How Does Digital Transformation Affect Total Factor Productivity: Firm-Level Evidence from China

Abstract

After the 2008 financial crisis, companies in China begun paying more attention to sustainable development. This article attempts to examine whether and how digital transformation affects total factor productivity (TFP) at the firm-level. Using 2913 listed companies in China from 2012 to 2018, this study finds that digital transformation is positively associated with corporate TFP in China. Our explanatory variable of firm-level digitalization index is constructed via text analysis methods. After a series of robustness checks and different attempts that mitigate endogeneity concerns, our findings remain valid. However, traditional information and communication technologies (ICT) cannot improve corporate TFP. Further analyses of three plausible channels indicate that digital technologies improve TFP primarily through cost reduction and human substitution rather than supervision advantage. The results indicate that firms achieve actual benefits from the digital transformation, and how digital transformation improve the sustainable development. This study could serve as a policy inspiration for other developing countries.

1. Introduction

Because of the severe downturn of the Chinese economy caused by the 2008 financial crisis, the Chinese government urgently sought to find ways to improve the economic quality and maintain sustainable development. In addition, digital transformation has become an important topic in China. The concept of a “digital economy” has been included in the Chinese Government Work Report since 2017. After USA, China has the second largest investment in emerging technologies, which reached GBP 18 billion during the 2015–2019 period. The enterprise digital transformation is a micro-phenomenon of the digital economy. Whether and how the enterprise digital transformation has an impact on total factor productivity (TFP) remains inconclusive. To answer these questions, we face four challenges:

First, the causal inference of technological shocks on economic consequences is difficult with macro-level measurement [1]. There are some studies concerning the impact of traditional information and communication technology (ICT) on productivity [2,3,4], but they are far from a consensus. At the firm level, we can draw more reliable conclusions due to the direct link between digital technologies and productivity and observe the behavior changes of companies after digital transformation.

Second, many studies focus on the impact of technological trends in developed countries. The potential impacts of digital technologies may be different in developing countries due to the poor technological foundation [5] and absorptive capacity [6]. Additionally, data are often scarce in developing countries.

Third, the relation between digital transformation and corporate performance is diverse. On the one hand, digital technologies can accomplish more complex tasks than ICT. Digital technologies include big data, analytics, cloud, social media, mobile platforms and intelligent solutions [7]. The powerful capabilities of information collection, data processing, and prediction help improve corporate performance [8,9]. On the other hand, the problems of adaptability and mismanagement arising from the shock of new technologies can lead to negative impacts [10]. Since there are two opposing views, evaluating the effect of digital transformation on a firm’s performance is difficult.

Fourth, the channels through which digital technologies affect corporate TFP are not clear. Some studies suggest that ICT can replace low-skilled workers [11] but offer no evidence of digital technologies’ ability to carry this out. In addition, can digital technologies influence TFP through other channels such as reducing the costs in production process and monitoring employees?

Accordingly, this paper aims to explore the relationship between enterprise digital transformation and corporate TFP by addressing the following four questions:

• Does enterprise digital transformation promote or inhibit the improvement of the TFP of Chinese companies?
• Are the effects of digital technologies and traditional ICT on TFP different?
• How do we measure the level of digital transformation and ICT intensity at firm-level?
• Through which channels does enterprise digital transformation affect TFP?

To conduct firm-level research, the key step is quantifying the enterprise digitalization index. Due to the lack of firm-level data on digital transformation, there are only a few empirical studies at the micro-level. In this study, we used the text analysis methods to construct an enterprise digitalization index. Text analysis has been used to measure some firm characteristics, such as investor sentiment, financing constraints [12]. Similar to Li et al. (2013) [13], this article selects the Management Discussion & Analysis (MD&A) section of each annual report as the corpus. We use word2vec technology, as described in Li et al.’s work (2021) [14], and further use the public corpus provided by Tencent to build a complete digitization dictionary with 3246 words. Then, we obtain large panel data with 11,136 firm-year observations from the 2012–2018 annual reports of 2913 listed companies in China. Using these data, we show that digital technologies are positively associated with corporate TFP. This finding remains valid after a series of robustness checks. Our work provides a feasible solution for empirical researches about enterprise digital transformation.

We then address the potential concerns of reverse causality, which means that firms with higher TFP may have sufficient capital to invest more in digital technologies. Three identification strategies are used. First, we use the difference-in-differences method by treating 4G services as an exogenous shock that is approved by the authorities [15]. This shock can directly contribute to the development of digital technologies. The results show that corporate TFP is improved when 4G services enter a city. Second, we retest the relation between digital technologies and corporate TFP on propensity-matched samples. The results are in line with the baseline model. Third, the concerns of reverse causality and omitted variables are addressed by using instrumental variables, i.e., the average digitization indices of other companies in each industry. The results are also robust in this analysis. The above evidences confirm that the positive digitization–TFP relation is unlikely to be driven by endogeneity concerns.

We also examine whether the impact of digital technologies and traditional ICT on TFP is different. The measure of ICT intensity is similar to the enterprise digitization index. The results suggest that only digital technologies and not ICT are able to improve TFP.

