How Digital-Economy Policy Boosts TFP: Evidence and Quadruple Mechanisms from China’s Manufacturing Sector

Abstract

Do China’s provincial digital-economy policies causally improve firm productivity and manufacturing sustainability? This paper addresses this question using a panel of Chinese manufacturers from 2008 to 2023. We first construct a novel, manually coded index of provincial policy intensity. We then use an instrumental-variable strategy, based on historical post-office density and governors’ STEM backgrounds, to identify causal effects. We find that digital-economy policy has a positive and significant impact on firm-level total factor productivity (TFP). Doubling the cumulative policy stock raises TFP by approximately 3%. This effect is transmitted through four key mechanisms: enhanced innovation quality, tax incentives, targeted digital subsidies, and knowledge spillovers. These channels support sustainable, innovation-led upgrading rather than mere input expansion. We also find the TFP gains are much larger in provinces with strong fiscal capacity and in firms with high digital absorptive capabilities. This paper contributes by providing clear causal evidence of the policy–TFP link and, crucially, by quantifying the four specific mechanisms that translate digital policy into durable, productivity-based sustainability in manufacturing.

1. Introduction

The digital economy has moved from a niche phenomenon to a central driver of global value creation. Between 2010 and 2021 the worldwide volume of cross-border data flows expanded by a factor of fifty, eclipsing trade in goods as the fastest-growing conduit of economic exchange. Advocates contend that digital general-purpose technologies—cloud computing, big-data analytics, and artificial intelligence—can reverse the post-2005 productivity slowdown documented for advanced economies [1,2]. Yet firm-level evidence remains mixed: while early adopters reap sizeable productivity premia [3,4], diffusion is uneven, and the public sector often intervenes to accelerate take-up. China provides a natural laboratory to study whether and how policy can convert digital hype into productivity gains. In 2023, China’s digital economy reached a scale of 53.9 trillion yuan, accounting for 42.8% of GDP, highlighting its significant role in the national economy [5]. A dense web of central and provincial initiatives backs this expansion, ranging from the 2015 Internet Plus strategy to the 2022 14th Five-Year Plan for Digital-Economy Development [6]. Provincial governments deploy a policy toolkit that combines tax rebates for software firms, earmarked subsidies for industrial Internet pilots, and the designation of digital-economy “pilot zones” that relax regulatory constraints [7].

In this context, we interpret firm-level total factor productivity (TFP) as an efficiency-centered indicator of the economic dimension of sustainable manufacturing. More specifically, TFP serves as a direct and rigorous measure of eco-efficiency. This concept quantifies a firm’s ability to ’produce more value with fewer or better-deployed inputs,’ which inherently includes optimizing the use of materials, energy, and capital. This conceptualization, which places efficiency and resource productivity at the center of sustainability, aligns firmly with the eco-efficiency concept established in the manufacturing and industrial-ecology literature [8,9]. Accordingly, our analysis informs sustainable development insofar as digital policy supports productivity-based upgrading and innovation capacity in industry, which are central to sustainable industry and infrastructure agendas [10]. Our scope is confined to the economic pillar of sustainability, consistent with the sustainable operations management perspective [11].

Despite the scale of this effort, two first-order questions remain unanswered. First, does the intensity of local digital-economy policy causally raise firm-level total factor productivity (TFP)? Second, through which specific mechanisms does any such effect materialize? The existing literature offers only partial answers. Micro-evidence links broadband penetration and cloud adoption to higher productivity for Chinese manufacturers [12,13], yet these studies treat the policy environment as an exogenous backdrop. Separate strands analyze individual instruments, tax relief [14], or R&D subsidies [15,16], but stop short of assessing their joint impact. Moreover, identification strategies commonly hinge on pooled OLS or difference-in-differences assumptions that may be invalid if policy adoption is endogenous to local economic conditions [17]. This paper fills the gap in three ways. First, we assemble a novel, province-year index of digital-economy policy intensity based on hand-coded regulations and guidelines. Second, we exploit two external shifters, historical post-office density interacted with national science-and-technology expenditure, and central policy attention interacted with governors’ STEM training, to construct Bartik-style instruments [18]. The approach isolates plausibly exogenous variation in local policy activism. Third, we open the black box of causality by quantifying four mechanisms suggested by theory and policy documents: innovation quality, tax incentives, digital subsidies, and knowledge spillovers.

Using a panel of A-share manufacturers from 2008 to 2023, we find that doubling the cumulative stock of provincial digital policies increases firm-level TFP by about 3 percent. Mechanism tests reveal that approximately one-third of the effect is mediated by higher-quality patents, one-quarter by preferential tax treatment, one-fifth by earmarked digital subsidies, and the remainder by inter-firm knowledge spillovers. Heterogeneity analysis shows stronger effects in provinces with ample fiscal resources and in digitally intensive industries, underscoring the complementarity between public capacity and private absorptive capability [19]. Taken together, these results are consistent with an efficiency-centered view of sustainable manufacturing on the economic dimension.

The contributions are threefold. Conceptually, we integrate four channels of industrial policy into a unified framework and measure their relative importance. Empirically, we deploy new instruments that address the simultaneity between policy activism and economic performance. Practically, our analysis provides evidence relevant to sustainable manufacturing policy by indicating when digital industrial policy is most likely to deliver productivity-based gains that are durable and broadly shared; the larger dividends in fiscally capable provinces and among firms with higher absorptive capacity suggest the importance of complementary public capacity and private-sector readiness. The rest of this paper is organized as follows. Section 2 discusses the institutional background and literature. Section 3 describes our data and empirical strategy. Section 4 presents the results. Section 5 provides the discussion, and Section 6 concludes the paper.

2. Institutional Background and Literature Review

2.1. China’s Digital-Economy Policy Landscape

China’s commitment to the digital economy dates back to the mid-2000s, when the State Council began issuing “informatization” guidelines. A break-through occurred in 2015 with the launch of the Internet Plus action plan and the Big-Data Development Outline, both of which pledged fiscal subsidies and VAT rebates to firms adopting cloud computing and industrial Internet solutions [20]. The 2017 13th Five-Year Plan for the Digital Economy shifted emphasis from infrastructure roll-out to data-driven business models, setting quantitative targets for the share of digital-core industries in GDP [21]. Policy instruments diversified: preferential corporate-income-tax rates for software enterprises, matching grants for AI pilot projects, and “digital transformation” vouchers redeemable by SMEs [22,23,24]. In 2019 the National Development and Reform Commission approved a first batch of provincial Digital-Economy Innovation and Development Pilot Zones, offering lump-sum subsidies and talent incentives to local champions [25]. Most recently, within the 14th Five-Year period (2021–2025), the State Council issued in March 2022 the 14th Five-Year Plan for the Development of the Digital Economy [6]. The plan makes “building a Digital China” (shuzi Zhongguo) a central objective. It sets clear tasks: develop data-factor markets; deepen industrial digitalization across manufacturing, services, and agriculture; raise the share of digital-core industries; expand inclusive digital public services; and improve governance through standards, competition policy, and security safeguards [6]. Provinces are encouraged to pilot regulatory sandboxes and use complementary fiscal tools. These objectives guide provincial implementation and the very policy toolkit studied here. These instruments align with the tax relief, subsidy support, innovation, and knowledge-spillover channels analyzed in this paper. Collectively, these overlapping programs provide a rich institutional setting in which firms are exposed to tax relief, earmarked subsidies, pilot preferential treatment, and intensified knowledge flows-precisely the four channels explored later in this paper.

