Agricultural blockchain readiness environment and carbon emission intensity in maize production: evidence from provincial panel data in major maize-producing areas of China

1 Introduction and literature review

Decarbonization of agricultural production is not solely a question of reducing factor inputs. In major maize-producing areas, emissions are influenced by production scale, input allocation, information flows, service provision, and the institutional capacity to coordinate dispersed producers. Digital technologies may improve these conditions by supporting information sharing, traceability, and more standardized production management. Blockchain-related applications are of potential relevance because they can provide tamper-resistant records and facilitate verification across supply-chain participants. Nonetheless, the empirical question is not whether blockchain technology is intrinsically low-carbon, but rather whether the regional conditions that make such applications feasible are associated with variations in agricultural carbon emission intensity.

This distinction matters for measurement. The ABRE Index used in this study is a composite measure of digital infrastructure, the agricultural industrial base, and policy support. It is therefore an index of regional readiness and enabling conditions, not a direct observation of the number, intensity, or effectiveness of blockchain applications actually deployed in maize production. The positive correlation between block_index and blockchain-related patent activity (r = 0.1883, p = 0.0114) provides supplementary evidence of blockchain relevance, but it does not validate the index as realized adoption. The empirical interpretation must consequently remain at the level of association between readiness conditions and emissions outcomes.

Prior studies have linked digital-economy development and agricultural carbon performance through several related channels. One strand emphasizes digital infrastructure and information access, which may improve the allocation of inputs and facilitate the diffusion of green technologies. Digital economy may reduce agricultural carbon intensity by promoting agricultural technology progress (Zhong et al., 2022). Provincial-level evidence shows rural digital economy exerts mitigating effects on agricultural carbon emissions (Jin et al., 2024). Rural digital economy can cut agricultural carbon emissions by stimulating green technology innovation (Su et al., 2024). Existing research confirms the emission reduction potential of digital economy in agricultural sectors (Wang et al., 2024). Moderated mediation models are widely used to identify the heterogeneous pathway of digital economy’s carbon effect (Xu et al., 2024). A second strand focuses on agricultural technological progress and innovation. A third examines agricultural socialized services, including machinery operations, input provision, extension, and production outsourcing. These services can serve as an organizational carrier through which information and production requirements are transmitted to farms. Existing evidence on digitalization and agricultural emissions is informative, but broad digitalization measures make it difficult to isolate the specific role of blockchain-enabling conditions. Research on blockchain in agriculture has also largely discussed traceability, transparency, and supply-chain efficiency, while comparatively less evidence links regional readiness conditions to production-level carbon-emission intensity in major maize-producing areas. Early reviews systematically summarize the application scenarios of blockchain in agricultural food supply chains (Kamilaris et al., 2019). Blockchain can reshape the sustainability performance of supply chains by improving information transparency and traceability (Saberi et al., 2019). Recent work constructs technical frameworks to implement blockchain traceability within agri-food systems (Baku et al., 2024). Most existing blockchain studies for agriculture focus on traceability design rather than carbon emission outcomes (Vern et al., 2025). European case evidence also confirms blockchain’s capacity to transform agricultural supply chain operation (Hurduzeu and Popescu, 2024).

The present study addresses this gap with a provincial panel of 12 major maize-producing provinces observed from 2010 to 2024, yielding 180 province-year observations. The empirical design uses province and year fixed effects and province-clustered standard errors, supplemented by alternative measures, leave-one-out estimates, a blockchain-relevance check, and a wild cluster bootstrap procedure. The analysis contributes in three bounded respects. (1) It operationalizes blockchain-related readiness as a multidimensional enabling environment rather than equating readiness with adoption. (2) It connects this construct to carbon emissions per unit of maize-sown area. Third, it evaluates agricultural socialized services as a possible organizational pathway and examines whether the association varies with economic development.

The evidence should not be read as a causal estimate of blockchain adoption. The two-way fixed-effects (TWFE) specification absorbs time-invariant provincial factors and common year shocks, but the observational design does not fully eliminate time-varying confounding, reverse causality, or measurement limitations in the readiness index. The mechanism results are likewise consistent with, rather than definitive proof of, mediation. Finally, the interaction coefficient is negative but does not reach the conventional 10% significance threshold (−0.1499, p = 0.103). These qualifications define the scope of the study’s contribution and the policy interpretation that follows.

2 Theoretical analysis and research hypotheses

2.1 Digital agriculture, blockchain readiness, and agricultural green development

Accordingly, the agricultural blockchain readiness environment examined in this study is not equivalent to actual blockchain adoption intensity or the intensity of blockchain use. Instead, it represents the extent to which regional technological infrastructure, organizational capacity, and institutional support provide favorable conditions for potential blockchain-enabled applications in agricultural production, socialized services, and carbon-data management.

This concept is primarily defined through the technology–organisation–environment (TOE) framework. In this study, the TOE framework is used to explain the formation of readiness conditions rather than to directly model blockchain adoption behavior. The technology acceptance model provides the theoretical foundation for understanding technology adoption decisions (Davis, 1989). The TOE framework was originally developed to explain organizational technology adoption decisions (Tornatzky and Fleischer, 1990). The TOE framework posits that the adoption and diffusion of a new technology depend jointly on the technological base, organizational capacity, and the external institutional environment. Building on this framework, the ABRE Index (block_index) is constructed along three dimensions. Digital infrastructure captures the basic digital connectivity conditions necessary for data transmission, information sharing, and potential blockchain-enabled coordination among agricultural actors. The agricultural industrial base reflects organizational and industrial capacity, including the capacity of agricultural organizations and service providers to generate, manage, and utilize digital production information. Policy support captures the fiscal and institutional backing provided by local governments for the diffusion of new technologies.

