Can Smart Manufacturing Pilot Policy Curb Firms’ Symbolic Digital Transformation? Evidence from China Using Double Machine Learning

Abstract

Symbolic digital transformation, whereby firms overstate digital initiatives through digital narratives without substantive upgrading, may undermine the developmental value of industrial digitalization. This study examines whether China’s Smart Manufacturing Pilot Policy (SMPP) curbs such behavior. Using panel data on Chinese A-share listed manufacturing firms from 2011 to 2024, we treat the staggered implementation of the SMPP as a quasi-natural experiment and estimate policy effects within a double machine learning framework. The baseline results show that the SMPP significantly reduces firms’ symbolic digital transformation (SDT), and this finding remains robust to alternative specifications and endogeneity tests. Dynamic effect analysis indicates that the policy generates a persistent restraining effect, although its marginal effect gradually declines as governance becomes more normalized over time. Mechanism analysis shows that the policy mainly works by easing financing constraints and reducing information asymmetry, while increased media attention creates a countervailing reputational incentive that may encourage SDT. Threshold analysis further reveals that the policy effect is stronger among firms with higher managerial myopia and is most pronounced when corporate opacity is moderate, but becomes insignificant once opacity exceeds a critical level. Overall, the SMPP promotes a shift from symbolic to substantive digital transformation and may provide indirect implications for sustainable manufacturing development.

1. Introduction

Sustainable manufacturing is widely recognized as a pathway for reconciling industrial growth with resource conservation, environmental protection, and long-term competitiveness. As manufacturing firms face increasing pressure to improve energy efficiency, reduce emissions, and strengthen supply-chain resilience, digital technologies have become an important driver of industrial upgrading [1]. By enabling real-time monitoring, data-driven decision making, and process optimization, digital transformation can support a more efficient, resilient, and low-carbon manufacturing system. For emerging economies undergoing industrial restructuring, the quality of digital transformation is therefore closely linked to broader sustainable development goals.

China has placed digital transformation at the center of its manufacturing upgrading strategy. Over the past decade, the country has promoted the integration of digital technologies with manufacturing through policies related to industrial internet applications, intelligent equipment, and demonstration programs. Within this broader framework, the Smart Manufacturing Pilot Policy (SMPP) is a particularly important instrument. Implemented in multiple batches, the SMPP identifies pilot firms and projects, promotes the diffusion of replicable smart manufacturing practices, and provides policy guidance and institutional support. In this sense, the policy is intended not only to accelerate technology adoption, but also to improve the substance of firm-level digital transformation.

However, digital transformation does not necessarily imply genuine upgrading. Some firms invest in digital infrastructure and organizational change, while others may strengthen digital narratives without comparable substantive improvements. This study defines the latter pattern as symbolic digital transformation (SDT): namely, a divergence between firms’ digital narratives and substantive digital practices. As a form of digital window dressing, SDT can distort external assessments of firm capability and weaken industrial policies aimed at sustainable manufacturing. Distinguishing it from substantive digital transformation is therefore essential for evaluating whether policy-driven upgrading delivers real economic and sustainability benefits.

This issue is especially salient in a policy-led context. Because the SMPP confers both policy recognition and public visibility, it may affect firms’ incentives in two opposite directions. On the one hand, it can encourage substantive digital transformation by improving access to resources and clarifying upgrading expectations. On the other hand, it may intensify firms’ incentives to manage external impressions if stakeholders respond more strongly to digital narratives than to actual transformation outcomes. Whether the SMPP ultimately restrains or induces SDT is therefore an open empirical question with important governance implications.

Although prior research has examined the productivity, innovation, and environmental effects of digital transformation, several gaps remain. Most studies focus on the outcomes of digital adoption, while less attention has been paid to symbolic responses. Research on smart manufacturing policy has mainly examined technological upgrading and operational performance, but rarely whether such policies curb SDT. In addition, the effects of policy intervention may depend on firms’ financing conditions, information environments, and managerial incentives. Financing constraints, information asymmetry, media attention, managerial myopia, and corporate opacity have not yet been integrated into a unified framework.

Against this backdrop, we investigate whether and through what channels the SMPP influences SDT among Chinese manufacturing firms. Drawing on panel data for manufacturing firms listed on China’s A-share market from 2011 to 2024, we leverage the staggered implementation of the SMPP as a quasi-natural experiment. The policy effect is estimated using a DML framework, which accommodates high-dimensional covariates and potential nonlinearities while reducing bias arising from model misspecification, thereby enhancing the credibility of the causal inference. We further conduct mechanism analysis and threshold analysis to examine the evolution, channels, and nonlinear conditions of the policy effect.

The empirical results show that the SMPP significantly inhibits SDT. This finding remains robust across alternative specifications and endogeneity checks. The dynamic analysis indicates that the restraining effect persists after implementation but gradually weakens over time. Mechanism analysis suggests that the SMPP curbs SDT mainly by easing financing constraints and reducing information asymmetry, while media attention creates a countervailing reputational incentive that partially offsets the policy’s disciplining effect.

Threshold analysis further reveals important heterogeneity. The inhibitory effect of the SMPP is stronger among firms with higher managerial myopia, indicating that policy intervention is particularly effective where short-term incentives would otherwise favor symbolic responses. The policy effect is most pronounced when corporate opacity is moderate, but becomes insignificant once opacity exceeds a critical threshold. This pattern suggests that smart manufacturing policy is less effective in excessively opaque information environments, where external monitoring and internal accountability are both weakened.

This study delivers three distinct contributions to the existing literature on industrial digital policies and corporate digital behavior. First, it extends the research scope of industrial digitalization evaluation by distinguishing SDT from substantive digital upgrading, establishing a differentiated measurement framework that captures the narrative–operational decoupling of corporate digital strategies, rather than merely measuring observable digital investment outcomes. Second, this paper provides rigorous causal evidence for the supervisory governance function of China’s SMPP. Leveraging staggered policy implementation and DML estimation, the research addresses high-dimensional confounding bias and endogenous pilot selection to reliably isolate the policy’s restraining effect on cosmetic digital signaling. Third, this work integrates financing frictions, information opacity, media scrutiny, managerial myopia and corporate opacity into an integrated analytical framework. By identifying both reinforcing and countervailing mediating channels together with state-contingent threshold effects, it enriches scholarly understanding of the heterogeneous transmission paths through which industrial policies shape firms’ opportunistic digital strategies.

2. Literature Review and Hypothesis Development

2.1. Smart Manufacturing Policy and Symbolic Digital Transformation

The evolution of the global industrial landscape has positioned smart manufacturing as a core driver for upgrading traditional production systems [2,3]. In emerging economies, governments frequently deploy pilot policies to stimulate industrial upgrading, utilizing the SMPP as a critical tool to guide enterprises toward Industry 4.0 integration [4,5]. While the technological benefits of such policies are widely acknowledged, the actual implementation at the micro-firm level often diverges from policy intentions. Driven by institutional pressures, firms may engage in SDT, which refers to the strategic decoupling of external digital rhetoric from internal substantive technological investment [6]. According to Institutional Theory, when faced with strong coercive pressures from government mandates, organizations with high legitimacy needs but insufficient capabilities tend to disclose symbolic cues rather than making substantive changes [7].

However, the SMPP is designed with rigorous evaluation mechanisms that fundamentally alter this dynamic. Unlike broad, non-binding digital advocacy, pilot policies typically involve strict ex ante screening, in-process monitoring, and ex post performance assessments. This structured governance acts as a disciplinary mechanism that realigns a firm’s strategic narrative with its actual operational execution [8]. Furthermore, the core objective of smart manufacturing is to achieve tangible improvements in production efficiency and environmental sustainability, leaving minimal room for superficial “window-dressing” [9]. When enterprises are embedded in the SMPP framework, they are compelled to integrate advanced algorithms and cyber–physical systems into their core operations, which inherently demands substantive resource commitment rather than mere strategic signaling [10]. Consequently, the stringent institutional constraints and performance verifications embedded in the SMPP significantly compress the opportunistic space for strategic decoupling, forcing firms to transition from symbolic posturing to substantive digital capability building [11]. Based on the above reasoning, this paper develops the subsequent hypothesis:

Hypothesis 1.

The SMPP significantly curbs firms’ symbolic digital transformation.

2.2. The Mediating Role of Financing Constraints

Resource Dependence Theory posits that substantive organizational transformation is highly contingent upon the continuous acquisition of external critical resources. Digital transformation is an inherently capital-intensive and highly uncertain endeavor, making Financing Constraints (FC) a primary bottleneck for manufacturing enterprises [12]. When firms face severe financial frictions, they lack the necessary capital to overhaul legacy systems or invest in core digital infrastructure [13]. Under such resource scarcity, management is highly incentivized to engage in digital transformation whitewashing—projecting a false image of digital advancement to appease stakeholders while avoiding the prohibitive costs of actual implementation [14].

The implementation of the SMPP effectively mitigates this resource dilemma through a powerful signaling effect. Government endorsement via pilot status acts as a credible certification of a firm’s technological trajectory and future profitability, which significantly reduces the risk premium demanded by external investors [15]. This policy-induced certification facilitates smoother access to both debt and equity financing channels [16]. With the alleviation of FC, enterprises are no longer forced to rely on symbolic disclosures to maintain market legitimacy; instead, they possess the financial slack required to execute substantive digital upgrades [17]. Moreover, the influx of capital allows firms to withstand the short-term performance shocks typically associated with deep technological restructuring [18]. As financial barriers are dismantled by the policy signal, the fundamental motivation for adopting symbolic strategies dissipates [19]. Derived from the aforementioned deduction, the second testable hypothesis is formalized as below:

Hypothesis 2.

The SMPP curbs firms’ symbolic digital transformation by alleviating financing constraints.

2.3. The Mediating Role of Information Asymmetry

The prevalence of SDT is deeply rooted in the principal-agent problem, which is exacerbated by Information Asymmetry (IA) between corporate insiders and external stakeholders. In traditional manufacturing environments, the opacity of operational processes allows managers to exploit information gaps, presenting exaggerated digital narratives without the risk of immediate detection [20]. High levels of IA provide a protective shield for opportunistic behaviors, enabling firms to reap the reputational benefits of digitalization while evading the substantive costs [21].