Finally, we explore three possible underlying economic mechanisms through which digital transformation affects corporate TFP. First, the cost in production process is considered. Specifically, the impact of digital transformation on corporate TFP is stronger in firms with higher initial costs. Second, we investigate whether digital transformation reduces the need for employees. Our analysis shows that digital transformation has a significant substitution effect on low-skilled workers and poses a potential threat to medium-skilled workers. Third, we seek to examine whether digital transformation helps improve TFP from the perspective of supervision advantage. The results show that the productivity effect is not more pronounced in complex firms, indicating that the channel of supervision is not supported. These results suggest that the main mechanisms are cost reduction and human replacement.

Our study speaks to three streams of the literature. First, we add to the burgeoning literature on digital transformation. Most of literature about digital transformation is based on case studies [16,17]. Additionally, there are no relevant data on digital technology investment of Chinese companies in the existing database. This paper constructs an enterprise digitization index of Chinese companies and carries out empirical researches based on large panel data possible. Second, our study extends the existing empirical literature on TFP. Many scholars explored the determinants of TFP with macro-level data in developed countries [2,3,18]. Using firm-level data of Chinese companies, we provide direct evidence that digital transformation plays an important role in improving the economic quality and maintain sustainable development. Third, it contributes to the relation between digital transformation and corporate TFP. Our study proves that only digital technologies and not traditional ICT can improve TFP and the underlying channels are cost reduction and human replacement.

The remainder is organized as follows. The second part is the theoretical framework and hypothesis development. The third part presents the data and the model. The fourth part shows the main results. The fifth part presents the channel analysis. The sixth part offers the conclusions.

2. Theoretical Framework and Hypothesis Development

2.1. Enterprise Digital Transformation

Enterprise digital transformation is built on the foundation of digital technologies, ushering in unique changes in business processes. Digital technologies include big data, analytics, cloud, social media, mobile platforms and intelligent solutions [7]. Most of the literature about digital transformation is based on case studies [19]. Some scholars discuss the drivers of the success of the digital transformation such as tangible and intangible resources, financial and human resources [20], and collaboration networks [21]. Some case studies find that digital technologies lead to new business model innovations [22]. However, there is a lack of empirical analysis to discuss the impact of digital transformation [19]. To conduct empirical researches, finding an appropriate measure of enterprise digitization index is the first step. This paper uses the text analysis methods to quantify and examine the relationship between enterprise digital transformation and sustainable development.

2.2. Total Factor Productivity

The TFP measures the efficiency and sustainability between inputs and outputs. Although firms have exactly the same quality and quantity of capital, labor and other factor inputs, their output levels are different if they have different TFP. Other indicators such as corporate earnings, ROE, growth, size, and risk are also used to measure corporate performance [23,24]. However, changes in these indicators may be the result of temporary increases in labor or capital inputs, temporary prosperity in the industry, or debt-financed operations. These indicators cannot serve as measures of enterprise sustainable development like TFP. Campisi et al. (2019) [23] also found that the results from financial indicators are not always congruent with TFP. Moreover, TFP is more suitable than single-factor productivity (SFP) in measuring production efficiency. SFP such as per capita output and labor productivity can only represent the output efficiency of a single production factor. In contrast, TFP reflects the output efficiency after combining various factor inputs; thus, it has a more comprehensive and realistic meaning than indicators of SFP. Based on the above analysis, this paper adopts TFP as the measure of sustainable development.

2.3. Digital Transformation and Corporate TFP

According to the resource-based view theory, the resources of digital technologies can improve corporate performance [7,24]. The digitization of a firm’s products and services enable new ways to capture value [25]. Data resources are the core of the digital economy and the enterprise digital transformation. Digital technologies can process unstructured data such as text data and image data [9], greatly expanding the available data resources.

Combining digital technologies and data can benefit a firm’s efficiency. Digital products are quickly updated in the virtual world with the help of data analysis [16]. Accurate price information of raw materials and intermediate products can be predicted by digital technologies [8], which help reduce purchase costs. Additionally, digital technologies can learn human skills such as communicating with people from massive amounts of data [26], which means that the need for employees is reduced. Furthermore, the data and information of market demand, inventory status, financial situation and employee performance are monitored by digital technologies in real time [27,28], which helps motivate employees and managers to work hard. Digital transformation promotes democratic decision-making to allow practitioner groups to decide jointly on substantial aspects of the firm’s future path [29] and automated decision-making [30], which improves the efficiency and quality of decision-making.

Digital transformation may also bring negative effects, that is, the “productivity paradox”. Soh and Markus (1995) [10] constructed a process theory model and proved that the investment in new technologies does not immediately improve company performance, but requires a process of absorption. Brynjolfsson et al. (2017) [31] think that the productivity effect does not appear unless related complementary innovations have been fully developed and implemented. Mismanagement can also lead to the investment in digital technologies becoming a burden [32,33]. Many digital transformation projects fail if there is no structured approach to manage the adoption of the technology innovation [30]. The negative significant impact of the failure is reflected in business operability and performances, such as disrupting the everyday business processes at great speeds in novel flows and customer-centric disturbances [30]. A developing country such as China is more prone to the productivity paradox due to poor technological foundation [5], absorptive capacity [6] and organizational learning ability [34].

Therefore, the relationship between digital transformation and corporate TFP can be diverse. On the one hand, digital technologies help efficiently use data to reduce costs, replace workers and supervise employees, which could improve TFP. On the other hand, the digital transformation of Chinese companies is likely to fail because of the underinvestment, mismanagement and adaptability problems. We thus hypothesize as follows:

Hypothesis 1a (H1a).