It should be noted that the three dimensions do not represent blockchain technologies themselves. Instead, they capture the technological, organizational, and institutional foundations required for potential blockchain-enabled agricultural applications. Blockchain implementation in agriculture depends not only on the availability of blockchain technology but also on digital infrastructure, reliable data generation, organizational coordination, and institutional support. Therefore, block_index is designed to measure the readiness environment for blockchain applications rather than realized blockchain adoption intensity.

Although these components are also related to agricultural development, the ABRE Index differs from conventional agricultural modernization measures by emphasizing their role as enabling conditions for data connectivity, organizational coordination, and institutional support required for blockchain applications.

It should be emphasized that block_index measures regional readiness and support conditions for blockchain applications rather than realized adoption intensity. Measuring realized blockchain adoption would require direct indicators such as on-chain transaction records, smart-contract applications, blockchain-based traceability coverage, or agricultural blockchain project participation, which are unavailable for a consistent provincial panel during the study period.

2.2 Theoretical mechanisms and research hypotheses

Maize production is characterized by dispersed producers, multiple input stages, and incomplete process records. In the absence of credible records and effective oversight, the use of chemical fertilizers, pesticides, diesel, and machinery may be driven by experience-based decisions, redundant operations, and excessive input application. A stronger support environment facilitates the deployment of applications such as input traceability, machinery-operation records, production-service management, and carbon-footprint accounting. Digital record-keeping, traceability-oriented management, and potential blockchain-enabled applications may reduce information asymmetry and coordination costs among governments, service organizations, and producers, thereby improving the allocation of agricultural inputs and the scheduling of field operations (Shen and Zhang, 2024).

H1: Improvement in the agricultural blockchain readiness environment may be negatively associated with the carbon emission intensity of maize production.

Agricultural socialized services constitute an organizational pathway through which blockchain-enabling conditions may potentially be incorporated into agricultural production processes. Because individual smallholders generally lack the technical capacity and economic incentives to directly deploy blockchain systems, agricultural socialized service organizations may serve as intermediary actors that translate digital information management and traceability-oriented practices into practical production management (Shen et al., 2023). Services such as input supply, machinery operations, agricultural extension, plant protection, and production outsourcing enable dispersed farmers to be incorporated into professionalized, standardized, and recordable production networks. As the support environment for agricultural blockchain readiness strengthens, data connectivity, process recording, and collaborative management between service providers and farmers also improve. Under these conditions, input use, machinery operations, and production outsourcing can generate more credible records and be linked more effectively to low-carbon certification, subsidy verification, and regulatory requirements, thereby reducing redundant operations, excessive input application, and managerial inefficiency (Shi et al., 2023; Chen and Liu, 2023).

H2: The agricultural blockchain readiness environment may be associated with lower carbon emission intensity of maize production by facilitating improvements in agricultural socialized services; the mediation pathway is examined empirically.

The extent to which the Agricultural Blockchain Readiness Environment translates into actual emission reductions is also contingent upon regional absorptive capacity and the prevailing conditions of production organisation. Economically advanced regions tend to possess superior digital infrastructure, more ample fiscal and technological endowments, better developed agricultural service systems, and a higher degree of producer organisation. Consequently, improvements in the support environment are more readily converted into documented production processes, effective constraints on agricultural inputs, and enhanced green management capacities. In contrast, less developed regions may be hindered by inadequate infrastructure, a shortage of skilled personnel, limited service coverage, and low digital proficiency among producers.

The blockchain-specific features that distinguish this study from general digitalisation analyses rest on three core properties: immutable records, process traceability, and multi-actor verification. Each of these properties corresponds to a distinct emission-reduction channel within the context of agricultural socialized services. First, immutable records enhance the credibility of data on input application and machinery operations. When service organizations and farmers recognise that input records cannot be altered retroactively, redundant operations and excessive use of inputs become more difficult to conceal, which in turn encourages more disciplined input management and reduces carbon-intensive input consumption. Second, process traceability enables the provenance of inputs and the accounting of production processes to be linked to low-carbon certification and green product premiums. Certification based on traceability can create market incentives for service organizations to adopt standardized, low-input production protocols, thereby lowering emissions per unit of output. Third, multi-actor verification mitigates information asymmetries among farmers, service providers, government agencies, and downstream buyers, reducing the coordination costs associated with carbon data reporting and subsidy verification. These properties do not require full blockchain deployment to be analytically relevant; rather, they delineate the directional pathways through which a blockchain-ready environment may shape agricultural production management. These pathways are empirically examined via the agricultural socialized services channel in Section 4.4.

H3: The negative association between the agricultural blockchain readiness environment and carbon emission intensity of maize production may be stronger in regions with higher levels of economic development.

3 Research design

3.1 Study area and data sources

The study covers 12 major maize-producing provinces. The selection rule is based on the provincial ranking of maize output: the sample includes provinces that consistently rank among the leading maize producers in China and jointly account for the overwhelming majority of national maize output, while also satisfying two feasibility criteria: (1) regional coverage, so that the Northeast, North China, the Huang–Huai–Hai region, and the Northwest are all represented; and (2) data continuity, so that the required indicators are available for the full 2010–2024 period. Provinces with frequent changes in statistical definitions or incomplete indicator series were excluded. This rule is intended to balance representativeness with data availability, although the resulting sample is not a random draw from the population of Chinese provinces, and the findings should be generalized with caution. Provinces with smaller maize production scales or discontinuous statistical series are not included. The final sample of 12 provinces is determined jointly by the importance of maize production and data availability (Table 1).

Table 1. Sample provinces and regional classification.

Regional classifications are based jointly on geographic location, climatic conditions, and maize-production characteristics.

