How to overcome the productivity pains of digital transformation? The U‑shaped TFP trajectory of Chinese agricultural firms, fixed‑asset constraints and subsidy threshold effects

Abstract

Digital transformation has profoundly reshaped agricultural production systems, yet its productivity effects remain insufficiently explored. Using panel data of Chinese listed agricultural firms from 2010 to 2022, this paper examines the effects, mechanisms, moderation, and heterogeneity of digital transformation on total factor productivity (TFP). First, digital transformation exerts a significant U-shaped nonlinear effect on TFP: the linear coefficient is −0.154 (p < 0.01), and the squared coefficient is 0.036 (p < 0.01), indicating that early-stage fixed asset accumulation, organizational restructuring, and system adaptation frictions drag down TFP, while after crossing the critical threshold, network synergy and factor allocation optimization drive sustained TFP growth. Second, the fixed asset-to-revenue ratio serves as a key transmission channel, forming a nonlinear mediation path of “digitalization → fixed asset-to-sales ratio → TFP.” Bootstrap tests show that the indirect effect turns from negative to positive as digitalization increases and is significant only at high levels. Third, government subsidies significantly moderate the U-shaped relationship, shifting the turning point leftward and increasing curvature, reflecting a “turning point forward shift and dividend amplification” mechanism. Fourth, the U-shaped relationship is more pronounced for non-state-owned, small-scale, and livestock enterprises. Theoretically, this paper constructs a nonlinear mediation framework of “digitalization–asset structure–TFP,” reveals the inverted U-shaped transmission of the fixed asset-to-revenue ratio, identifies the nonlinear moderating role of government subsidies, and enriches research on digital transformation and agricultural TFP. Practically, it suggests establishing a “pain period” tolerance mechanism, shifting subsidies toward soft support, and precisely supporting non-state-owned, small-scale, and livestock enterprises to release productivity dividends.

Introduction

Digital technologies are reshaping agricultural production modes, industrial organization, and factor allocation patterns. They exert profound influences on the productivity and long‑term resilience of the agricultural sector. As leading entities within China’s agricultural industrial chain, listed agricultural firms perform core functions. These functions include technological innovation, standardized production, and industrial upgrading. Accordingly, they act as critical micro‑level carriers for agricultural digitalization. Their digital practices represent the industry’s advanced level, and their transformation experience exerts demonstration and spillover effects on a wide range of agricultural operators. Digitalization spans digital technology, networks, and data, but its productivity effects are not automatic. Insufficient TFP gains often stem from delays caused by mismatches between digital investment and complementary intangible assets such as organizational and human capital (Brynjolfsson et al., 2021).

Agriculture features long production cycles, strong biological constraints, high asset specificity, and pronounced policy sensitivity. For this reason, productivity gains from digital transformation do not merely reflect automatic technological dividend release. Substantial frictions often arise during the transformation process. Simply put, two practically relevant questions arise: whether digitalization can sustainably lift total factor productivity (TFP), and whether this process exhibits stage‑driven nonlinearities. Answers to these questions help consolidate the micro‑foundation of agricultural modernization and stabilize food systems.

Existing literature identifies dual effects of digital technologies on agricultural development across macro and micro contexts. At the macro level, digital and ICT applications lift agricultural productivity and boost factor output efficiency (Bocean, 2024; Rajkhowa and Baumüller, 2024; Shamshiri et al., 2024; Ronzhin et al., 2025; He et al., 2025). Micro‑level evidence tells a complementary story (Khanna et al., 2022; Nath, 2024; Sharma et al., 2022). Farm automation and improved producer digital literacy raise technical efficiency and food output (Čechura et al., 2021; Pagliuca et al., 2025; Zhang et al., 2025).

Notably, digital agriculture constitutes systemic change rather than a simple stack of separate technologies (Walter et al., 2017). Its performance depends heavily on supporting infrastructure, digital literacy, and cross‑chain collaboration (Huyen et al., 2025; Xia et al., 2026). In other words, its empowering effects are not unconditional (Cano et al., 2026; McFadden et al., 2023). Institutional arrangements and actors’ adaptive capacity set binding limits. In extreme cases, digital adoption may even widen developmental gaps within agriculture (Bronson, 2018; Klerkx and Rose, 2020; Ehlers et al., 2021; MacPherson et al., 2022).

When it comes to transmission channels, most existing work explains digitalization‑driven efficiency gains through soft‑side factors. These soft‑side factors cover innovation capacity, human capital, and inclusive finance (Hu et al., 2024; Mengfang et al., 2026). Firm‑oriented studies add further insights. Digitalization can lift TFP via multiple practical pathways. These pathways include optimized capital allocation, reduced redundant labor, expanded R&D spending, green innovation, and scale expansion (Jin et al., 2024; Peng et al., 2026). Even so, digital dividends will not materialize on their own. External institutional and market environments impose real constraints (Dong and Xu, 2025; Li and Yu, 2025; Zhao et al., 2025).

Prior scholarship has laid a solid analytical foundation. Still, three notable research gaps deserve attention. The first concerns the assumed relationship between digitalization and TFP. Most papers postulate a straightforward positive linear link. A small subset detects U‑shaped nonlinear patterns. Yet these studies mostly stop at documenting curve shapes. They offer limited depth on the root causes of sunk costs and organizational frictions in early‑stage transformation.

The second pertains to mechanism identification. Existing mechanism tests concentrate heavily on soft channels such as innovation and human capital. They largely ignore hard constraints brought by fixed‑asset restructuring. Such restructuring matters greatly for agricultural firms given their high asset specificity. Few studies run targeted mediation tests from the fixed‑asset structure perspective. Moreover, most models implicitly assume constant mediation effects across all transformation phases.

The third relates to policy evaluation practice. Relevant work mostly estimates linear moderating effects generated by government subsidies. It pays insufficient attention to subsidy intensity thresholds. As a result, existing evidence cannot answer a key policy question. Under what conditions can subsidies help firms overcome transformation bottlenecks?

Against these documented gaps, this paper delivers three distinct contributions. First, it moves beyond standard linear presumptions. It characterizes the U‑shaped trajectory of agricultural listed firms’ digital transformation, where performance first declines and then improves. Rather than only confirming the existence of a U‑shape, this work unpacks the internal logic behind transformation‑related pains.

Second, it builds upon unique industrial attributes of agriculture. The fixed‑asset‑to‑sales ratio (FAS_ratio) is introduced as a sector‑specific mediating variable. This paper tests its inverted‑U transmission pathway. Unlike widely‑used generic mediators such as innovation and human capital, this variable highlights how asset‑structure restructuring shapes transformation frictions.

Third, it explores nonlinear moderation arising from government subsidies. It pinpoints meaningful boundaries for effective subsidy intensity. Low subsidy levels cannot offset sunk transformation costs. Subsidies need to cross a certain threshold to help firms pass the U‑shaped inflection point. By contrast, excessively generous subsidies may spur over‑investment in capital and worsen asset misallocation.

This paper therefore develops a nonlinear analytical framework: “digital transformation → fixed‑asset‑to‑sales ratio → TFP”. The analysis draws on panel data for Chinese A‑share listed agricultural firms over 2010–2022. It applies two‑way fixed‑effect specifications alongside nonlinear mediation and moderation models. Four core research questions guide the empirical analysis: whether digital transformation produces a U‑shaped nonlinear influence on agricultural firms’ TFP; whether the fixed‑asset‑to‑sales ratio operates as a nonlinear mediating channel; whether government subsidies exert nonlinear moderation over this U‑shaped relationship; and whether these effects display heterogeneity across ownership, firm size, and agricultural sub‑sectors.

In short, this study disentangles the dynamic productivity trajectory of agricultural digitalization. It illuminates asset‑structure‑based transmission mechanisms and clarifies boundary conditions for subsidy effectiveness. Its conclusions advance theoretical understanding of sustainable digital agriculture. They also supply empirical support for multiple stakeholders. Firms can reference these findings to design rational digital investment strategies. Policymakers can draw evidence to refine agricultural digital subsidy schemes and advance food security objectives.

Theoretical analysis and research hypotheses

Digital transformation and the U‑shaped trajectory of agricultural TFP

Based on organizational change and adjustment cost theories, digital technology exerts a non‑linear impact on firm productivity. Digital transformation entails coexisting adjustment costs and efficiency gains, yet the two effects do not synchronize over time. Adjustment costs dominate the early transformation stage, whereas productivity dividends emerge only after digital accumulation crosses a critical threshold, jointly forming a U‑shaped evolutionary trajectory of TFP. China’s A‑share agricultural listed firms cover the full industrial chain, including seed breeding, large‑scale cultivation, livestock breeding, feed production and agricultural processing. Leading firms in each segment have implemented diverse digital practices such as AI‑assisted breeding, intelligent field management and full‑chain digital governance, though transformation performance varies substantially across enterprises. Such heterogeneity is primarily determined by whether firms’ digital accumulation has transcended the cost‑dominated stage to capture efficiency dividends. As a radical organizational restructuring and factor reallocation process with high sunk costs, digital transformation inevitably generates pronounced time lags in productivity improvement.

This U‑shaped evolutionary pattern is more prominent in the agricultural sector than in manufacturing. Agricultural production is inherently constrained by biological growth cycles, climatic uncertainties in cultivation, and high asset specificity of breeding equipment. Coupled with fragmented industrial chain layout and insufficient digital literacy among grassroots practitioners, the transmission from digital investment to productivity growth is subject to more severe frictions and longer lag periods. These industry‑specific characteristics amplify adjustment costs in agricultural digitalization, prolonging the downward phase of the U‑curve and intensifying short‑term TFP losses.

Early transformation stage: cost frictions and TFP inhibition

Early digital construction requires substantial sunk costs in hardware and system deployment, platform iteration, and staff training. These upfront investments occupy limited working capital, technical talent, and managerial bandwidth, crowding out resources allocated to routine agricultural production and operation. This substitution bite is particularly sharp in agriculture, where long production cycles and rigid capital turnover leave little room for transitional slack.

Beyond resource crowding‑out, digital renovation conflicts with traditional agronomic practices and inherent operational routines, triggering organizational restructuring frictions. The reshaping of business processes and institutional frameworks disrupts stable production order in the short run. Meanwhile, insufficient cross‑subject data connectivity and immature industrial data assets hinder the in‑depth embedding of digital technologies. The absence of precise digital governance and factor synergy further delays efficiency improvement, while the learning costs of frontline employees exacerbate transformation pains. Collectively, overlapping cost frictions and delayed benefits suppress agricultural TFP during the initial transformation period.