2.2. Related Literature

Our paper connects three distinct strands of literature: (i) the link between digitalization and productivity; (ii) the effectiveness of industrial policy, particularly in the digital realm; and (iii) the specific mechanisms driving firm-level performance gains.

Digitalization and productivity. A large empirical literature links digital technologies to firm-level performance. Early studies on IT adoption in the United States showed positive, but heterogeneous, effects on productivity once organizational capital is accounted for [26,27]. More recent evidence confirms that specific technologies like cloud computing, AI, and e-commerce raise TFP, especially for multi-plant exporters [28,29,30].

In the Chinese context, this is a very active area of research. Early studies focused on infrastructure, finding that broadband penetration increases labor productivity through skill upgrading [31,32]. Recent work now looks at specific platforms, such as industrial Internet platforms, which are shown to boost TFP by improving supply-chain visibility and optimizing resource allocation [33,34,35]. Other research confirms a direct link between a firm’s overall digital transformation and its productivity [36,37], including green TFP [38,39]. Several recent papers, including those suggested by our reviewers, have studied the specific topic of the digital economy’s impact on manufacturing TFP [40,41], generally finding a positive association. Meta-analyses also suggest that digital adoption explains up to 20% of cross-firm productivity dispersion in manufacturing [42,43].

Industrial policy and firm behavior. The second area is industrial policy. Theoretically, targeted state intervention can correct learning and coordination externalities [17]. Empirically, Ref. [44] documents that China’s “Strategic Emerging Industries” program increased patent counts and TFP, largely via R&D subsidies. Recent work confirms that specific digital policies can also spur innovation [45,46], and that even government “digital attention” can have an effect [5]. The type of policy instrument matters. Tax instruments are a key lever: Ref. [47] shows preferential VAT rebates spur green innovation, while other studies find that corporate-income-tax holidays accelerate equipment investment [48,49,50]. More recent work confirms that R&D tax credits specifically accelerate a firm’s digital transformation [51,52]. Direct government subsidies are another key tool, shown to promote both digitalization [14] and the move toward green-digital systems [53]. Evidence outside China echoes these themes: Japan’s subsidies increased export competitiveness [54], while Brazil’s PITCE program raised private R&D [55]. However, critics correctly warn that these policies can fail, leading to misallocation [56,57] or wasteful rent-seeking [58]. This debate highlights the strong need for credible, causal identification strategies, such as the one we employ in this paper.

Mechanism-specific insights. Finally, our study builds on research into how digital policies might work. We focus on four channels that are often discussed separately. First, the innovation channel. Digital infrastructure can enlarge patent impact by lowering search costs and enabling recombinant innovation [59,60]. Recent studies confirm that firms that digitalize tend to produce higher-quality patents [61,62]. Second, the financial channel. Tax incentives are known to help firms overcome cash constraints [63,64,65]. This can encourage them to take on the high-risk, high-return R&D projects that are common in digital innovation [66]. Third, the direct support channel. Direct digital subsidies signal government priorities. This signal can “crowd in” private investment and help new technologies spread faster [67,68,69]. Fourth, the spillover channel. When digital and traditional firms co-locate, knowledge can diffuse between them, raising citation flows and merger activity [70,71,72,73]. The digital economy, in particular, has been shown to accelerate this kind of knowledge sharing [74,75]. While these channels are often studied in isolation [76], our paper contributes by placing all four within a single, unified empirical framework.

2.3. Hypothesis Development

Taken together, policy documents emphasize productivity gains through digital adoption, while extant research outlines four plausible transmission mechanisms. This leads to the following testable statements. First, we posit a main effect:

H1 (Main Effect):

A stronger provincial digital-economy policy environment increases firm-level total factor productivity.

Next, we articulate mechanism-specific expectations:

H2 (Innovation Quality):

Digital-economy policy raises TFP by improving the average quality of firm patents.

H3 (Tax Incentives):

Digital-economy policy enhances productivity by providing firms with direct tax incentives.

H4 (Digital Subsidies):

Digital-economy policy translates into higher TFP by providing firms with earmarked digital subsidies.

H5 (Knowledge Spillovers):

Digital-economy policy boosts TFP by facilitating knowledge diffusion from local digital enterprises to traditional firms.

3. Data and Empirical Strategy

3.1. Data Sources and Sample Construction

Our analysis relies on a firm-year panel constructed by merging four distinct datasets for A-share listed manufacturing firms in China from 2008 to 2023. First, firm-level financial fundamentals, ownership structure, and corporate governance data are sourced from the China Stock Market & Accounting Research (CSMAR) database. Second, detailed invention patent records, including forward citations, are drawn from the IncoPat database provided by the China National Intellectual Property Administration (CNIPA). Third, province-level statistics on fiscal capacity and education are obtained from the China Statistical Yearbook. Finally, we build a comprehensive corpus of policy documents by retrieving all central and provincial-level regulations, plans, and guidelines related to the digital economy from the Lawdata legal repository, a database maintained by Peking University.

3.2. Variables and Measurement

3.2.1. Dependent Variable

Our primary outcome is firm-level total factor productivity (TFP). It is estimated using the semi-parametric procedure developed by Levinsohn and Petrin (2003) [77], which uses intermediate inputs as a proxy for unobserved productivity shocks to address the simultaneity problem. We use firm-level capital stock, number of employees, and intermediate inputs as production factors, and deflate firm output using industry-specific Producer Price Indices (PPIs).

3.2.2. Core Explanatory Variable

The key independent variable, policy, is a province-year index measuring the intensity of local digital-economy initiatives. The construction follows a rigorous four-step process. First, we construct a comprehensive keyword lexicon related to the digital economy by referencing academic literature, national planning documents (e.g., the “14th Five-Year Plan”), and local regulations. Second, using this lexicon, we perform a broad search of the Lawdata database, yielding an initial set of potentially relevant provincial policy documents. Third, and crucially, each of these documents was manually reviewed to verify that it contained specific, substantive support measures for the digital economy. This rigorous screening, which distinguishes our index from simple keyword counts, yielded a final set of authentic digital-economy policies. Finally, for each province-year, we calculate the cumulative stock of policies issued to date and apply a log-transformation, $\log (1+\mathrm{PolicyCount})$, to arrive at our final index. To ensure coder reliability, the manual review and coding were conducted independently by two research assistants, and any discrepancies were resolved by a third senior coder, achieving an inter-rater reliability

Table A1. Comprehensive keyword lexicon for digital-economy policy index.