Data on maize output, sown area, and agricultural inputs are drawn primarily from the China Statistical Yearbook, the China Rural Statistical Yearbook, provincial statistical yearbooks, and provincial statistical communiqués. Data on internet penetration, rural broadband access, fiscal expenditure on agriculture, forestry, and water affairs, total agricultural machinery power are obtained from publicly available statistical materials, provincial statistical communiqués, and information released by relevant government departments. The level of agricultural socialized services is calculated as the ratio of the output value of professional and auxiliary activities in agriculture, forestry, animal husbandry, and fisheries to the total output value of these sectors. During data processing, units were standardized, statistical definitions were cross-checked across sources, and observations were matched by province and year.

Small gaps in the panel arose for a limited number of province-year observations because of changes in statistical definitions or the unavailability of source materials. Linear interpolation was applied when missing values did not exceed two consecutive periods. Missing values at the beginning or end of the sample period were filled using the nearest available observation, carried backward or forward as appropriate. The data-processing records identify 24 manually flagged province-year observations, representing 13.33% of the full sample. The final values of the checked variables are nonmissing for these observations, and none is marked as extrapolated or as belonging to a partial panel. Because the available flags identify manual mapping or adjustment rather than definitive linear interpolation, these observations are not described as imputed observations. Their potential influence is examined separately through a sensitivity analysis that excludes the flagged observations.

To distinguish data-processing procedures from substantive measurement concepts, the dataset includes observation-level flags for selected manual mappings and adjustments. Specifically, 12 province-year observations are flagged for manual mapping of the 2010 internet variable, and a further 12 province-year observations are flagged for manual processing of the 2023 maize data, with no overlap between these two groups. No observations are flagged as extrapolated for the 2010 internet variable, nor are any flagged as belonging to a partial 2024 panel. These indicators are employed solely for data-quality diagnostics and for the sensitivity analysis presented below; they are not included as explanatory variables in the baseline specification.

3.2 Variable construction

The accounting boundary is restricted to the maize-production stage and excludes storage, processing, transportation, sales, and consumption. Six carbon sources are considered: chemical fertilizer, pesticides, agricultural plastic film, diesel, irrigation, and tillage. Because provincial statistics do not provide continuous field-level observations, soil nitrous oxide (N₂O) emissions and other in-field biogenic emissions are not included in the accounting boundary. The resulting measure therefore represents an estimated input-related carbon emission intensity rather than a complete maize carbon footprint, and it is interpreted accordingly throughout the study. Provincial statistics generally report only aggregate agricultural input use and do not provide continuous maize-specific input data. Provincial fertilizer, pesticide, plastic-film, and diesel use is therefore allocated to maize production according to maize’s share of total crop-sown area.

Limitations and sensitivity of input allocation. The allocation procedure is an approximation imposed by data constraints and implicitly assumes that per-unit-area input intensity is broadly comparable across crops. Potential error arises from two sources. First, maize and other crops differ in their use of fertilizer, pesticides, plastic film, and diesel, so area-based allocation may overestimate or underestimate actual maize inputs. Second, if regional cropping structures or the degree of scale and intensification in maize production are correlated with block_index, allocation error may not be random, creating nonclassical measurement error and potential endogeneity. In the absence of maize-specific input data, this approach is a second-best means of maintaining long-run interprovincial comparability. As a sensitivity test, the study replaces per-unit-area carbon emissions (ce_area) with output-based maize carbon emission intensity (ce); the direction of the core result is unchanged. Changing the denominator, however, cannot fully remove allocation error in the numerator of total emissions, and this issue is therefore retained as a study limitation. The carbon accounting framework adopted in this study follows the classic agricultural carbon flux estimation framework proposed by (West and Marland, 2002). Farm management activities such as tillage and fertilization are important sources of agricultural carbon emissions (Lal, 2004). We refer to existing research on the carbon footprint of China’s crop production when selecting emission factors for maize cultivation (Cheng et al., 2011).

Total carbon emissions from maize production are calculated using the emission-factor method:

C E it = Σ k ( A kit × E F k ) (1)

In Equation 1, CE_it denotes total carbon emissions from maize production in province i and year t, A_kit denotes the quantity of input or production activity k, and EF_k is the corresponding emission factor. Output-based carbon emission intensity and carbon emissions per unit of maize-sown area are calculated as follows: All emission indicators are expressed in carbon (C) units rather than carbon-dioxide equivalents. ce_area is measured in kg C per hectare (kg C/ha), and ce is measured in kg C per tonne of maize output (kg C/t). Following the convention of the agricultural carbon-accounting literature (Kamilaris et al., 2019; Saberi et al., 2019), the emission factors in Table 2 are specified in kg C per unit of input, and no conversion to CO₂-equivalent is applied. Taken together, CE_it, ce, and ce_area constitute an estimated input-related carbon emission measure for maize production.

c e it = C E it / Y it (2)

ce _ are a it = C E it / S it (3)

Table 2. Carbon sources and emission factors for maize production.

Emission factors follow values commonly used in the agricultural carbon-emission literature; units and sources are reported in the table.

In Equations 2, 3, Y_it is maize output and S_it is maize-sown area. The core explanatory variable is the ABRE Index (block_index), which measures regional blockchain readiness rather than actual on-chain use intensity. The index comprises digital infrastructure, the agricultural industrial base, and policy support; the specific indicators are presented in Table 3. The logarithm of the number of blockchain-related patents in each province (ln_patent) is also collected as an alternative proxy for blockchain technological activity. Because a substantial share of province-year observations records zero patents, ln_patent is defined as ln(1 + patents) to retain zero observations in the logarithmic transformation.