The SMPP fundamentally disrupts this opaque environment by mandating the deployment of interconnected cyber–physical systems and transparent data architectures. The transition toward smart manufacturing inherently requires the digitization of the entire supply chain and production lifecycle, which converts previously tacit operational knowledge into explicit, traceable data streams [22]. This technological shift drastically enhances accounting information transparency and operational visibility [23]. As the firm’s internal data becomes more standardized and accessible, external monitors—such as auditors, investors, and regulators—can more accurately assess the authenticity of the firm’s digital claims [24]. Furthermore, the reduction in IA limits the ability of management to manipulate project data or engage in the misrepresentation of digital responsibilities [25]. When the true state of a firm’s digital infrastructure is easily verifiable, the reputational risks and potential penalties associated with being exposed as a “digital washer” outweigh the short-term benefits of symbolic disclosure [26]. Consequently, the transparency enforced by smart manufacturing practices eliminates the informational blind spots that SDT relies upon [27]. Thus, we hypothesize:

Hypothesis 3.

The SMPP curbs firms’ symbolic digital transformation by reducing information asymmetry.

2.4. The Mediating Role of Media Attention

While the SMPP generally promotes substantive upgrading, its interaction with external public scrutiny introduces a complex behavioral paradox. As a high-profile national strategy, smart manufacturing initiatives inevitably attract intense Media Attention (MA) [28]. According to Legitimacy Theory, media acts as a powerful social magnifying glass; heightened coverage places immense pressure on firms to conform to societal and political expectations rapidly [29]. Although media scrutiny typically serves as a governance mechanism, in the context of emerging technological trends, excessive MA can generate overwhelming short-term performance pressure [30].

When a firm is selected for a pilot policy, the sudden surge in MA demands immediate and visible proof of digital progress [31]. However, substantive digital transformation is a protracted, path-dependent process that rarely yields instant results [32]. Caught between the media’s demand for rapid success stories and the slow reality of technological integration, management faces a severe legitimacy crisis [33]. To quickly satisfy public expectations and maintain market valuation, firms may be induced to prioritize superficial digital rhetoric and cosmetic technological adoption over deep, time-consuming structural changes [34]. The digital media landscape, characterized by rapid information dissemination and fragmented reporting, further amplifies this conformity pressure, incentivizing firms to construct a “transformation illusion” to feed the news cycle [35]. Therefore, while the policy itself aims for substantive change, the intense media spotlight it attracts inadvertently creates a high-pressure environment that can trigger short-term, symbolic coping mechanisms. Accordingly, we propose the final hypothesis:

Hypothesis 4.

The SMPP raises media attention, and this channel may partially offset the policy’s restraining effect by inducing firms’ SDT.

2.5. Summary of Research Hypotheses and Mechanism Framework

From the above theoretical reasoning, this work forms an integrated mechanism framework to elaborate the causal chain linking the SMPP to firms’ SDT. As shown in Figure 1, the SMPP may shape firms’ digital transformation behavior through both a direct policy effect and several underlying transmission channels. The central argument is that, by providing policy guidance, resource support, and external governance pressure, the SMPP can reduce firms’ incentives to rely on symbolic digital narratives or superficial digital investments, thereby restraining SDT. Accordingly, this study proposes that the SMPP has a significant inhibitory effect on firms’ SDT.

The SMPP affects SDT through two types of competing transmission channels with opposite directional impacts. (1) Two core supportive governance channels: alleviated financing constraints and lowered information asymmetry. These two paths fundamentally reduce firms’ incentives for digital window dressing and are the key mechanisms explaining why the policy restrains symbolic transformation. (2) One countervailing adverse spillover channel: elevated media attention. Greater public scrutiny creates short-term reputational incentives for superficial digital rhetoric, which acts only as a risk factor partially weakening the policy’s disciplinary power rather than a supportive governance mechanism. The restraining force from the two core channels dominates the mild offsetting media effect, yielding a significantly negative total policy effect without logical contradiction.

In sum, this study posits that the effect of the SMPP on SDT is shaped by multiple causal mechanisms rather than a single linear pathway. Financing constraints and information asymmetry represent restraining channels through which the SMPP suppresses SDT, whereas media attention represents a partially offsetting channel that may induce SDT.

SMPP exerts a direct negative restraining effect on SDT (H1). H2 and H3 are reinforcing mediation channels: SMPP alleviates financing constraints and reduces information asymmetry; higher financing constraints and information asymmetry positively promote SDT, so the two channels indirectly strengthen the policy’s inhibitory effect. H4 is a countervailing mediation channel: SMPP raises media attention, and greater media attention positively boosts cosmetic digital signaling, partially offsetting the policy’s overall curbing effect.

3. Research Design

3.1. Data Sources and Sample Selection

This study takes Chinese A-share listed manufacturing firms over the period 2011–2024 as the initial research sample. Information on firms participating in the SMPP is manually collected from the demonstration project lists released by the Ministry of Industry and Information Technology (MIIT) and then matched to listed firms. Firm-level financial and corporate governance data are mainly drawn from the China Stock Market and Accounting Research (CSMAR) database and supplemented with information disclosed in firms’ annual reports. Regional-level variables are obtained from the China City Statistical Yearbook.

To enhance the reliability of the empirical analysis, several screening procedures are applied to the original sample. First, firms subject to delisting risk warnings or other abnormal trading conditions are excluded, as such firms are more likely to exhibit distorted financial reporting and atypical operating behavior. Second, observations from the year of initial public offering are removed in order to reduce disturbances associated with substantial changes in capital structure, financing arrangements, and cash flow conditions during the listing stage. Third, firm-year records lacking complete data on core financial and governance indicators are eliminated to guarantee the integrity and uniformity of the panel sample. Finally, all continuous variables are winsorized at the 1st and 99th percentiles to alleviate the influence of extreme values on the empirical results.

After these procedures, the final sample comprises 28,825 firm-year observations. Data cleaning, variable construction, and baseline econometric analysis are conducted in Stata 19.0, while Python 3.12 is used for high-dimensional algorithmic estimation and the visualization of dynamic treatment effects.

3.2. Variable Measurement

3.2.1. Dependent Variable: Symbolic Digital Transformation (SDT)

Drawing on prior relevant research [36], this section employs institutional theory, organizational legitimacy theory, impression management theory and signaling theory to define symbolic digital transformation (SDT), while demarcating its theoretical boundaries relative to substantive digital transformation and other analogous corporate strategic behaviors.

From the institutional theory perspective, substantive digital transformation refers to sustained and substantial corporate resource allocation to digital infrastructure construction, production process upgrading and internal organizational restructuring. Firms implement long-run digital upgrading initiatives to comply with industry-wide digitalization institutional norms, achieve sustained productivity gains and acquire substantive operational legitimacy. Under such circumstances, corporate resource input, daily operational practices and public disclosure information maintain internal consistency.

In contrast, SDT represents a typical decoupling practice between public disclosure and actual business operations, driven by firms’ demand for superficial institutional legitimacy. To cater to policy and market expectations, enterprises release digital-related statements and textual disclosures without matching tangible digital capital investment or internal operational adjustments, leaving core production and operation models fundamentally unchanged. From the organizational legitimacy theory perspective, the primary objective of SDT is to obtain superficial recognition from regulators and market participants rather than realize fundamental improvements in operational efficiency.

To resolve conceptual ambiguity, this study further differentiates SDT from five closely interrelated constructs based on impression management and signaling theories. First, digital disclosure is a neutral umbrella concept encompassing all public information regarding corporate digital development, and SDT constitutes an opportunistic subset of digital disclosure lacking substantive digital resource backing. Second, impression management denotes a general theoretical framework covering all behaviors designed to shape external stakeholders’ perceptions, whereas SDT is a targeted impression management strategy tailored to industrial digital policy supervision and capital market investors. Third, greenwashing refers to decoupled environmental information disclosure induced by environmental regulation; digital washing shares identical decoupling logic with SDT yet is confined to environmental governance contexts, while SDT centers on industrial digitalization policy shocks, thereby forming distinct domain boundaries between the two constructs. Fourth, classical strategic signaling theory posits that credible signals rely on costly, observable real inputs to reflect firms’ intrinsic quality. By contrast, SDT generates low-cost, unsubstantiated pseudo-signals that distort the standard transmission logic of credible information.

Four core theoretical paradigms jointly constitute the theoretical foundation of SDT. Institutional theory explains the external institutional pressure that motivates firms to decouple digital disclosures from real operational activities. Organizational legitimacy theory identifies the core incentive underlying SDT: the pursuit of surface compliance legitimacy. Impression management theory elaborates the information-manipulation mechanism through which firms reshape external stakeholders’ perceptions. Signaling theory further characterizes SDT as a distorted, noncredible form of symbolic pseudo-signaling.

Specifically, the construction of SDT involves two components. The first is corporate digital disclosure (CDD), which captures the intensity of digital transformation signals released by firms. Following Zhao et al. (2021) [37], this study measures firms’ digitalization based on the frequency of 99 digitalization-related terms across four dimensions. The specific procedure is as follows. First, we collected the annual reports of manufacturing listed firms from 2011 to 2024 and converted them into machine-readable text. We then used Python to extract the text from the Management Discussion and Analysis section. Second, using a segmentation dictionary (see Appendix B 

Table A1. Details on the construction of the CDD index and keyword selection.