Post‑threshold stage: digital empowerment and TFP promotion

Once digital investment accumulates beyond the critical threshold, firms enter the mature transformation stage. Mature digital infrastructure, unified data standards, and improved organizational adaptability turn early digital inputs into productive stock capital. This capital stock reduces marginal application costs and generates increasing marginal returns, while short‑term resource crowding‑out and organizational frictions gradually dissipate, driving sustained TFP growth through multiple efficiency‑enhancing channels.

Digital technologies penetrate core agricultural production and operational links to release systematic efficiency dividends. Precision monitoring and intelligent breeding technologies mitigate biological and climatic uncertainties, stabilizing agricultural output and industrial resilience. Cross‑link digital connectivity breaks the information barriers of the traditional decentralized agricultural chain, alleviating the information asymmetry between small‑scale production and large market demand and reducing transaction and matching costs. In addition, traceable and dynamic operational data support scientific production planning and resource scheduling, optimizing the allocation efficiency of core factors including capital, labor, and land. Continuous digital iteration further forms cumulative learning and scale effects, constructing a self‑reinforcing positive cycle of digital empowerment.

At this stage, digital efficiency gains fully offset residual adjustment costs, realizing a structural shift from TFP inhibition to TFP promotion. Driven by the increasing marginal returns of data factors, the productivity dividends of digital transformation continue to expand with the improvement of digital maturity.

Overall, the non‑linear impact of digital transformation on agricultural TFP stems from the dynamic shift between short‑term negative frictional effects and long‑term positive empowerment effects. Early digital transformation generates dominant cost frictions that restrain TFP, while digital development beyond the critical threshold delivers persistent productivity promotion. Accordingly, this study proposes the following hypothesis:

H1: Digital transformation exhibits a significant U‑shaped nonlinear relationship with the TFP of agricultural enterprises. Digitalization suppresses TFP at low transformation levels but positively promotes TFP after crossing a critical threshold.

The mediating role of fixed‑asset‑to‑revenue ratio: an inverted‑U channel

Agricultural firms face not only capital scarcity but also severe asset irreversibility: once sunk into highly specific facilities such as breeding bases and cold‑chain warehouses, capital cannot be easily redeployed. This lock‑in effect shapes how digital transformation transmits to productivity—not as a monotonic improvement, but as an inverted‑U mediated process via the fixed‑asset‑to‑revenue ratio (FAS_ratio). Drawing on adjustment cost theory and the resource‑based view, we argue that digital transformation exerts stage‑specific and asymmetric effects on TFP through this asset‑structure channel. Constrained by highly specific agricultural assets, the interplay between digital inputs and physical capital sequentially passes through two phases—investment misallocation and synergy release—and the mediating role of FAS_ratio correspondingly shifts from productivity suppression to productivity improvement.

Early‑stage transformation: rigid investment expansion and resource crowding‑out effect

In the initial phase of digital transformation, firms prioritize building digital infrastructure, deploying intelligent terminals, and implementing business information systems, alongside digital retrofitting of conventional production equipment. Compared with industrial sectors, digital retrofitting in agriculture features higher asset specificity, greater supporting modification costs, and longer adaptation cycles for real‑world production scenarios, leading to concentrated upfront capital outlays. On the basis of existing production facilities, firms install integrated digital systems including environmental monitoring, automated feeding, and intelligent control modules, which directly drive rapid expansion of physical capital stock.

Nevertheless, at the early transformation stage, digital technologies are yet to be deeply embedded into core production workflows, and increasing returns to scale from data factors have not materialized. Revenue growth remains heavily reliant on traditional operational modes and established sales channels and therefore fails to keep pace with the expansion of fixed assets. Rapid accumulation of fixed assets paired with lagging revenue growth raises capital lock‑in per unit of revenue and pushes up the FAS_ratio, rendering firms more capital‑intensive.

From a micro resource‑allocation perspective, a rising FAS_ratio indicates that substantial liquid capital is immobilized within highly specific physical assets, accompanied by declining asset turnover. For agricultural enterprises—which typically face financing frictions and high adjustment costs—this rising FAS_ratio significantly crowds out resources that could otherwise fund R&D, flexible factor deployment, and operational upgrading. Furthermore, early‑stage frictions such as data fragmentation and workflow incompatibility between newly deployed digital systems and legacy production systems persist. Firms bear dual capital burdens from maintaining conventional assets and funding newly‑added digital facilities, which further exacerbates maturity mismatch in investment and financing as well as inefficient capital lock‑in. In this phase, FAS_ratio operates as a binding constraint: digital‑driven asset expansion raises the ratio, diverting liquidity and managerial attention away from productivity‑enhancing activities—yielding a significant negative indirect effect on TFP.

Deepened transformation: systematic relaxation of efficiency constraints through digital‑asset integration

Once digital inputs cross a certain capability threshold, digital systems operate stably, sustained data accumulation drives iterative optimization of business processes, and scale as well as network synergy effects gradually emerge. The relationship between digital technologies and physical fixed assets evolves from capital accumulation and direct substitution in early‑stage transformation toward intelligent empowerment and stock‑asset revitalization, triggering a fundamental reversal in the evolutionary trend of FAS_ratio.

This transition is enabled by three mutually intertwined synergistic pathways.

First, digital tools—IoT sensors, predictive analytics, and remote diagnostics—enable lightweight retrofitting of existing machinery and cold‑chain assets. This improves utilization frequency and output efficiency without massive new capital outlays, mitigating the idleness caused by agriculture's seasonal cycles.

Second, removal of information barriers and optimized capital allocation. Mature digital systems break information silos across production, logistics, and sales chains and mitigate internal and external information asymmetry. On one hand, enhanced operational transparency improves corporate credit standing, expands external financing access, and optimizes capital structures. On the other hand, real‑time visibility over asset performance via digital platforms dismantles scheduling barriers across seasons, regions, and internal departments. Highly specific fixed assets can be coordinated and deployed across scenarios and time horizons, substantially lowering asset idleness rates.

Third, digital ecosystems foster leaner operational models. With the maturation of agricultural industrial internet platforms and socialized digital service providers, new business forms such as equipment sharing, capacity collaboration, and cloud‑based scheduling have proliferated. These developments help firms break away from the capital‑intensive path dependence of expanding production capacity via asset acquisition. Certain crop‑growing and seed‑breeding enterprises no longer need to continuously acquire proprietary fixed assets for capacity expansion. Instead, they leverage digital ecosystems to integrate external socialized resources and separate ownership from usage rights of fixed assets, thereby curbing excessive lock‑in of in‑house physical capital at the source. Meanwhile, high‑quality revenue growth driven by digital empowerment reduces capital lock‑in per unit revenue and yields a sustained decline in FAS_ratio.

Under the combined influence of the above three mechanisms, FAS_ratio enters a downward trajectory after crossing its inverted‑U inflection point. Capital misallocation pressures stemming from earlier capital‑intensive expansion are alleviated, and rigid constraints imposed by asset structures on production efficiency are gradually removed. The indirect transmission effect of FAS_ratio on TFP accordingly shifts: its negative influence diminishes continuously until asset structures no longer constitute core bottlenecks restraining digital efficiency gains, converging to a neutral state of fundamental support. This suggests that the mediating value of digital transformation on the asset side lies not in directly boosting productivity, but in reshaping asset structures to mitigate efficiency drag accumulated under traditional capital‑intensive operations.

In summary, digital transformation imposes an inverted‑U effect on agricultural firms’ fixed‑asset‑to‑sales ratio, transmitting to TFP via this asset‑structure channel through a stage‑contingent nonlinear mediating mechanism. At the early transformation phase, rigid investment raises the fixed‑asset‑to‑sales ratio. The mediating pathway theoretically generates efficiency suppression. As transformation deepens, the ratio falls once firms cross the inflection point. Integration between digital and tangible assets, alongside legacy asset revitalization, delivers productivity gains, so the mediating pathway yields efficiency enhancement in theory.

That said, how the mediating effect appears empirically depends on where the bulk of sampled firms sit along the inverted‑U curve. When most firms have moved past the inflection point, our sample will mostly capture the rising transmission on the curve’s right segment: the indirect effect switches from negative to positive and may become statistically significant at high digitalization levels. If most firms have yet to reach the inflection point, the declining transmission on the left segment dominates, and the indirect effect is expected to be largely negative. Should most firms cluster near the inflection point, the indirect effect approaches zero, and the mediating mechanism loses statistical significance.

No matter where the sample aggregate lies on the inverted‑U curve, the inverted‑U association between digital transformation and the fixed‑asset‑to‑sales ratio remains our core theoretical contention. The specific sign and magnitude of the indirect effect, by contrast, hinge on sample distribution. Accordingly, we formulate Hypothesis H2:

H2: The fixed‑asset‑to‑revenue ratio acts as a nonlinear mediator between digital transformation and TFP. Digital transformation exerts an inverted‑U influence on the fixed‑asset‑to‑sales ratio, and the sign and magnitude of the indirect effect change conditional on digitalization levels.

The moderating role of government subsidies

Digital transformation exerts a U‑shaped nonlinear influence on the total factor productivity (TFP) of agricultural enterprises. Conceptually, digital transformation refers to the integration of digital technologies, network infrastructure, and data resources into production, operations, management, and business models, thereby reshaping value creation and factor allocation. TFP improvement refers to efficiency gains beyond factor accumulation, reflecting technological progress, allocative efficiency, organizational capability, and scale effects. Government subsidies refer to resource transfers and policy signals provided through fiscal grants, special support, and tax incentives, which affect firms’ investment, financing, and risk-taking. Resource constraints refer to structural limitations arising from financing constraints, cash-flow pressures, and shortages of complementary assets such as organizational capital, human capital, and data governance capacity. Firms first experience an efficiency‑erosion phase before crossing a critical threshold to enter an efficiency‑gain regime. Fundamentally, this process reflects a dynamic trade‑off between sunk‑cost absorption and critical‑scale formation. Nevertheless, agricultural enterprises are generally characterized by low profit margins, severe financing constraints, and lengthy investment payback periods. During the efficiency‑erosion stage, cash‑flow pressures may force firms to discontinue digital‑related investments, making it difficult for them to traverse the U‑shaped inflection point. These financing constraints, cash-flow pressures, and shortages of complementary assets constitute the resource constraints that shape the digitalization–TFP relationship for agricultural enterprises. Resource constraints therefore act as a boundary condition: the stronger they are, the less likely digitalization is to translate into TFP gains. Government subsidies relax these constraints, and in doing so they not only directly moderate the digitalization–TFP link but also indirectly shape it through the resource-constraint channel. As a key industrial policy instrument, government subsidies moderate this relationship primarily by altering firms’ capacity to bear the fixed costs of digital transformation, thereby reshaping the inflection‑point position, curve slope, and transition conditions of the underlying U‑shaped link between digital transformation and TFP. Specifically, subsidies operate through three distinct channels.