Table 3. Indicator system for the ABRE index.

All indicators are positively oriented. The entropy-weighting procedure consists of five steps: (1) range-standardize the raw indicators to remove dimensional effects; (2) calculate the share of each province-year observation in the full sample for each indicator; (3) calculate the entropy value of each indicator; (4) obtain the divergence coefficient as one minus the entropy value and normalize the coefficients to derive indicator weights; and (5) multiply the standardized values by their weights and sum them to obtain dimension scores and block_index. Because the entropy method assigns weights according to sample variation, it reduces the influence of subjective judgment on index construction. Detailed information about each variable can be found in Tables 4, 5.

Table 4. Definitions of main variables.

Table 5. Data sources and processing.

To assess the blockchain relevance of the core index, we examine its correlation with blockchain-related patent activity and compare the baseline results with an alternative patent-based measure. These checks provide supplementary construct-relevance evidence, but they cannot fully establish discriminant validity from broader digitalization or agricultural development. First, the correlation between block_index and the logarithm of blockchain-related patents (ln_patent) is 0.1883 (p = 0.0114), suggesting that regions with a stronger support environment tend to display higher levels of blockchain-related technological activity. When ln_patent replaces block_index as the core explanatory variable in the baseline two-way fixed-effects specification, its coefficient is positive but statistically insignificant (β = 0.000613, robust SE = 0.005261, p = 0.909; 95% CI = [−0.010967, 0.012193]). This result does not reproduce the associational pattern observed for the multidimensional ABRE Index. The positive pairwise correlation between the two measures (r = 0.1883, p = 0.0114) still supports the technological relevance of the index, but patent counts cannot substitute for the combined readiness environment in explaining production-level carbon emission intensity.

3.3 Model specification and testing strategy

The empirical sample comprises 12 major maize-producing provinces from 2010 to 2024. The baseline specification is a two-way fixed-effects model that controls for unobserved time-invariant provincial characteristics and common year shocks:

ce _ are a it = α + βblock _ inde x it + γ X it + μ i + λ t + ε it (4)

In Equation 4, ce_area_it denotes carbon emissions per unit of maize-sown area in province i and year t, block_index_it is the ABRE Index, and X_it is the vector of control variables. Province and year fixed effects are represented by μ_i and λ_t, respectively. The robustness tests replace the dependent variable with output-based maize carbon emission intensity and replace the core explanatory variable with the logarithm of blockchain-related patents (ln_patent). The mechanism analysis combines a bootstrap mediation test based on 1,000 resamples with the Sobel test to assess the statistical significance of the indirect effect. The Dumitrescu–Hurlin panel Granger test is further used to examine the temporal predictive direction between the variables.

Because the ABRE Index incorporates macro indicators of the agricultural industrial base, it may mechanically overlap with variables such as agricultural output and mechanization. To avoid over-controlling and exacerbating multicollinearity, the baseline model does not include additional controls constructed from the same underlying components. Robustness tests encompass replacing the dependent variable, excluding the year 2020, lagging the core explanatory variable by one period, estimating Driscoll–Kraay standard errors, and applying a wild cluster bootstrap correction for the small number of clusters. Supplementary analyses also replace block_index with ln_patent, a patent-based proxy for blockchain-related technological activity. This specification is interpreted as a construct-comparison exercise rather than as corroborating robustness evidence. The mechanism analysis centers on agricultural socialized services and employs a two-step regression framework, supplemented by bootstrap and Sobel tests of the indirect effect. Heterogeneity is examined by dividing provinces into high- and low-development groups based on the annual median of per capita GDP across provinces in each year, and testing group differences through interaction terms. We follow the standard Stata implementation guide for difference and system generalized method of moments (GMM) estimation (Roodman, 2009). The difference GMM estimator was first proposed for dynamic panel data estimation (Arellano and Bond, 1991). System GMM extends difference GMM by adding additional moment conditions for the level equation (Blundell and Bond, 1998). Given that the sample size of N = 12 does not meet the large-N, small-T conditions required for system GMM, system GMM is used solely as a methodological reference and is excluded from the core evidence.

3.4 Model diagnostics and additional robustness considerations

A Hausman test comparing fixed- and random-effects specifications yields χ² = 58.28 and p = 0.0000, rejecting the null and supporting the fixed-effects model. The mean VIF is 3.15, below the conventional threshold of 10, indicating no serious multicollinearity. The Dumitrescu–Hurlin panel Granger test is also used to examine the direction of the relationship between the agricultural blockchain readiness environment and carbon emissions, thereby providing supplementary evidence regarding temporal ordering and potential reverse-direction relationships.

4 Empirical results and analysis

4.1 Descriptive statistics

The balanced panel contains 180 province-year observations for 12 major maize-producing provinces over 2010–2024. Table 6 reports descriptive statistics for the main variables. Carbon emissions per unit of maize-sown area (ce_area), measured in kg C per hectare, have a mean of 0.2468 and a standard deviation of 0.2418, indicating substantial cross-provincial variation. The distribution is right-skewed, with the maximum (1.1584 kg C/ha) exceeding the 75th percentile by a wide margin, driven mainly by provinces with high fertilizer and plastic-film intensity.

Table 6. Descriptive statistics.

ce_area is measured in kg C per hectare of maize-sown area; ce is measured in kg C per tonne of maize output; service_ratio is a ratio scaled between 0 and 1; block_index is a normalized index between 0 and 1. Carbon indicators are reported in carbon (C) units rather than CO₂-equivalent units.