Dimension	Category Terms	Dictionary Terms
Digital Technology Applications	data, digital, digitization	data management, data mining, data networks, data platforms, data centers, data science, digital control, digital technology, digital communications, digital networks, digital intelligence, digital terminals, digital marketing, digitization, big data, cloud computing, cloud IT, cloud ecosystem, cloud services, cloud platforms, blockchain, Internet of Things, machine learning
Internet Business Model	Internet, e-commerce	mobile internet, industrial internet, industry internet, Internet solutions, Internet technology, Internet thinking, Internet action, Internet business, Internet mobility, Internet applications, Internet marketing, Internet strategy, Internet platforms, Internet model, Internet business model, Internet ecosystem, e-commerce, electronic commerce, Internet, “Internet+”, online and offline, online-to-offline, online-to-online, O2O, B2B, C2C, B2C, C2B
Intelligent Manufacturing	intelligence, intelligentization, automation, digitization, integration	artificial intelligence, advanced intelligence, industrial intelligence, mobile intelligence, intelligent control, intelligent terminals, intelligent mobility, intelligent management, smart factories, intelligent logistics, intelligent manufacturing, intelligent warehousing, intelligent technology, intelligent equipment, intelligent production, industrial IoT, intelligent systems, intelligentization, automatic control, automatic monitoring, automatic inspection, automated production, digital control, integration, integrated solutions, integrated control, integrated systems, industrial cloud, future factory, intelligent fault diagnosis, lifecycle management, manufacturing execution system, virtualization, virtual manufacturing
Modern Information Systems	information, informatization, networking	information sharing, information management, information integration, information software, information systems, information networks, information terminals, information centers, informatization, networking, industrial information, industrial communications

for details on the construction of the CDD index and keyword selection), we applied the Jieba word segmentation tool to all sample texts and counted the disclosure frequency of keywords from four dimensions: digital technology application, internet-based business models, intelligent manufacturing, and modern information systems. Based on the resulting term-frequency data, we standardized the indicators and employed the entropy weighting method to determine the weight of each indicator, thereby constructing the CDD index.

The second component is corporate digital performance (CDP), which is intended to capture firms’ substantive digital transformation rather than their disclosure behavior. In contrast to CDD, which is typically measured through text analysis of annual reports or other disclosure documents, CDP is not constructed from textual data. Instead, this study adopts the corporate digital transformation index provided by the CSMAR database. The index is calculated as a weighted composite measure based on six dimensions: strategic guidance, technology-driven development, organizational empowerment, environmental support, digital outcomes, and digital application. Accordingly, it offers a more comprehensive and objective assessment of firms’ actual digital transformation efforts and achievements. The specific indicators and corresponding weights are reported in Appendix B 

Table A2. Breakdown of CDP Sub-indicators and Respective Weight Values.

Primary Indicator	Primary Indicator Weight	Secondary Indicator	Secondary Indicator Weight
Strategic Leadership	34.72%	Establishment of Digital Management Roles at the Management Level	23.82%
		Forward-looking Nature of Management-Level Digital Innovation Guidance	27.88%
		Continuity of Management-Level Digital Innovation Guidance	18.79%
		Breadth of Management-Level Digital Innovation Guidance	12.83%
		Intensity of Management-Level Digital Innovation Guidance	16.68%
Technology-Driven	16.20%	Artificial Intelligence Technology	55.04%
		Blockchain Technology	12.98%
		Cloud Computing Technology	18.32%
		Big Data Technology	13.66%
Organizational Empowerment	9.69%	Digital Capital Investment Plan	50.22%
		Digital Human Capital Investment Plan	25.53%
		Digital Infrastructure Development	12.06%
		Technology Innovation Base Development	12.19%
Environmental Support	3.42%	Number of Invention Patents in the Industry	19.23%
		R&D Activity in the Industry	17.79%
		New Product Development and Sales in the Industry	14.98%
		Intensity of Digital Technology in the Industry	11.57%
		Intensity of Digital Capital Input in the Industry	11.40%
		Intensity of Human Capital Input in the Industry	7.89%
		Fiber Optic Density in the City	4.77%
		Mobile Switching Capacity in the City	4.03%
		Scale of Fixed Internet Broadband Access Users in the City	4.00%
		Scale of Mobile Internet Users in the City	4.34%
Digital Achievements	27.13%	Digital Innovation Standards	36.68%
		Digital Innovation Papers	11.74%
		Digital Invention Patents	23.54%
		Digital Innovation Qualifications	14.73%
		Digital National Awards	13.31%
Digitalization Application	8.84%	Technological Innovation	63.42%
		Process Innovation	23.78%
		Business Innovation	12.80%

.

Finally, SDT is measured as the standardized difference between CDD and CDP. Because the textual disclosure indicator and the composite digital transformation index are constructed on different scales, they are standardized before the difference is calculated. The measurement model is specified as follows:

$$SD{T}_{it}=\left({\displaystyle \frac{CD{D}_{it}-\overline{CDD}}{{\sigma }_{CDD}}}\right)-\left({\displaystyle \frac{CD{P}_{it}-\overline{CDP}}{{\sigma }_{CDP}}}\right)$$

(1)

where SDT_it denotes the level of SDT of firm i in year t; CDD_it represents the firm’s digital disclosure intensity; and CDP_it captures its substantive digital transformation performance. $\overline{CDD}$ and $\overline{CDP}$ are the sample means of the corresponding variables, while ${\sigma }_{CDD}$ and ${\sigma }_{CDP}$ denote their standard deviations. A positive SDT value indicates that digital disclosure exceeds substantive digital performance, suggesting a stronger tendency toward SDT. By contrast, a negative SDT value indicates that substantive digital performance exceeds digital disclosure. This should not be interpreted simply as a lower degree of symbolic transformation. Rather, it reflects a disclosure lag relative to substantive progress, implying that firms’ actual digital upgrading is not fully reflected in their external communication. In this sense, SDT captures asymmetry between disclosure and performance, with positive values reflecting disclosure surplus and negative values reflecting disclosure deficiency.

3.2.2. Independent Variable: Smart Manufacturing Pilot Policy (SMPP)

As a core supporting initiative under the national strategy Made in China 2025, the Intelligent Manufacturing Pilot Demonstration Program was officially launched by China’s Ministry of Industry and Information Technology (MIIT) in 2015. Between 2015 and 2018, MIIT publicly released one batch of pilot intelligent manufacturing entities each year through official circulars, selecting enterprises with mature digital transformation foundations as demonstration benchmarks to promote the nationwide rollout of smart manufacturing technologies and models. For empirical analysis in this paper, we manually compile the complete roster of pilot firms announced across all four batches. We then match these pilot entities with the dataset of A-share listed corporations by full enterprise name to identify treated sample firms. After data cleaning and matching procedures, we finally obtain a balanced sample consisting of 85 listed pilot companies. Specifically, the first batch includes 12 listed pilot firms, the second batch contains 19, the third batch covers 26, and the fourth batch comprises 28 listed pilot enterprises. Based on this procedure, the smart manufacturing pilot policy variable (SMPP) is constructed as a policy indicator. Specifically, for firm i, SMPP equals 1 from the year in which the firm is first selected into the pilot program onward, and 0 otherwise. Thus, firms that are never selected remain in the control group throughout the sample period, while treated firms switch from 0 to 1 beginning in the year of policy entry.

3.2.3. Mediators: Financial Constraints (FC)

Financial constraints are measured using the SA index. Compared with conventional proxies constructed from potentially endogenous financial variables, the SA index is less vulnerable to reverse causality and is therefore more suitable for the present analysis. A larger absolute value of the SA index indicates more severe external financing constraints faced by the firm [38].

3.2.4. Mediators: Information Asymmetry (IA)

Information asymmetry is measured following the market microstructure literature and related studies [39]. Using high-frequency stock trading data, this study applies principal component analysis to extract the first principal component from three indicators: the illiquidity ratio, liquidity ratio, and return reversal measure. The resulting composite index is used to proxy for the degree of information asymmetry between the firm and outside investors. A higher value indicates a more pronounced information asymmetry problem.

3.2.5. Mediators: Media Attention (MA)

Media attention is measured following Wang (2022) [40]. Specifically, the annual number of news reports related to a listed firm in an authoritative news database is used to capture the intensity of external media scrutiny. To address the severe right skew of the raw indicator, we apply a log transformation by taking the natural log of one plus the total volume of news coverage in all regression analyses.

3.2.6. Control Variables

To reduce risks stemming from omitted variable bias, we incorporate multiple firm and regional characteristic controls following the measurement frameworks documented in prior relevant studies [4,41,42].

Firm-level control variables cover a variety of financial operational and corporate governance indicators. Specifically, firm size (size) equals the natural logarithm of firms’ year-end total assets; total asset turnover (tat) is computed as operating revenue divided by average total assets; financial leverage (lev) is defined as the ratio of total liabilities to total assets; and liquidity ratio (liq) reflects the proportion of current assets relative to current liabilities. Revenue growth (growth) represents the annual growth rate of operating revenue. Furthermore, return on equity (roe) proxies for corporate profitability, while Tobin’s Q (tobin) captures firms’ market valuation and future investment potential. In terms of governance features, board size (boardsize) is calculated as the natural logarithm of the total number of board members, and board independence (indep) refers to the percentage of independent directors within the board of directors.

To mitigate spatial spillover and endogenous sorting of high-capability firms into pilot industrial clusters, the high-dimensional covariate set of our DML model incorporates a series of time-varying city-level regional indicators with linear and quadratic terms. For regional attributes, two additional controls are incorporated: industrial structure (industry), calculated as the proportion of secondary and tertiary industry aggregate value-added within regional GDP, and regional economic quality (gdp_per), represented by local per capita GDP. Combined with firm and year two-way fixed effects, these regional covariates absorb time-varying locational advantages and unbalanced regional development gaps that attract high-quality firms to agglomerate in pilot zones, partially balancing systematic capability gaps between treated and untreated firms across pilot and non-pilot cities.