First, subsidies defray rigid investment burdens and incentivize more intensive digital investment in the early stage. In the initial phase of digital transformation, hardware acquisition, system deployment, and equipment retrofitting entail substantial sunk costs, continuously depleting corporate cash flow before efficiency gains materialize. Government subsidies introduce external capital, alleviating cash-flow constraints and improving the feasibility of digital investment; they also lower the marginal cost of such investment, encouraging firms to expand their scale. Under these conditions, firms are more willing to make larger, more concentrated digital investments during early transformation. Yet more intensive early-stage investment also intensifies frictions in resource reallocation. Compatibility costs between traditional assets and digital systems, together with pressures to reengineer organizational processes, are released in a concentrated manner over the short run, making the efficiency-loss phase steeper and deeper. At the same time, larger-scale investment accelerates data accumulation and system integration, enabling firms to reach critical scale sooner. Thus, while steepening the downward segment, subsidies shift the U-shaped inflection point to the left.

Second, subsidies supply patient capital to facilitate firms’ attainment of critical scale. Productivity spillovers from digital technologies do not increase linearly with investment inputs. Large‑scale efficiency dividends from network synergy, data‑driven decision‑making, and process reengineering emerge only after data accumulation, intelligent equipment penetration, and system interconnection reach certain thresholds. Bound by tight financing frictions, agricultural enterprises relying solely on internal resources often fail to sustain investment until such critical conditions are met. Their digital initiatives consequently remain confined to localized pilot projects, without generating systematic, large‑scale value from digital‑physical integration. Government subsidies function as risk‑tolerant patient capital that delivers supplementary resource support, raising the probability that firms will cross the U‑shaped inflection point and move into the efficiency‑gain phase.

Third, subsidies transmit policy signals to mobilize external resources and enhance market credibility. Beyond direct financial transfers, fiscal subsidies signal to financial institutions and supply‑chain partners that the recipient represents a high‑quality digital‑transforming entity. Such policy signals reduce information‑screening costs for external investors, broaden firms’ access to external financing, and amplify the leveraging effect of public funds. Meanwhile, government subsidies enhance firms' risk‑taking capacity for digital experimentation and capability‑building investments. This encourages managerial decision‑making to shift from short‑term survival orientation toward long‑term growth objectives, optimizes factor allocation, and further unlocks productivity gains once digital transformation matures (Figure 1).

Conceptual theoretical framework. Solid lines indicate formal hypothesized paths (H1–H3), whereas dashed lines denote exploratory heterogeneity analyses without prespecified hypotheses. H1 assumes a U-shaped relationship between digital transformation and TFP. H2 identifies the inverted-U nonlinear mediating role of the fixed-asset occupancy rate. H3 posits that government subsidies moderate this U-shaped association by shifting the inflection point leftward and increasing curvature.

In sum, government subsidies operate as a nonlinear moderator: they encourage earlier and more intensive digital investment, shift the U-shaped inflection point leftward, and steepen the declining segment of TFP, while amplifying productivity gains along the rising segment. This moderating effect, however, is intensity‑contingent. Low‑intensity subsidies cannot fully offset initial sunk costs and are insufficient to help firms cross the inflection point; as a result, the U-shaped evolutionary process is unlikely to emerge. Only when subsidies reach a certain threshold can firms overcome the critical barriers and fully realize the U‑shaped productivity evolution. We therefore propose the following research hypothesis:

H3: Government subsidies exert a nonlinear moderating effect on the U-shaped relationship between digital transformation and TFP in agricultural enterprises. In the declining phase, they encourage firms to cross the inflection point earlier, accelerating the rate of productivity suppression; in the rising phase, they amplify productivity gains.

Materials and methods

Data sources and sample processing

Our initial sample consists of Chinese A‑share listed firms classified under agriculture, forestry, animal husbandry, fisheries, and agricultural support services, following the China Securities Regulatory Commission (CSRC) industry codes. We set the sample period to 2010–2022, a time window covering the early deployment and subsequent rapid expansion of digital practices across the agricultural sector.

We apply two standard sample‑screening filters. Firms flagged as ST (special treatment) or *ST are removed to avoid estimation distortions induced by severe financial distress. We also drop observations with substantial missing values for core financial variables. These screening procedures yield an unbalanced panel of 48 firms across 13 years.

Financial accounting data are sourced from the CSMAR database. To construct our digital transformation measure, we retrieve the Management Discussion and Analysis (MD&A) sections from firm annual reports published on the CNINFO platform and conduct textual analysis using the dictionary‑based approach described in the subsequent section.

Variable selection and measurement

Dependent variable: total factor productivity

We estimate firm-level total factor productivity (TFP) of agricultural listed firms using a Cobb–Douglas production function. Output is measured by operating revenue, labor by the number of employees, and capital by net fixed assets. Intermediate inputs are constructed as the sum of operating costs, selling expenses, administrative expenses, and financial expenses, minus depreciation and amortization and cash paid to and for employees. Investment is measured by cash paid for fixed assets, intangible assets, and other long-term assets. All monetary variables are expressed in ten thousand yuan. For data preparation, missing values of depreciation and amortization are replaced with zero. Output, labor, capital, intermediate inputs, and investment are log-transformed after adding one, and winsorized annually at the 1st and 99th percentiles.

We estimate the production function and compute firm-level TFP using the LP (Levinsohn and Petrin, 2003) and OP (Olley and Pakes, 1996) methods. Both methods follow two-stage semi-parametric estimation and are implemented using the levpet and opreg commands in Stata, respectively. For the LP estimator, intermediate inputs serve as the proxy variable, labor as the free variable, and capital as the state variable. For the OP estimator, investment serves as the proxy variable, while capital and firm age are specified as state variables. Labor is treated as a free variable, and industry, year, and province dummies are included as additional free variables. An exit variable is included to correct for sample selection bias, and ownership type and export participation are further controlled for. Both methods use two-stage semi-parametric estimation. The LP-based TFP, denoted TFP_LP, serves as our baseline measure, and the OP-based TFP, denoted TFP_OP, is used for robustness checks.

Core explanatory variable: digital transformation

We measure corporate digital transformation using a text-mining word-frequency approach. Following the quantification methods of Wu et al. (2021) and Zhao et al. (2021), we extract textual information from corporate annual reports to construct our digital transformation measures. Following Wu et al. (2021), we first collect the full text of annual reports, perform text cleaning and word segmentation, and count the frequencies of keywords related to artificial intelligence, blockchain, cloud computing, big data, and digital technology applications. We then take the natural logarithm of the total keyword frequency plus one to obtain the baseline digital transformation index, Digital, which is used in the main regression analysis.

For robustness checks, following Zhao et al. (2021), we first extract the full text of annual reports and calculate keyword frequencies across four dimensions: digital technology application, internet business models, intelligent manufacturing, and modern information systems. We then use the entropy weight method to construct a composite digital transformation score as an alternative proxy, denoted DigitalA.

Mediating variable: fixed‑asset‑to‑sales ratio

We employ the fixed‑asset‑to‑sales ratio (FAS_ratio), defined as net fixed assets divided by operating revenue, as the mediating variable. This indicator measures fixed‑asset intensity per unit of revenue and captures capital lock‑in, rising asset specificity, and rigid fixed costs accompanying hardware investment expansion in agricultural digitalisation. It constitutes a key structural variable that underpins the crowding‑out and constraint‑relaxation mechanisms hypothesized in our theoretical framework.

Moderating variable: government subsidies

Subsidy intensity is measured as the ratio of total annual government subsidies to total assets, which mitigates size‑induced measurement distortions. To examine the nonlinear moderating effect of government subsidies on the U‑shaped relationship between digital transformation and TFP, including changes in curve shape, inflection‑point position, and interval effects, we construct interaction terms between subsidy intensity and both the linear and quadratic terms of the digital transformation index.

Control variables

We include the following firm‑level controls, guided by prior literature and the specific operational features of agricultural enterprises: firm size (Size), leverage ratio (Lev), financing constraints (SA), return on assets (ROA), and ownership concentration (Shrcr1). These variables account for confounding effects on productivity arising from resource endowments, financial structures, financing frictions, profitability, and governance heterogeneity.

Definitions, symbols, and measurement specifications for all variables are summarized in Table 1.

Variable definitions and measurements.

Variable category | Variable name | Symbol | Variable definition
Dependent variable | Total factor productivity | TFP | Firm-level total factor productivity estimated using the LP method.
Independent variable | Digital transformation | Digital | ln(1+total frequency of digital transformation keywords extracted from firm annual reports).
Mediating variable | Fixed-asset-to-sales ratio | FAS_ratio | Net fixed assets divided by operating revenue.
Moderating variable | Government subsidy | Subsidy | Total government subsidies divided by total assets.
Control variable | Firm size | lnAssets | Natural logarithm of total assets.
Asset-liability ratio | Lev | Total liabilities divided by total assets.
Financing constraints | SA_index | SA index for firm financing constraints, Hadlock and Pierce (2010).
Return on assets | ROA | Net profit divided by total assets.
Ownership concentration | Shrcr1 | Shareholding ratio of the largest controlling shareholder.

Model specification

Baseline model: U‑shaped effect

To test the hypothesized U‑shaped effect of digital transformation on agricultural listed firms' TFP, we specify the following two‑way fixed‑effects model with a quadratic term for digitalization: TF P i , t = α 0 + β 1 Digita l i , t + β 2 Digita l i , t 2 + ∑ γ j Contro l j ( i , t ) + u i + μ t + ε i , t (1)

Where TFPi,t denotes total factor productivity; Digitali,t and Digital2i,t are the digital transformation index and its squared term; ∑γj Controlj(i,t) denotes a vector of firm‑level control variables; and ui, μt, and εi,t capture firm fixed effects, year fixed effects, and the error term, respectively. Equation 1 serves to verify Hypothesis H1.

Nonlinear mediation model

Conventional stepwise mediation tests are ill‑suited for nonlinear transmission mechanisms. Following Hayes and Preacher (2010), we adopt a bias‑corrected Bootstrap procedure to trace the inverted‑U‑shaped mediation path through the fixed‑asset‑to‑sales ratio (FAS_ratio).

The mediation equation is specified as: FAS _ rati o i , t = α 0 + a 1 Digita l i , t + a 2 Digita l i , t 2 + ∑ γ j Contro l j , ( i , t ) + u i + μ t + ε i , t (2)

where FAS_ratioi,t refers to the fixed-asset-sales ratio of firm i in year t, and a1 and a2 are the core coefficients for digital transformation. All other settings mirror those of the baseline regression. A pattern with a1 > 0 and a2 < 0 confirms an inverted U-shaped relationship (rising then falling). Equation 2 serves to verify Hypothesis H2.