The ABRE Index has a mean of 0.2827 and a maximum of 0.8049, revealing marked differences in digital infrastructure, the agricultural industrial base, and policy support across provinces. The maximum is approximately 25 times the minimum value of 0.0318. Because the two-way fixed-effects model identifies the coefficient from within-province variation, we decompose the total variation of block_index: the within-province standard deviation is 0.0797 and accounts for approximately 22.7% of the total variation, while the between-province component accounts for the remainder. The presence of within-province variation provides the variation required to estimate the fixed-effects association, although it does not by itself establish causal identification. The logarithm of blockchain-related patents (ln_patent), defined as ln(1 + patents), has a mean of 1.5649 and a standard deviation of 1.7686; approximately 38.9% of province-year observations record zero patents, indicating that blockchain-related patent output was limited in most provinces and years.

4.2 Baseline regression results

Table 7 reports the baseline estimates from the two-way fixed-effects model. Model (1) includes only the core explanatory variable together with province and year fixed effects, while Models (2) through (4) sequentially add the urbanization rate, the logarithm of maize-sown area, and the logarithm of per capita GDP. In the full specification shown in column (4), the coefficient on block_index is −0.5957 and exhibits marginal significance (statistically significant at the 10% level) (p = 0.063) based on conventional cluster-robust standard errors. Given that the panel contains only 12 cross-sectional units, conventional cluster-robust standard errors may be downward biased. We adopt the Driscoll–Kraay standard error method to address cross-sectional dependence in panel data (Driscoll and Kraay, 1998). Wild cluster bootstrap is used to obtain reliable statistical inference under small sample clustering (Cameron et al., 2008); two supplementary inference procedures are therefore employed. First, the Driscoll–Kraay covariance estimator, which accounts for cross-sectional dependence, heteroskedasticity, and serial correlation, yields the same coefficient of −0.5957, with a robust standard error of 0.221 and a 95% confidence interval of [−1.071, −0.120] (p = 0.012), indicating significance at the 5% level; the estimator is implemented with the default bandwidth of 2 lags (floor (T^(1/3))), which is standard for this panel dimension. Second, a wild cluster bootstrap test with 3,000 replications, Rademacher weights, and the null imposed yields a bootstrap standard error of 0.287, a test statistic of −2.0768, and a 95% confidence interval of [−1.169, −0.022] (p = 0.0415), indicating significance at the 5% level. The consistent negative coefficient across the conventional and supplementary inference procedures supports the stability of the estimated association, although the conventional cluster-robust result is significant only at the 10% level. In economic terms, a one-standard-deviation increase in the ABRE Index (approximately 0.17 units) is associated with a decline of roughly 0.0995 units in carbon emissions per unit of maize-sown area. Full results of the two supplementary inference procedures are reported in Table A1 in the Appendix.

Table 7. Baseline regression results.

Robust standard errors clustered by province are reported in parentheses; *, , and * denote p < 0.10, p < 0.05, and p < 0.01, respectively. The dependent variable ce_area is measured in kg C per hectare of maize-sown area.

4.3 Robustness tests

As an additional robustness check, the baseline specification is re-estimated with ln(ce_area) as the dependent variable. The coefficient on block_index remains negative but is no longer statistically significant (β = −0.2273, robust SE = 0.3074, p = 0.475). This attenuation is consistent with the right-skewed distribution of ce_area and the compression of large values by the logarithmic transformation; the level specification therefore remains the primary specification, while the logarithmic transformation provides directional but not statistically significant support. Among the control variables, ln_corn_area is significantly negative (β = −0.7129, p = 0.001), indicating that larger maize-sown areas are associated with lower emission intensity per unit of area.

Table 8 summarizes the robustness analyses, including the alternative dependent-variable specification, the exclusion of 2020, the one-period lag specification, and the patent-based explanatory-variable specification.

Table 8. Robustness tests of the baseline results.

Robust standard errors clustered by province are reported in parentheses. All specifications include province fixed effects and year fixed effects. *, , and * denote significance at the 10, 5, and 1% levels, respectively. Column (1) uses ce, measured in kg C per tonne of maize output, as the dependent variable. Column (4) replaces block_index with ln_patent as the core explanatory variable while retaining the same controls and fixed effects; the coefficient on ln_patent is not statistically significant (p = 0.909).

The patent-based specification produces a different result. When ln_patent replaces block_index as the core explanatory variable in the same two-way fixed-effects specification, the coefficient on ln_patent is positive but statistically insignificant (β = 0.000613, robust SE = 0.005261, t = 0.12, p = 0.909; 95% CI = [−0.010967, 0.012193]). This result indicates that patent activity does not reproduce the same associational pattern with carbon-emission intensity as the multidimensional ABRE Index. It is consistent with a conceptual distinction between technological output and the institutional, infrastructural, and organizational conditions captured by the ABRE Index, but it should not be interpreted as definitive evidence of the index’s discriminant validity.

To assess the validity of the inference procedures, we first examine the error structure of the baseline specification. The Wooldridge test for first-order serial correlation in panel data rejects the null hypothesis of no serial correlation (F(1, 11) = 66.567, p < 0.001), indicating the presence of first-order autocorrelation in the residuals. In addition, Pesaran’s test of cross-sectional independence (Pesaran, 2004) yields a test statistic of −2.342 (p = 0.019), rejecting the null hypothesis of cross-sectional independence and indicating the presence of cross-sectional dependence. These two findings jointly motivate the use of cluster-robust standard errors and the Driscoll–Kraay covariance estimator, both of which are robust to serial correlation and cross-sectional dependence, in the baseline and robustness analyses.