Enterprise digital transformation is positively associated with corporate TFP.

Hypothesis 1b (H1b).

Enterprise digital transformation is negatively associated with corporate TFP.

3. Data and Model

3.1. Data Source

We use text analysis methods to construct enterprise digitization indices from the annual reports of listed companies in China. Chinese companies widely adopted traditional technologies such as ERP in 2012, followed by the widespread adoption of digital technologies in 2015. In order to compare the impact of digital technologies and traditional technologies on productivity, our sample starts in 2012. Additionally, earlier annual report documents are difficult to parse with Python programs. The data of firm-level variables are mainly from the China Stock Market & Accounting Research (CSMAR) database and supplemented from Wind. The provincial data such as the price deflator are from the National Bureau of Statistics of China. To exclude the confounding effects of the COVID-19 pandemic and Ukraine crisis, we cut the TFP data to 2019 and the data of enterprise digitization indices to 2018. On the one hand, the COVID-19 pandemic and rise of commodity prices due to the Ukraine crisis have led to a stagnation in production activities since 2020. On the other hand, the COVID-19 pandemic accelerated the digital transformation of companies [19]. Such disruptive events can lead to huge biases in the estimation results. Additionally, there is a lack of annual price deflators for investment after 2019, which also leads to a bias in the estimating TFP. Therefore, we cut the TFP data to 2019 and the data of enterprise digitization indices to 2018. Hence, our sample period is between 2012 and 2018. All continuous variables are winsorized at the 1st and 99th percentiles to mitigate the potential impact of outliers.

We cleaned the data according to the following procedures to obtain our final sample. First, we deleted the companies on the Science and Technology Innovation Board because of their abnormal liquidity. Second, we excluded financial companies. Third, the special treatment (ST) and particular transfer (PT) firms with abnormal financial systems were excluded. Fourth, we deleted the observations of the IPO year. Fifth, abnormal observations with negative revenue or zero total assets were excluded. Sixth, we deleted samples with no control variables. The final sample included 11,136 firm-year observations for 2913 unique firms.

3.2. Empirical Specification

To explore the relation between digital transformation and corporate TFP, we estimate the following panel regression model:

$${TFP}_{i,t+1}={\alpha }_0+{\alpha }_1{Digital}_{i,t}+{\alpha }_2{Control}_{i,t}+{Size decile}_{i,t}\times {Year}_t+{Firm}_i+{\epsilon }_{i,t+1}$$

(1)

Here, i denotes firms, and t denotes years. The explained variable TFP is the total factor productivity of a firm. We forward TFP by one year to partially mitigate endogeneity concerns. The explanatory variable Digital is the enterprise digitization index constructed using text analysis methods. Control represents a vector of characteristics correlated with corporate TFP. See Section 3.3, Section 3.4 and Section 3.5 for definitions of the dependent and independent variables. We include Firm, the firm fixed effects, to absorb firm-specific time-invariant factors. Following Acemoglu et al. (2022) [35], we include Size decile × year, the size decile-year interaction fixed effects, to absorb™ economic shocks related to firm size. The Size decile is the decile ranking of natural logarithm of total assets. Standard errors are clustered at the firm level for all estimates.

3.3. Measuring Enterprise Digitization Index

The measurement of technological development in existing literature is mostly carried out at the macro level. Acemoglu et al. (2014) and Jorgenson and Stiroh (2000) [2,4] measured ICT-related investment based on computer hardware, software, and communication equipment. Choi and Hoon Yi (2009) and Czernich et al. (2011) [36,37] used broadband investment and Internet penetration rate to proxy the development of Internet.

At the micro-level, studies before 2010 used firm-level ICT investment from ComputerWorld (CW) database and InformationWeek (IW) database, which are no longer updated and the panels are small [24,38]. Recently, a few scholars have made meaningful explorations on the measurement of investment or adoption of digital technologies. Brynjolfsson et al. (2018) [39] measured how easily an employee can be replaced by AI according to occupation level. Chen et al. (2019) [40] measured the level of digital innovation based on digital technology patents. However, traditional ICT investment, number of AI-related employees, and digital innovation can only be used as proxy variables for the adoption of digital technologies.

In this paper, we directly measured the enterprise digitization index through word frequency. The usage of words in the annual reports reflects the company’s strategic deployment and future planning [13]. Therefore, it is scientific to quantifying the level of digitization from the perspective of statistics on word frequency involving “enterprise digital transformation” in annual reports.

The specific construction method of enterprise digitization index is as follows. First, based on digital policy documents issued by the Chinese government, digitization-related reports provided by investment banks and research firms, and annual reports from 10% of the Chinese listed companies, we artificially selected words related to digitization and then obtain 1048 seed words. The reason we selected annual reports from 10% of listed companies is to reduce unnecessary manual work. The other words related to digital transformation in annual reports of the remaining 90% of the listed companies can be found through the natural language processing methods in the next steps.

Second, we used Word2vec method to obtain word vectors of the seed words and words in annual report corpus. According to the cosine similarity between the seed words and the words in annual report corpus, we took the top five words that are most similar to each seed word as synonyms. Then, we removed words unrelated to digitization and words that repeat with the seed words. Using the remaining words as the extensions of the seed words, we obtained 966 synonyms in annual reports.