3.3. Model Specification

3.3.1. Baseline Model

This paper adopts a partially linear DML framework to quantify the causal impact exerted by the SMPP on firms’ SDT [43]. Compared with conventional linear estimators, the DML approach is better suited to high-dimensional settings and can flexibly account for complex nonlinear relationships between covariates and the outcome variable. Therefore, to mitigate concerns about non-random policy selection and high-dimensional confounding, this study employs a DML framework to flexibly control for a rich set of observed covariates. However, DML is used here as a bias-reduction tool for observed confounders rather than as a stand-alone identification strategy. Meanwhile, we adopt orthogonal K-fold cross-fitting for nuisance function training. All folds are partitioned by firm rather than individual firm-year observations; all time-series data of a single firm stay within one fold. This design prevents within-firm information leakage caused by splitting one firm’s multi-year observations across training and validation samples. For variance estimation, this paper uniformly computes firm-clustered standard errors to address serial correlation and heteroskedasticity within each firm. The baseline model is specified as follows:

$$SD{T}_{it}=\beta SMP{P}_{it}+f\left(X_{it}\right)+U_{it}$$

(2)

$$E\left(U_{it}\left|SMP{P}_{it},X_{it}\right.\right)=0$$

(3)

In Equation (2), SDT_it denotes the level of SDT for firm i in year t, and SMPP_it is the treatment indicator for the SMPP. The coefficient β is the parameter of primary interest and captures the average treatment effect (ATE) of the policy on firms’ SDT. A significantly negative estimate of β indicates that the policy effectively curbs firms’ SDT. X_it stands for a comprehensive collection of high-dimensional concurrent controls with linear and quadratic specifications, whose inclusion eliminates confounding factors that simultaneously correlate with policy participation decisions and firms’ digital strategic behaviors. The baseline model further incorporates two-way fixed effects, namely year fixed effects and firm fixed effects, to control for time-invariant unobserved heterogeneity at the firm level and common temporal shocks. Accordingly, the high-dimensional control variables are transformed through the within-estimator demeaning procedure before estimation, so that identification is obtained from within-firm variation net of year-specific common factors. To further alleviate potential endogeneity concerns, this paper carries out an extra robustness test by lagging all control variables one period in the baseline regression specification. This alternative specification helps mitigate the possibility that contemporaneous adjustments in firm characteristics may be correlated with current outcomes, thereby reducing concerns regarding reverse causality and simultaneity. The unknown function f(∙) captures complex nonlinear effects of the covariates, while U_it is the stochastic error term.

A direct application of machine-learning algorithms to Equation (2) may introduce regularization bias. Although regularization improves predictive performance and helps avoid overfitting in high-dimensional contexts, it may also bias the estimation of the target parameter. To address this issue, the DML framework further models the treatment assignment process through the following auxiliary equation:

$$SMP{P}_{it}=m\left(X_{it}\right)+V_{it}$$

(4)

$$E\left(V_{it}\left|X_{it}\right.\right)=0$$

(5)

In Equation (4), m(X_it) represents the conditional expectation of treatment assignment given the high-dimensional covariates, which can be interpreted as the propensity score or its nonlinear prediction component. V_it is the orthogonalized residual. Estimation proceeds following the Neyman orthogonalization principle. Specifically, the sample is partitioned using K-fold cross-fitting. In the training subsamples, random forests are used to estimate f(∙) and m(∙). In the corresponding validation subsamples, the predicted nonlinear components are partialled out from both the outcome and treatment variables to obtain orthogonalized residuals. The residualized outcome is then regressed on the residualized treatment by ordinary least squares, yielding an estimate of the treatment effect that is robust to regularization bias and asymptotically normal.

3.3.2. Causal Mediation Analysis

To further unpack the “black box” through which the SMPP restrains firms’ SDT, it is necessary to examine the underlying transmission mechanisms. Conventional mechanism tests commonly rely on the stepwise regression approach, often referred to as the three-step method. However, this strategy suffers from inherent methodological limitations. On the one hand, it is not well suited to addressing endogeneity bias arising from omitted unobservable factors. On the other hand, it relies heavily on strict linear and additive assumptions, which may distort the identification of the underlying causal channels.

To overcome these limitations, this study draws on recent advances in the literature [44,45,46] and constructs a more general causal mediation analysis framework based on the factual and potential outcomes approach. This framework makes it possible to decompose the total effect of policy intervention into an indirect effect transmitted through the mediator and a direct effect that operates independently of the mediating pathway. In this way, the analysis provides a more rigorous basis for identifying how the SMPP affects firms’ SDT through different causal channels.

Let SMPP_i denote the binary treatment status indicating whether firm i is selected into the smart manufacturing pilot program, where SMPP_i ∈ {0, 1}. Let M_i (SMPP) denote the potential value of the mediator under treatment status SMPP. The potential outcome of SDT is denoted by SDT_i (SMPP,m), which depends on both treatment status SMPP and mediator status m. Under this framework, the average treatment effect (ATE), denoted by Δ, is defined as:

$$\Delta =E\left[SD{T}_i(1,M_i(1))-SD{T}_i(0,M_i(0))\right]$$

(6)

If Δ is statistically significant, the policy exerts a meaningful effect on firms’ SDT. Building on this, the average causal mediation effect (ACME), denoted by δ(SMPP), is defined as the change in the outcome induced solely by the shift in the mediator from M_i (0) to M_i (1), while holding treatment status constant:

$$\delta (SMPP)=E\left[SD{T}_i(SMPP,M_i(1))-SD{T}_i(SMPP,M_i(0))\right]$$

(7)

Similarly, the average direct effect (ADE), denoted by θ(SMPP), is defined as the change in the outcome attributable solely to the variation in treatment status, while holding the mediator constant at M_i(SMPP).

$$\theta (SMPP)=E\left[SD{T}_i(1,M_i(SMPP))-SD{T}_i(0,M_i(SMPP))\right]$$

(8)

Accordingly, the total effect can be decomposed as Δ = θ(1) + δ(0) = θ(0) + δ(1). In conventional mediation analysis, the counterfactual distributions involved in these quantities are often simulated using quasi-Bayesian Monte Carlo methods. To further account for high-dimensional and nonlinear confounding, this study incorporates DML into the mediation framework. Under the partially linear specification adopted here, the treatment and mediator channels are assumed to be separable, without explicitly modeling their nonlinear interactions. This specification helps avoid the curse of dimensionality and facilitates a clearer decomposition of the policy effect.

3.3.3. Panel Threshold Model

To further examine whether the effect of the SMPP on firms’ SDT exhibits nonlinear heterogeneity, this study adopts a panel threshold model based on the classical threshold regression framework [47]. This approach allows the policy effect to vary across different regimes defined by firm-level characteristics. This paper takes the single-threshold setting for illustration and constructs the corresponding regression model below.

$$SD{T}_{it}=\alpha +{\beta }_{1}SMP{P}_{it}\times I(q_{it}\le \gamma )+{\beta }_{2}SMP{P}_{it}\times I(q_{it}>\gamma )+{\displaystyle \sum \theta }Control{s}_{it}+{\epsilon }_{it}$$

(9)

In Equation (9), q_it denotes the threshold variable, which in this study includes managerial myopia (MM) and corporate opacity (CO). The parameter γ represents the threshold value, which is estimated endogenously from the data. I(∙) is an indicator function that equals 1 when the condition in parentheses is satisfied and 0 otherwise. Controls_it includes year fixed effects, firm fixed effects, and a set of time-varying control variables. ε_it is the error term. If multiple thresholds are identified, the model can be extended accordingly by dividing the sample into additional regimes. In this study, the threshold variable is measured contemporaneously to capture the firm’s state at the time when the policy effect materializes. This specification is consistent with the threshold regression framework, in which the threshold variable functions as a state-contingent conditioning factor rather than a post-treatment outcome. Accordingly, the estimated threshold effects should be interpreted as heterogeneity in the policy impact across different firm states, rather than as a causal effect of the threshold variable itself. It should be emphasized that, to mitigate concerns about endogenous regime classification and omitted-variable bias, the empirical specification further incorporates firm and year fixed effects, together with a full suite of time-varying covariates to improve estimation accuracy.

4. Empirical Results and Analysis

4.1. Descriptive Statistics

Table 1. Descriptive Statistics.

Variable	Obs	Mean	SD	Min	Max
SDT	28,825	−0.0003	1.0915	−4.5418	27.2991
SMPP	28,825	0.0219	0.1463	0.0000	1.0000
FC	28,825	−3.8511	0.2673	−4.5479	−3.2948
IA	28,825	−0.3537	0.5705	−2.1485	1.2178
MA	28,825	4.9618	1.3137	0.0000	8.0737
size	28,825	22.092	1.1757	17.6413	27.6377
tat	28,825	0.6205	0.3880	−0.0583	7.6092
lev	28,825	0.3845	0.1957	0.0071	0.9958
liq	28,825	2.9895	4.0829	0.0947	204.7421
growth	28,825	0.2885	8.3294	−1.4449	944.0996
boardsize	28,825	2.0990	0.1941	1.3863	2.8904
indep	28,825	37.739	5.4992	14.2900	80.0000
tobin	28,825	2.0421	1.6554	0.6212	122.1895
roe	28,825	0.0627	4.4431	−66.5353	713.2036
industry	28,825	53.3423	10.2609	29.7000	85.3000
gdp_per	28,825	92,840.02	40,734.57	16,413	228,167

reports the descriptive statistics for the main variables used in this study. Several features of the data merit attention. First, SDT shows substantial cross-firm variation. Although its mean is close to zero, the standard deviation reaches 1.0915, and the sample displays notable divergence across extreme values. This pattern suggests that listed manufacturing firms differ markedly in the extent to which they engage in SDT, thereby providing sufficient variation for identifying the governance effect of the SMPP. Second, the mean value of SMPP is 0.0219, indicating that 2.19% of the full-sample firm-year observations are covered by the policy during the sample period. This proportion is broadly consistent with the selective and gradual rollout of pilot-based industrial policies in China. Finally, the control variables generally fall within plausible ranges and exhibit adequate dispersion, suggesting that the sample is suitable for subsequent empirical analysis.

4.2. Correlation Analysis and Multicollinearity Diagnostics

To examine the pairwise relationships among the main variables and assess potential multicollinearity, this study reports a correlation heatmap and variance inflation factors (VIFs). As shown in Figure 2, SMPP is negatively correlated with SDT, and the correlation is statistically significant at the 1% level. This pattern provides preliminary evidence consistent with the baseline expectation that the SMPP may help restrain firms’ SDT. In addition, several control variables are significantly correlated with SDT, indicating that firms’ SDT is associated with a broad set of firm-level and regional characteristics. These patterns further support the inclusion of a rich set of controls in the empirical analysis.

