Digital transformation and enterprise total factor productivity

Digital transformation enhances enterprise Total Factor Productivity (TFP) through three major mechanisms. First, regarding the organizational coordination and integration effect, digital applications enhance the intensity of information disclosure, build financing platforms to alleviate financing constraints [5], and guide investor attention by releasing positive signals, thereby improving stock liquidity [10]. Second, regarding the technological innovation effect, digitization strengthens enterprise connections and network integration, accelerating knowledge spillovers [14]; it constructs unified supply chain platforms to achieve efficient collaboration, optimize resource allocation, and improve downstream productivity [15]; simultaneously, it enhances information analysis capabilities, aiding in the excavation and conversion of potential consumer demands [16]. Finally, regarding the management efficiency effect, digitization reduces internal communication and information acquisition costs, enhancing hierarchical synergy [17]; it promotes the evolution of production organization towards flattening and automation, precisely allocating labor, systematically reducing costs, and improving efficiency and quality stability [18]. In summary, digital transformation has a significant positive impact on TFP.

However, the application effects of different digital technologies vary significantly. [3] and [19] analyzed the impact of transformation on financial performance from an overall perspective; [20] and [21] focused on the specific impacts of Blockchain and Artificial Intelligence, respectively. [11] found that except for Blockchain, all other types of technologies can significantly improve Return on Total Assets; [22] pointed out that the integrated application of different technologies has a “1+1>2” multiplier effect on TFP; [23] also confirmed that except for Blockchain, the enhancement effects of other technology dimensions on high-quality development are significant. Based on this, we propose the hypothesis:

• H1: Digital transformation has a significant positive impact on enterprise Total Factor Productivity.
• H1a: Different digital technology directions have different effects on enterprise Total Factor Productivity.

The mediating role of dynamic capabilities

[24] and [25] defined dynamic capabilities as the ability of an enterprise to integrate, build, and reconfigure internal and external resources to cope with environmental changes. Subsequent scholars [26,27] further emphasized their behavioral orientation in innovating competitive advantages and reallocating resources. Comprehensively, dynamic capabilities are the core abilities for an enterprise to identify and allocate resources, improve efficiency through technological innovation, and optimize management to support long-term development.

Regarding their dimensions, although there are various classifications in academia, [25] proposed three dimensions: sensing opportunities, seizing opportunities, and transforming/reconfiguring; [27] defined them as adaptive, absorptive, and innovative capacities; [28] constructed a four-dimensional system of environmental insight, change renewal, technological flexibility, and organizational flexibility; [29] divided them into a five-dimensional framework of environmental adaptation, organizational change, resource integration, learning, and strategic isolation mechanisms. However, the core essentially boils down to resources, change/innovation, and learning. Given that existing research generally adopts the three-dimensional framework of “Organizational Coordination and Integration,” “Change and Reconstruction,” and “Learning and Absorption,” this paper also follows this classification: First, digital transformation enhances organizational coordination and integration capabilities. It helps enterprises break path dependence, optimize governance structures and processes, and reduce internal costs; simultaneously, it promotes effective management and collaboration of data resources, alleviating information asymmetry, enabling enterprises to respond precisely to stakeholder needs [30,31]. Second, digital transformation improves change and reconstruction capabilities. By introducing advanced IT and data analysis tools, it drives enterprises towards intelligent manufacturing and builds dedicated digital platform capabilities [32]. This innovation capability drives the reallocation of resources, thereby achieving the reconstruction of the value chain [33]. Finally, digital transformation strengthens learning and absorption capabilities. As the foundation of dynamic capabilities, it promotes the generation, transmission, and institutionalization of new knowledge [34]. This not only supports the construction of efficient digital decision-making processes [35] but also drives the organization towards platform-based and agile management transformation, reshaping capability structures and driving business model innovation, significantly enhancing organizational flexibility and response speed in dynamic environments [36,37].

In summary, the improvement of dynamic capabilities enables enterprises to use digital technologies to open up resource circulation channels, precisely match human, financial, and material resources, reduce misallocation and idleness, thereby reducing costs and improving production efficiency. Based on this, we propose the hypotheses:

• H2: Digital transformation enhances enterprise Total Factor Productivity by improving dynamic capabilities.
• H2a: Digital transformation enhances enterprise Total Factor Productivity by improving organizational coordination and integration capabilities.
• H2b: Digital transformation enhances enterprise Total Factor Productivity by improving change and reconstruction capabilities.
• H2c: Digital transformation enhances enterprise Total Factor Productivity by improving learning and absorption capabilities.