Third, in order to make the digitization dictionary more versatile, it should be not only suitable for the annual report analysis in this project but also for text analysis projects such as conference calls and online community comments. Therefore, we further used a public corpus of over 8 million Chinese words provided by Tencent to find more synonyms. In this step, we obtained 1232 synonyms and integrated them into a complete digitization dictionary. There were 3246 digitization-related words in total. Therefore, the complete digitization dictionary in this article consisted of three parts: seed words, synonyms in annual reports, and synonyms of 8 million words provided by Tencent. The processes involved in building the digitization dictionary are shown in Figure 1.

Fourth, we used Python to count the absolute frequency of digitization-related keywords in the annual reports of each Chinese listed company, and, respectively, divide it by the total number of sentences and the total words in each annual report. The standardized variables, namely Digital1 and Digital2, are used to measure the level of enterprise digitization.

3.4. Measuring Corporate Total Factor Productivity

To measure corporate TFP, this paper adopts the LP method [41], which is widely used in other literature [42,43]. The data envelopment analysis (DEA) method is also used to measure corporate TFP [23,44], but it leads to many unobserved values when applied to panel data of Chinese listed companies. When the panel data of Chinese listed companies are converted into balanced panel data to calculate TFP, a large number of firm samples are excluded, resulting in the failure to objectively reflect the overall changes in TFP of all Chinese listed companies.

The LP method has the following advantages. Compared to the OLS method, the LP method can smoothly control for unobserved productivity shocks by using intermediate inputs as proxies when estimating TFP [43]. The advantage of the LP method relative to the OP method [45] in estimating TFP is that intermediate inputs respond more fully to the productivity shock than investment [41]. Compared to DEA method, the LP method does not lose many firm samples when estimating TFP. Therefore, we adopt the LP method to measure corporate TFP.

The data for measuring corporate TFP using the LP method are as follows. Output is constructed using the natural logarithm of operating income. The labor input is measured using the natural logarithm of the number of employees. Capital investment is defined as the natural logarithm of the net fixed assets. Intermediate input is measured via the logarithm of the cash paid for goods and services. All other variables except labor input are deflated using annual price deflators for output, input and investment, which are normalized with 2010 deflators.

To ensure the robustness of empirical results, we remeasured corporate TFP using the OP method [45], and the GMM method [46].

3.5. Control Variables

Following the prior literature, we controlled for a set of variables that impact corporate TFP [35]. Firm age (Age) is measured from the year of initial public offering. Return on total assets (ROA) is defined as earnings divided by total assets. Cash flow (Cash) is the ratio of operating cash flows to total assets. We also control for debt ratio (Debt) defined as the ratio of interest-bearing liabilities to total assets, Tobin’s Q (TQ), the natural logarithm of fixed assets (Fassets), and ownership type (SOE) which is equal to one if a firm belongs to a state-owned enterprise, and zero otherwise. In addition, the real GDP (GDP) and Herfindahl index (HHI) are also included to, respectively, account for economic development of each province and the concentration in each industry-year.

3.6. Summary Statistics

Table 1. The summary statistics of relevant variables.

Variables	N	Mean	SD	Min	Max
TFP	11,136	12.7372	1.0715	10.4779	15.7823
Digital1	11,136	0.2134	0.2718	0.0000	1.4868
Digital2	11,136	0.3158	0.3942	0.0000	2.1221
Age	11,136	2.1521	0.7757	0.6931	3.2189
Cash	11,136	0.0046	0.0069	−0.0161	0.0234
ROA	11,136	0.0403	0.0508	−0.1639	0.1885
Debt	11,136	0.0814	0.0181	0.0253	0.1000
TQ	11,136	0.0021	0.0013	0.0009	0.0085
Fassets	11,136	20.4339	1.6573	16.1565	25.0261
SOE	11,136	0.4055	0.4910	0.0000	1.0000
GDP	11,136	3.9678	2.3819	0.2759	9.5815
HHI	11,136	0.0409	0.0781	0.0044	0.4234
Size decile	11,136	5.4952	2.8755	1.0000	10.0000

shows the summary statistics of the variables used in our baseline analysis. The enterprise digitization index Digital1 (Digital2) has a mean value of 0.2134 (0.3158), which is lower than the standard deviation, implying that the level of digitization differs significantly across firms. The average value and standard deviation of TFP are, respectively 12.7372 and 1.0715, indicating that the value of TFP is stable.

4. Main Results

4.1. Baseline Results

Table 2. Digital transformation and corporate TFP.