The outcome equation takes the form: TF P i , t = α 0 + c 1 ' Digita l i , t + c 2 ' Digita l i , t 2 + bFixed _ Asse t i , t + ∑ γ j Contro l j , ( i , t ) + u i + μ t + ε i , t (3)

where variables follow the definitions above; b captures the linear impact of the fixed-asset occupancy rate on TFP; c1' and c2' denote the direct-effect coefficients conditional on the mediator. The remaining specifications are consistent with the baseline model.

Using the nonlinear mediation system (Equations 2, 3), we compute instantaneous indirect and direct effects and test the statistical significance of the nonlinear mediating mechanism. These estimations test Hypothesis H3.

Nonlinear moderation model

To examine whether government subsidies reshape the U‑shaped curve and shift its inflection point, we introduce interaction terms between subsidy intensity and both the linear and quadratic terms of digitalization, as specified in Equation 4: TF P i , t = α 0 + ω 1 Digita l i , t + ω 2 Digita l i , t 2 + ω 3 Subsid y i , t + ω 4 Digita l i , t × Subsid y i , t + ω 5 Digita l i , t 2 × Subsid y i , t + ∑ γ j Contro l j , ( i , t ) + u i + μ t + ε i , t (4)

where Digitali,t × Subsidyi,t and Digital2i,t × Subsidyi,t are the interaction terms linking (squared) digital transformation to government subsidy intensity, which identify the nonlinear moderating effect of subsidies. All other variable definitions and model settings follow the baseline regression.

Results

Descriptive statistics

All continuous variables are winsorized at the 1st and 99th percentiles to reduce the influence of outliers. Table 2 reports the summary statistics. The mean (standard deviation) of TFP_LP is 7.856 (0.839), indicating considerable productivity dispersion across agricultural listed firms. The primary digital‑transformation index (Digital) averages 2.381, with a maximum of 4.595, suggesting that while most firms remain at early digitalization stages, a few have progressed substantially. The fixed‑asset‑to‑sales ratio (FAS_ratio) exhibits a mean of 0.665 and ranges from 0.092 to 3.010, confirming the high capital‑intensity typical of agricultural operations. The subsidy variable has fewer observations (N = 449) than the full sample (N = 465) due to missing subsidy records for certain firm‑year pairs. Overall, the sample displays adequate variation in key variables to identify the hypothesized nonlinear relationships.

Descriptive statistics of variables.

Variable | Observations | Mean | Std. dev. | Min | Max
TFP_LP | 465 | 7.856 | 0.839 | 3.998 | 10.664
TFP_OP | 465 | 6.245 | 0.654 | 2.917 | 8.250
Digital | 465 | 2.381 | 0.951 | 0.000 | 4.595
DigitalA | 465 | 0.872 | 0.937 | 0.000 | 3.367
FAS_ratio | 465 | 0.665 | 0.576 | 0.092 | 3.010
Subsidies | 449 | 0.010 | 0.019 | 0.000 | 0.118
lnAssets | 465 | 21.866 | 1.011 | 19.478 | 25.986
Lev | 465 | 0.430 | 0.195 | 0.057 | 0.9801
ROA | 465 | 0.018 | 0.077 | −0.284 | 0.270
Shrcr1 | 465 | 34.306 | 15.543 | 3.966 | 74.950
SA_index | 465 | −3.809 | 0.232 | −4.430 | −3.197

The number of observations for government subsidies is 449 because data on subsidy amounts for 16 observations are missing from the annual reports of sample firms.

Baseline regression and U‑shaped relationship test

To rigorously examine the nonlinear impact of digital transformation on TFP among agricultural listed firms and validate the theoretical mechanisms and research hypotheses developed above, this section estimates the baseline regression Equation 1. Table 3 reports the detailed estimation outcomes.

Baseline regression results.

Variable | (1) | (2) | (3) | (4)
Linear only & without controls | Without controls | Linear only | Baseline
Digital | 0.103** | −0.242** | −0.005 | −0.154***
(0.061) | (0.119) | (0.033) | (0.053)
Digital2 | – | 0.083*** | – | 0.036***
– | (0.025) | – | (0.012)
lnAssets | – | – | 0.552*** | 0.530***
– | – | (0.055) | (0.056)
Lev | – | – | 0.440** | 0.416*
– | – | (0.208) | (0.209)
ROA | – | – | 1.539*** | 1.566***
– | – | (0.336) | (0.326)
Shrcr1 | – | – | 0.001 | 0.001
– | – | (0.004) | (0.004)
SA_index | – | – | 0.630** | 0.422
– | – | (0.276) | (0.306)
Constant | 7.611*** | 7.889*** | −2.045 | −2.252
(0.144) | (0.156) | (1.541) | (1.580)
Firm FE | Yes | Yes | Yes | Yes
Year FE | Yes | Yes | Yes | Yes
Observations | 465 | 465 | 465 | 465
Adjusted R2 | 0.823 | 0.811 | 0.904 | 0.890

*p < 0.1, **p < 0.05, ***p < 0.01. Robust firm‑level clustered standard errors in parentheses. This convention holds for all subsequent tables unless noted otherwise.

Baseline regression results

We estimate Equation 1 with two‑way fixed effects to test the nonlinear effect of digital transformation on agricultural firms' TFP. Table 3 reports the results. Column (1) reports the regression results with only the linear term of digital transformation, while Column (2) includes the quadratic term of digital transformation. Columns (3)–(4) present the regression results after incorporating control variables: Column (3) contains only the linear term of digital transformation, and Column (4) adds its quadratic term. Taken together, these results demonstrate that the specification incorporating the quadratic term of digital transformation is necessary.

In Column (4), the linear coefficient of digital transformation is −0.154 (p < 0.01) and the quadratic coefficient is 0.036 (p < 0.01), meeting the necessary condition for a U‑shaped relationship. This provides initial support for Hypothesis H1, which predicts an initial decline followed by a subsequent rise in TFP as digitalization deepens.

The control variables generally behave as expected. Firm size, leverage, and profitability (ROA) all exhibit significantly positive effects on TFP, suggesting that scale economies, prudent leverage, and operational efficiency underpin productivity in agricultural firms. Ownership concentration and financing constraints carry the anticipated signs but are statistically insignificant, implying that governance structure and financial frictions play a more limited marginal role in our sample. The adjusted R2 of 0.890 indicates a good model fit.

Validation of the U‑shaped pattern

Following Haans et al. (2016), we conduct a formal U‑shape test. The estimated inflection point is 2.14, which falls well within the observed range of the digital transformation [0, 4.595]. The marginal slope is significantly negative at the lower bound (digitalization = 0) and significantly positive at the upper bound (digitalization = 4.595), confirming the statistical validity of the U‑shape.

Figure 2 plots the predicted TFP against digital transformation, along with the 95% confidence interval and the inflection point coordinates (2.14, 7.82). The distribution of observations relative to the turning point is noteworthy: 60.43% of firm‑year observations lie to the right of the inflection point, indicating that the majority of sample firms have already entered the efficiency‑enhancing phase of digitalization. The remaining 39.57% are located on the left side, still trapped in the transition‑pain stage.

The U-shaped relationship between digital transformation and TFP of agricultural listed firms. The solid navy line denotes predicted TFP values. Grey vertical bars indicate the 95% confidence interval derived from the delta method with firm‑clustered standard errors. The red dashed vertical line marks the estimated turning point of the U‑shaped curve. Rug ticks at the bottom illustrate the sample distribution of the digital‑transformation index.

Taken together, these results strongly support Hypothesis H1. Digital transformation exerts a significant U‑shaped nonlinear effect on agricultural TFP, characterized by an early suppression phase followed by a subsequent productivity dividend. The inflection‑point and slope tests jointly corroborate the robustness of this finding.

Endogeneity treatment

Endogeneity concerns arise from reverse causality between firms’ digital‑transformation choices and total factor productivity, as well as sample‑selection bias. We mitigate these threats using the control function approach, instrumental‑variable two‑stage least‑squares (IV‑2SLS), and propensity‑score matching (PSM). The corresponding estimation outputs are reported in Columns (1)– (3) of Table 4.

Endogeneity test results.

Variable | (1) | (2) | (3)
CF approach | IV-2SLS | PSM
Digital | −0.138** | −0.152*** | −0.220**
(0.067) | (0.056) | (0.0843)
Digital2 | 0.033** | 0.036*** | 0.0499***
(0.014) | (0.013) | (0.0172)
v ̂ | 0.194 | – | –
(0.251) | – | –
v ̂ 2 | 2.309 | – | –
(4.425) | – | –
Controls | Controlled | Controlled | Controlled
Firm FE | Yes | Yes | Yes
Year FE | Yes | Yes | Yes
First‑stage F‑statistic | 7909.150 | 2872.450 | –
Kleibergen-Paap rk LM | – | 15.830 | –
Kleibergen-Paap Wald F | – | 2872.450 | –
Observations | 465 | 465 | 441

The instruments are the same-year peer mean of digital transformation and its square. Column (1) reports the control function approach, treating both Digital and Digital2 as endogenous; the instruments are the same-year peer mean of digital transformation and its square. Column (2) reports an auxiliary 2SLS specification with only Digital treated as endogenous; the equation is exactly identified, so Hansen J cannot be computed. Column (3) reports PSM with kernel matching; 24 unmatched observations are dropped, leaving 441 valid firm‑year observations. *p < 0.1, **p < 0.05, ***p < 0.01.

Control function approach

The linear and quadratic terms of corporate digital transformation are both theoretically susceptible to endogeneity concerns. Following Wooldridge (2015), this study employs the control function approach as the primary method to address potential endogeneity, with corresponding estimation results presented in Column (1) of Table 4. In the first-stage regression, we construct instrumental variables using the annual average digital transformation level of other peer firms and its quadratic term to estimate the Digital variable and extract the corresponding residual series. In the second stage, we incorporate this residual and its quadratic form into the baseline specification as extra control variables, so as to capture the unobserved confounding factors that may jointly affect the endogenous digital transformation indicators.

The first-stage F-statistic reaches 7909.15, strongly validating the effectiveness of the selected instrumental variables. The second-stage estimation results show that the linear coefficient of digital transformation is −0.138 (Bootstrap SE = 0.067, p = 0.039) and the quadratic coefficient is 0.033 (Bootstrap SE = 0.014, p = 0.022). These results are highly consistent with the baseline findings. Meanwhile, both the residual and its quadratic term show no statistical significance, which implies that endogeneity bias in the original model is negligible and further confirms the robustness of our core conclusions. To alleviate estimation bias induced by the generated regressor issue, we compute standard errors based on 1000 bootstrap replications with firm-level clustering.