Data sources

This paper uses A-share listed companies from 2010 to 2023 as the sample and performs the following screening and processing on the initial sample:

1. Delete samples where total assets are less than total liabilities;
2. Delete financial enterprises;
3. Delete ST and *ST enterprises;
4. Delete samples with an enterprise age of less than 1 year;
5. Exclude samples with missing key variables and apply winsorization at the 1% level on both sides for continuous variables.

Construction of digital transformation indicators

Given that traditional dictionary methods struggle to mine deep text semantics and suffer from insufficient measurement validity, this paper employs the LDA (Latent Dirichlet Allocation) topic model to perform natural language processing on the annual report texts of listed companies. We construct a measurement path of “data collection—semantic analysis—element extraction” to accurately depict the “diverse direction” characteristics of enterprise digital transformation. The overall process is shown in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. LDA-based digitization level indicator construction process.

https://doi.org/10.1371/journal.pone.0347212.g001

LDA topic model training and identification.

The Gensim package is used to build the LDA topic model, with Topic Coherence Score as the optimization criterion [38]. By traversing different numbers of topics and evaluating their coherence scores, the optimal number of topics is determined to be 50 (as shown in Fig 2). Subsequently, the 50 initial topics identified are manually semantically encoded to extract the Top-N keywords and their probability distributions for each topic.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Corporate Annual Reports – Change in Number of Themes.

https://doi.org/10.1371/journal.pone.0347212.g002

Topic mapping mechanism based on “TF-IDF weighted cosine similarity”.

To accurately classify the 50 initial LDA topics into specific digital technology directions, this paper constructs a mapping mechanism based on the Vector Space Model, with the following steps:

Step 1: Construction of Direction Feature Vectors (Seed Lexicon Vectorization). Drawing on the research of [10] and [23], this paper deconstructs digital transformation into five specific directions: Artificial Intelligence, Big Data, Cloud Computing, Blockchain, and Digital Technology Application. For each direction, a “seed lexicon” comprising high-frequency technical terms is compiled. The TF-IDF (Term Frequency-Inverse Document Frequency) algorithm is then employed to transform the seed lexicon of each direction into a feature vector, (j = 1,...,5), within a high-dimensional space. This vector not only accounts for the frequency of keyword occurrences but also reduces the weight of generic terms through inverse document frequency, thereby highlighting the unique technical semantics inherent to each direction.

Step 2: Topic Vectorization and Similarity Calculation. For the 50 initial topics identified by the LDA model, the Top-N keywords and their respective probability distributions are extracted for each topic. Similarly, TF-IDF weighting is applied to generate topic feature vectors, (k = 1,...,50). Subsequently, the Cosine Similarity algorithm is utilized to calculate the cosine of the angle between each topic vector and the five direction feature vectors, serving as a measure of semantic proximity. The calculation formula is as follows:

(1)

Where S_k,j denotes the similarity score between the k-th topic and the j-th digital direction, with a value range of [0, 1]. A value closer to 1 indicates a higher semantic alignment between the topic and that specific digital direction. This process generates a 50×5 similarity matrix.

Step 3: Mapping Rules and Threshold Screening. For the generated 50×5 similarity matrix, this paper implements the following mapping and denoising protocols. It is important to note that rather than relying on subjective “intermediate-level clustering” or simple hard-threshold elimination methods, we adopt a more robust mechanism termed “Maximum Similarity Matching + Sigmoid Weight Decay”:

Maximum Similarity Matching Principle: For each LDA topic k, we traverse its similarity scores across the five digital directions and assign it to the direction j^* yielding the highest score, defined as:

(2)

Noise Suppression via Sigmoid Function: Recognizing that certain topics may exhibit low semantic correlation with all digital directions (i.e., “noise topics”), direct weighting could introduce interference. To address this, drawing on the concept of activation functions in neural networks, this paper introduces a Sigmoid transformation to perform non-linear mapping on the original similarity scores, thereby achieving soft denoising. The specific formula is as follows:

(3)

Step 4: Indicator Aggregation and Final Calculation Formula. After determining the directional attribution of each valid topic, we aggregate the digital intensity for enterprises across each direction. For the annual report of enterprise i in year t, the degree of digital transformation belonging to direction j, denoted as Digital_i,t,j, is defined as the sum of the topic probabilities for all topics assigned to direction j within that annual report:

(4)

Where P_i,t,k represents the topic distribution probability; is an indicator function ensuring that only topics assigned to direction j are accumulated; and denotes the weight coefficient adjusted by the Sigmoid function, which automatically filters out noise interference with low relevance.