Variables	TFP	TFP
	(1)	(2)
Digital1	0.1399 ***
	(3.2734)
Digital2		0.0787 ***
		(2.6604)
Age	0.1828 ***	0.1855 ***
	(4.8191)	(4.8956)
Cash	0.3992	0.3984
	(0.3753)	(0.3750)
ROA	1.9102 ***	1.9108 ***
	(11.7933)	(11.7854)
Debt	0.0026 ***	0.0026 ***
	(2.5868)	(2.6003)
TQ	0.4169	0.3222
	(0.0617)	(0.0476)
Fassets	0.0772 ***	0.0775 ***
	(4.1902)	(4.2020)
SOE	0.0455	0.0442
	(0.7627)	(0.7375)
GDP	0.0124	0.0138
	(0.7965)	(0.8851)
HHI	0.1956	0.1798
	(0.9986)	(0.9190)
Constant	10.4270 ***	10.4149 ***
	(29.5208)	(29.4890)
Firm Effect	Yes	Yes
Size decile-Year	Yes	Yes
Observations	11,136	11,136
Adjusted R2	0.2488	0.2481

Notes: *** is significant at the levels of 1%.

shows the results of model (1). In columns (1) and (2), we use Digital1 and Digital2, respectively, to measure the level of enterprise digitization. The coefficients of both enterprise digitalization indices are positive and significant at the 1% level. Specifically, an increase in Digital1 of 1% is on average associated with a 3.5% (=0.1399 × 0.2718/1.0715) increase in TFP. The productivity paradox is rejected and the hypothesis H1a is supported. Next, we briefly analyzed the coefficients of control variables. The coefficients of age and ROA are significantly positive, indicating that old corporations and companies with stronger profitability have a higher TFP.

4.2. Endogeneity

4.2.1. Difference-in-Differences Analysis Using 4G Services

To address the potential endogeneity problem, we re-estimated the effect of digital transformation on corporate TFP using the difference-in-differences (DID) method by treating 4G services as an exogenous shock. Briefly, 4G allows data transmission to speed up dramatically through mobile networks [15], which leads to the rapid development of AI, big data, and Internet of Things (IoT). Therefore, we argue that this shock can directly improve the level of digital transformation.

Following Ding et al. (2022) [15], we created a dummy variable 4G that equals 1 if the 4G network of city i is implemented in year t, and 0 otherwise. The cities Shanghai, Hangzhou, Nanjing, Guangzhou, Shenzhen and Xiamen are implemented since 2013; Beijing, Qingdao, Wenzhou, Tianjin, Shenyang, Chengdu and Fuzhou are implemented since 2014; the rest of the cities are implemented since 2015.

Table 3. Difference-in-differences approach.

Variables	TFP
4G	0.0489 ***
	(2.9627)
Age	0.1900 ***
	(5.0284)
Cash	0.2948
	(0.2769)
ROA	1.9048 ***
	(11.7535)
Debt	0.0026 ***
	(2.6328)
TQ	−0.3034
	(−0.0445)
Fassets	0.0782 ***
	(4.2170)
SOE	0.0454
	(0.7460)
GDP	0.0178
	(1.1535)
HHI	0.1070
	(0.5563)
Constant	10.3998 ***
	(29.2669)
Firm Effect	Yes
Size decile-Year	Yes
Observations	11,136
Adjusted R2	0.2477

Notes: *** is significant at the levels of 1%.

presents the DID estimate of the effect of digital transformation on corporate TFP. The coefficient of 4G is significantly positive, which is in line with the baseline model.

4.2.2. Instrumental Variables Approach

The two stage least-squares (2SLS) method is used to further address the concerns of reverse causality and omitted variables. When corporate TFP grows higher, a company may invest more in digital technologies. This potential decision driven by performance would spuriously reduce the estimated relation between digital technologies and corporate TFP.

Similar to Acemoglu et al. (2022) [35], we used the average digitization indices of other companies in each industry as the instrumental variables, namely IndDigi1 and IndDigi2. A company will increase their investment in digital technologies after observing other companies in this industry investing in digital technologies. Additionally, the average digitization indices have no impact on a firm’s TFP except through their influence on that firm’s digitization index. Therefore, our instrumental variables satisfy the requirements of relevance and exogeneity.

Table 4. Instrumental variables approach.

	First Stage	Second Stage	First Stage	Second Stage
Variables	Digital1	TFP	Digital2	TFP
	(1)	(2)	(3)	(4)
IndDigi1	0.5091 ***
	(9.6543)
Digital1		0.7015 **
		(2.4221)
IndDigi2			0.5018 ***
			(9.8563)
Digital2				0.4755 **
				(2.3117)
Age	0.0393 ***	0.1513 ***	0.0408 **	0.1598 ***
	(2.8850)	(3.7194)	(2.0754)	(3.9981)
Cash	−0.3908	0.6932	−0.7242 *	0.7685
	(−1.4660)	(0.6316)	(−1.9065)	(0.6992)
ROA	0.0236	1.9104 ***	0.0222	1.9132 ***
	(0.4488)	(11.6093)	(0.2797)	(11.5682)
Debt	0.0004	0.0024 **	0.0005	0.0024 **
	(1.3195)	(2.3463)	(1.2902)	(2.3770)
TQ	−2.4003	1.9000	−3.0526	1.6985
	(-0.9761)	(0.2762)	(−0.8456)	(0.2464)
Fassets	0.0106 **	0.0740 ***	0.0128 *	0.0752 ***
	(2.0762)	(4.0203)	(1.8954)	(4.1182)
SOE	−0.0058	0.0472	0.0080	0.0397
	(−0.2595)	(0.8243)	(0.2535)	(0.6862)
GDP	0.0107 *	0.0010	0.0047	0.0066
	(1.8877)	(0.0633)	(0.6024)	(0.4201)
HHI	−0.3435 ***	0.4622 *	−0.5076 ***	0.4343 *
	(−3.2676)	(1.8520)	(−3.3935)	(1.7570)
Constant	−0.2425 **	10.5010 ***	−0.2313 *	10.4444 ***
	(−2.2879)	(29.6567)	(−1.6631)	(30.1168)
Firm Effect	Yes	Yes	Yes	Yes
Size decile-Year	Yes	Yes	Yes	Yes
Observations	11,126	11,126	11,126	11,126
Adjusted R2	0.2463	0.2208	0.2429	0.2205

Notes: ***, **, and * are significant at the levels of 1%, 5%, and 10%, respectively.

shows that the regression coefficients of the instrumental variables are significantly positive, which means our conclusions are still valid.