Cat. ID	Category	English Keywords
I	Core AI Tech	AI, Artificial Intelligence, Generative AI, AIGC, Large Models (LLMs), Foundational Models, Reinforcement Learning, Computer Vision, OCR, Intelligent Quality Inspection, Digital Twin, Smart Manufacturing
II	Big-Data Tech	Big Data, Data Mining, Text Mining, Data Visualization, Data Middle Platform, Data Lake, Data Warehouse, Master Data Management, Knowledge Graph, Spatio-temporal Big Data, ETL/ELT, Federated Learning
III	Cloud/Edge Comp.	Cloud Computing, IaaS, PaaS, SaaS, Kubernetes, Microservices, Serverless, DevOps, CI/CD, Multi-cloud, Hybrid Cloud, Object Storage, API Gateway, Cloud-Edge Collaboration, Edge Computing, Cloud Platform
IV	Blockchain/Trusted	Blockchain, Digital Currency, Distributed Ledger (DLT), Alliance Chain, Cross-chain, Zero-Knowledge Proof (ZKP), Homomorphic Encryption, Decentralized Identity (DID), Traceability, Digital RMB (e-CNY), Smart Financial Contracts
V	Thematic/General	Digital Economy, Digital Transformation, Industry Digitalization, Digital Industrialization, Digital Technology, Smart City, Intelligentization, Intelligent Traffic, Intelligent Medical, E-commerce
VI	Industrial Apps	Industrial Internet, IIoT, MES/ERP/PLM/SCADA, Supply Chain Platform, Industry Internet Identification, Two-Wheel Integration, Intelligent Energy, Smart Wearables, Smart Agriculture, Unmanned Retail
Infrastructure, Governance, and Methodological Categories
VII	Network and Compute	5G/5G Private Network, 6G, Gigabit Optical Network, IPv6, Compute Power, Compute Network/Hub, “East Data West Computing” (Dong Shu Xi Suan), Data Center (IDC/Green DC/Dual-Carbon DC)
VIII	Data Elements and Governance	Data Elements, Data Element Market, Data Licensing/Operation, Data De-identification, Data Classification/Grading, Cross-border Data Flow, Public Data Openness, Data Exchange Platform, Data Assetization, Data Rights Confirmation
IX	Security and Compliance	Network Security, Classified Protection, Data Security Law, PIPL (Personal Information Protection Law), Critical Information Infrastructure, Privacy Computing/Secure MPC, Trusted Computing, Zero Trust, Security Sandbox
X	Policy Tools (Inclusion)	Guidance/Implementation Document, Action Plan, Special Program/Plan, Management Measures, Funding/Subsidies/Grants, Innovation/Cloud Vouchers, Tax Incentives/Super Deduction, Pilot/Demonstration/Benchmark, Recognition, Procurement
XI	Exclusion Words	Interpretations/Explanations, News Releases, Dynamic Information, Summary/Overview, Speeches/Addresses, Interviews, Pictures, Reprints, Briefing, Meeting Minutes, Popular Science

.

3.2.3. Instrumental Variables

To address potential reverse causality and omitted-variable bias, we follow the shift-share logic of [18] and construct two external instruments for policy. These instruments leverage geographically predetermined conditions interacted with time-varying national shocks, a methodology validated and widely adopted in recent empirical studies [78,79,80,81].

IV 1: Post-Office × National S&T Expenditure. We interact the cross-sectional stock of local post offices in 1984 with the time-varying national science-and-technology outlay. Historically, a denser postal network facilitated telecommunications infrastructure and, by extension, the diffusion of digital-economy programs, satisfying the relevance criterion. Because the 1984 post-office count is predetermined and, by itself, unlikely to affect modern cross-boundary digital innovation in manufacturing, the exclusion restriction is plausible.

IV 2: Central Attention × Governor STEM Background. Guided by the “central agenda → local execution” logic of Chinese industrial policy, we construct the interaction

$$\ln \left(\mathrm{Central}\mathrm{DE}{\mathrm{attention}}_t+1\right)\times \left[\ln \left(\mathrm{education}{\mathrm{years}}_{pt}+1\right)\times \mathbf{1}\left\{\mathrm{STEM}\mathrm{major}\right\}\right],$$

where Central DE attention is the annual count of State-Council documents that contain the term “digital economy”. The bracketed term captures the governor’s technical policy capacity: it is the natural logarithm of the governor’s years of schooling, multiplied by a dummy that equals 1 for engineering or natural-science majors and 0 otherwise. Provinces led by technically trained and highly educated governors are more likely to translate central digital-economy directives into local actions, ensuring instrument relevance, while the historical education profile of the governor is orthogonal to contemporaneous firm-level productivity shocks, satisfying the exclusion restriction.

3.2.4. Mechanism Variables

To trace the channels through which digital-economy policy affects productivity, we construct four mediators—each expressed in logarithms to reduce skewness—using firm-level data from CSMAR and CNIPA. Innovation Quality is $log\left(1+{\overline{\mathrm{Cites}}}_{it}\right)$, where ${\overline{\mathrm{Cites}}}_{it}$ is the average forward citation count of invention patents granted to firm i in year t (CNIPA). Tax Incentives is $log\left(1+{\mathit{TaxRebate}}_{it}\right)$; ${\mathit{TaxRebate}}_{it}$ equals the cash value of tax refunds and preferential-rate credits disclosed in firms’ cash-flow statements (CSMAR). Digital Subsidies is $log\left(1+{\mathit{Subsidy}}_{it}\right)$, where ${\mathit{Subsidy}}_{it}$ denotes the government grant amount explicitly earmarked for digital R&D or digital-transformation projects, reported under “specific project subsidies” in CSMAR. Knowledge Spillovers is $log\left(1+{\mathit{SpilloverShare}}_{it}\right)$. ${\mathit{SpilloverShare}}_{it}$ is the proportion of firm i’s patent citations in year t that reference patents owned by local digital enterprises. Digital enterprises are identified according to the Statistical Classification of the Digital Economy and Its Core Industries (2021), and citation matching at the IPC-4 level uses CNIPA data. Higher values indicate greater absorption of digital knowledge by traditional manufacturers.

3.2.5. Control Variables

Following the literature, we include a standard set of time-varying firm-level controls: firm size (log total assets), leverage, firm age, and ownership concentration. We also control for province-level factors: GDP per capita, industrial structure, and education expenditure.

Table 1. Variable definitions and data sources.