We further examine the stationarity properties of the main variables using the Levin–Lin–Chu (LLC) and Im–Pesaran–Shin (IPS) panel unit-root tests. We employ the Levin–Lin–Chu test to examine the stationarity of panel variables (Levin et al., 2002). The Im–Pesaran–Shin test is further applied to accommodate heterogeneous panel unit root properties (Im et al., 2003). Pesaran’s CD test is used to detect cross-section dependence across panel units (Pesaran, 2004). As reported in Table 9, the estimation outcomes diverge between the two testing procedures. Under the LLC test, which imposes a common autoregressive parameter across panels, the null hypothesis of a unit root is rejected for the control variables, namely the urbanization rate, the logarithm of maize-sown area, and the logarithm of per capita GDP. By contrast, under the IPS test, which permits panel-specific autoregressive parameters, none of the five variables rejects the unit-root null. This discrepancy is not unexpected in a short panel comprising only 15 periods, where unit-root tests typically possess low power and exhibit considerable sensitivity to assumptions about cross-sectional homogeneity. Given the limited time dimension, the common temporal patterns captured by year fixed effects, and the mixed supplementary evidence reported in Table 10 (the Dumitrescu–Hurlin test points to predictive ordering, whereas the Kao test does not reject the null of no cointegration), we retain the level specification and interpret the resulting estimates with due caution. Accordingly, the findings should be viewed as suggestive of an association rather than as conclusive evidence of a causal relationship.

Table 9. Panel unit root tests.

The null hypothesis of both tests is that the relevant panels contain unit roots. Lag length is selected by AIC with a maximum of three lags. LLC refers to the Levin–Lin–Chu test, which assumes a common autoregressive parameter; IPS refers to the Im–Pesaran–Shin test, which allows for panel-specific autoregressive parameters. The panel includes 12 provinces observed over 15 years.

Table 10. Supplementary tests of temporal predictive direction and cointegration.

To examine whether the baseline result is driven by any particular province, we conduct a leave-one-out analysis in which the baseline specification is re-estimated 12 times, each time excluding one province. As reported in Table 11, the coefficient on block_index remains negative in all 12 specifications, ranging from −0.330 to −0.884. The coefficient exhibits marginal significance (statistically significant at the 10% level) or better in 10 of the 12 specifications, including one specification significant at the 5% level (p = 0.018 when province 9 is excluded). The two non-significant specifications are obtained when provinces 3 and 7 are excluded; however, their coefficients remain negative. Thus, excluding any single province does not reverse the direction of the estimated association, although statistical significance is not uniform across all specifications.

Table 11. Leave-one-out robustness check.

Each row reports the block_index coefficient from the baseline specification (column (4) of Table 7) estimated after excluding the indicated province. Standard errors are clustered by province. The coefficient remains negative in all 12 specifications and is significant at the 10% level or better in 10 of 12 cases (including one at the 5% level), indicating that excluding any single province does not reverse the negative sign of the estimated association, although statistical significance is not uniform across all specifications.

Table 12 provides two supplementary pieces of evidence. First, the correlation between block_index and ln_patent is positive and statistically significant (r = 0.1883, p = 0.0114), indicating that regions with stronger readiness conditions tend to exhibit higher levels of blockchain-related patent activity. This result supports the blockchain relevance of the constructed index but does not establish that the index measures realized blockchain adoption. Second, the wild cluster bootstrap result remains statistically significant at the 5% level (p = 0.0415), supporting the stability of the baseline inference under a small number of clusters.

Table 12. Robustness test using wild cluster bootstrap and blockchain-relevance validation.

** indicates significance at the 5% level.

4.4 Data-processing sensitivity analysis

As an additional data-processing sensitivity check, we exclude the 24 province-year observations carrying manual mapping or adjustment flags. These observations represent 13.33% of the full sample, reducing the estimation sample from 180 to 156 observations. Using the same two-way fixed-effects specification as the baseline model, the coefficient on block_index changes only slightly, from −0.5957 (SE = 0.2877, p = 0.063) in the full sample to −0.6011 (SE = 0.2924, p = 0.064) in the reduced sample. The negative sign, coefficient magnitude, and 10% statistical significance are preserved. These results support the stability of the estimated association with respect to the exclusion of manually flagged observations. Full estimation results of this sensitivity analysis are presented in Table A2 in the Appendix. Because the flags identify manual mapping or adjustment rather than definitive linear interpolation, this analysis is interpreted as a sensitivity test for data-processing decisions rather than as a conventional interpolation-exclusion test.

4.5 Mechanism analysis

Building on the distinction between blockchain readiness conditions and actual blockchain adoption intensity, this study further examines whether agricultural socialized services constitute an organizational pathway. In this context, agricultural socialized services are viewed as an organizational carrier that may facilitate the transmission of digital and blockchain-enabling conditions into agricultural production management.

The mechanism is examined by estimating whether the agricultural blockchain readiness environment is associated with lower carbon emission intensity through improvements in agricultural socialized services. A bootstrap mediation test based on 1,000 resamples and a 95% percentile confidence interval, combined with the Sobel test, is employed to evaluate the statistical significance of the indirect effect. The mediation analysis in this paper draws on the resampling-based inference framework developed by (Preacher and Hayes, 2008). Given the small number of clusters, the bootstrap resamples provinces rather than province-year observations to preserve the panel structure and the dependence within provinces. This procedure provides a supplementary assessment of the indirect effect, although inference remains subject to the limitations imposed by the small number of provincial clusters. Agricultural socialized services encompass input supply, machinery operations, agricultural extension, and production outsourcing, and therefore provide an organizational platform through which digital information management, traceability-oriented practices, and potential blockchain-enabled applications may provide additional support for agricultural production processes. Table 13 reports the estimation results of the mechanism test with agricultural socialized services as the mediating pathway.

Table 13. Mechanism test: agricultural socialized services.