The advantages of this method are twofold: First, the automated vector matching implemented via code eliminates the subjectivity and arbitrariness inherent in manual coding. Second, by leveraging TF-IDF and Cosine Similarity, the method precisely captures deep semantic associations, thereby ensuring the objectivity, reproducibility, and transparency of the indicator construction.

Dependent variable

This paper sets Enterprise Total Factor Productivity (TFP) as the dependent variable. Referencing the ACF method proposed by [39], enterprise TFP is calculated in the form of value added. Specifically, operating revenue represents total output, cash paid for purchasing goods and accepting services measures intermediate inputs, and capital and labor inputs are measured by net fixed assets and the number of employees, respectively. Value added is calculated as Total Output minus Intermediate Inputs. Additionally, in subsequent sections, the LP method [40] and OP method [41] will be used to recalculate enterprise TFP to replace the dependent variable for robustness checks.

Independent variables

Theoretical basis and connotation of dimension division.

The division of digital transformation directions in this study is mainly based on the “Technology-Application” dual-layer framework proposed by [10], referencing the expansion and refinement of digital technology keywords by [23] and [11]. Specifically, the four underlying technologies of Artificial Intelligence, Big Data, Cloud Computing, and Blockchain (i.e., the “ABCD” architecture) and the fusion layer of Digital Technology Application together constitute the five digital transformation directions of this paper (see Table 1).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Business Transformation Directions and Theme Words.

https://doi.org/10.1371/journal.pone.0347212.t001

This five-dimensional classification is adopted primarily based on two considerations. First, the need for policy alignment and practical representativeness. This classification system not only closely echoes the background of the rapid development of the global digital economy but also effectively captures the real digital strategic orientations reflected in formal information disclosures such as corporate annual reports. However, existing research indicates that different digital technologies have significant heterogeneity in their impact on corporate performance [10]; therefore, measuring digital levels solely with a composite index can easily obscure the asymmetry of the action mechanisms of various technology paths, leading research into an “average number trap.” Second, aiming to crack the black box of the “Solow Productivity Paradox.” This paper retains and operationalizes this multi-dimensional structure with the core purpose of identifying the differentiated roles of different technology directions in activating enterprise dynamic capabilities and thereby enhancing TFP. Specifically, Artificial Intelligence may drive innovation more through “intelligent manufacturing integration,” while Big Data focuses on optimizing management efficiency. Thus, this subdivided perspective helps provide finer empirical evidence for understanding the complex economic consequences of digital transformation.

Mediating variables

As a core organizational routine for enterprises to perceive environmental changes, integrate and reconfigure resources, and continuously create competitive advantages [25], dynamic capability is essentially a high-order, implicit, and highly context-dependent management process. Its intrinsic complexity determines that direct and comprehensive empirical measurement faces significant challenges. Although questionnaire surveys can capture managers’ subjective perceptions of capabilities to some extent, they are difficult to apply to causal inference research with large samples and long panels due to limitations in sample size, common method bias, and data timeliness [42]. Mainstream existing literature generally uses observable objective indicators as proxy variables for dynamic capabilities. While unable to fully restore their procedural essence, they offer important advantages in controlling endogeneity and improving external validity. Therefore, this paper constructs panel data structures and, referencing the research ideas of [43], realizes the measurement of dynamic capabilities based on financial indicators.