Table 2. VIF Results.

Variable	VIF	1/VIF
gdp_per	2.31	0.43
Industry	2.30	0.43
lev	1.68	0.60
boardsize	1.62	0.62
indep	1.51	0.66
size	1.45	0.69
liq	1.40	0.72
tat	1.07	0.94
tobin	1.07	0.94
SMPP	1.03	0.97
roe	1.00	1.00
growth	1.00	1.00
Mean VIF	1.45

reports the VIF results. The maximum VIF is 2.31, the minimum is 1.00, and the mean VIF is 1.45, all of which are well below the conventional threshold. Although some moderate correlations exist among individual covariates—particularly between Industry and gdp_per, and between boardsize and indep—the overall results suggest that multicollinearity is not severe and is unlikely to materially affect coefficient estimation. Therefore, the baseline regression results should not be driven by collinearity among the explanatory variables.

At the same time, Figure 2 also suggests that the relationships among SDT and the covariates are not uniformly linear. For this reason, this study adopts the DML framework in the baseline analysis. By orthogonalizing the treatment variable and outcome variable with respect to high-dimensional controls, the DML approach helps reduce regularization bias and relaxes restrictive functional-form assumptions in conventional linear regressions. This is particularly useful in the present context, where firms’ SDT may be shaped by complex interactions across firm attributes, governance conditions, and regional environments.

4.3. Baseline Regression Results

A partially linear DML model is applied to identify the benchmark effect of the SMPP on SDT. We follow the above workflow, utilizing random forest and five-fold cross-validation to compute out-of-sample predictions and orthogonal treatment effects, with all estimation results summarized in

Table 3. Baseline regression results.

Variables	(1)	(2)	(3)	(4)
	SDT	SDT	SDT	SDT
SMPP	−0.669 ***	−0.211 ***	−0.164 ***	−0.148 ***
	(−16.483)	(−5.231)	(−3.946)	(−3.584)
Linear Control Terms	No	Yes	Yes	Yes
Quadratic Control Terms	No	No	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes
Firm Fixed Effects	Yes	Yes	Yes	Yes
Observations	28,825	28,825	28,825	25,558

Notes: *** denotes statistical significance at the 1% level. Values in parentheses are t-statistics.

. Column (1) merely controls firm and year fixed effects in residualization. The policy coefficient is negative and significant at the 1% threshold, revealing that the program effectively curbs symbolic digital transformation for targeted manufacturers.

To further mitigate potential confounding bias, Columns (2) and (3) progressively expand the high-dimensional feature set used in the machine learning stage. Specifically, Column (2) introduces linear terms of all control variables into the regression, whereas Column (3) further supplements corresponding quadratic terms. As the specification becomes more stringent, the absolute magnitude of the estimated coefficient declines, but the coefficient on SMPP remains significantly negative at the 1% level throughout. In the most demanding specification shown in Column (3), the estimated treatment effect is −0.164. This finding suggests that the SMPP exerts a robust restraining effect on firms’ SDT, thereby providing support for Hypothesis 1.

Taken together, the stepwise estimation results show that the negative effect of the SMPP on SDT is not sensitive to the enrichment of controls within the orthogonalized DML framework. This pattern strengthens the credibility of the baseline finding and suggests that the policy’s governance effect is not merely driven by omitted observable characteristics. In substantive terms, the evidence indicates that the SMPP helps curb firms’ tendency to engage in symbolic rather than substantive digital transformation.

To mitigate endogeneity risks stemming from simultaneous correlation between covariates and the explained variable, we re-run the baseline specification with all control variables lagged by one year. As shown in Column (4) of Table 3, the estimated coefficient on SMPP remains significantly negative and economically meaningful, indicating that the inhibitory effect of SMPP on SDT is robust to the use of lagged covariates.

This result suggests that the baseline finding is not driven by the simultaneous adjustment of firm characteristics and digital transformation outcomes in the same year. By using lagged controls, the specification better captures the dynamic influence of prior firm conditions on current SDT while simultaneously lowering estimation bias driven by reverse causality and unobserved omitted factors. Therefore, the persistence of the main effect under this more conservative setting provides further evidence that the estimated relationship is stable and not sensitive to the timing of the control variables.

4.4. Robustness Checks

4.4.1. Parallel Trend Test

We adopt an event-study regression to verify the parallel trend premise and unpack the dynamic evolution of policy impacts, with the year right before policy rollout (t = −1) set as the benchmark reference window. Figure 3 illustrates that all coefficients corresponding to pre-policy time points lack statistical significance, whose 95% confidence intervals all contain the value zero. This outcome confirms no persistent gaps in symbolic digital transformation levels existed between pilot and non-pilot firms prior to policy launch, which validates the parallel trend condition required for causal identification.

Despite insignificant standalone pre-event coefficients, separate inspection of each lead term cannot fully rule out systematic pre-existing divergent trends. For this reason, we implement a joint significance test across all pre-policy coefficients. The test result cannot reject the null hypothesis that all lead coefficients collectively equal zero (F = 1.4791, p = 0.2182). This result delivers supplementary evidence that the treatment and control groups shared comparable development trajectories before the policy took effect.

After policy implementation, the estimated coefficients become negative and remain so in the subsequent periods, indicating that the SMPP exerts a persistent restraining effect on firms’ SDT. The magnitude of the effect increases during the early years after implementation and reaches its strongest level around the third year, suggesting that the policy effect is not only immediate but also cumulative to some extent. In later periods, however, the absolute magnitude of the coefficients declines and the estimates gradually lose statistical significance. One reasonable interpretation for this dynamic pattern lies in the gradual evolution of policy impacts: after the policy matures and enterprises finish adjusting their strategic layouts, the marginal restraining effect brought by the policy gradually fades and plateaus.

4.4.2. Placebo Test

To eliminate the concern that the baseline findings are driven by unobservable omitted variables or random disturbances, this study follows prior research [48] and conducts a placebo test. Specifically, the treatment status indicating whether a firm is selected as a smart manufacturing pilot is randomly reassigned 500 times, and the policy effect is re-estimated in each iteration using the DML model. If the estimated effect in the benchmark analysis is not generated by chance, the placebo coefficients obtained from the random permutations should be centered around zero and remain statistically insignificant.

Figure 4 presents the distribution of the placebo estimates. The density curve and scattered placebo coefficients are tightly concentrated around zero, forming a bell-shaped distribution with no systematic departure from the null. Most of the associated p-values are well above conventional significance thresholds, indicating that the randomly assigned pseudo-treatment effects are generally insignificant. In stark contrast, the true coefficient derived from benchmark regression, marked by the red dashed vertical line, falls within the far-left tail of the simulated placebo distribution and deviates distinctly from the cluster of random estimated coefficients. This empirical pattern demonstrates that the negative policy impact captured in baseline regressions cannot be attributed to stochastic errors or arbitrary sample grouping. Accordingly, the placebo test results solidify the causal validity and statistical robustness of this study’s core empirical findings.

4.4.3. Alternative K-Fold Splits

Because the DML estimator relies on K-fold cross-validation to reduce overfitting in the nuisance-function estimation, the choice of K may affect the estimates through incidental variation in sample splitting. To assess the sensitivity of the baseline results to this issue, this study resets the number of folds to K = 3 and K = 8, respectively. As reported in Columns (1) and (2) of

Table 4. Robustness Tests.

Variables	Dependent Variable: SDT
	Alternative K-Fold Splits		Alternative DML Algorithms			Alternative DML Specification
	(1)	(2)	(3)	(4)	(5)	(6)
	3-Fold	8-Fold	Lasso	GradBoost	SVM	Interactive Model
SMPP	−0.170 ***	−0.171 ***	−0.276 ***	−0.201 ***	−0.285 ***	−0.453 ***
	(−4.164)	(−4.142)	(−6.954)	(−4.989)	(−7.833)	(−20.996)
Linear Control Terms	Yes	Yes	Yes	Yes	Yes	Yes
Quadratic Control Terms	Yes	Yes	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes
Firm Fixed Effects	Yes	Yes	Yes	Yes	Yes	Yes
Observations	28,825	28,825	28,825	28,825	28,825	28,825

Notes: *** denotes statistical significance at the 1% level. Values in parentheses are t-statistics.

, the estimated coefficients on SMPP are −0.170 and −0.171, both significant at the 1% level. All three specifications yield significantly negative policy effects at the 1% level, fully proving our baseline estimates and standard errors are insensitive to the choice of cross-fitting folds.

4.4.4. Alternative DML Algorithms

We swap the baseline random forest for Lasso, GradBoost and SVM to eliminate bias caused by algorithm specificity, with updated results listed in Columns (3)–(5) of Table 4. The SMPP coefficient stays significantly negative at the 1% level for all three substitutes. The uniform findings confirm the policy effect is robust and unaffected by different machine learning specifications.

4.4.5. Alternative DML Specification

The benchmark estimation relies on a partially linear double machine learning model, which restricts the treatment effect to follow a linear additive form conditional on all covariates. To loosen this functional-form constraint, we adopt an interactive estimation framework that permits richer nonlinear interaction relationships between control variables and the policy treatment term. As shown in Table 4, the magnitude of the estimated policy coefficient varies across alternative machine learning specifications relative to the random forest baseline. Most notably, Column (6) generates a considerably larger coefficient magnitude. This pattern suggests that the policy’s restraining effect on SDT may exhibit pronounced non-linearity and is conditional on heterogeneous firm-level attributes.

Standard OLS imposes a uniform linear treatment effect across all sample firms, which restricts the ability to capture differentiated policy responses tied to firms’ distinct operational and digital backgrounds. By contrast, the interactive DML framework embeds covariate-policy interactions within nuisance function estimation and relaxes such rigid linearity constraints. The conditional heterogeneous effect uncovered here delivers clear methodological rationale for departing from conventional single linear specifications and adopting the more flexible DML framework throughout our analysis. As reported in Column (6) of Table 4, the coefficient on SMPP remains significantly negative after we adopt this fully flexible interactive specification. This result further confirms that the baseline finding is not driven by the linear additivity assumption embedded in the partially linear framework.