Robust standard errors clustered by province are reported in parentheses; *, , and * denote p < 0.10, p < 0.05, and p < 0.01, respectively. The results in this table are consistent with the hypothesized organizational pathway: the ABRE Index promotes agricultural socialized services, which in turn reduces the carbon emission intensity of maize production.

Overall, the evidence is consistent with agricultural socialized services representing a possible organizational pathway through which stronger blockchain-readiness conditions may be associated with lower carbon-emission intensity in maize production. This interpretation remains suggestive rather than causal. The bootstrap mediation test with 1,000 province-level resamples yields an indirect effect of −0.250, with a 95% percentile confidence interval of [−0.487, −0.061] (p = 0.018), supporting the statistical significance of the organizational pathway.

4.6 Heterogeneity analysis

The association between the ABRE Index and carbon emission intensity may exhibit regional heterogeneity, as levels of economic development influence digital infrastructure, organizational capacity, and the ability of agricultural actors to absorb and utilize emerging technologies. To avoid treating economic development as a time-invariant provincial characteristic, the variable high_pgdp is constructed separately for each year. Specifically, for each year, provinces with per capita GDP strictly above the cross-provincial median are assigned a value of one, and the remaining provinces are assigned a value of zero. Consequently, high_pgdp is defined at the province-year level, allowing a province to switch between the high- and low-development groups over time. During the period from 2010 to 2024, seven of the 12 provinces in the sample switch groups at least once. This confirms that high_pgdp contains within-province temporal variation and is therefore not a time-invariant provincial characteristic that would be subsumed by province fixed effects.

We estimate an interaction specification between the ABRE Index and high_pgdp. The coefficient on the ABRE Index represents the association in the low-development group, whereas the interaction coefficient measures the difference between the high- and low-development groups. Given the limited number of provincial units, the results are interpreted as evidence of heterogeneous associations rather than definitive causal differences.

Table 14 reports a coefficient of −0.5307 for the ABRE Index in the low-development group (SE = 0.2227, p = 0.036). The interaction coefficient is −0.1499 (SE = 0.0839, p = 0.103). Although the interaction coefficient is negative, it does not reach the conventional 10% significance threshold, and the evidence for a statistically different association between the two groups remains limited. The linear combination of the ABRE Index coefficient and the interaction coefficient yields an estimated association of −0.6806 in the high-development group (SE = 0.2930, p = 0.040; 95% CI: −1.3255 to −0.0357). This result indicates a negative association within the high-development group, but it should not be interpreted as definitive evidence that the association differs significantly between the two groups.

Table 14. Heterogeneity analysis by economic development level measured by Per Capita GDP.

Robust standard errors clustered at the provincial level are reported in parentheses. high_pgdp is a province-year indicator equal to one when a province’s per capita GDP is strictly above the cross-provincial median in the same year, and zero otherwise. The indicator is reclassified annually; seven of the 12 sample provinces switch between the high- and low-development groups at least once during 2010–2024. The coefficient on the ABRE Index represents the low-development group, while the interaction coefficient represents the difference between the two groups. The implied high-development-group association is obtained from a linear-combination test. The dependent variable is ce_area, measured in kg C per hectare of maize-sown area.

4.7 Additional tests: temporal predictive direction and long-run relationships

We perform panel Granger causality tests to examine the predictive relationship between variables, following (Granger, 1969). Pedroni panel cointegration test is applied to check long-run equilibrium relationships among variables (Pedroni, 1999). The Kao test is also implemented as a supplementary residual-based panel cointegration test (Kao, 1999). To further examine temporal predictive direction and potential reverse-order relationships, the analysis supplements the baseline results with the Dumitrescu–Hurlin panel Granger test and the Kao panel cointegration test.

As reported in Table 10, the Dumitrescu–Hurlin test rejects the null hypothesis that block_index does not Granger-cause ce_area (Z-bar = 3.1664, p = 0.0015). This indicates that the agricultural blockchain readiness environment contains significant predictive information for subsequent changes in carbon emissions. The reverse-direction test yields a p-value of 0.0727, showing marginal significance (statistically significant at the 10% level), while it does not reach the conventional 5% statistical threshold. This result does not constitute strong evidence that reverse predictive ordering dominates the relationship, although reverse-direction predictive effects cannot be fully ruled out. Overall, the results provide supplementary evidence consistent with temporal ordering from the agricultural blockchain readiness environment to subsequent carbon emission changes, but they do not establish a strict causal effect.

5 Conclusions and policy implications

5.1 Main findings

Using 180 province-year observations from 12 major maize-producing provinces during 2010–2024, this study examines whether the Agricultural Blockchain Readiness Environment Index is associated with carbon emissions per unit of maize-sown area. The index combines digital infrastructure, the agricultural industrial base, and policy support. Its meaning is deliberately narrower than actual blockchain adoption: it captures the regional conditions that may facilitate future blockchain-enabled information, traceability, and verification practices.

The baseline estimates indicate a statistically significant negative association between the block_index and ce_area. This pattern is not reversed in the leave-one-out sensitivity analysis; the coefficient remains negative across all 12 specifications, ranging between −0.330 and −0.884, and is statistically significant at the 10% level or higher in 10 of the specifications. The wild cluster bootstrap test, employing 3,000 Rademacher-weighted replications, yields a p-value of 0.0415. The positive correlation between the index and blockchain-related patent activity (0.1883, p = 0.0114) is consistent with the index’s intended relevance to blockchain; nevertheless, this finding does not imply that provinces have adopted blockchain technology in maize production, nor that such adoption has driven the observed differences in emission intensity.