IV‑2SLS estimation results

We first attempted 2SLS estimation treating both Digital and Digital2 as endogenous variables. Unfortunately, the instruments lacked sufficient explanatory power for Digital2, with the Kleibergen-Paap rk Wald F statistic equal to 4.325. This value falls below the Stock-Yogo 10% critical value of 7.03, pointing to a weak-instrument problem that renders the estimates unreliable. To further verify whether our conclusions hold under strong instruments, we therefore present an auxiliary IV specification here. In this setup, only Digital is treated as endogenous, while Digital2 enters the main equation as an exogenous regressor; the corresponding results are shown in Column (2) of Table 4. Admittedly, this specification is not theoretically ideal. Even so, its strong instruments deliver supporting evidence that our core findings remain unchanged under a strong-instrument framework.

Our instrument is the average digital transformation level of other firms within the same year. This measure is neither an industry-year average nor a lagged variable. A single firm can hardly affect this yearly average of other peers. Moreover, year fixed effects are included in the model, absorbing common aggregate shocks across years. The first-stage Kleibergen-Paap rk Wald F statistic stands at 2872.45, far exceeding the Stock-Yogo 10% critical value of 16.38, which confirms strong instruments. The coefficient on Digital is −0.152 (p < 0.01), and that on Digital2 is 0.036 (p < 0.01). These estimates align closely with those from the baseline regression and the control function approach.

Given that the number of instruments equals the number of endogenous regressors, the Hansen J statistic equals zero, so the overidentification test cannot be implemented. To address this limitation, we further incorporate the one-period lagged value of digital transformation as an additional instrument. This yields two instruments for one endogenous variable, leaving one degree of freedom for the overidentification test. The resulting Hansen J statistic is 2.093 (p = 0.148), so we fail to reject the null hypothesis that the instruments are exogenous. Still, the lagged term may be contaminated by the persistent lagged effects of digital transformation, making the exogeneity argument less convincing. For this reason, we do not adopt it as our primary instrument. To further address possible industry-level time-varying shocks, we add industry-year trend terms in a robustness check; the results remain unchanged. Taken together, we report the 2SLS results where only Digital is endogenous, as supplementary evidence alongside the control function estimates.

Propensity‑score matching

To alleviate self‑selection bias stemming from firms’ endogenous digital‑transformation decisions, we split the sample into treatment and control groups based on the median value of Digital, and perform kernel matching using the full set of control variables together with year dummies as covariates. Regression results based on the matched sample are shown in Column (2) of Table 4. The coefficient of Digital is significantly negative at the 5% significance level, while its squared term Digital2 is significantly positive at the 1% level. The significance of the control variables remains largely consistent with the baseline regression.

We further adopt alternative matching algorithms, namely one‑to‑one nearest‑neighbour matching and radius matching, and obtain stable results across specifications. In all three PSM exercises, Digital carries a significantly negative coefficient and Digital2 a significantly positive coefficient. The implied inflection points range from 2.204 to 2.452, all lying within the empirical sample range. This range brackets the baseline inflection point of 2.14, with the slightly higher values consistent with the expectation that correcting for self‑selection raises the threshold required for digital dividends to materialize. This evidence suggests that our baseline results are not an artefact of sample‑selection bias.

Robustness checks

We conduct three sets of robustness checks—alternative measurement of the dependent variable, alternative measurement of the core explanatory variable, and sample period adjustment—to assess the sensitivity of our baseline findings. Table 5 reports the results.

Robustness checks of digital transformation on agricultural firms' TFP.

Variable | (1) | (2) | (3)
Alternative dependent variable | Alternative independent variable | Adjusted sample window
Digital | −0.178*** | – | −0.151**
(0.063) | – | (0.056)
Digital2 | 0.049*** | – | 0.037***
(0.017) | – | (0.013)
DigitalA | – | −0.132** | –
– | (0.058) | –
DigitalA2 | – | 0.053** | –
– | (0.022) | –
Controls | Controlled | Controlled | Controlled
Firm FE | Yes | Yes | Yes
Year FE | Yes | Yes | Yes
Observations | 465 | 465 | 355
Adjusted R2 | 0.796 | 0.890 | 0.882

Alternative dependent‑variable measure

We re‑estimate Equation 1 using TFP computed via the Olley‑Pakes (OP) method (TFP_OP) as the dependent variable. Column (1) of Table 5 presents the results. The coefficient on Digital is −0.178 (p < 0.01) and on Digital2 is 0.049 (p < 0.01), compared with −0.154 and 0.036 in the baseline LP specification. The slightly larger magnitudes under the OP estimator are consistent with the fact that OP uses investment as a proxy for productivity shocks—which, given the lumpy nature of agricultural capital expenditure, may magnify the magnitude of the estimated U‑shaped association. The inflection point (1.82) lies within the sample range, confirming that the U‑shape is robust to the choice of TFP estimator.

Alternative core explanatory‑variable measure

To address measurement‑error concerns in the construction of our digitalization index, we replace Digital (based on Wu et al., 2021) with DigitalA (based on Zhao et al., 2021), which employs an expanded keyword dictionary. Column (2) of Table 5 shows that the coefficient on DigitalA is −0.132 (p < 0.05) and on DigitalA_sq is 0.053 (p < 0.05). Although the magnitudes differ somewhat from the baseline (−0.154 and 0.036), the sign pattern and statistical significance remain unchanged, and the implied inflection point (1.25) falls inside its sample range [0, 3.367]. This alleviates concerns that our main results are driven by the specific choice of the keyword dictionary.

Adjustment of the sample time window

The COVID‑19 pandemic (2020–2022) constituted an exogenous shock to agricultural production, supply chains, and firm operations, which could confound the digitalization‑TFP relationship. To isolate this effect, we exclude observations from 2020–2022 and re‑estimate the baseline model. Column (3) of Table 5 reports the results. The coefficient on Digital is −0.151 (p < 0.01) and on Digital2 is 0.037 (p < 0.01), close to the baseline estimates. The inflection point in this subsample is 2.04, slightly lower than the baseline value of 2.14, suggesting that the pandemic may have raised the threshold for digital dividends to materialize; once this crisis period is removed, firms reach the turning point somewhat earlier. Crucially, the U‑shape persists even when this period is excluded.

Overall, the U‑shaped relationship between digital transformation and agricultural TFP remains stable across alternative variable definitions and sample compositions. Control variable estimates show no material shifts. Collectively, these results reinforce the reliability and generalizability of our baseline conclusions.

Mechanism and moderating effect analysis

Mechanism analysis: the mediating role of fixed‑asset‑to‑sales ratio

As theorized in Section 2.2, the fixed‑asset‑to‑Sales ratio (FAS_ratio) serves as the core mediating channel through which digital transformation transmits stage‑dependent effects on TFP. Existing studies document that fixed‑asset investment acts as a critical mediating linkage connecting digital transformation and firm‑level TFP (Zhao, 2026; Chen et al., 2026). Prior literature documents significant U-shaped patterns between digital technologies and TFP in service industries, rendering conventional linear mediation methods (e.g., Baron and Kenny, 1986) inappropriate for our nonlinear research setting. Accordingly, following Hayes and Preacher (2010), we adopt a bias-corrected Bootstrap procedure to compute instantaneous indirect effects across different levels of digital transformation.

We estimate Equations 2, 3 using mean‑centered variables to mitigate multicollinearity. The results are reported in Table 6. Column (1) reports a linear coefficient of digitalization of −0.067, which is not statistically significant, and a quadratic coefficient of −0.055 (p < 0.05). The inflection point lies at c_Digital = −0.61, falling within the sample range, thus confirming an inverted-U relationship between digital transformation and FAS_ratio. This implies that firms first accumulate fixed assets during early‑stage digital investment, and then rationalize their asset structures through sharing, leasing, and more efficient utilization after crossing the inflection point.

Nonlinear mediation effect test results.

Variable | (1) | (2)
FAS_ratio | TFP
c_Digital | −0.067 | −0.036
(0.050) | (0.026)
c_Digital2 | −0.055** | −0.009
(0.021) | (0.012)
c_ FAS_ratio | — | −0.823***
— | (0.095)
Controls | Controlled | Controlled
Firm FE | Yes | Yes
Year FE | Yes | Yes
Observations | 465 | 465
Adjusted R2 | 0.824 | 0.962

Bootstrap test for nonlinear mediating transmission mechanism
Digital level | Indirect effect | BC 95% CI | Direct effect | BC 95% CI
Low level (mean − 1SD) | −0.031(0.036) | [−0.100, 0.041] | −0.019(0.020) | [−0.060, 0.018]
Average level (mean) | 0.055(0.039) | [−0.031, 0.123] | −0.035(0.025) | [−0.085, 0.011]
High level (mean + 1SD) | 0.140**(0.059) | [0.015, 0.18] | −0.052(0.044) | [−0.158, 0.024]

Bootstrap-derived standard errors are reported in parentheses. To mitigate multicollinearity arising from quadratic‑term specifications, all core explanatory variables and the mediator are mean‑centred prior to estimation. *p < 0.1, **p < 0.05, ***p < 0.01.

Column (2) reveals that the coefficient on FAS_ratio is −0.823 (p < 0.01), indicating that a higher asset‑to‑revenue ratio substantially suppresses TFP. This negative effect is economically intuitive: excessive fixed‑asset holdings inflate depreciation and maintenance costs, reduce capital turnover and operational flexibility, and generate frictions that drag down productivity.

To formally test the mediation significance, we compute the instantaneous indirect and direct effects at three representative levels of digital transformation: low (one standard deviation below the mean), average, and high (one standard deviation above the mean). We implement 1,000 Bootstrap replications clustered at the firm level and use bias‑corrected 95% confidence intervals (BC 95% CI) to judge statistical significance.

Bootstrap test results show that the instantaneous indirect effects at low and mean levels of digital transformation are −0.031 and 0.055, respectively, with 95% confidence intervals that include zero and are therefore insignificant. At the high level, the instantaneous indirect effect increases to 0.140, with a confidence interval that excludes zero, and is significant at the 5% level. For the direct effects, the coefficients at low, medium, and high levels of digital transformation are −0.019, −0.035, and −0.052, respectively; all confidence intervals include zero, and none passes the significance test. This indicates that, within the sample period, the direct effect of digital transformation on firm TFP is insignificant, and its enabling effect is transmitted entirely through the nonlinear mediating channel of the fixed-asset-to-revenue ratio.