Variable	Definition and Measurement	Source
Panel A: Dependent Variable
TFP	Total factor productivity, estimated using the Levinsohn–Petrin (2003) [77] method.	CSMAR, author’s calc.
Panel B: Core Explanatory and Instrumental Variables
Policy	$log(1+\mathrm{CumulativeCount})$ of provincial digital-economy policies.	Lawdata, author’s calc.
Post Office IV	Interaction of log post offices (1984) and log national S&T expenditure.	Historical records, CSY
Governor IV	Interaction of central government attention and governor’s STEM-weighted education level.	State council, Governor CVs
Panel C: Mechanism Variables
Innovation Quality	$log(1+\mathrm{Average}\mathrm{forward}\mathrm{patent}\mathrm{citations})$.	IncoPat (CNIPA)
Tax Incentives	$log(1+\mathrm{Cash}\mathrm{value}\mathrm{of}\mathrm{tax}\mathrm{rebates})$.	CSMAR
Digital Subsidies	$log(1+\mathrm{Earmarked}\mathrm{digital}\mathrm{R\&D}\mathrm{subsidies})$.	CSMAR
Knowledge Spillovers	$log(1+\mathrm{Share}\mathrm{of}\mathrm{citations}\mathrm{to}\mathrm{local}\mathrm{digital}\mathrm{firms})$.	IncoPat, author’s calc.
Panel D: Control Variables
Firm Size	Log of total assets.	CSMAR
Leverage	Total liabilities divided by total assets.	CSMAR
Firm Age	Log of (current year - founding year + 1).	CSMAR
Ownership Conc.	Shareholding ratio of the largest shareholder.	CSMAR
GDP per capita	Log of provincial GDP per capita.	CSY
Industry Structure	Share of tertiary industry in provincial GDP.	CSY
Education Exp.	Share of education spending in provincial fiscal expenditure.	CSY

Notes: CSY refers to the China Statistical Yearbook. Author’s calc. indicates calculations by the authors based on raw data.

summarizes the definitions, measurements, and data sources for all variables used in this study.

3.3. Empirical Strategy

3.3.1. Baseline Model

Our baseline specification is the following two-way fixed-effects (TWFE) model:

$${\mathrm{TFP}}_{ipt}={\beta }_1{\mathit{Policy}}_{pt}+{\mathbf{X}}_{it}^{\prime }\delta +{\mu }_i+{\lambda }_t+{\pi }_p+{\epsilon }_{ipt}.$$

(1)

where ${\mathrm{TFP}}_{ipt}$ is the productivity of firm i in province p at year t. ${\mathbf{X}}_{it}^{\prime }$ is the vector of controls. $\delta$ is the associated coefficient vector; ${\mu }_i$, ${\lambda }_t$, and ${\pi }_p$ denote firm, year, and province fixed effects, respectively. Standard errors are clustered at the province level.

3.3.2. Identification Strategy: 2SLS

To address endogeneity concerns, we employ a 2SLS approach using the two instruments described above. The second-stage equation is

$${\mathrm{TFP}}_{ipt}={\gamma }_1{\widehat{\mathrm{Policy}}}_{pt}+{\mathbf{X}}_{it}^{\prime }\delta +{\mu }_i+{\lambda }_t+{\pi }_p+{\nu }_{ipt},$$

(2)

where ${\widehat{\mathrm{Policy}}}_{pt}$ is the predicted value from the first-stage regression. Because each of our specifications is exactly identified, the validity of the exclusion restriction rests on the theoretical arguments provided above, rather than on statistical overidentification tests. The instrument’s strength is confirmed by the Kleibergen–Paap F-statistic.

3.3.3. Mechanism and Heterogeneity Analysis

We test our four proposed mechanisms using a standard two-step mediation analysis. First, we regress the mediator on the policy variable. Second, we include both the policy and the mediator in the main TFP regression. The extent of mediation is gauged by the attenuation of the policy coefficient. Heterogeneity is explored by re-estimating our baseline model on subsamples.

4. Results

4.1. Baseline Two-Way Fixed-Effects Estimates

Table 2. Digital-economy policy and firm productivity: baseline results.

	(1) Basic	(2) + Controls	(3) + Ind × Year FE
Policy	0.0409 ***	0.0282 **	0.0265 **
	(0.0141)	(0.0117)	(0.0129)
Controls	No	Yes	Yes
Firm FE	Yes	Yes	Yes
Year FE	Yes	Yes	Yes
Province FE	Yes	Yes	Yes
Industry×Year FE	No	No	Yes
Observations	17,300	17,300	17,300
R-squared	0.2781	0.3156	0.3056

Notes: Standard errors clustered at the province level in parentheses. The dependent variable is TFP (Levinsohn–Petrin). Controls include firm size, leverage, firm age, ownership concentration, provincial GDP per capita, industry structure, and education expenditure. * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

reports the benchmark TWFE estimates. Because the key regressor is the natural logarithm of one plus the provincial policy count, each coefficient measures the absolute change in firm-level TFP produced by a one-unit increase in that log index (i.e., almost a three-fold rise in the raw number of policies). Controlling only for firm and year fixed effects in column (1), the coefficient on Policy is 0.0409. Interpreting this semi-elasticity, a doubling of the policy count, corresponding to a 0.693 log-point increase, raises TFP by approximately $0.0409\times 0.693\approx 0.028$ units, or about 2.5% relative to the sample mean value of 1.12. Column (2) adds a comprehensive set of firm- and province-level controls. The policy coefficient decreases to 0.0282 but remains highly significant, indicating that observable characteristics explain part, yet not all, of the baseline association. Column (3) further absorbs industry–year fixed effects to purge common sector-specific shocks. The estimate stabilizes at 0.0265. The minimal change between columns (2) and (3) confirms that the positive relationship is not driven by time-varying industry trends. Taken together, the baseline results provide robust evidence that stronger digital-economy policy environments are associated with materially higher firm productivity, even after accounting for an extensive battery of controls and fixed effects.

4.2. Addressing Endogeneity: 2SLS Results

Our baseline TWFE estimates are likely biased by endogeneity. To identify the causal effect of digital-economy policy on TFP, we employ a 2SLS (Two-Stage Least Squares) strategy using the two instruments. Before presenting the causal estimates, we must first establish the validity of our instruments. As supporting evidence for the exclusion restriction (R1C1), we conduct a balance test [18]. We test whether the cross-sectional components of our instruments are correlated with pre-treatment provincial observables from the baseline year (2008). The results are presented in

Table 3. Balance tests: IV components vs. pre-treatment provincial characteristics.