The mechanism analysis provides suggestive evidence for an organizational pathway. The ABRE Index is positively associated with agricultural socialized services (0.1099, p < 0.01), while agricultural socialized services are negatively associated with ce_area (−2.2765, p < 0.05) in the outcome equation. These estimates are compatible with the interpretation that service organizations may help translate digital and blockchain-enabling conditions into production-management practices. They should not, however, be described as definitive causal mediation evidence. The heterogeneity analysis offers weaker evidence: the interaction between block_index and the high-development indicator is −0.1499 (p = 0.103), which does not reach the conventional 10% significance threshold. The implied association in the high-development group is −0.6806 (SE = 0.2930, p = 0.040), based on a linear-combination test. Thus, the data provide only marginal support for a stronger negative association in more developed provinces.

The data-processing sensitivity analysis yields similar results. After excluding 24 manually flagged observations, the coefficient on block_index changes only slightly from −0.5957 (SE = 0.2877, p = 0.063) to −0.6011 (SE = 0.2924, p = 0.064). The negative sign, coefficient magnitude, and 10% statistical significance are preserved. This result supports the stability of the estimated association with respect to the exclusion of manually flagged observations.

5.2 Policy implications

The policy implications follow from these findings without assuming immediate or universal blockchain adoption. Regional governments may prioritize the foundational conditions measured by the ABRE Index, including digital connectivity, agricultural organizational capacity, and appropriately targeted policy support. Investments in agricultural socialized services may be particularly relevant because service providers can potentially aggregate demand, standardize operations, and transmit production information across fragmented farms. Blockchain pilots should be tied to clearly defined information, traceability, or verification needs rather than treated as an independent guarantee of emissions reduction. Evaluation should distinguish infrastructure and readiness from actual deployment and should report adoption, usage, and implementation outcomes separately.

Three boundaries qualify these implications. The sample covers 12 major maize-producing provinces, so external validity beyond this setting is limited. The ABRE Index is an enabling-environment measure and may still capture broader digital or agricultural development. Finally, the fixed-effects results are associational rather than strictly causal, and the mechanism and heterogeneity tests have limited inferential strength given the small number of provincial clusters. Future work should combine more direct measures of blockchain use with finer production or farm-level data and designs better suited to causal identification.

5.3 Limitations and future research

This study has several limitations that also suggest directions for future research.

• (1) The ABRE Index is a readiness-environment measure rather than a direct measure of realized blockchain adoption. The patent-based proxy also has limitations. Although block_index is positively correlated with blockchain-related patent activity (r = 0.1883, p = 0.0114), the patent-based specification is statistically insignificant (β = 0.000613, robust SE = 0.005261, p = 0.909). Patent counts therefore provide only limited supplementary information and should not be treated as a direct measure of realized blockchain adoption or as definitive evidence of the discriminant validity of the ABRE Index. This contrast suggests that patent activity and the broader readiness environment should not be treated as equivalent constructs. Future research should incorporate more direct measures of realized blockchain applications, such as traceability coverage, on-chain transaction records, smart-contract applications, or participation in agricultural blockchain projects.
• (2) The sample comprises only 12 major maize-producing provinces over 2010–2024, selected based on maize production scale and data availability rather than random sampling from all Chinese provinces. This may lead to potential sample selection bias: the negative association may be more prominent in major producing regions with better agricultural foundations, and the findings cannot be directly generalized to non-major maize-producing provinces or the whole country. Although the sample covers the three core maize-producing regions and provides meaningful within-province variation, the small cross-sectional dimension limits statistical power and external validity. Future research could expand the sample and adopt appropriate methods to address selection bias.
• (3) The observational two-way fixed-effects design does not fully resolve endogeneity. The Dumitrescu–Hurlin test indicates that block_index contains predictive information for subsequent changes in ce_area (p = 0.0015), but this result does not establish a structural causal relationship or eliminate time-varying confounding. No instrumental-variable strategy is employed in the present study. Future research should consider theoretically justified instruments, policy shocks, or quasi-experimental designs such as difference-in-differences based on blockchain pilot programs.
• (4) The estimation of maize-related carbon emissions relies on allocating provincial agricultural inputs according to maize’s share of total crop-sown area. This is a practical approximation necessitated by data availability and may introduce measurement error because input intensity differs across crops. The use of output-based carbon-emission intensity as an alternative outcome does not fully eliminate possible measurement error in the numerator. Future research should prioritize crop-specific input-use data.
• (5) Agricultural socialized services are measured using a provincial-level macroeconomic proxy. Agricultural socialized services are important intermediaries to help smallholders adopt green production technologies (Shi et al., 2023). Farm size heterogeneity shapes the effectiveness of agricultural socialized services in reducing chemical input (Chen and Liu, 2023). The mechanism results are therefore suggestive of an organizational pathway but cannot establish how individual farmers, service providers, or production processes respond to blockchain-related information and verification practices. Micro-level data on production outsourcing, machinery operations, input records, or traceability participation would allow this pathway to be examined more directly.
• (6) The heterogeneity classification relies on annual provincial medians within a small sample. The interaction coefficient is negative but does not reach the conventional 10% significance threshold (−0.1499, p = 0.103). The implied association in the high-development group is statistically distinguishable from zero based on a linear-combination test, but the difference between the two groups is not estimated precisely enough to reject equality at conventional significance levels. The evidence of a stronger association in more economically developed regions is therefore tentative and should be re-examined using a larger sample and alternative measures of absorptive capacity.

Taken together, these limitations imply that the estimates should be interpreted as an association between blockchain-related readiness conditions and carbon-emission outcomes within the observed sample, rather than as the causal effect of blockchain deployment itself.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Jiawei Liu, Xi’an Jiaotong-Liverpool University, China

Reviewed by: Julian Hoxha, American University of the Middle East, Kuwait

Yutong Sun, Chinese Academy of Sciences (CAS), China