Across all machine learning variants in Table 4, the policy coefficient remains negative and statistically significant at the 1% level, consistent with our baseline findings. The cross-model disparity in coefficient size stems from different assumptions regarding effect heterogeneity rather than conflicting causal inferences.

4.4.6. Additional Robustness Tests

Several additional robustness checks are also performed. Specifically, this study replaces the baseline measure of the dependent variable, excludes observations that may be affected by other contemporaneous policy interventions, and further controls for province-by-year interaction effects. All modified estimation specifications yield qualitatively identical outcomes, which further reinforces the robustness of this paper’s core findings. Owing to space constraints, the detailed estimates are not reported here but are available from the authors upon request.

4.5. Addressing Endogeneity

4.5.1. Instrumental Variable Approach

Given that selection into the pilot program is not entirely random, this study further addresses potential endogeneity by re-estimating the effect using a double machine learning partial linear instrumental variable (DML-PLIV) model. Following the research design of Zhao et al. (2020) [49], the interaction between the number of fixed telephones per 100 persons in 1984 and the time trend is employed as a panel instrumental variable.

This instrument is motivated by both relevance and exogeneity considerations. On the one hand, early fixed-line telephone penetration reflects the initial endowment of regional information infrastructure, whose path-dependent evolution may have provided part of the technical foundation for the subsequent rollout of smart manufacturing policy. On the other hand, as a historical regional-level variable, it is unlikely to directly affect firms’ contemporary SDT, except through policy exposure. As reported in Column (1) of

Table 5. Endogeneity Tests.

Variables	(1)	(2)	(3)	(4)	(5)
	DML-PLIV	Heckman-DML	PSM-DML	PSM-DML	PSM-DML
			1:1 Nearest Neighbor	Caliper	Kernel
SMPP	−5.089 ***	−0.660 ***	−0.193 ***	−0.207 ***	−0.209 ***
	(−3.783)	(−3.120)	(−3.267)	(−3.484)	(−5.286)
IMR		0.211 **
		(2.26)
Cragg-Donald Wald F Statistic	93.284 ***
Controls	Yes	Yes	Yes	Yes	Yes
Year Fixed Effects	Yes	Yes	Yes	Yes	Yes
Firm Fixed Effects	Yes	Yes	Yes	Yes	Yes
Observations	28,825	24,023	1784	1771	28,825

Notes: **, *** denote significance at the 5% and 1% levels, respectively. Values in parentheses are t-statistics.

, the coefficient on SMPP remains significantly negative at the 1% level after instrumenting for policy assignment. Moreover, the Cragg–Donald Wald F statistic equals 93.284, which greatly surpasses the common weak instrument cutoff and verifies the strong relevance of the chosen instrument. Such evidence consolidates the baseline result that the SMPP effectively curbs firms’ symbolic digital transformation.

Notably, the DML-PLIV coefficient differs markedly from the baseline ATE. This numerical gap does not imply flawed identification or economically implausible outcomes for three reasons. First, the two estimators capture separate causal parameters: baseline DML computes full-sample ATE, while DML-PLIV yields LATE only for instrument-compliant firms. Raw coefficients cannot be directly compared, and the larger IV value merely reflects stronger policy responses among this subgroup. Second, IV estimation relies on fully residualized variables, whose rescaled metrics make raw IV coefficients incomparable to unadjusted baseline results. Third, the Cragg–Donald F-statistic of 93.284 far surpasses Stock–Yogo thresholds, largely alleviating weak-instrument bias. A 500-round permutation placebo test yields coefficients concentrated near zero, with almost no draws close to −5.089, ruling out LATE inflation driven by random noise. All specifications consistently confirm that the SMPP significantly curbs firms’ SDT.

4.5.2. Heckman Two-Stage Correction

To further account for potential self-selection into the pilot program, this study implements a Heckman two-stage correction within the DML framework. The exclusion restriction used in the selection equation is the same as that adopted in the IV analysis. Specifically, the first stage estimates a Probit model to obtain the inverse Mills ratio (IMR), and the second stage incorporates the IMR into the orthogonalized residual regression of the DML model.

Column (2) of Table 5 presents a positive and statistically significant IMR term at the 5% level, confirming severe sample selection bias within the sample. Even after correcting this bias, the SMPP coefficient remains negative and significant at the 1% level. It follows that the negative policy impact detected in baseline estimations is not caused by corporate self-selection into pilot projects, which strengthens the robustness of our primary findings.

4.5.3. Propensity Score Matching

Kernel density distributions of propensity scores are graphed to examine pre-policy balance across treated and untreated firms. Figure 5 shows that the propensity scores of the control group are concentrated in the 0–0.05 range, while those of the treatment group are shifted to the right and span a wider interval. Substantial overlap is observed mainly in the 0–0.08 range, indicating limited common support and suggesting that policy assignment may be correlated with firms’ pre-treatment characteristics. To reduce potential bias arising from this non-random selection, we implement two complementary PSM-DML robustness checks. First, we use truncation matching, including 1:1 nearest-neighbor matching with a 0.01 caliper, to exclude observations outside the common support and retain only comparable samples for regression. Second, we apply Epanechnikov kernel matching to down-weight observations outside the support region while preserving the full sample. Across the three matching settings displayed in Columns (3)–(5) of Table 5, the SMPP coefficient is significantly negative at the 1% level. It demonstrates that the policy impact remains robust when we adopt alternative matching standards and sample selection steps.

5. Further Analysis

5.1. Mediation Mechanism Test

To examine potential transmission mechanisms, this study conducts mediation analyses following the conventional causal interpretation framework. It should be noted, however, that causal mediation relies on the sequential ignorability assumption, which requires that there be no unobserved confounders affecting the mediator and outcome simultaneously, and no treatment-induced confounding. Accordingly, mediation outcomes merely offer preliminary evidence of potential mechanisms rather than full nonparametric causal decomposition. We also add comprehensive time-variant firm controls plus firm and year fixed effects to curb omitted-variable bias.

5.1.1. Causal Mediation Analysis Based on DML

• The Resource Empowerment Channel: Alleviating Financing Constraints

Column (1) of

Table 6. Causal Mediation Analysis Results.

Variables	(1)	(2)	(3)
	Average Treatment Effect (ATE)	Average Direct Effect (ADE)	Average Causal Mediation Effect (ACME)
FC	−0.208 ***	−0.198 ***	−0.010 ***
	(−5.275)	(−5.025)	(−4.579)
IA	−0.211 ***	−0.210 ***	−0.001 *
	(−5.343)	(−5.311)	(−1.658)
MA	−0.208 ***	−0.210 ***	0.002 **
	(−5.266)	(−5.324)	(2.200)
Linear control terms	Yes	Yes	Yes
Quadratic control terms	Yes	Yes	Yes
Year fixed effects	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes
N	28,825	28,825	28,825

Notes: *, **, *** denote significance at the 10%, 5%, and 1% levels, respectively. Values in parentheses are t-statistics.

shows that the SMPP significantly reduces firms’ SDT, as reflected in the significantly negative average treatment effect (ATE). More importantly, the average causal mediation effect (ACME) through financing constraints (FC) is also significantly negative at the 1% level. This result indicates that the SMPP not only sends a positive policy signal to external capital providers, but also improves firms’ access to financial resources. As financing constraints are alleviated, firms face less pressure to rely on symbolic digital disclosure or rhetorical exaggeration to attract external support. The evidence therefore supports Hypothesis 2.

• The Information Governance Channel: Reducing Information Asymmetry

Column (2) of Table 6 reports a significantly negative mediation effect through information asymmetry (IA). This finding suggests that the SMPP improves firms’ information environment while promoting substantive digital upgrading. In practice, policy implementation often requires firms to strengthen digital integration in production, operations, and data management, which enhances the transparency and verifiability of firm-level information. As information asymmetry declines, external stakeholders are better able to distinguish substantive digital transformation from symbolic compliance, thereby reducing managerial room for opportunistic digital signaling. Accordingly, Hypothesis 3 is supported.

• The “Reputational Game” Channel: Increasing Media Attention

Column (3) of Table 6 identifies a countervailing mediation channel. Specifically, the ACME transmitted via media attention (MA) carries a positive value and passes the 5% significance test. By contrast, the policy’s total and direct effects are both significantly negative. This pattern indicates that greater media exposure induced by the SMPP may generate reputational pressure and heightened external expectations. Under such circumstances, some firms may become more inclined to use rhetorical embellishment or symbolic digital signaling to maintain an innovation-oriented image. In this sense, media attention partially offsets the policy’s restraining effect on SDT. Therefore, Hypothesis 4 is supported.

5.1.2. Traditional Stepwise Mediation Test with Sobel and Bootstrap Robustness Check

To provide correlational cross-validation for the causal mediation results derived from DML in Section 5.1.1, this subsection implements the classical three-step stepwise mediation framework supplemented with Sobel Z-test and Bootstrap 95% confidence interval test, and the full estimation results are presented in

Table 7. Traditional Stepwise Mediation Test with Sobel and Bootstrap Cross-Validation.

Variables	Resource Empowerment		Information Governance		Reputational Game
	(1) FC	(2) SDT	(3) IA	(4) SDT	(5) MA	(6) SDT
SMPP	−0.007 **	−0.160 ***	−0.048 **	−0.161 ***	0.140 ***	−0.166 ***
	(−2.29)	(−3.85)	(−2.40)	(−3.88)	(2.25)	(−3.99)
FC		0.542 ***
		(7.15)
IA				0.057 ***
				(4.69)
MA						0.013 ***
						(3.20)
Linear control terms	Yes	Yes	Yes	Yes	Yes	Yes
Quadratic control terms	Yes	Yes	Yes	Yes	Yes	Yes
Year fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
Firm fixed effects	Yes	Yes	Yes	Yes	Yes	Yes
N	28,825	28,825	28,825	28,825	28,825	28,825
Sobel test (Z-statistic)	−2.150 **		−2.135 **		1.841 *
Bootstrap test	[−0.0071, −0.0013]		[−0.0053, −0.0007]		[0.0005, 0.0036]

Notes: *, **, *** denote significance at the 10%, 5%, and 1% levels, respectively. Values in parentheses are t-statistics.