	(1) Post Office IV Component	(2) Governor IV Component
Pre-treatment covariates
GDP per capita	0.0437	0.0274
	(0.0993)	(0.0715)
Industry Structure	0.0011	$-0.0006$
	(0.0039)	(0.0031)
Education Exp.	$-0.0025$	0.0013
	(0.0062)	(0.0047)
Constant	0.2138	0.4526
	(0.4061)	(0.3394)
Observations	31	31
$R^2$	0.0473	0.0361
Joint F-test (all covariates $=0$): p-value	0.7421	0.8647

Notes: Province-level OLS with heteroskedasticity-robust (HC3) standard errors in parentheses. Dependent variables are the components of the two instruments defined in the paper: (1) the cross-sectional share term for Post Office IV (log of 1984 post offices per capita); (2) the cross-sectional ability term for Governor IV (STEM-major indicator interacted with education years, evaluated at 2008). Pre-treatment covariates are log provincial GDP per capita, industry structure, and education expenditure share.

.

Table 3 shows that all individual coefficients on the 2008-level covariates are statistically insignificant. The joint F-tests yield p-values of 0.7421 and 0.8647, meaning the null hypothesis that these pre-treatment characteristics are jointly zero cannot be rejected. This result supports the exogeneity assumption. Second, we test the relevance assumption.

Table 4. First-stage regressions for Policy.

	(1) IV1 Only	(2) IV2 Only	(3) Both IVs
Dependent Variable: Policypt
IV1: Post Offices1984×Nat. S&T	0.3874 ***		0.2289 ***
	(0.0321)		(0.0410)
IV2: Central DE Attention×Governor STEM		0.4152 ***	0.2431 ***
		(0.0355)	(0.0442)
Controls	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes
Year FE	Yes	Yes	Yes
Province FE	Yes	Yes	Yes
Observations	17,300	17,300	17,300
Angrist–Pischke partial $R^2$ (excluded IVs)	0.1183	0.1027	0.1842
Kleibergen–Paap $rk$ F (excluded IVs, joint)	5255.7800	4952.3300	3987.4160
Instrument-specific F (AP, robust)	1845.1230	1622.4870

Notes: Panel fixed-effects regressions at the firm×year level. Standard errors clustered at the province level in parentheses. All columns include the same controls and fixed effects as the main model. IV₁ is the interaction of the cross-sectional 1984 post-office density (per capita) with national S&T outlays; IV₂ is the interaction of central “digital economy” attention with the governor’s STEM-weighted education index. Reported diagnostics include Angrist–Pischke partial $R^2$ for excluded instruments, the Kleibergen–Paap $rk$ F for joint significance, and instrument-specific robust AP F in columns (1)–(2). * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

presents the first-stage regression results.

As Table 4 shows, both instruments are statistically significant predictors of the Policy variable, both individually (Columns 1 and 2) and jointly (Column 3). The Kleibergen–Paap F-statistics are above conventional thresholds, mitigating concerns about weak instruments. Having established instrument exogeneity (Table 3) and relevance (Table 4), we now present the main 2SLS causal estimates in

Table 5. 2SLS estimates using external instruments.

	(1)	(2)	(3)
Dependent Variable: TFP	IV1Only	IV2Only	IV1and IV2
Panel A — Second-stage (2SLS) coefficients
Policy	0.0482 ***	0.0515 ***	0.0502 ***
	(0.0151)	(0.0154)	(0.0153)
Controls	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes
Year FE	Yes	Yes	Yes
Province FE	Yes	Yes	Yes
Panel B — Instrument diagnostics and weak-IV-robust inference
Anderson–Rubin LM p-value	0.0000	0.0000	0.0000
AR 95% CI for Policy	[0.0235,0.0764]	[0.0251,0.0793]	[0.0242,0.0778]
Kleibergen–Paap $rk$ F	5255.7800	4952.3300	3987.4160
Number of instruments	1	1	2
Hansen J (over-ID) p-value	n/a	n/a	0.4321
Observations	17,300	17,300	17,300

Notes: Standard errors clustered at the province level in parentheses. Column (3) uses both instruments jointly, yielding an over-identified model; the Hansen J test does not reject the over-identifying restrictions. We report Anderson–Rubin (AR) confidence intervals for the Policy coefficient as weak-IV-robust inference. The Kleibergen–Paap $rk$ F statistics correspond to the joint significance of the excluded instruments in the first stage. * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

.

Table 5 presents second-stage results using each instrument separately and both jointly. Across specifications, the estimated effect of Policy on firm-level TFP is stable: 0.0482 (IV₁ only), 0.0515 (IV₂ only), and 0.0502 (IV₁ and IV₂). The over-identified model passes the Hansen J test ($p=0.4321$), suggesting no evidence against the over-identifying restrictions. Anderson–Rubin 95% confidence intervals exclude zero in all cases, e.g., $[0.0242,0.0778]$ with both instruments, providing weak-IV-robust support for a positive causal effect. This result provides statistical evidence that the instruments satisfy the exclusion restriction. Economically, the estimate from Column (3) (0.0502) implies that doubling the number of digital-economy measures increases firm TFP by about 3.5% ($0.0502\times \ln \left(2\right)\approx 0.035$). This robust and positive causal impact provides strong support for our main hypothesis H1.

4.3. Robustness Checks

We conduct several tests to verify the robustness of our main findings. First, we perform a placebo test to ensure our identification strategy is not capturing pre-existing productivity trends. We test whether our instrumented policy variable affects an outcome it should not: the firm’s pre-treatment TFP. We define this outcome ($TF{P}_i^{pre}$) as the firm’s average TFP over the 2008–2010 period, before major digital-economy policies were enacted.

Table 6. Placebo test: pre-treatment TFP (mean over 2008–2010).

	(1) RF: IV1 Post Office IV	(2) RF: IV2 Governor IV	(3) 2SLS: Policy (IV1 and IV2)
Dependent Variable: ${\mathit{TFP}}_{\mathit{i}}^{\mathit{pre}}$
Post Office IV	0.0008 (0.0016)
Governor IV		−0.0003 (0.0012)
Policy (instrumented)			0.0041 (0.0087)
Controls	Yes	Yes	Yes
Firm FE	No	No	No
Year FE	No	No	No
Province FE	No	No	No
Observations	1108	1108	1108
$R^2$	0.0127	0.0114	0.0102
First-stage/weak-IV diagnostics for col. (3)
Kleibergen–Paap rk F	8.4321
Anderson–Rubin LM p-value	0.5912

Notes: Firm-level cross-section. $TF{P}_i^{pre}$ is the mean TFP for firm i over 2008–2010. All specifications include the same firm-level and province-level controls as in the main models, measured at baseline. Standard errors are clustered at the province level and reported in parentheses. In column (3), policy is instrumented jointly by the two external instruments (IV₁ and IV₂).

presents these results. Columns (1) and (2) show the reduced-form (RF) estimates, indicating that neither of our instruments has a statistically significant association with pre-treatment TFP. Column (3) reports the 2SLS result, where the coefficient on the instrumented Policy variable is small (0.0041) and statistically insignificant. Furthermore, the weak-IV-robust Anderson–Rubin LM p-value is 0.5912, which confirms this null finding. These results support our causal interpretation by showing that our instruments do not pick up spurious pre-existing productivity trends.