These results verify Hypothesis H2, namely that the fixed-asset-to-revenue ratio plays a nonlinear mediating role between digital transformation and TFP. Because about 75% of firms in the sample have crossed the inflection point (located at c_Digital = −0.61, with the 25th percentile of c_Digital at −0.59, the median at 0.015, and the mean at 0), they lie on the right half of the inverted-U curve. This study therefore mainly observes the declining segment: the higher the digitalization level, the lower the FAS_ratio. In this case, a decline in the FAS_ratio instead produces a positive indirect effect, and the faster the decline, the larger the indirect effect, manifesting as incremental transmission. This result is consistent with the Bootstrap test, in which the indirect effect is significantly positive at the high level. The negative effects at low and medium levels are insignificant, possibly because few low-digitalization firms are present in the sample, resulting in insufficient statistical power.

Moderating effect of government subsidies on the nonlinear U‑shaped relationship

Government subsidies, as a key policy instrument for industrial digitalization, directly offset sunk costs, ease financing constraints, and absorb part of the experimentation costs associated with digital transformation. To test whether subsidies reshape the U‑shaped digitalization‑TFP relationship, we introduce government subsidies along with their interaction terms with both the linear and quadratic terms of digital transformation Equation 4. The estimation results are reported in Table 7.

Moderation‑effect estimates of government subsidies.

Variable | (1) | (2) | (3)
Full sample | Low subsidy group | High subsidy group
c_Digital | 0.024 | −0.027 | 0.054
(0.027) | (0.045) | (0.036)
c_Digital2 | 0.037*** | 0.034 | 0.050***
(0.012) | (0.021) | (0.016)
c_subsidy | −3.463** | – | –
(1.612) | – | –
c_Digital#c_Subsidy | 2.902** | – | –
(1.676) | – | –
c_Digital2#c_Subsidy | 1.193** | – | –
(0.590) | – | –
Controls | Controlled | Controlled | Controlled
Firm FE | Yes | Yes | Yes
Year FE | Yes | Yes | Yes
Observations | 449 | 220 | 237
Adjusted R2 | 0.921 | 0.922 | 0.924

All variables are mean‑centered. The high‑ and low‑subsidy groups are divided based on the median value of government subsidy intensity. *p < 0.1, **p < 0.05, ***p < 0.01.

Column (1) of Table 7 presents the full‑sample regression with interaction terms. The coefficient on c_Digital is 0.024, which is not statistically significant, and the coefficient on c_Digital2 is 0.037 (p < 0.01), confirming that the baseline U‑shape remains robust. Among the interaction terms, c_Digital#c_Subsidy is significantly positive (2.902, p < 0.1) and c_Digital2#c_Subsidy is significantly positive (1.193, p < 0.05). These signs indicate that subsidies steepen the U‑curve and shift the inflection point leftward. Evaluated at the mean subsidy level, the inflection point is [2.06] (compared with 2.14 in the baseline), confirming that policy support enables firms to cross the turning point at a lower digitalization threshold.

To further unpack the conditional nature of this moderating effect, we split the sample into low‑ and high‑subsidy subsamples using the median subsidy intensity. Columns (2) and (3) report the group‑wise results. In the low‑subsidy group, the coefficients on c_Digital and c_Digital2 are both insignificant, indicating that without sufficient policy funding, the nonlinear effect of digital transformation on TFP does not emerge. In the high-subsidy group, the coefficient on c_Digital2 is 0.050 and significantly positive at the 1% level, whereas c_Digital is insignificant, indicating that under higher subsidy intensity, the U-shaped nonlinearity of digital transformation on TFP is significantly present. These subsample results suggest that the moderating effect of subsidies depends on the intensity of policy support; only when subsidy scale crosses a certain threshold can the U-shaped efficiency gains from digitalization be released.

Taken together, these results support Hypothesis H3. Government subsidies perform a dual moderating function in the TFP evolution of agricultural firms undergoing digital transformation: shifting the inflection point leftward and amplifying the resulting gains. By restructuring the risk-return profile of firm transformation and altering the efficiency transmission path of asset allocation, subsidies reshape the nonlinear relationship between digital transformation and TFP. On the one hand, they shift the U-shaped inflection point leftward, enabling firms to cross the efficiency-loss phase earlier; on the other hand, they increase the curvature of the curve, making the efficiency gains after crossing the inflection point more pronounced.

Heterogeneity analysis: systematic differences across firm groups

Following Haans et al. (2016), we estimate full‑sample models that interact group dummies with both the linear and quadratic terms of digital transformation. As their framework suggests, group differences in U‑shaped relationships cannot be reliably determined based solely on the statistical significance of individual interaction terms. We therefore perform a joint Wald test to assess whether the full coefficient vector of the nonlinear specification differs systematically across subgroups. Notably, a significant Wald test only confirms overall statistical divergence between two nonlinear curves and does not necessarily imply distinct U‑shaped patterns across groups. We thus further supplement the full‑sample interaction results with split‑sample regressions to compare curve configurations across subgroups.

Heterogeneity across ownership types

Table 8 reports the ownership‑based heterogeneity results. Column (1) presents the full‑sample interactive‑model estimates, while Columns (2) and (3) show split‑sample results for non‑state‑owned enterprises (non‑SOEs) and state‑owned enterprises (SOEs), respectively. The joint Wald test yields a p‑value of 0.048, rejecting the null hypothesis that the two groups share identical nonlinear effects at the 5% significance level. This indicates statistically meaningful differences in the nonlinear productivity trajectories between SOEs and non‑SOEs.

Ownership‑based heterogeneity test results.

Variable | (1) | (2) | (3)
Full sample | Non‑SOEs | SOEs
c_Digital | 0.064* | 0.044* | −0.090
(0.036) | (0.042) | (0.056)
c_Digital2 | 0.053*** | 0.047** | −0.002
(0.015) | (−0.019) | (−0.024)
SOE | −0.053 | – | –
(0.125) | – | –
SOE#c.c_Digital | −0.113** | – | –
(0.052) | – | –
SOE#c.c_Digital2 | −0.045* | – | –
(0.026) | – | –
Controls | Controlled | Controlled | Controlled
Firm FE | Yes | Yes | Yes
Year FE | Yes | Yes | Yes
Observations | 464 | 253 | 209
Adjusted R2 | 0.891 | 0.926 | 0.908
Wald test p‑valuefor between‑group differences | 0.048

SOE is a property‑right dummy variable. It equals 1 for state‑owned enterprises and 0 for non‑state‑owned enterprises. *p < 0.1, **p < 0.05, ***p < 0.01.

The split‑sample results are displayed in Columns (2) and (3). For non‑SOEs, the linear term of digital transformation is significantly negative, and the quadratic term is significantly positive, confirming a robust U‑shaped relationship. In contrast, for SOEs, neither the linear nor the quadratic coefficient achieves statistical significance, so no meaningful U‑shaped pattern emerges within this subsample. Taken together, these results suggest that the U‑shaped digitalization‑TFP trajectory exists among non‑SOEs but is statistically absent in the SOE sample.

Heterogeneity across firm size

Table 9 reports the firm‑size heterogeneity results. Column (1) presents the full‑sample interactive‑model estimates, while Columns (2) and (3) present split‑sample results for large and small agricultural firms, respectively. The joint Wald test yields a p‑value of 0.021, rejecting the null hypothesis that the two groups share identical nonlinear effects at the 5% significance level. This indicates statistically meaningful differences in the nonlinear productivity trajectories between large and small agricultural firms.

Firm‑size heterogeneity test results.

Variable | (1) | (2) | (3)
Full sample | Large firms | Small firms
c_Digital | 0.091** | 0.003 | 0.079*
(0.038) | (0.043) | (0.040)
c_Digital2 | 0.049*** | 0.066** | 0.051***
(0.014) | (0.0274) | (0.014)
Is_large | −0.006 | |
(0.074) | |
is_large#c.c_Digital | −0.134** | |
(0.027) | |
is_large#c.c_Digital2 | 0.011 | |
(0.027) | |
Controls | Controlled | Controlled | Controlled
Firm FE | Yes | Yes | Yes
Year FE | Yes | Yes | Yes
Observations | 465 | 227 | 229
Adjusted R2 | 0.918 | 0.926 | 0.860
Wald test p‑valuefor between‑group differences | 0.021

is_large is a firm‑size dummy variable. It equals 1 for large‑scale enterprises and 0 for small‑scale enterprises. *p < 0.1, **p < 0.05, ***p < 0.01.

The split‑sample results are displayed in Columns (2) and (3). For large firms, the coefficient on c_Digital is 0.003 and is not statistically significant, while the coefficient on c_Digital2 is 0.066 and significant at the 5% level. For small firms, the coefficient on c_Digital is 0.079 and significant at the 10% level, and the coefficient on c_Digital2 is 0.051 and significant at the 1% level. A U-shaped relationship holds for both large and small firms, yet the linear coefficient is markedly lower for large firms than for small firms. The estimated inflection points are 1.61 for small firms and 2.36 for large firms. This implies that large firms require a higher level of digitalization to cross the turning point and unlock productivity gains—a pattern consistent with the higher adjustment costs and organizational rigidities inherent in large agricultural firms’ digital transformation.

Heterogeneity across sub‑sectors

The agricultural sector features long industrial chains and diverse business formats, with marked disparities across sub‑sectors in production‑operation modes, digital infrastructure, and transition pathways. We divide our sample into two groups: livestock‑and‑support‑service firms and crop‑planting firms. Table 10 reports the heterogeneity estimates. Column (1) presents the full‑sample interactive‑model results, while Columns (2) and (3) show split‑sample regressions for the two groups.

Subsector heterogeneity test results.

Variable | (1) | (2) | (3)
Full sample | Livestock & ancillaryservices | Crop farming
c_Digital | −0.062 | 0.065* | −0.032
(0.073) | (0.037) | (0.045)
c_Digital2 | 0.022 | 0.039** | 0.019
(0.021) | (0.017) | (0.025)
Is_livestock | −0.109 | – | –
(0.068) | – | –
is_livestock#c.c_Digital | 0.148** | – | –
(0.059) | – | –
is_livestock#c.c_Digital2 | 0.017 | – | –
(0.030) | – | –
Controls | Controlled | Controlled | Controlled
Firm FE | Yes | Yes | Yes
Year FE | Yes | Yes | Yes
Observations | 465 | 248 | 216
Adjusted R2 | 0.892 | 0.922 | 0.892
Wald test p‑valuefor between‑group differences | 0.039

is_livestock is an industry dummy variable. It equals 1 for listed agricultural firms operating in livestock‑raising and supporting service sectors, corresponding to industry codes A03‑A04 and A05; it equals 0 for listed agricultural firms engaged in crop planting, corresponding to industry codes A01‑A02. *p < 0.1, **p < 0.05, ***p < 0.01.