. All regressions uniformly incorporate linear and quadratic control variables as well as firm and year two-way fixed effects to absorb unobservable time-invariant firm characteristics and macro yearly shocks, with a balanced sample of 28,825 firm-year observations maintained across all specifications.

Three parallel mediating channels are verified. First, the resource empowerment channel via financing constraints: the policy significantly eases financing frictions, which further restrains SDT. Second, the information governance channel via information asymmetry: the pilot program mitigates information friction and curbs cosmetic digital disclosure. Both channels pass the Sobel test and Bootstrap significance check at the 5% level. Third, a countervailing reputational channel through media attention exists: pilot accreditation raises media coverage, which in turn incentivizes firms’ superficial digital signaling, with its indirect effect significant at the 10% level.

Every regression setup generates a significantly negative direct restraining impact of SMPP on SDT at the 1% significance level. The identical coefficient direction and stable statistical significance across the three mediating paths are highly consistent with the earlier DML causal mediation results. These findings supply extra correlational evidence to validate the multi-path competing mechanism proposed in this paper.

5.1.3. Summary of the Mediation Results

Mediation results confirm two core restraining transmission channels via financing constraints (FC) and information asymmetry (IA), which jointly underpin the overall inhibitory effect of the SMPP. In contrast, media attention (MA) constitutes an adverse spillover path that partially offsets the policy’s disciplinary impact; this countervailing channel represents a passive side risk instead of a mechanism supporting the policy’s governance function.

Based on the average causal mediation effects (ACME) in Table 6, we quantify each channel’s explanatory share of the total average treatment effect (ATE): the FC channel accounts for roughly 4.8% of total restraint, the IA channel explains 0.5%, and the MA reverse channel offsets around 0.96% of the policy’s inhibitory force. All three indirect channels achieve conventional statistical significance, even if their economic magnitudes appear modest. While these indirect mechanisms only account for a small fraction of the total policy effect, they uncover the micro underlying transmission logic linking industrial policy to firms’ symbolic digital behavior and deliver clear theoretical implications for multi-channel competing mechanisms. The classical stepwise mediation test in Table 7 further cross-validates the consistent sign and significance of all three transmission paths, reinforcing the reliability of our identified channels under an alternative empirical framework.

The aggregated negative indirect restraint generated by the two core channels exceeds the small positive indirect effect of media attention, consistent with the baseline finding that the policy significantly curbs SDT.

The difference in the relative roles of competing channels can be further interpreted from theoretical logic. Media coverage mainly creates short-term reputational pressure on listed firms. Increased public exposure may temporarily drive firms to release rhetorical digital statements, yet such pressure is temporary and fails to form stable long-term governance binding force.

In contrast, eased financing constraints and improved information transparency generate long-term structural institutional restrictions. First, sustained access to long-term capital hinges on tangible digital investment; empty digital publicity cannot maintain continuous credit and equity financing support. Second, the SMPP enhances corporate information transparency and behavioral traceability, which gradually raises regulatory and compliance costs for SDT over the years. Accordingly, structural institutional restrictions with long-run binding attributes produce lasting governance influences, which differ from the transient reputational disturbance triggered by media exposure.

Although elevated media attention generates partial countervailing pressure that incentivizes superficial digital rhetoric, this opposing force is weak and temporary. The combined restraining effect from eased financing constraints, reduced information asymmetry and the policy’s built-in supervision rules dominates this reputational incentive, leading to the significantly negative total policy effect consistent with our baseline regression.

5.2. Threshold Effect Analysis

To further examine whether the governance effect of the SMPP varies across firm characteristics, this study estimates the panel threshold model developed in the previous section. For threshold identification, we implement Bootstrap resampling with 300 random draws to compute critical values for the F-statistic threshold tests, which ensures the validity of threshold number selection and threshold value estimation. The results are reported in

Table 8. Threshold Number Identification Tests.

Threshold Variable	Number of Thresholds	F-Statistic	p-Value	Critical Value
				10%	5%	1%
Managerial Myopia	Single	31.489 ***	0.007	17.837	21.799	30.357
	Double	32.521	0.723	44.907	47.674	51.431
	Triple	7.389	0.327	13.359	17.115	32.686
Corporate Opacity	Single	11.554 **	0.040	8.661	11.234	14.489
	Double	17.251 ***	0.010	11.341	13.865	17.038
	Triple	9.785	0.173	11.333	12.750	17.075

Notes: **, *** denote significance at the 5% and 1% levels, respectively. Values in parentheses are t-statistics.

,

Table 9. Threshold Estimates.

Threshold Variable	Threshold Specification	Estimated Threshold Value	95% Confidence Interval
Managerial Myopia	First Threshold	0.0174	[0.0174, 0.0193]
Corporate Opacity	First Threshold	0.0551	[0.0412, 0.0605]
	Second Threshold	0.0686	[0.0638, 0.0870]

and

Table 10. Threshold Regression Results.

Panel A: Managerial Myopia Threshold
Threshold Variable	Regime	Threshold Regression Coefficient	t-Statistic	Controls	N
Managerial Myopia	SMPP (q ≤ 0.0174)	0.0398	0.2155	YES	28,825
	SMPP (q > 0.0174)	−0.5781 ***	−11.7028
Panel B: Corporate Opacity Threshold
Threshold Variable	Regime	Threshold Regression Coefficient	t-Statistic	Controls		N
Corporate Opacity	SMPP (q < 0.0551)	−0.1903 **	−2.5737	YES		28,825
	SMPP (0.0551 < q ≤ 0.0686)	−0.8574 ***	−5.6119
	SMPP (q > 0.0686)	0.1356	0.9394

Notes: **, *** denote significance at the 5% and 1% levels, respectively. t-statistics in parentheses.

.

5.2.1. Threshold Effect of Managerial Myopia

Managerial myopia can shape firms’ responses to industrial policy by influencing the temporal horizon of managerial decision-making. From an agency-theoretic perspective, managers facing strong short-term incentives tend to prioritize near-term performance signals over long-horizon capability building, thereby distorting strategic choices and weakening substantive value creation [50,51]. Such short-termism increases the likelihood that firms will engage in SDT, as managers may place greater emphasis on visible and immediately legible actions that help sustain external legitimacy rather than on less observable but more fundamental upgrading efforts. The attention-based view lends extra credibility to this logic. It posits managerial attention as a limited resource; myopic executives favor visible, reportable projects over sustained long-term investments [52]. Conversely, firms run by forward-looking managers actively engage in long-cycle innovative investments, diminishing the marginal disciplinary effect of the policy. Accordingly, the governance effect of the SMPP is expected to vary nonlinearly with managerial myopia and to become more pronounced when short-termism is stronger.

Following Hu et al. (2021) [53], we construct the managerial myopia indicator via textual content analysis: the index equals the proportion of short-term horizon keyword occurrences relative to the full text word count, multiplied by 100. Larger values of this metric correspond to stronger corporate managerial short-termism. As shown in Table 8 and Table 9, managerial myopia passes the single-threshold test at the 1% significance level, and the estimated threshold value is 0.0174. Table 10 further shows that when managerial myopia is below 0.0174, the coefficient of SMPP is not statistically significant. Once managerial myopia exceeds this threshold, however, the coefficient becomes significantly negative at −0.5781.

This result indicates that the restraining effect of the SMPP is more pronounced in firms with stronger short-term managerial orientation. A plausible explanation is that firms with relatively weak agency distortions are already more inclined toward substantive digital upgrading and therefore have less incentive to engage in SDT. By contrast, when managerial myopia is more pronounced, the pilot policy can exert stronger disciplinary effects through project screening, policy supervision, and resource allocation, thereby more effectively curbing SDT.

5.2.2. Threshold Effect of Corporate Opacity

Corporate opacity can be understood through the lenses of information asymmetry, signaling theory, and legitimacy theory. When firms exhibit greater opacity, external stakeholders face more difficulty distinguishing substantive digital upgrading from rhetorical disclosure, which creates more room for symbolic transformation. From the perspective of legitimacy theory, when actual transformation progress is difficult to verify, firms may rely more heavily on symbolic narratives to maintain external legitimacy [54]. In this sense, corporate opacity affects the extent to which external stakeholders can observe, evaluate, and discipline firms’ strategic behavior. When disclosure quality is relatively high, market participants and regulators can more readily detect inconsistencies between digital rhetoric and substantive upgrading, leaving less room for symbolic transformation. As opacity increases, however, weak disclosure quality may conceal firms’ actual digital efforts and reduce the effectiveness of external monitoring, thereby amplifying opportunities for opportunistic behavior. Prior research also suggests that opaque reporting environments impair information quality and weaken external governance, which further supports the expectation that the policy’s disciplinary effect is likely to vary nonlinearly across opacity levels and may become less effective when opacity is excessively high [55,56].

To examine whether the effect of the SMPP depends on firms’ disclosure environment, this study uses corporate opacity as the threshold variable. Bhattacharya et al. (2003) [57] argue that the main factors affecting accounting opacity include earnings aggressiveness, loss avoidance, and earnings smoothing. This suggests that earnings management is a key source of corporate opacity. Discretionary accruals are widely used as a measure of earnings management; therefore, we assess corporate opacity from the perspective of discretionary accruals. If a firm’s discretionary accruals exhibit substantial fluctuations and remain persistently high in absolute value, the firm is more likely to engage in earnings manipulation, and its corporate opacity is correspondingly higher. Following Wang et al. (2009) [58], corporate opacity is measured as the sum of the absolute values of discretionary accruals over the previous three years. A higher value indicates greater corporate opacity. This measure is conceptually distinct from information asymmetry (IA), which is used as the mediating variable in the mechanism analysis. Specifically, IA captures market-level information frictions between firms and external investors based on high-frequency trading data, whereas corporate opacity reflects the quality and transparency of firm-level financial reporting. Although both variables relate to the information environment, they represent different dimensions and serve different analytical purposes in this study.