4.3. Other Robustness Tests

In addition, we conducted extensive robustness tests. We used a dummy variable to regress again, that is, assign the previous digitalization index to 1,2,3,4 and 5 from low to high. The result is still robust. The corporate TFP is also re-measured using the OP method and GMM method, and the results remain valid. According to the 2012 version of the guidance on industry classification of listed companies issued by the China Securities Regulatory Commission (CSRC), we exclude the manufacturing industry of computers, communication and other electronic equipment and the service industry of information transmission, software and information technology, and the re-regression model (1) based on samples from other industries that are not directly related to digital technologies. We still find a significant and positive relationship between digitization and corporate TFP, indicating that the regression result is not entirely determined by firms in directly related industries. We also use the OLS method and control for industry fixed effects and province fixed effects, and the results remain valid. These empirical results are not presented due to the limited nature of this article.

4.4. Traditional ICT

Having established a robust positive effect of digital transformation on corporate TFP, we next examine whether the impact of digital technologies and traditional ICT on TFP is different. Digital technologies can accomplish more complex tasks than ICT. In addition to traditional data, digital technologies can also collect unstructured data such as images and text from mobile devices and social media [9]. The powerful capabilities of data analysis and prediction assist firms in making informed decisions based on big data [8]. We expect digital technologies to play a more important role in TFP than traditional ICT.

Similar to the construction of the enterprise digitalization index, we use 1019 words related to traditional ICT to measure the ICT intensity. We did not find a productivity effect of ICT in

Table 5. Traditional ICT and corporate TFP.

Variables	TFP	TFP
	(1)	(2)
ICT1	0.0962
	(1.4317)
ICT2		0.0215
		(0.3704)
Age	0.1927 ***	0.1915 ***
	(5.1164)	(5.0951)
Cash	0.3419	0.3267
	(0.3208)	(0.3067)
ROA	1.9110 ***	1.9111 ***
	(11.7754)	(11.7744)
Debt	0.0027 ***	0.0027 ***
	(2.6846)	(2.6527)
TQ	0.1608	0.0896
	(0.0236)	(0.0131)
Fassets	0.0779 ***	0.0779 ***
	(4.2051)	(4.1964)
SOE	0.0467	0.0452
	(0.7733)	(0.7452)
GDP	0.0147	0.0150
	(0.9488)	(0.9705)
HHI	0.0850	0.1140
	(0.4393)	(0.5819)
Constant	10.3988 ***	10.4074 ***
	(29.2974)	(29.2788)
Firm Effect	Yes	Yes
Size decile-Year	Yes	Yes
Observations	11,136	11,136
Adjusted R2	0.2473	0.2468

Notes: *** is significant at the levels of 1%.

, consistent with the productivity paradox of ICT found by [3,18]. This confirms that only digital technologies and not ICT are able to improve TFP.

5. Channel Analysis

5.1. Cost Reduction

We posit that cost reduction through digital transformation could cause higher TFP for a firm. Firm’s products are quickly updated in the virtual world with the help of digital technologies [16], which greatly saves the cost of experiments in the real world. The powerful information collection and forecasting capabilities provided by digital technologies enable companies to accurately predict price information of raw materials and intermediate products [8], reducing purchase costs. The productivity effect is expected to be more pronounced in firms with higher initial costs.

Following Bharadwaj (2000) [24] and Ai et al. (2021) [47], the ratio of main operation cost to sales is used to measure the cost in the production process. Our main variables of interest are the interaction terms between the digitization indices and initial costs. We find that the interaction terms Cost × Digi1 and Cost × Digi2 are both positive and significant in

Table 6. Digital transformation and corporate TFP: Cost reduction.

Variables	TFP	TFP
	(1)	(2)
Cost × Digi1	0.4483 **
	(2.1976)
Cost × Digi2		0.3996 ***
		(2.8505)
Cost	0.0487 ***	0.0452 ***
	(2.9744)	(3.2232)
Digital1	−0.1519
	(−1.1027)
Digital2		−0.1870 *
		(−1.9408)
Age	0.1828 ***	0.1852 ***
	(4.6516)	(4.7202)
Cash	0.2342	0.2075
	(0.2159)	(0.1916)
ROA	1.8953 ***	1.8943 ***
	(11.2458)	(11.2137)
Debt	0.0028 ***	0.0028 ***
	(2.7104)	(2.7237)
TQ	8.4035	8.6691
	(1.2101)	(1.2455)
Fassets	0.0695 ***	0.0706 ***
	(3.2925)	(3.3438)
SOE	0.0267	0.0253
	(0.3934)	(0.3743)
GDP	0.0109	0.0124
	(0.6857)	(0.7803)
HHI	0.1936	0.1598
	(0.9110)	(0.7612)
Constant	10.5544 ***	10.5317 ***
	(26.3625)	(26.3829)
Firm Effect	Yes	Yes
Size decile-Year	Yes	Yes
Observations	10,408	10,408
Adjusted R2	0.2627	0.2628

Notes: ***, **, and * are significant at the levels of 1%, 5%, and 10%, respectively.