Acknowledging the aforementioned methodological limitations, and based on dual considerations of theoretical logic and empirical feasibility, this paper deconstructs dynamic capabilities into three dimensions: Organizational Coordination and Integration Capability, Change and Reconstruction Capability, and Learning and Absorption Capability. Proxy variables with sufficient theoretical support are selected for measurement, with the specific construction logic as follows:

First, Organizational Coordination and Integration Capability reflects the ability of an enterprise to efficiently allocate and coordinate internal and external resources to optimize operational efficiency [29]. The effectiveness of this capability is ultimately manifested in the improvement of resource allocation efficiency and the optimization of cost structures. Accordingly, this paper selects Return on Assets (ROA) and Operating Cost Ratio (OCR) as its proxy variables. Among them, ROA measures the efficiency of an enterprise in using total assets to create net profit, serving as a comprehensive financial representation of resource integration effectiveness; OCR directly reflects the level of internal process coordination and supply chain management, embodying the refinement of operational management. Together, from the two complementary perspectives of “output efficiency” and “cost control,” they constitute effective explicit indicators for this dimension [21,44].

Second, Change and Reconstruction Capability emphasizes the initiative of an enterprise to achieve strategic transformation through technological innovation and business model updates [33]. The formation of this capability relies on continuous R&D investment and is ultimately condensed into identifiable knowledge outputs. Therefore, this paper adopts R&D Expenditure and Total Patent Grants to construct an “Input-Output” dual-dimensional measurement system. R&D expenditure represents the resource commitment and strategic intent of the enterprise in technological change [45], while the number of patent grants is a key marker of the institutionalization and marketization of innovation outcomes [46]. This combination effectively captures the complete chain of change and reconstruction capability from resource input to value realization.

Third, Learning and Absorption Capability refers to the potential of an organization to identify, acquire, and digest external knowledge, and apply it to business practices. Its core foundation lies in the reserve of high-quality human capital. Given that the organizational learning process is difficult to observe directly, this paper, drawing on [34] and [47], uses the Proportion of Employees with Bachelor’s Degree or Above (DH) as a proxy variable. Highly educated employees usually possess stronger information processing capabilities, cross-domain knowledge integration abilities, and a willingness for continuous learning, serving as key carriers for organizational absorption of external knowledge. Although this indicator focuses on the “potential foundation” of capability rather than the “dynamic process,” under the constraint of lacking micro-behavioral data, as a structural antecedent of learning and absorption capability, it has full theoretical rationality and empirical operability.

In summary, although the dynamic capability measurement system constructed in this paper is an indirect approximation of the theoretical construct, every proxy variable is rooted in clear theoretical logic and widely supported by existing literature. By focusing on observable results after capabilities act on enterprise operations, this scheme effectively realizes the operationalization of the multi-dimensional connotation of dynamic capabilities under the premise of ensuring the feasibility of large-sample empirical research, providing a reliable data foundation for subsequently testing the mechanism path where digital “diverse direction” transformation affects TFP through dynamic capability reconstruction.

Control variables

Drawing on [11,23], and [48], this paper selects a series of enterprise-level control variables to control the impact of other potential factors on the dependent variable. The specific meanings of relevant variables are shown in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Variable Definitions.

https://doi.org/10.1371/journal.pone.0347212.t002

Baseline regression model

First, to test the direct impact of digital transformation on TFP, this paper constructs a two-way fixed effects model as the baseline regression:

(5)

Where the subscripts i and t denote the enterprise and year, respectively; TFP_it represents the Total Factor Productivity of enterprise i in year t, estimated using the ACF method proposed by [39]; and Digital_it signifies the level of digital transformation for enterprise i in year t. Controls_it refers to a set of control variables covering dimensions such as enterprise size, age, financial leverage, equity structure, and corporate governance. Their selection follows common practices in empirical research within this field [10,48] to account for the potential impact of enterprise heterogeneity. Furthermore, and represent enterprise individual fixed effects and year fixed effects, respectively, employed to control for time-invariant enterprise characteristics and macroeconomic cycle shocks; is the random error term. The estimated coefficient in the model measures the marginal impact of digital transformation on enterprise TFP. If is significantly positive, it indicates that digital transformation has a significant enhancing effect on enterprise TFP, thereby supporting the research hypothesis of this paper.