The threshold test results in Table 8 and Table 9 show that corporate opacity passes both the single-threshold and double-threshold tests, with estimated threshold values of 0.0551 and 0.0686, respectively. The regression results in Table 10 further indicate that the effect of the SMPP differs markedly across opacity regimes. When corporate opacity is below 0.0551, SMPP significantly restrains SDT, with a coefficient of −0.1903. When corporate opacity falls between 0.0551 and 0.0686, the restraining effect becomes substantially stronger, with the coefficient reaching −0.8574. When corporate opacity exceeds 0.0686, however, the coefficient becomes statistically insignificant.

These findings suggest that the governance effect of the SMPP is strongest when firms exhibit a moderate degree of reporting opacity. When opacity is relatively low, firms are already subject to stronger disclosure discipline, leaving limited room for the policy to generate additional governance gains. When opacity reaches an intermediate level, the pilot policy appears better able to identify weak substantive transformation incentives and impose effective external discipline. By contrast, when opacity becomes excessively high, poor reporting quality and distorted disclosure may make it more difficult for external oversight mechanisms to observe firms’ actual digital transformation progress, thereby reducing the effectiveness of policy monitoring and external governance.

5.3. Batch Heterogeneity and Temporal Pattern Analysis

To further examine the robustness of the policy effect across different rounds of pilot selection, this study progressively incorporates successive batches of pilot firms within the DML framework. As shown in Figure 6, the estimated effect of the SMPP on SDT remains significantly negative across all specifications, while its magnitude varies across batches. When the estimation is restricted to the first batch, the coefficient reaches −0.226 (p < 0.05), indicating a relatively strong restraining effect at the initial stage of implementation. As the second and third batches are successively included, the estimated coefficients adjust to −0.203 and −0.186, respectively; when all batches are incorporated, the coefficient further declines to −0.164, while remaining statistically significant at the 1% level.

Importantly, these batch-wise estimates should not be interpreted mechanically as evidence of a monotonically weakening policy effect over time. Because the progressive inclusion of later batches changes the composition of the treated sample, the observed coefficient differences may reflect heterogeneity in firm characteristics, selection criteria, industry affiliation, regional settings, and implementation intensity across batches, rather than a pure temporal attenuation of the policy impact. Accordingly, this section is intended to characterize batch-level heterogeneity in policy effectiveness, not to infer a simple time trend.

From an interpretive perspective, the stronger effect observed for the first batch may be associated with the more concentrated policy attention and stricter implementation environment surrounding the initial rollout. By contrast, later batches were introduced under a broader and more heterogeneous policy context, which may naturally lead to differences in estimated effect sizes. However, such variation does not imply that the policy effect necessarily fades out in a structural sense, nor does it support a mechanical inference that firms revert to symbolic window-dressing after an initial audit wave. Rather, the evidence suggests that the policy’s governance impact is contingent on the rollout stage and treatment composition.

It should be noted that the batch-based heterogeneity examined in this section is conceptually different from the event-time dynamics reported above. The former compares the estimated effects across different rounds of pilot selection, whereas the latter captures how the treatment effect evolves with the number of years elapsed since policy implementation. Therefore, the batch-level variation observed here should be understood as evidence of cross-batch heterogeneity, not as a substitute for or contradiction to the event-study results.

6. Conclusions and Discussion

6.1. Main Findings

This study examines whether the SMPP can curb firms’ SDT. Using Chinese A-share listed manufacturing firms as the research sample and exploiting the staggered implementation of the SMPP as a quasi-natural experiment, we find that the policy significantly restrains SDT. This conclusion remains robust across a range of additional tests, including event-study estimation, parallel trend tests, placebo tests, alternative machine learning specifications, different sample split ratios, alternative model specifications, and multiple endogeneity treatments, indicating that the SMPP exerts a substantive governance effect on firms’ symbolic responses to digital transformation. Mechanism analyses further suggest that the policy mainly restrains SDT by alleviating financing constraints and reducing information asymmetry, while media attention weakens this effect to some extent.

The threshold regression results further indicate that the governance effect of the SMPP on firms’ SDT is conditional on both managerial myopia and corporate opacity. With respect to managerial myopia, the inhibitory effect of the policy becomes more pronounced when managerial myopia exceeds the estimated threshold. This suggests that firms with more short-term-oriented managers are more likely to engage in SDT, and therefore have greater room for external policy intervention to correct such behavior. In this sense, the SMPP can serve as an external governance mechanism that disciplines managerial short-termism and encourages firms to shift from symbolic digital narratives toward more substantive digital transformation. Regarding corporate opacity, the results show a nonlinear pattern: the policy effect is strongest under moderate opacity, remains significant but weaker under low opacity, and becomes statistically insignificant under high opacity. This finding implies that moderate information frictions may provide room for policy supervision and external governance to play a corrective role, whereas excessive opacity may obstruct policy signal transmission and weaken external monitoring. Overall, the threshold evidence suggests that smart manufacturing pilot policies are particularly effective in curbing SDT among firms with stronger managerial short-termism and manageable levels of information opacity.

The analysis of batch heterogeneity and temporal evolution reveals that the smart manufacturing pilot policy exerts its most prominent restraining influence in the initial implementation phase, with marginal effects decaying steadily in subsequent years. This pattern suggests that the initial policy shock generates stronger external discipline and compliance incentives, whereas firms progressively adapt to a more normalized regulatory environment as the policy continues, leading to a decline in the marginal governance effect. Therefore, the dynamic evidence not only reinforces the baseline conclusion from a temporal perspective, but also reveals the life-cycle evolution of policy effectiveness.

6.2. Policy Implications

First, the evaluation of the SMPP should place greater emphasis on the substantive content of digital transformation rather than on firms’ rhetorical claims. In addition to policy declarations and strategic narratives, assessment criteria should assign more weight to observable and verifiable indicators, such as digital equipment investment, process integration, intelligent manufacturing applications, and organizational restructuring. This would help reduce firms’ incentives to engage in SDT and improve the alignment between policy objectives and implementation outcomes.

Second, supporting measures should be better aligned with firms’ financing and information conditions. Since the policy effect operates partly through alleviating financing constraints and reducing information asymmetry, regulators and financial institutions should improve the targeting of subsidies, credit support, and disclosure requirements toward firms undertaking genuine digital upgrading. For firms with high corporate opacity or pronounced managerial myopia, follow-up supervision and disclosure verification should be strengthened to prevent symbolic compliance and enhance the penetrability of policy oversight.

Third, the implementation of the SMPP should adopt a dynamic governance framework. Given that the restraining effect is strongest in the early stage and weakens over time, the policy should be accompanied by periodic evaluation, phased supervision, and timely policy recalibration. Such arrangements would help sustain regulatory effectiveness throughout the policy life cycle and reduce the likelihood that firms revert to SDT once the policy environment becomes normalized.

Finally, policymakers need to strike a reasonable balance between mandatory information transparency and the risk of corporate symbolic digital window-dressing. These two objectives are not inherently contradictory, and a set of targeted, multi-layer governance measures can help coordinate them effectively. For example, disclosure rules can be designed to require verifiable quantitative reporting on substantive digital investment, thereby improving information transparency while encouraging market participants to assess transformation quality on the basis of hard operational data rather than narrative claims. In addition, the pilot evaluation mechanism can place greater weight on rolling multi-year assessments, which would reduce firms’ short-term incentives to generate superficial digital achievements for media exposure. At the same time, differentiated and strengthened supervision can be directed toward firms with high managerial myopia and high corporate opacity, so as to curb symbolic transformation without imposing unnecessary burdens on all manufacturing firms. Finally, closer coordination with financial media and related information intermediaries may help standardize reporting norms, alleviate one-sided short-term performance pressure, and reduce firms’ incentives for cosmetic digital signaling.

6.3. Limitations and Future Research

Despite the robustness of the empirical design and the consistency of the main findings, this study has several limitations that should be acknowledged. First, the analysis is based on Chinese A-share listed manufacturing firms, which are characterized by relatively standardized disclosure practices and stronger regulatory visibility. While this setting is appropriate for identifying the governance effect of the SMPP, the external validity of the findings may be constrained when extended to unlisted firms, small and medium-sized enterprises, or firms in other institutional contexts. Future research may examine whether the restraining effect of the SMPP on SDT differs across broader firm populations or under alternative regulatory environments.

Second, although this study constructs SDT and several key explanatory dimensions using well-established textual and financial measures, some underlying organizational motives remain difficult to observe directly. In particular, symbolic strategic behavior may reflect a combination of impression management, managerial cognition, and external institutional pressure, which cannot be fully disentangled using archival data alone. This is a common constraint in empirical studies of symbolic corporate behavior. Future research may combine quantitative identification with survey data, interviews, or case-based evidence to further unpack the micro-level drivers behind firms’ symbolic digital responses.

Third, the policy effect identified in this study is evaluated within the observed policy window, and the longer-term evolution of firms’ strategic responses may still depend on subsequent regulatory adjustment, technological change, and market competition. Although the dynamic analysis provides evidence on the attenuation of the policy effect over time, the full life-cycle consequences of the SMPP may require a longer observation horizon. Future studies may therefore further track whether the policy generates persistent behavioral correction or merely temporary compliance in the long run.

Moreover, although this study controls for a rich set of firm-level covariates and fixed effects, it cannot completely rule out potential spatial spillovers or the sorting of high-capability firms into pilot areas. In particular, because the SMPP was implemented in staggered waves across selected cities and industrial clusters, firms located in pilot regions may be exposed to broader local industrial upgrading dynamics that are difficult to fully separate from the policy treatment itself. Future research could address this issue by incorporating more granular geographic data, exploiting more localized spatial variation, or applying spatial econometric methods to more precisely isolate spillover effects and regional selection patterns. These extensions would help strengthen the causal interpretation of policy effects in similar policy evaluation settings.

Overall, these limitations do not undermine the main conclusions of this study, but rather point to several promising directions for extending the research on industrial digital policy and symbolic corporate behavior.