Table 7. Robustness checks.

	(1) L1.Policy	(2) ln(TFP)	(3) Winsorized	(4) Competitiveness
L1.policy	0.0300 *** (0.0123)
Policy		0.0032 ** (0.0013)		0.3120 *** (0.0579)
Wpolicy			0.0272 *** (0.0109)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Province FE	Yes	Yes	Yes	Yes
Observations	14,198	17,200	17,200	17,300
R-squared	0.2992	0.2906	0.3035	0.0384

Notes: Standard errors clustered at the province level in parentheses. Column (1) uses lagged policy to test for reverse causality; Column (2) uses log TFP as the dependent variable; Column (3) applies 1% winsorization to both TFP and policy variables; Column (4) uses firm competitiveness as the alternative outcome variable. * p < 0.10, ** p < 0.05, *** p < 0.01.

demonstrates that the positive policy–productivity nexus is insensitive to a battery of alternative specifications. Column (1) replaces the contemporaneous regressor with its first lag, $\ln ({\mathit{policy}}_{p,t-1}+1)$. The coefficient remains positive and highly significant ($0.0300$ and p < 0.01), indicating that the effect persists when policy precedes productivity by at least one year, which helps to rule out reverse causality. To ensure results are not driven by the functional form assumption, column (2) re-estimates the baseline with $\left(\ln \right(\mathrm{TFP}\left)\right)$ as the outcome. In this log-log specification, the coefficient on the policy index represents an elasticity. The estimate of 0.0032 implies that a 10% increase in policy intensity lifts TFP by approximately 0.032% ($10\%\times 0.0032$). This result is consistent in sign and significance with our main findings. Column (3) trims the top and bottom 1% of both the policy and TFP distributions. The policy coefficient remains at $0.0272$, confirming that extreme observations do not drive the baseline relationship. Finally, column (4) replaces TFP with a firm-level competitiveness index constructed from profit margins and market share. The estimated effect, $0.3120$, is not only statistically significant but also economically large, suggesting that digital-economy policies enhance performance on broader strategic dimensions beyond productivity.

Our baseline productivity measure is Levinsohn–Petrin (LP) value-added TFP using industry PPIs to deflate outputs and inputs. To probe sensitivity to production-function assumptions, we construct two additional firm-year TFP series: (i) Olley–Pakes (OP), using investment as the control proxy and dropping zero-investment observations; and (ii) Ackerberg–Caves–Frazer (ACF), using materials as the proxy.

Table 8. Correlation of alternative TFP measures.

	TFP (LP)	TFP (OP)	TFP (ACF)
TFP (LP)	1.0000	0.9234	0.9056
TFP (OP)	0.9234	1.0000	0.9378
TFP (ACF)	0.9056	0.9378	1.0000

Notes: Pearson correlations computed on the largest common sample across TFP series. LP = Levinsohn–Petrin; OP = Olley–Pakes; ACF = Ackerberg–Caves–Frazer.

shows that the alternative TFP measures are highly correlated with LP TFP ($\rho \in [0.90,0.94]$).

Table 9. Policy effect across alternative TFP estimators.

	(1) TFP (LP)	(2) TFP (OP)	(3) TFP (ACF)
Policy	0.0265 ** (0.0129)	0.0287 ** (0.0124)	0.0251 ** (0.0118)
Controls	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes
Year FE	Yes	Yes	Yes
Province FE	Yes	Yes	Yes
Observations	17,300	16,948	17,112
R-squared	0.3056	0.3021	0.3042

Notes: Standard errors clustered at the province level in parentheses. Column (1) replicates the baseline TWFE with LP TFP. Column (2) uses OP TFP. Column (3) uses ACF TFP. All columns include the same controls and fixed effects as the baseline model. * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

replicates the baseline specification with OP and ACF TFP as outcomes; the policy coefficient remains positive, stable in magnitude, and statistically significant under province-clustered inference.

4.4. Mechanism Analysis

Table 10. Mechanism analysis results.

Dep. Var.	M1: Innov. Quality		M2: Tax Incentives		M3: Digital Subsidies		M4: Know. Spillovers
	(1) Innov.	(2) TFP	(3) Tax	(4) TFP	(5) Subs.	(6) TFP	(7) Spill.	(8) TFP
Policy	0.5250 *** (0.1650)	0.0180 ** (0.0091)	0.8100 *** (0.2700)	0.0200 *** (0.0077)	0.4500 *** (0.1650)	0.0210 *** (0.0081)	0.0750 *** (0.0220)	0.0240 *** (0.0091)
Innovation Quality		0.0170 *** (0.0054)
Tax Incentives				0.0080 *** (0.0022)
Digital Subsidies						0.0120 *** (0.0044)
Knowledge Spillovers								0.0450 ** (0.0220)
Controls	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Province FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Observations	17,300	17,300	17,300	17,300	17,300	17,300	17,300	17,300
R-squared	0.2392	0.3057	0.4059	0.3057	0.3285	0.3056	0.2738	0.3056

Notes: Standard errors clustered at the province level in parentheses. All columns include the same controls and fixed effects as the baseline specification. * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

investigates four channels through which digital-economy policy may raise firm-level productivity: innovation quality, tax incentives, digital subsidies, and knowledge spillovers. In each block, column (A) estimates the impact of policy on the mediator, while column (B) re-introduces the mediator into the TFP regression to gauge how much of the baseline policy effect it absorbs.

4.4.1. (M1) Innovation Quality

A one-unit increase in the log policy index raises the (log) average citation count of patents by 0.525. When innovation quality is added to the TFP equation, its coefficient is positive and highly significant ($0.017$, p < 0.01), while the policy sem-elasticity falls from 0.0265 (baseline, Table 2) to 0.018. The 32% reduction indicates that higher-quality innovation is the single largest conduit.

4.4.2. (M2) Tax Incentives

Policy intensity strongly increases firms’ tax rebates ($0.810$). Including tax incentives lowers the policy coefficient further to 0.020, implying that preferential taxation explains about one-quarter of the total effect.

4.4.3. (M3) Digital Subsidies

The amount of digital-related government subsidies rises with policy strength ($0.450$, s.e. 0.165). When digital subsidies enter the TFP regression, they are significant ($0.012$, s.e. 0.004) and the policy estimate shrinks to 0.021. Fiscal support therefore accounts for roughly one-fifth of the overall impact.

4.4.4. (M4) Knowledge Spillovers

Policy also enlarges the share of patents by traditional firms that cite local digital enterprises ($0.075$, s.e. 0.022). Adding knowledge spillovers reduces the policy coefficient only slightly, to 0.024, suggesting a modest yet non-negligible channel (10% of the baseline effect).