Mediating effect model

To further reveal the internal mechanism by which digital transformation affects TFP, this paper, based on the perspective of dynamic capability reconstruction, adopts the three-step method for testing mediating effects proposed by [49]. Given that the mediating variable in this paper—dynamic capability—includes five specific measurement indicators (i.e., ROA and OCR for organizational coordination and integration capability; LnRD and PN for change and reconstruction capability; DH for learning and absorption capability), to simplify expression and unify the model form, the following uses MED_it to collectively refer to the above five mediating variables. In specific empirical testing, these five indicators will be substituted sequentially for regression analysis. The mediating effect models are set as follows:

(6)

(7)

In the equations above, with the exception of the mediating variable MED_it, all other variables are defined consistently with the baseline regression model (Eq 1). The coefficient measures the impact of digital transformation on dynamic capabilities (the mediating variable); quantifies the effect of dynamic capabilities on Total Factor Productivity (TFP) after controlling for the level of digital transformation; and represents the direct effect of digital transformation on TFP after the inclusion of the mediating variable.

The logic for testing the mediating effect is as follows: First, the coefficient in Eq (1) must be statistically significant to confirm the existence of the total effect. Second, both the coefficient in Eq (2) and the coefficient in Eq (3) must be simultaneously significant, indicating that digital transformation can indirectly influence TFP by shaping dynamic capabilities. If, under these conditions, the direct effect coefficient in Eq (3) remains significant but its magnitude is smaller than , it implies that dynamic capabilities play a partial mediating role. Conversely, if becomes insignificant, it suggests a complete mediating role. Through this series of models, this paper aims to open the “black box” regarding how digital transformation affects productivity and to empirically verify whether the reconstruction of dynamic capabilities serves as a critical transmission mechanism between the two.

Descriptive statistics

Table 3 presents the descriptive statistics of the main variables. The mean of enterprise total factor productivity (TFP_ACF) is 7.873, ranging from 0 to 12.872, indicating significant TFP differences across enterprises. The mean of digital transformation intensity (Digital) is 0.073, with a range of 0.007 to 0.261, suggesting that most enterprises have a low level of digital transformation, while a few achieve high intensity, highlighting the need for further guidance to promote enterprise digital transformation.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Descriptive Statistics of Main Variables.

https://doi.org/10.1371/journal.pone.0347212.t003

Baseline estimation results of main effects

Table 4 reports the fixed effects regression results of digital transformation on enterprise TFP. This paper adopts a progressive regression strategy: In Column (1), the coefficient of Digital Transformation (Digital) is 0.997, significantly positive at the 1% level; in Column (2), after adding enterprise-level control variables, the coefficient drops to 0.665 but remains highly significant. This result is consistent with the findings of [3] and [48], supporting Hypothesis H1 of this paper, indicating that enterprise digital transformation has a robust promoting effect on TFP. From an economic significance perspective, taking Column (2) as the standard, for every 1 unit increase in the degree of digitization, TFP increases by an average of 0.665; relative to the sample mean of 7.873, this increase is approximately 8.45% (i.e., 0.665/7.873 × 100%), indicating that digital transformation not only has statistical significance but also possesses important economic value.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Baseline Regression of Main Effects.

https://doi.org/10.1371/journal.pone.0347212.t004

Testing of “diverse direction” dimensions of digital transformation

To avoid biases caused by differences in sample selection, this paper first includes these five variables simultaneously in a joint regression on the unified full sample to ensure a consistent comparison baseline. Table 5 reports the estimation results of the impact of five digital transformation directions—Artificial Intelligence (AI), Big Data (BD), Digital Technology Application (DTA), Cloud Computing (CC), and Blockchain (BC)—on enterprise TFP (TFP_ACF). On this basis, a Wald coefficient constraint test (test command) is formally used to examine the statistical significance of the differences in the impact of various digital transformation paths on enterprise productivity.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Testing of Five Dimensions of Digital Transformation.

https://doi.org/10.1371/journal.pone.0347212.t005

The results show that the promoting effect of Artificial Intelligence (AI) is the most significant, with a coefficient of 0.222 (p < 0.01), indicating that enterprise investment in the field of AI has a strong positive driving effect on improving productivity; Big Data (BD) and Digital Technology Application (DTA) also show robust positive effects, with coefficients of 0.272 (p < 0.05) and 0.188 (p < 0.01) respectively, indicating that data-driven innovation and technology integration are important pathways for efficiency improvement. However, the coefficient for Cloud Computing (CC) is −0.238 (p > 0.1) and fails the significance test, implying that at the current stage, the application of cloud computing may not yet bring productivity improvements, or there exists a “digital divide” effect—i.e., enterprises have only achieved infrastructure upgrades but have not yet completed business process reconstruction and value conversion. The coefficient for Blockchain (BC) is 0.170 (p > 0.1), also insignificant, possibly because this technology is still in an early exploration stage, and its economic value has not been fully released.

