, which uses 2013 to 2020 of China’s listed A-share manufacturing firms and constructs the digital transformation index from their crawled annual reports by text analysis method.

The study finds that manufacturing digital transformation does obviously enhance their total factor productivity; moreover, as digital transformation deepens, the promotion role of digitalization gradually weakens after reaching the inflection point of maximum value until it turns to inhibition, that is, the two have a nonlinear inverted U-shape relationship, which stands up several robustness tests and sensitivity analysis. We propose three mechanisms through which digital transformation can exert influence on corporate total factor productivity. First, it can promote the level of service-oriented manufacturing, especially the embedded one that is closely related to the core business of firms, to enhance their competitiveness. Second, it can improve the input–output efficiency of technological innovation, as well as the quality and efficiency of their operations. Third, it can expand the boundaries of production possibilities and achieve economies of scale, which all contribute to their total factor productivity. Moreover, in firms with different ownership properties, pollution attributes, and company size, the effect on manufacturing total factor productivity exerted by digital transformation is significantly heterogeneous. Specifically, the digital transformation of state-owned, non-key pollution monitoring and large firms exploit a better effect on total factor productivity enhancement. In addition, there is an asymmetry in the positive promotion influence digital transformation exerts on corporate total factor productivity; that is, there is a threshold value (digital = 4.094), and the positive effect reaches the maximum when digital transformation reaches this point, while the positive effect gradually decreases after exceeding this threshold, showing the law of diminishing marginal effect.

At present, the overall digital transformation degree in China’s manufacturing firms has been in the rising area of that inverted U-shape curve, and most of them have not yet reached the threshold of the greatest promotion effect, so the current should strengthen the investment in digital infrastructure such as artificial intelligence and industrial Internet and improve the serviceability of digital infrastructure for manufacturing enterprises. The effect of digital transformation lies in the capability to obtain and comprehensively utilize data. Constructing digital infrastructure can expand data sources and promote data flow velocity over different nodes so as to promote the quality and efficiency of enterprise data utilization, which is an important condition for corporate digital transformation. At present, manufacturing digitalization in China is still generally not high, and there is still a lot of room for development in the future. As far as manufacturing enterprises themselves are concerned, they should follow the principle of differentiation, carry out digital transformation abiding by their own business characteristics, and pay attention to the powerful role digital transformation plays in enhancing embedded service-oriented manufacturing, technological innovation, and economies of scale. In the process of digital transformation, manufacturing companies can continuously learn and explore the integration and adaptation of embedded service-oriented manufacturing, technological innovation, and digitalization to elevate their products’ quality and operation efficiency. As far as we are concerned, firms that belong to the key pollution monitoring units do not implement adequate transformation and have more disadvantages compared with the non-key ones to conquer. So, the government may create more friendly policies to help the key ones to grow stronger to accelerate their digital transformation. In addition, the government and society also should pay attention to guide and help small firms to acquire the essential resources they urgently need in their digital transformation. Since many small and medium-sized firms are not strong enough, they can start with informatization and intelligence, then steadily promote transformation and enhance efficiency. In addition, this paper also finds evidence that excessive digitalization may indeed damage the productivity of enterprises, so some large and already high-transformed firms should not blindly pursue extensive excessive digital transformation and be vigilant against the occurrence of digital paradox.

Although this paper portrays the whole extent of corporate digital transformation using the text analysis method, it is still necessary to further consider how to measure the digitalization of each process within an enterprise more accurately to study the consequences of digital transformation more specifically. However, considering the availability of data, our paper is based on 2013 to 2020 of A-share listed manufacturing firms and does not include unlisted manufacturing firms whose data are difficult to obtain, so it is worth exploring whether the results obtained are also applicable to unlisted manufacturing enterprises in the future.