A back-of-the-envelope decomposition of coefficient reductions across columns (2), (4), (6), and (8) attributes roughly 35% of the total policy effect to innovation quality, 25% to tax incentives, 20% to digital subsidies, and 10% to knowledge spillovers, leaving a residual direct effect of about 10%. These findings corroborate the paper’s “quadruple-mechanism” interpretation: digital-economy reforms ignite manufacturing productivity through complementary improvements in innovation, taxation, fiscal support, and inter-firm knowledge flows. These findings confirm our mechanism hypotheses, H2 (Innovation Quality), H3 (Tax Incentives), H4 (Digital Subsidies), and H5 (Knowledge Spillovers).

A methodological limitation of our decomposition is the challenge of maintaining causal consistency. While the main effect (Policy → TFP) is identified using 2SLS, the mechanism tests rely on the TWFE baseline. This is necessary because conducting a full 2SLS causal mediation model with multiple, potentially correlated mediators in a non-linear panel setting is computationally intractable. We acknowledge that the TWFE-based decomposition may suffer from residual endogeneity. Secondly, regarding the decomposition method, we note that the four channels are estimated separately, and the reduction in the Policy coefficient for each is not statistically independent. This means that our ’back-of-the-envelope’ decomposition should be interpreted as an approximate allocation of the total TWFE effect across the four mechanisms, and there is an inherent risk of double-counting given the strong correlation (e.g., between tax incentives and subsidies). Our goal is to provide a relative picture of the importance of these channels, not a precise measure of their mutually exclusive contributions.

4.5. Heterogeneity Analysis

Digital-economy policy may not affect all firms and regions equally; its impact is expected to vary with institutional resources, firm capacity, and spatial development gaps. To probe these differences, we re-estimate the baseline specification for a series of mutually exclusive subsamples.

Table 11. Heterogeneity analysis: institutional environment.

	Fiscal Self-Sufficiency		Education Expenditure
	(1) High	(2) Low	(3) High	(4) Low
Policy	0.0414 *** (0.0153)	−0.0216 (0.0170)	0.0429 *** (0.0138)	−0.0077 (0.0240)
Controls	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes
Province FE	Yes	Yes	Yes	Yes
Observations	8427	8519	8264	8334
R-squared	0.3192	0.2936	0.3163	0.2856

Notes: Standard errors clustered at the province level in parentheses. Fiscal self-sufficiency is measured by the ratio of local fiscal revenue to expenditure; education expenditure is provincial spending on education as a share of fiscal expenditure. High/Low groups are defined by a province-level median split within the sample. * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

focuses on the institutional environment—fiscal self-sufficiency and education expenditure—while

Table 12. Heterogeneity analysis: firm and regional characteristics.

	Firm Size		Digital Intensity		Geographic Region
	(1) Large	(2) Small	(3) High	(4) Low	(5) East	(6) Non-East
Policy	0.0401 *** (0.0135)	0.0384 *** (0.0129)	0.0315 ** (0.0153)	0.0138 (0.0170)	0.0465 * (0.0241)	0.0191 (0.0132)
Controls	Yes	Yes	Yes	Yes	Yes	Yes
Firm FE	Yes	Yes	Yes	Yes	Yes	Yes
Year FE	Yes	Yes	Yes	Yes	Yes	Yes
Province FE	Yes	Yes	Yes	Yes	Yes	Yes
Observations	8453	8360	8884	8182	3347	13,679
R-squared	0.1912	0.2199	0.3003	0.3192	0.3171	0.3098

Notes: Standard errors clustered at the province level in parentheses. Firm size is based on a median split of log total assets; digital intensity is proxied by the log frequency of core digital-technology keywords in each firm’s annual report; geographic regions are Eastern vs. Non-Eastern provinces. High/Low groups are defined by median split. * $p<0.10$, ** $p<0.05$, *** $p<0.01$.

examines firm size, industry-level digital intensity, and geographic region.

Columns (1) and (4) of Table 11 reveal that the productivity payoff to digital-economy policy is highly contingent on local public-finance conditions. In provinces with high fiscal self-sufficiency the policy semi-elasticity is 0.0414 (s.e. 0.0153, $p<0.01$), whereas it turns negative and insignificant in low-capacity provinces (−0.0216). A similar split appears for education expenditure: the coefficient is 0.0429 ($p<0.01$) in high-spending regions but virtually zero in low-spending ones. These patterns suggest that abundant local fiscal resources and human-capital investment are pre-requisites for turning policy signals into productivity gains.

Table 12 examines within-province heterogeneity. Firm size does not qualitatively alter the effect—both large and small enterprises benefit—but the magnitude is slightly higher for large firms (0.0401 vs. 0.0384). By contrast, digital-intensive industries register a markedly stronger response (0.0315, $p<0.05$) than low-intensity sectors, where the coefficient is positive yet statistically indistinguishable from zero. Finally, the impact is pronounced in the eastern region (0.0465, $p<0.10$) but much weaker and insignificant in the central-western provinces (0.0191).

The evidence indicates that digital-economy reforms yield the largest productivity dividends when (i) local governments have fiscal headroom to co-finance digital infrastructure and (ii) firms possess greater absorptive capacity—either through scale advantages or an existing digital footprint. The weaker effects observed in fiscally constrained and less digitalized settings highlight potential bottlenecks that policymakers must address to ensure broad-based productivity growth.

5. Discussion

This paper set out to answer whether and through which channels China’s provincial digital-economy policies raise firm-level productivity. Our baseline two-way fixed-effects estimates reveal that doubling the cumulative stock of provincial digital measures lifts TFP by about 3 percent, a magnitude comparable to the worldwide firm-level returns to cloud migration [82,83] and to the early broadband effects reported for Chinese firms [84]. Instrumental-variable results, based on historical post-office density and governors’ STEM backgrounds, suggest that OLS understates the true impact, consistent with negative selection mechanisms stressed by [85] in their misallocation framework. Viewed through a sustainability lens, these TFP gains reflect the efficiency-centered facet of sustainable manufacturing (creating more value with fewer or better-deployed inputs) and therefore speak to the economic pillar of sustainability. This is consistent with the literature on innovation-induced resource productivity and eco-innovation as a bridge between performance and sustainability goals [86,87].

5.1. How Do the Findings Enrich the Digital-Productivity Debate?

The mainstream view holds that digital technologies are a general-purpose catalyst whose payoff depends on complementary intangible capital. Our quadruple-mechanism evidence supports this conjecture: policy raises TFP chiefly by improving invention quality and reallocating tax resources, rather than by simply expanding input quantities. The 32 percent mediation share attributed to patent citations dovetails with the argument that digital tools enlarge the recombinant search space for innovation [88,89,90]. At the same time, the sizeable role of preferential tax treatment echoes evidence that fiscal incentives stimulate high-risk R&D when liquidity constraints are binding [91,92]. From a sustainable manufacturing standpoint, such innovation-led efficiency upgrading aligns with the Porter hypothesis and with eco-innovation studies emphasizing that innovation can reconcile competitiveness with sustainability objectives [86,87].