The joint Wald test yields a p‑value of 0.039, rejecting the null hypothesis of identical nonlinear effects at the 5% significance level. This indicates statistically meaningful differences in the digitalization‑TFP relationship across the two sub‑sectors. The split‑sample results are displayed in Columns (2) and (3). For livestock and agricultural support‑service firms, the linear coefficient is 0.065 (p < 0.1), whereas the quadratic coefficient is 0.039 and significant at the 5% level. This pattern suggests that these firms —mainly those in the livestock and auxiliary services sector—have largely moved beyond the downward-sloping segment of the U-curve, with the positive marginal-return phase predominantly detectable within our sample window. In contrast, neither the linear nor the quadratic term is statistically significant for crop-planting firms, indicating that no nonlinear productivity effect is detectable in this subsample. In contrast, neither the linear nor the quadratic term is statistically significant for crop‑planting firms, indicating that no nonlinear productivity effect is detectable in this subsample.

These results suggest that the nonlinear TFP effect of digital transformation does not hold universally across all agricultural sub‑sectors; its presence is highly contingent on sector‑specific production attributes. Specifically, the positive marginal‑return pattern emerges in livestock and support‑service sectors, where production cycles are relatively shorter, and digital‑adaptation costs may be more readily absorbed. By contrast, crop‑planting firms—characterized by longer production cycles and stronger natural‑environment constraints—show no statistically identifiable nonlinear response.

Discussion

Our findings reveal that digital transformation does not generate immediate productivity gains for agricultural firms. Instead, it follows a U‑shaped trajectory, with transitional efficiency losses giving way to productivity dividends only after firms cross a critical threshold. This nonlinear pattern is transmitted through asset‑structure adjustments, asymmetrically moderated by government subsidies, and varies markedly across firm ownership types, scales, and sub‑sector attributes. The following section interprets these empirical outcomes, situates them within existing scholarly debates, and elaborates on their practical and policy implications.

Nonlinear mechanism in baseline regressions: cost‑benefit trade‑offs behind the “initial‑suppression‑later‑promotion” pattern

Baseline results confirm a significant U‑shaped relationship between digital transformation and the TFP of agricultural listed firms, indicating that agricultural digitalization is not an instant efficiency‑enhancing upgrade, but a staged evolutionary process involving transitional efficiency frictions and subsequent dividend release. Accordingly, evaluating the productivity implications of digital transformation cannot rely solely on average effects; the dynamic transitional stage of individual firms serves as a critical precondition for understanding its economic consequences.

From a cost‑benefit perspective, early-stage digital transformation entails substantial sunk costs and coordination frictions arising from specialized hardware and software procurement, organizational restructuring, and workforce skill adaptation. The resultant managerial frictions and resource misallocation inevitably suppress operational productivity. As digitalization advances beyond the inflection point, network synergies and increasing returns to scale gradually emerge, breaking inter‑organizational information barriers and optimizing factor allocation. These improvements offset upfront transformation costs and ultimately promote TFP growth. Given the inherent characteristics of agricultural production—including spatial dispersion, long production cycles, and strong natural constraints—the productivity value of digital technologies relies heavily on iterative operational adaptation and internal organizational adjustment.

These findings enrich the literature on the digitalization‑productivity nexus. Macro-level studies have documented significant agricultural productivity improvements driven by ICT penetration and digital technology adoption (Bocean, 2024; Rajkhowa and Baumüller, 2024). Nevertheless, such evidence predominantly captures average net effects, potentially masking the transitional efficiency losses incurred by micro-level firms in the early digitalization stage—a nuanced distinction highlighted by our U-shaped trajectory. Furthermore, while existing manufacturing studies generally identify positive linear productivity effects of digital transformation (Wu et al., 2025), our agricultural context yields a distinct U-shaped pattern. This discrepancy suggests that industrial attributes fundamentally reshape the cost‑benefit structure of digital transformation, indicating that manufacturing-based conclusions cannot be directly generalized to the agricultural context. For policy practice, these outcomes underline the necessity of tolerating short-term efficiency losses and establishing inclusive institutional arrangements to accommodate transitional digitalization frictions in agriculture.

Mediating mechanism: fixed‑asset configuration reshapes nonlinear efficiency profiles

The mediation results further demonstrate that the U-shaped digitalization‑TFP relationship is largely shaped by firms’ fixed-asset restructuring behavior. In the initial digitalization stage, firms purchase substantial smart equipment and digital infrastructure, pushing up the fixed-asset-to-revenue ratio. Heavy-asset expansion generates a negative indirect effect and becomes an important source of efficiency pains. At this stage, digitalization is manifested more in asset scale expansion, while production processes and operating models are not simultaneously optimized. Asset accumulation instead squeezes liquidity resources and intensifies cost pressures.

As the level of digitalization rises further, the pace of fixed-asset expansion slows, and the negative pressure from heavy assets gradually weakens and turns into a positive effect. Notably, within our sample period, nearly 75% of agricultural firms have already crossed the fixed-asset expansion stage. At high levels of digital transformation, a decline in the fixed-asset-to-revenue ratio instead produces a significant positive indirect effect, and the faster the decline, the larger the indirect effect, manifesting as incremental transmission. The direct effect remains insignificant throughout, indicating that the direct technology-enabling effect of digitalization is relatively limited. The rebound of the U-shaped curve stems more from efficiency recovery after the pressure of heavy-asset expansion subsides than from output leaps directly brought by digital technology. This finding suggests that digitalization for agricultural firms is not equivalent to equipment stacking and asset expansion; hardware investment lacking process innovation, scenario implementation, and production structure optimization is difficult to translate into genuine productivity gains. Firms need to avoid the “heavy-asset trap” to release the long-term value of digital technology.

Prior studies mostly attribute digital productivity improvement to soft-factor upgrades such as innovation capability and human capital accumulation (Hu et al., 2024; Mengfang et al., 2026; Jin et al., 2024). Such literature generally assumes stable and linear mediating effects across transformation stages. By contrast, this study reveals a unique asset-structure reconstruction channel in agricultural firms and demonstrates the phase-dependent nature of mediating mechanisms throughout digital transformation.

Moderating mechanism: asymmetric phase‑specific empowerment of government subsidies

Government subsidies exert a distinct asymmetric moderating effect on the U-shaped digitalization‑TFP relationship, effectively shifting the inflection point leftward and altering curve curvature. However, its moderating effectiveness is highly contingent on firms’ digital transformation stages. In the early transition period, subsidy policies predominantly target digital equipment procurement, which inadvertently incentivizes excessive fixed-asset investment, exacerbates asset stockpiling and resource crowding-out, and ultimately intensifies short-term productivity suppression. This undesirable outcome stems from hardware-biased subsidy design, which lacks systematic support for organizational reform, operational optimization, and workforce training, thereby generating noticeable policy distortion. Only after firms cross the transformation inflection point can the productivity-enhancing advantages of subsidies be fully released, accelerating TFP improvement.

These results indicate that subsidy empowerment exhibits prominent stage thresholds. Uniform, one-size-fits-all financial support cannot effectively resolve transitional frictions; instead, differentiated subsidy schemes tailored to firms’ dynamic digitalization stages are essential. Moreover, this finding differs from the inverted‑U-shaped moderating effect identified in the logistics industry (Yang et al., 2026). This divergence suggests that subsidy effectiveness is highly sector-contingent, implying that policy formulation must fully incorporate industry-specific production characteristics and institutional environments.

Interpretation of heterogeneous mechanisms

Ownership‑type heterogeneity

Ownership-based heterogeneity stems from divergent operational objectives, governance mechanisms, and institutional constraints. Non-state-owned firms are market-oriented and face intense competitive pressure, with flexible organizational structures that enable efficient adaptation during digital restructuring. After overcoming transitional frictions, digital technologies can be deeply embedded throughout production and operational processes, generating sustained productivity spillover effects. In contrast, state-owned enterprises undertake multiple policy objectives and are constrained by institutional inertia and rigid performance assessment mechanisms. Their digital transformation often exhibits formalistic tendencies characterized by excessive hardware procurement and insufficient practical scenario application, thereby making it difficult to translate digital investment into substantive productivity improvement. Consequently, the U-shaped efficiency reversal is significantly weakened among state-owned firms. These heterogeneous outcomes align with existing evidence, which suggests that non-state-owned firms are more responsive to digital technological progress (Cao and Tang, 2025; Cui et al., 2024).

Firm‑size heterogeneity

The size-based heterogeneity results are consistent with the resource-based view (Barney, 1991). Large firms possess abundant capital and technological resources, enabling large-scale digital infrastructure renovation. Nevertheless, their complex hierarchical structures and strong organizational path dependence incur higher restructuring and trial-and-error costs in the early transformation stage, resulting in more pronounced short-term efficiency suppression. Once organizational adaptation is completed, large firms fully leverage scale economies and resource integration advantages, accelerating the release of digital productivity dividends. By comparison, despite limited resources, small firms are organizationally flexible and possess strong dynamic adjustment capabilities, enabling them to translate digital investments into efficiency gains more quickly. Consequently, their inflection point occurs at a lower level of digitalization, and marginal returns are realized earlier.

Sub‑sector heterogeneity

Sub-sector heterogeneity is fundamentally determined by differences in asset specificity, production cycles, and natural constraint intensity. Livestock and agricultural service industries feature standardized production procedures, high compatibility between digital systems and business processes, and strong digital asset versatility. These characteristics shorten organizational adaptation cycles and enable rapid productivity dividend release after breaking through cost thresholds. In contrast, crop planting is highly constrained by natural conditions and seasonal cycles. Precision agricultural technologies require long-term data accumulation, multi-cycle model iteration, and continuous technical calibration, resulting in high uncertainty and delayed productivity returns. Accordingly, crop planting faces a higher threshold for efficiency reversal, with limited digital dividend release observed within the sample period. Consistent with this logic, crop-planting firms in our sample show no statistically detectable progress toward the inflection point, as neither the linear nor the quadratic term is significant in this subsample, which explains the absence of significant nonlinear productivity patterns in crop planting.

In sum, our results suggest that navigating the transition‑pain phase of agricultural digitalization requires more than simply scaling up technology inputs. Prudent management of fixed‑asset expansion, combined with phase‑tailored subsidy policies that balance hardware investment with organizational and operational support, is critical for helping agricultural firms escape the transitional efficiency trap and enter a self‑reinforcing phase of digital productivity dividends.

Conclusion

This study systematically examines the nonlinear evolutionary pattern, internal transmission mechanisms, and boundary conditions of digital transformation's impact on agricultural enterprises' total factor productivity. Its core theoretical contribution lies in its contextual differentiation from existing industrial evidence. Whereas manufacturing sectors tend to generate immediate, linear productivity gains from digitalization, agricultural digital transformation features prominent stage‑specific cost thresholds and delayed dividend release. The productivity reversal of digitalization in agriculture is fundamentally shaped by firms’ fixed‑asset expansion trajectories and dynamic transformation stages, with external subsidy policies further conditioning this nonlinear evolutionary process.