, indicating that cost reduction through digital transformation could cause higher corporate TFP.

5.2. Human Replacement

Digital technologies can already accomplish some complex tasks such as face recognition, chatting with humans and managing unmanned stores [26]. Due to the superior performance and no need for breaks, companies may prefer to use digital technologies rather than hire human employees. Bresnahan et al. (2002) [11] shows that the substitution of employees with technology can increase corporate productivity.

Following Bartel and Sicherman (1999) [48], we use two indicators, the ratio of workers without a bachelor degree to the total employees (Lowedu) and the ratio of production workers to the total employees (Pworkers), to measure the employment of low-skilled workers. Similar to Frey and Osborne (2017) [49], we define office and administrative support workers as middle-skilled employees, measured by the ratio of the sum of financial and administrative workers to the total employees (Mworkers). The results in

Table 7. Digital transformation and corporate TFP: Human replacement.

Variables	Lowedu	Pworkers	Mworkers	Lowedu	Pworkers	Mworkers
	(1)	(2)	(3)	(4)	(5)	(6)
Digital1	−2.0521 **	−2.8306 **	−1.0092 *
	(−2.1281)	(−2.5583)	(−1.6541)
Digital2				−1.4358 **	−1.7917 **	−0.6722
				(−2.2489)	(−2.0097)	(−1.5639)
Age	0.6593	0.4532	4.8018 ***	0.6374	0.4106	4.7888 ***
	(0.7366)	(0.4173)	(7.5726)	(0.7132)	(0.3772)	(7.5784)
Cash	38.2552	49.5092 *	−24.6413	37.9919	49.3305 *	−24.7376
	(1.4179)	(1.8715)	(−1.5917)	(1.4083)	(1.8645)	(−1.5995)
ROA	−4.1763	−5.8238	−2.2456	−4.1842	−5.8348	−2.2496
	(−1.1579)	(−1.5839)	(−1.0521)	(−1.1600)	(−1.5872)	(−1.0541)
Debt	0.0320	0.0445 *	0.0018	0.0319	0.0443 *	0.0018
	(1.5557)	(1.7229)	(0.1224)	(1.5533)	(1.7123)	(0.1188)
TQ	−280.3551	−81.5657	85.7967	−279.9069	−80.3161	86.1305
	(−1.5734)	(−0.4256)	(0.8316)	(−1.5691)	(−0.4181)	(0.8351)
Fassets	2.3139 ***	3.1145 ***	−0.4836 **	2.3108 ***	3.1091 ***	−0.4853 **
	(5.4711)	(6.5534)	(−2.3173)	(5.4579)	(6.5345)	(−2.3245)
SOE	−0.0115	−0.2737	1.3720 *	0.0110	−0.2447	1.3827 **
	(−0.0081)	(−0.1852)	(1.9478)	(0.0077)	(−0.1651)	(1.9625)
GDP	0.3321	0.5071	0.1258	0.3166	0.4826	0.1176
	(0.9546)	(1.2082)	(0.4907)	(0.9107)	(1.1504)	(0.4589)
HHI	−15.0612 ***	−15.4582 ***	0.5269	−15.0074 ***	−15.2653 ***	0.5747
	(−3.3687)	(−2.6778)	(0.1826)	(−3.3440)	(−2.6420)	(0.1992)
Constant	27.7093 ***	−18.0360 *	11.5649 ***	27.8713 ***	−17.8021 *	11.6464 ***
	(3.4371)	(−1.9080)	(2.7751)	(3.4556)	(−1.8829)	(2.7951)
Firm Effect	Yes	Yes	Yes	Yes	Yes	Yes
Size decile-Year	Yes	Yes	Yes	Yes	Yes	Yes
Observations	11,136	11,136	11,136	11,136	11,136	11,136
Adjusted R2	0.0836	0.0412	0.0752	0.0835	0.0409	0.0752

Notes: ***, **, and * are significant at the levels of 1%, 5%, and 10%, respectively.

show that digital transformation causes a significant decline in the employment of low-skilled occupations and threatens the employment of middle-skilled workers.

5.3. Cost Reduction and Human Replacement versus Supervision Advantage

In the previous two sections, we discussed that digital transformation improves TFP by saving costs in the production process and replacing some workers. However, one can also argue that the positive association is due to the supervision advantage of digital technologies. On the one hand, digital technologies can quickly collect and process massive amounts of data, allowing one to efficiently monitor the changes of market demand, inventory status and financial situation [28]. The agency conflicts are mitigated by reducing information asymmetry between principals and agents [50]. On the other hand, some digital technologies are specifically designed to help firms track and manage employee performance accurately and transparently, and therefore guarantee the effectiveness of incentive practices to motivate employees to work hard [27]. If supervision advantage is the underlying channel, we expect to see a more prominent effect of digital transformation on corporate TFP for more complex firms, which are more likely to face principal agent problems [51].