More importantly, the joint test result of these five coefficients is highly significant (F(4, 3050)=2.67, p = 0.0304), indicating that there are significant differences in the impact of different digital transformation paths on enterprise TFP. This finding supports the necessity of subdivided analysis by technology type and shows that the economic returns of digital transformation have obvious technological heterogeneity. Hypothesis H1a is established.

This finding deepens and refines the conclusions of [2] and [3] regarding the overall positive effect of digital transformation, revealing the heterogeneous characteristics of digital technology impacts. Particularly, the insignificant results of Cloud Computing and Blockchain provide new micro-evidence for explaining the ‘Solow Productivity Paradox’, which is consistent with the views of [8] and [50], i.e., there is a long ‘time lag’ from the deployment of general-purpose technologies to the generation of economic benefits, and enterprises may still be in the ‘installation period’ rather than the ‘harvest period’. In addition, this also confirms the findings of [20], i.e., the performance manifestation of emerging technologies such as blockchain highly depends on the coordination of specific governance structures, warning us not to simply equate all digital technologies.

Robustness checks

Instrumental variable method.

In constructing the instrumental variables for this study, we followed the approach of [23] and [51] by using the number of employed persons (in 10,000 persons) in urban units in the information transmission, software, and information technology services industries of the province where the enterprise is located for verification. The instrumental variable test results show that the F-statistic is 56.988, which is significantly greater than the critical values at the 10% and 15% levels, indicating the absence of weak instruments; the Kleibergen-Paap rk LM statistic is 15.890 (p = 0.0001), significantly rejecting the null hypothesis of underidentification. Table 6 reports the regression results of the instrumental variables. The first-stage regression results show that the coefficient of the instrumental variable is significantly positive, indicating that the instrumental variable is significantly positively correlated with enterprise digital transformation; the second-stage regression results show that the coefficient of digital transformation is significantly positive, indicating that after using instrumental variables to weaken the impact of endogeneity problems, the research conclusion still holds.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Heckman Two-Stage Regression Method.

https://doi.org/10.1371/journal.pone.0347212.t006

Heckman two-stage regression.

Considering that sample selection bias may also affect estimation results, this paper further uses the Heckman two-step method to correct possible bias issues. Drawing on [52], iin the first stage, we take whether a firm undertakes digital transformation as the dependent variable (1 for transformation, 0 otherwise), control for firm characteristics such as age, size, leverage, and estimate the Inverse Mills Ratio (IMR) via Probit regression. In the second stage, we include the IMR in the regression model. The results of the Heckman second stage in Column (2) of Table 7 show that after correcting for self-selection bias, the main conclusions of this paper remain valid.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Instrumental Variable Method.

https://doi.org/10.1371/journal.pone.0347212.t007

Lagging of explanatory variables.

First, considering that the impact of enterprise digital transformation may have a time lag [53], this study lags the level of digital transformation by one period and two periods respectively for regression testing. As shown in Columns (1) and (2) of Table 8, the coefficient estimates for Lag One Period (l_Digital) and Lag Two Periods (l2_Digital) are 0.362 (p < 0.05) and 0.265 (p < 0.1) respectively, both remaining statistically significant, indicating that the promoting effect of digital transformation on enterprise TFP remains significantly positive.

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Results of Lagged Explanatory Variables and PSM Regression.

https://doi.org/10.1371/journal.pone.0347212.t008

Replacing the dependent variable.

To test the robustness of the baseline regression results to TFP calculation methods, referencing studies by [23] and [54], the LP method and OP method are used to recalculate TFP. The results in Columns (1) and (2) of Table 9 show that the Digital coefficients are both significantly positive.

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Replacing Dependent and Explanatory Variables.

https://doi.org/10.1371/journal.pone.0347212.t009

Impact mechanism testing

Existing studies mainly measure dynamic capabilities via questionnaire cross-sectional data, which cannot show their evolution [42]. Thus, following [43], this paper uses panel data and financial indicators to measure dynamic capabilities as follows.