These findings yield clear policy implications. To advance high-quality agricultural digitalization, policy support should move beyond the traditional hardware-oriented subsidy model toward a refined governance framework centered on stage diagnosis and tiered, phase-tailored assistance. In particular, for firms in the period of efficiency pains, cost-sharing support should be provided to help them cross the U-shaped inflection point; for digitally mature firms, the policy focus should shift toward industrial collaboration and operational optimization. Complementary institutional reforms are also needed to alleviate problems such as asset rigidity and inefficient digital investment in state-owned agricultural enterprises, and to narrow the digital development divide among different types of agricultural operating entities.

Limitations and future prospects

Drawing on a sample of A‑share‑listed agricultural firms, this study identifies the nonlinear relationship between digital transformation and total factor productivity, alongside the mediating role of asset‑structure adjustment and the moderating boundary of government subsidies. Nevertheless, several limitations remain, which open avenues for future research.

First, endogeneity concerns persist. Although instrumental‑variable and PSM approaches are adopted to mitigate reverse causality and omitted‑variable bias, the non‑experimental nature of observational data prevents these strategies from fully eliminating potential confounding influences from unobserved time‑varying factors. Examples include managerial cognitive biases toward digital technologies and gradual organizational restructuring within firms. Accordingly, our findings should be interpreted primarily as robust statistical associations between digital transformation and firm productivity, rather than strict causal evidence. Future work may exploit exogenous policy shocks to construct quasi‑natural‑experiment settings and achieve more credible causal identification.

Second, the mechanism analysis admits further refinement. Constrained by data availability, this study unpacks efficiency frictions arising from digital transformation mainly from the perspective of fixed‑asset stock structure. It does not, however, disentangle heterogeneous effects between hardware‑oriented digital investment and software‑driven data‑capacity building. Subsequent research could compare divergent evolutionary trajectories for these two categories of digital inputs.

Third, policy‑related analyses admit further depth. This paper focuses on the aggregate asymmetric moderating effect of government subsidies without differentiating subsidy spending purposes and usage structures. Subsidies earmarked for distinct objectives exert substantially different influences on corporate asset restructuring and productivity dynamics. Follow‑up studies may disaggregate policy‑instrument types, compare the synergistic effects of alternative policy mixes, and inform better‑targeted governance frameworks for digital agriculture.

Finally, there are constraints on sample representativeness. The sample consists of A‑share‑listed agricultural firms, which are mostly large‑scale, market‑oriented leading enterprises. While they capture digital‑development patterns among modern agricultural corporations, they cannot represent grassroots operators such as farmer cooperatives and small‑ and medium‑sized agribusinesses. Future research could broaden sample coverage to examine digital‑transformation patterns across agricultural agents of varying scales and business forms, so as to further enrich the micro‑level theoretical framework for digital‑agriculture productivity.

References

References

BarneyJ. B. (1991). Firm resources and sustained competitive advantage. J. Manag. 17, 99–120. doi: 10.1177/014920639101700108

BaronR. M. KennyD. A. (1986). The moderator-mediator variable distinction in social psychological research: conceptual, strategic, and statistical considerations. J. Pers. Soc. Psychol. 51, 1173–1182. doi: 10.1037//0022-3514.51.6.1173

BoceanC. G. (2024). A cross-sectional analysis of the relationship between digital technology use and agricultural productivity in EU countries. Agriculture 14:519. doi: 10.3390/agriculture14040519

BronsonK. (2018). Smart farming: including rights holders for responsible agricultural innovation. Technol. Innov. Manag. Rev. 8, 7–14. doi: 10.22215/timreview/1135

BrynjolfssonE. RockD. SyversonC. (2021). The productivity J-curve: how intangibles complement general purpose technologies. Am. Econ. J. Macroecon. 13, 333–372. doi: 10.1257/mac.20180386

CanoP. MandriniG. BuckmasterD. ArnallB. CarrollM. ShardaA. . (2026). Farmer perspectives on digital agriculture in the U.S. Midwest. Agron. J. 118:e70268. doi: 10.1002/agj2.70268

CaoY. TangL. (2025). Digitalization and agricultural businesses’ environmental sustainability. Financ. Res. Lett. 86:108442. doi: 10.1016/j.frl.2025.108442

ČechuraL. Žáková KroupováZ. KostlivýV. (2021). Productivity and efficiency of precision farming: the case of Czech cereal production. Agris On-Line Pap. Econ. Inf. 13, 15–24. doi: 10.7160/aol.2021.130302

ChenJ. LiuY. LiangX. YaoH. (2026). The impact of digital inclusive finance on the development of agricultural producer services: a case study of Tiandong County. Front. Sustain. Food Syst. 10:1773976. doi: 10.3389/fsufs.2026.1773976

CuiY. ZhaoC. ZhangQ. (2024). Impact of digital transformation in agribusinesses on total factor productivity. Int. Food Agribus. Manag. Rev. 27, 843–857. doi: 10.22434/ifamr1064

DongJ. XuJ. (2025). Driving agricultural strength through digital transformation. Sci. Rep. 16:446. doi: 10.1038/s41598-025-29878-3, 41326544.

EhlersM.-H. HuberR. FingerR. (2021). Agricultural policy in the era of digitalisation. Food Policy 100:102019. doi: 10.1016/j.foodpol.2020.102019

HaansR. F. J. PietersC. HeZ.-L. (2016). Thinking about U: theorizing and testing U- and inverted U-shaped relationships in strategy research. Strateg. Manag. J. 37, 1177–1195. doi: 10.1002/smj.2399

HadlockC. J. PierceJ. R. (2010). New evidence on measuring financial constraints: Moving beyond the KZ index. The Review of Financial Studies, 23, 1909–1940. doi: 10.1093/rfs/hhq009

HayesA. F. PreacherK. J. (2010). Quantifying and testing indirect effects in simple mediation models when the constituent paths are nonlinear. Multivar. Behav. Res. 45, 627–660. doi: 10.1080/00273171.2010.498290, 26735713

HeJ. WeiZ. LeiX. (2025). Unveiling the digital revolution: catalyzing total factor productivity in agriculture. PLoS One 20:e0318333. doi: 10.1371/journal.pone.0318333

HuY. LiuJ. ZhangS. LiuY. XuH. LiuP. (2024). New mechanisms for increasing agricultural total factor productivity: analysis of the regional effects of the digital economy. Econ. Anal. Policy 83:766. doi: 10.1016/j.eap.2024.07.017

HuyenN. T. T. ThanhT. T. HaL. T. (2025). Developing sustainable and resilient agriculture system of European countries amid the rise of digital governance. Sustain. Futur. 10:101292. doi: 10.1016/j.sftr.2025.101292

JinS. R. TangS. Y. RenZ. J. (2024). Can digital transformation improve the total factor productivity of agricultural enterprises? Reform 2, 131–148.

KhannaM. AtallahS. S. KarS. SharmaB. WuL. YuC. . (2022). Digital transformation for sustainable agriculture in the United States: opportunities and challenges. Agric. Econ. 53, 924–937. doi: 10.1111/agec.12733

KlerkxL. RoseD. (2020). Dealing with the game-changing technologies of agriculture 4.0: how do we manage diversity and responsibility in food system transition pathways?Glob. Food Secur. 24:100347. doi: 10.1016/j.gfs.2019.100347

LevinsohnJ. PetrinA. (2003). Estimating production functions using inputs to control for unobservables. Rev. Econ. Stud. 70, 317–341. doi: 10.1111/1467-937X.00246

LiJ. YuY. (2025). Impact of government public data openness on agricultural total factor productivity. Int. Rev. Econ. Financ. 104:104561. doi: 10.1016/j.iref.2025.104561

MacPhersonJ. Voglhuber-SlavinskyA. OlbrischM. SchöbelP. DönitzE. MouratiadouI. . (2022). Future agricultural systems and the role of digitalization for achieving sustainability goals. A review. Agron. Sustain. Dev. 42:70. doi: 10.1007/s13593-022-00792-6, 35818482

McFaddenJ. NjukiE. GriffinT. (2023). Precision agriculture in the digital era: Recent adoption on U.S. farms (Economic Information Bulletin No. EIB-248). U.S. Department of Agriculture, Economic Research Service. Available online at: https://www.ers.usda.gov/publications/105893 (Accessed June 25, 2026).

MengfangX. I. E. YangyangZ. YichunL. I. WeiJ. I. A. (2026). The impact of digital transformation on total factor productivity of agricultural-related enterprises. Res. Agric. Mod. 47, 611–622. doi: 10.13872/j.1000-0275.2025.0812

NathS. (2024). A vision of precision agriculture: balance between agricultural sustainability and environmental stewardship. Agron. J. 116, 1126–1143. doi: 10.1002/agj2.21405

OlleyG. S. PakesA. (1996). The dynamics of productivity in the telecommunications equipment industry. Econometrica 64, 1263–1297. doi: 10.2307/2171831

PagliucaL. G. CarrerM. J. DamascenoR. de Mello Brandão VinholisM. de Souza FilhoH. M. (2025). Precision agriculture technologies adoption and technical efficiency of soybean farms in São Paulo, Brazil. Precis. Agric. 27:11. doi: 10.1007/s11119-025-10308-3

PengS. WangH. ZhangJ. HuW. HuangM. (2026). How does agricultural digital infrastructure construction empower high-quality agricultural development? The mediating roles of green technology innovation and farmland scale management. Agriculture 16:1037. doi: 10.3390/agriculture16101037

RajkhowaP. BaumüllerH. (2024). Assessing the potential of ICT to increase land and labour productivity in agriculture: global and regional perspectives. J. Agric. Econ. 75, 477–503. doi: 10.1111/1477-9552.12566

RonzhinA. FigurekA. SurovtsevV. DibirovaK. (2025). Digital transformation and precision farming as catalysts of rural development. Land 14:1464. doi: 10.3390/land14071464

ShamshiriR. R. SturmB. WeltzienC. FultonJ. KhoslaR. SchirrmannM. . (2024). Digitalization of agriculture for sustainable crop production: a use-case review. Front. Environ. Sci. 12:1375193. doi: 10.3389/fenvs.2024.1375193

SharmaV. TripathiA. K. MittalH. (2022). Technological revolutions in smart farming: current trends, challenges & future directions. Comput. Electron. Agric. 201:107217. doi: 10.1016/j.compag.2022.107217

WalterA. FingerR. HuberR. BuchmannN. (2017). Smart farming is key to developing sustainable agriculture. Proc. Natl. Acad. Sci. 114, 6148–6150. doi: 10.1073/pnas.1707462114, 28611194.

WooldridgeJ. M. (2015). Control function