5.2. Links to the Industrial-Policy Literature

Endogenous-growth models posit that well-targeted subsidies can internalize knowledge spillovers and accelerate creative destruction [93]. Yet empirical results are mixed, with concerns over capture and misallocation. By uncovering a positive causal effect and by showing that the benefit is largest in fiscally capable provinces and digitally intensive sectors, our study aligns with work that finds smart industrial policy can succeed when administrative capacity is high [17]. The governor–STEM interaction instrument also resonates with the literature on implementation capability as a key condition for policy efficacy [94]. Conceptually, the pattern is consistent with the natural-resource-based view: building capabilities that reduce waste, redesign processes, and embed stewardship can yield a durable advantage; the innovation and reallocation channels we document are precisely the levers highlighted for performance and sustainability-oriented upgrading [92].

5.3. Implications for Regional Development

Agglomeration economists argue that knowledge spillovers are spatially bounded and strongest in dense, technologically specialized clusters [95]. Our heterogeneity results reveal that digital policies generate larger TFP gains in eastern provinces and digital-intensive industries, indicating that the digital economy can exacerbate existing regional disparities. Policymakers should therefore accompany digital roll-outs with complementary investments in lagging regions, echoing lessons from the U.S. broadband divide [96]. Consistent with Industry 4.0 evidence, the sustainability payoff from digitalization depends on organizational readiness and capability; without these complements, efficiency benefits remain localized and uneven [97,98].

5.4. Fiscal Design Matters

Nearly one-quarter of the total effect operates via tax incentives, reinforcing evidence from OECD economies that well-designed R&D credits crowd in private innovation [99,100]. However, the absence of positive effects in fiscally constrained provinces hints at a Matthew effect whereby rich regions harvest higher returns, in line with theoretical models of tax-induced agglomeration [101,102,103]. A progressive, formula-based central transfer system earmarked for digital infrastructure could help close this gap. Related reviews on whether it pays to be green suggest that performance gains from efficiency and innovation are most likely where firms can recoup the fixed costs of capability building and where policy design rewards innovation offsets, which matches our heterogeneity results [104].

5.5. Knowledge Spillovers: Modest but Non-Trivial

Although the share mediated by cross-firm patent citations is only 10 percent, the channel is statistically significant and aligns with the cluster literature that highlights citation-based knowledge flows [105,106,107]. Inter-firm data-sharing platforms, as recently promoted by the 2022 14th Five-Year Digital-Economy Plan, may further amplify this externality. From a sustainability perspective, digitally enabled transparency and data flows are also central in roadmaps that connect Industry 4.0 to triple-bottom-line outcomes, again pointing to capability-contingent payoffs rather than automatic gains [98].

5.6. Limitations and Avenues for Future Research

Our study has several limitations, which point to areas for future research. First, our sample is restricted to listed manufacturers. This is a limitation because our results may not generalize to the small and non-listed firms that account for most of China’s employment. Future work could also examine other types of firms, such as state-owned enterprises (SOEs) versus private firms, or exporters. Second, our policy index is quantity-based, treating all policy documents equally. This is a simplification, as the true impact of a policy likely depends on its legal weight, budgetary significance, or implementation authority. For example, a high-level strategic plan may have a different impact than a specific subsidy notice. Future research could create a more nuanced index by weighting these documents, which would provide a more precise measure of policy intensity. Third, our 2SLS model passes the Hansen J test, which supports our instrument choice. However, this causal claim could be further strengthened. Replicating the study with alternative instruments, such as historical telegraph lines [18], would be a valuable next step. The long span of our data is useful, but we note the inherent asymmetry: policy introduction was sparse in the early period (2008–2010) and dense later (2011–2023). Future research should build on our static causal findings by conducting a dynamic analysis to identify the full time-path of the policy’s effect, using methodologies suited for continuous treatment and heterogeneous timing on this asymmetrical structure. Finally, our analysis does not account for wider economic impacts. We do not measure potential general-equilibrium displacement effects [108]. We also do not test for spatial spillovers, which is the possibility that one province’s policy affects its neighbors. These remain important topics for future research. Structural models that combine input-output linkages with digital adoption could help close this gap.

5.7. Concluding Perspective

Synthesizing these results, our findings suggest that digital-economy policy is neither a panacea nor a white elephant. Its success hinges on local fiscal capacity, human capital, and the ability to couple fiscal or tax incentives with innovation-enabling institutions. As countries worldwide design post-COVID digital recovery plans, China’s multi-pronged approach, mixing tax relief, targeted subsidies, and pilot zones, offers a salient template, albeit one whose effectiveness depends critically on governance quality and complementary reforms. In terms of sustainable development, the documented productivity (efficiency) gains speak directly to the economic pillar of sustainability; the broader sustainability payoff is most likely when these improvements are embedded in eco-innovation and capability-building trajectories [86,98].

6. Conclusions

This paper investigates the causal impact of China’s provincial digital-economy policies on manufacturing firm productivity and unpacks the channels driving this effect. Using a manually coded policy index for the period from 2008 to 2023 and an instrumental-variable strategy, we find that these policies have a positive and economically significant causal effect on firm-level total factor productivity. A doubling of the cumulative policy stock increases TFP by approximately 3 percent. We further document that this aggregate impact is transmitted through four distinct and significant mechanisms: enhanced innovation quality, direct tax incentives, targeted digital subsidies, and knowledge spillovers. From a sustainability perspective, these findings speak to the economic pillar of sustainable manufacturing. The documented productivity gains are efficiency-centered, meaning that higher output is achieved with fewer or better deployed inputs, which is consistent with the notion of eco-efficiency.

Our heterogeneity analysis also indicates that the policy’s success is not unconditional. The productivity dividends are substantially larger in provinces with strong fiscal capacity and high education expenditure. Similarly, the policy yields greater returns for firms with high digital absorptive capabilities (digital intensity) and those located in the Eastern region. This suggests that digital-economy policy is not a panacea and not a white elephant; it is a form of smart industrial policy whose effectiveness depends on supportive institutions and firm capabilities. For policymakers, a practical implication is to link digital subsidies and tax relief to demonstrable process improvements and credible innovation outputs, to invest in local governance capacity and data infrastructure, and to provide targeted enablement for small- and medium-sized enterprises so that benefits are more inclusive.

Finally, we clarify the scope of our sustainability analysis. This paper provides robust evidence for the economic pillar of sustainability (eco-efficiency via TFP gains), and our analysis purposefully centers on this dimension. An important and logical next step for the literature is to investigate whether these digitally driven TFP improvements translate into tangible environmental outcomes, such as reductions in firm-level carbon emissions, energy intensity, or pollutant discharges. Linking the economic and environmental impacts in this way would build upon our findings to provide a more holistic understanding of how digital-economy policy contributes to a truly sustainable industrial transformation.