Organizational Coordination and Integration Capability. Referencing [29] and [44], Return on Assets (ROA) and Operating Cost Ratio (OCR) are selected as proxy variables for resource integration management capability. The test results in Table 10 show: First, digital transformation significantly improves enterprise ROA and significantly reduces OCR, indicating its preliminary effect of optimizing internal resource allocation. Further mediating effect tests show that after introducing Digital and ROA simultaneously in the regression, the coefficient of ROA is significantly positive; similarly, after introducing OCR, its coefficient is significantly negative. These results satisfy the identification conditions for mediating effects, confirming that digital transformation enhances organizational coordination and integration by increasing ROA and reducing OCR, which partially mediates its positive effect on TFP, thus supporting Hypothesis H2a.

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Impact Mechanism Test of Digital Transformation on Enterprise TFP.

https://doi.org/10.1371/journal.pone.0347212.t010

Change and Reconstruction Capability. This study conducts tests from two dimensions: innovation input and output. The logarithm of R&D expenditure (LNRD) and total patent grants (PN) are used as proxy variables for enterprise R&D capability and technological innovation level, respectively. As shown in Table 11, the coefficient of digital transformation on LNRD is 1.384 (p < 0.01), and the coefficient on PN is 0.279 (p < 0.05), both of which are significantly positive. This indicates that digital transformation can directly enhance firms’ R&D investment and innovation output. Further mediating effect tests show that after adding LNRD and PN to the model, their coefficients on TFP_ACF are 0.0125 and 0.0258 respectively, and the coefficient of Digital remains significant. This confirms that digital transformation effectively strengthens enterprise change and reconstruction capabilities by promoting R&D investment and improving innovation output, thereby driving productivity improvement, verifying hypothesis H2b.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Impact Mechanism Test of Digital Transformation on Enterprise TFP.

https://doi.org/10.1371/journal.pone.0347212.t011

Learning and Absorption Capability. Referencing [47], this study uses the proportion of employees with a bachelor’s degree or above (DH) as a core indicator of enterprise knowledge reserves and learning conversion potential. This indicator first reflects the proportion characteristics of the highly educated group in the enterprise’s human resource structure, and its core points to the organization’s potential knowledge reserve level and learning conversion capability. From the connotation of learning and absorption capability, this indicator indirectly reflects the enterprise’s efficiency in acquiring new knowledge and technologies, the depth of understanding, and the potential for application: a high proportion usually means the organization possesses sufficient knowledge carriers, making it easier to complete the internal transmission, integration, and institutionalization of knowledge, thereby providing a human resource foundation and execution guarantee for the exertion of learning and absorption capability in dynamic capabilities. On this basis, the knowledge formed through learning and absorption can be effectively transformed and applied to business practices, and the powerful knowledge chain built through continuous learning may also bring competitive advantages to the enterprise [29]. As shown in Table 11, the coefficient of digital transformation (Digital) on DH is 3.631 (p < 0.05), which is significantly positive. This suggests that digital transformation significantly optimizes enterprises’ human capital structure and improves the organizational potential foundation for knowledge acquisition and absorption. In the mediating test, the coefficient of DH on TFP_ACF is 0.641 (p < 0.01), and the coefficient of Digital remains significant, indicating that learning and absorption capability plays a mediating role between digital transformation and productivity improvement. Therefore, digital transformation promotes the improvement of TFP by expanding high-knowledge human resources and enhancing organizational learning and absorption capability, thus establishing Hypothesis H2c.

In summary, Hypothesis H2 is established. This result supports the views of [4] and [29] regarding resource integration and innovation chain drivers. More importantly, unlike studies such as [42] that mostly use questionnaire surveys, the empirical results of this paper based on objective financial indicators overcome common method bias and more rigorously confirm the “Digital Transformation—Dynamic Capability—Performance” chain proposed by [44].

Heterogeneity analysis

Given the significant differences among enterprises in terms of factor endowments, industry attributes, and production characteristics, the enabling effect of digital transformation on total factor productivity (TFP) may be asymmetric. To further investigate this heterogeneity, this paper conducts heterogeneity tests from two dimensions: enterprise factor intensity characteristics and industry classification, with the results presented in Table 12.

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Heterogeneity Analysis by Property Rights.

https://doi.org/10.1371/journal.pone.0347212.t012