Similar to Simunic (1980) [52], the firm complexity is measured according to the scope of a firm’s operation (Scope) and the number of subsidiaries (Subsidiaries). Our main variables of interest are the interaction terms Scope × Digi and Sub × Digi. The interaction terms in

Table 8. Digital transformation and corporate TFP: Cost reduction and human replacement versus supervision advantage.

Variables	Digital1		Digital2
	TFP	TFP	TFP	TFP
	(1)	(2)	(3)	(4)
Sub × Digi	−0.0009		−0.0001
	(−0.8219)		(−0.1665)
Subsidiaries	0.0036 ***		0.0034 ***
	(4.6750)		(4.4527)
Scope × Digi		0.0066		0.0030
		(0.3924)		(0.2706)
Scope		0.0253 ***		0.0258 ***
		(3.1131)		(3.1371)
Digital1	0.1508 ***	0.1250 **
	(2.9149)	(2.4252)
Digital2			0.0748 **	0.0714 **
			(2.0929)	(1.9890)
Age	0.2004 ***	0.1750 ***	0.2039 ***	0.1775 ***
	(5.2475)	(4.6086)	(5.3449)	(4.6770)
Cash	0.6236	0.4542	0.6251	0.4515
	(0.5798)	(0.4275)	(0.5824)	(0.4253)
ROA	1.8966 ***	1.9162 ***	1.8981 ***	1.9165 ***
	(11.5475)	(11.8725)	(11.5461)	(11.8630)
Debt	0.0021 **	0.0026 **	0.0021 **	0.0026 ***
	(2.1300)	(2.5769)	(2.1219)	(2.5915)
TQ	1.3109	0.4591	1.2680	0.3574
	(0.1924)	(0.0683)	(0.1858)	(0.0531)
Fassets	0.0635 ***	0.0746 ***	0.0635 ***	0.0748 ***
	(3.3822)	(4.0849)	(3.3778)	(4.0947)
SOE	0.0414	0.0400	0.0400	0.0389
	(0.6986)	(0.6790)	(0.6725)	(0.6565)
GDP	0.0073	0.0102	0.0084	0.0115
	(0.4727)	(0.6575)	(0.5429)	(0.7449)
HHI	0.1833	0.1924	0.1800	0.1757
	(0.9421)	(0.9824)	(0.9250)	(0.8979)
Constant	10.6511 ***	10.4470 ***	10.6469 ***	10.4351 ***
	(29.6847)	(29.8949)	(29.6390)	(29.8572)
Firm Effect	Yes	Yes	Yes	Yes
Size decile-Year	Yes	Yes	Yes	Yes
Observations	10,942	11,136	10,942	11,136
Adjusted R2	0.2580	0.2516	0.2572	0.2508

Notes: *** and ** are significant at the levels of 1% and 5%, respectively.

are not statistically significant at conventional thresholds. The results indicate that the supervision channel does not explain the positive association between digital transformation and corporate TFP, leaving the cost reduction and human replacement as the main drivers.

6. Conclusions

6.1. Conclusions and Policy Implications

In recent years, the business world has shown great interest in whether the enterprise digital transformation can promote sustainable development in response to the growing recognition of the importance of sustainable economic development. However, how to quantify the impact of digital transformation is the major challenge due to the lack of data on the adoption or investment in digital technologies. To conduct this empirical research, we used text analysis methods to construct the enterprise digitization indices of 2913 listed companies in China. Therefore, we used a large panel data to empirically study the relationship between enterprise digital transformation and corporate TFP.

The results of this study are as follows. First, enterprise digital transformation improve the TFP of Chinese companies. After a series of robustness checks and different attempts that mitigate endogeneity concerns, our findings remain valid. Second, traditional ICT does not show the productivity effects that digital technologies do on corporate TFP. Furthermore, this paper further examines and verifies the plausible channels through which the digital transformation improves TFP. Our results conclude that supervision advantage does not explain the positive association between digital transformation and corporate TFP and the cost reduction and human replacement are the two main drivers.

Our findings provide three policy implications. First, digital transformation has brought significant benefits to the total factor productivity, providing more confidence and reasons to encourage digital transformation. Second, this appeal applies not only to China but also to other developing countries. Developing countries are at a disadvantage in global competition and need to focus more on sustainable economic development. The digital transformation could help to narrow the productivity gap between developing countries and developed countries in the long run. Third, the government should train low-skilled workers and middle-skilled workers to use digital technologies to avoid unemployment problem due to technology shocks.

In terms of management practices, multiple suggestions are provided for enterprises. First, investing in digital technologies is encouraged because digital technology is more powerful than traditional ICT and firms achieve actual benefits from the digital transformation. Second, the role of digital technology in supervision and governance might be ignored by managers. Using the data of market demand, inventory status and financial situation to motivate employees in timely could further enhance the productivity effects of digital technology. Third, our findings help to improve the probability of success of enterprise digital transformation. Only less than half of companies around the world have benefited from digital transformation and use digital technologies to reduce labor costs and production costs rather than other purposes, contributing to the success of digital transformation.

6.2. Limitations and Future Potentials

Research on the relationship between enterprise digital transformation and corporate TFP is an important topic for sustainable economic development and it needs further study. Future research could explore whether the productivity effect of digital transformation is strengthened or weakened by other factors, such as industry type and ownership type. In addition, more precise measurements of firm-level digitalization index can be considered. All of the above limitations could potentially become the focus of future research investigations.