2.1. Policy background

China’s agricultural industry has witnessed a marked escalation in chemical fertilizer use over recent decades. In 2000, the total consumption of chemical fertilizers (in pure terms) reached 46.37 million tons, peaking at 60.23 million tons in 2015 (https://data.stats.gov.cn/). Although the use of chemical fertilizers has since declined to 58.59 million tons in 2016, it remains significantly higher than the global average [1]. This excessive reliance on chemical fertilizers poses a significant threat to environmental sustainability and human health, contributing to soil acidification, water eutrophication, and ecological deterioration [48].

To address this issue, China’s Ministry of Agriculture and Rural Affairs introduced the “Action Plan for Organic-Substitute-Chemical-Fertilizer (OSCF) for Fruits, Vegetables and Tea” in 2017. This policy recognizes the potential of organic fertilizers, such as animal manure and agricultural by-products, to improve soil health and reduce environmental contamination. It aims to replace chemical fertilizers with organic alternatives in the production of fruits, vegetables, and tea, emphasizing resource recycling, cost-effectiveness, quality enhancement, and the exploration of a modern agricultural development path that is efficient, safe, resource-conserving, and environmentally friendly. The number of pilot counties expanded from 100 in 2017 to 238 in 2021. The policy calls for a 15% reduction in chemical fertilizer use in core production areas and well-known brand production bases within the pilot counties, with the aim of achieving zero growth in overall fertilizer use. Additionally, the policy emphasizes the development of circular agriculture by promoting the use of organic fertilizers, aiming to increase their use by over 20% in core production areas. To achieve these targets, the Chinese government allocated one billion CNY from its central budget in 2018 to support the OSCF, primarily focusing on subsidies for commodity and compost organic fertilizer, integration of water and fertilizer (including equipment and water-soluble fertilizer), soil testing, fertilizer formula test and fertilizer effectiveness evaluation, and technical training.

2.2. Theoretical analysis and research hypothesis

Profit maximization is a well-established analytical framework for understanding farmers’ behaviors and evaluating the effectiveness of policy instruments [1]. For rational farmers, the influence of agricultural subsidies on farm production can be conceptualized as the strategic allocation of agricultural inputs within a set of constraints, with the ultimate goal of maximizing profit. Two primary constraints faced by farmers are liquidity constraints and the scarcity of arable land [49,50]. Liquidity constraints hinder farmers’ ability to invest in environmentally-friendly agricultural inputs and technologies [51]. Moreover, the significant barrier to the adoption of chemical fertilizer substitutes, such as organic fertilizer, is the high costs in terms of agricultural inputs expenditure, investment in equipment and labor demand. Consequently, there is a lack of economic incentive for farmers to use chemical fertilizer substitutes [4,16]. The OSCF policy, however, can alleviate these liquidity constraints. It directly reduces chemical fertilizer use by incentivizing farmers to adopt more organic fertilizers as substitutes. Additionally, it encourages farmers to adopt advanced techniques and transition towards more intensive and large-scale agricultural production practices, further reducing chemical fertilizer use. Fig 1 illustrates how the OSCF policy drives reductions in chemical fertilizer use.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Influence mechanism of the OSCF policy on chemical fertilizer use.

https://doi.org/10.1371/journal.pone.0334751.g001

Firstly, the high costs of organic fertilizers significantly dampen the incentive for farmers to transition away from chemical fertilizers [52]. The OSCF policy partially mitigates this barrier by directly reducing the cost of organic fertilizers [37,53], acting as a form of income subsidy [54]. This reduction improves farmers’ profit margins and expected farm income, enabling them to purchase and use more organic fertilizers to replace chemical fertilizers, thereby reducing chemical fertilizer use [55]. The subsidy for organic fertilizer, whether based on the quantity of organic fertilizer applied or the cultivated area, encourages increased organic fertilizer usage even when farmers are already using chemical fertilizers efficiently [4,56].

Secondly, Advanced equipment, production skills, and knowledge are often required for the application of environmentally-friendly chemical fertilizer substitutes and production techniques [57]. However, farmers in China typically face liquidity constraints, making it difficult to invest in human capital or advanced equipment [58]. The OSCF policy addresses this issue by supporting the adoption of environmentally-friendly techniques and providing technical training, reducing the risks associated with adopting new technologies and improving farmers’ ability to absorb new knowledge and technologies [16,59]. The adoption of new technologies promotes the rational allocation of agricultural inputs and improves agricultural productivity, thereby optimizing and reducing chemical fertilizer use [60]. These interventions align with Farrell’s efficiency framework [61], which posits that optimizing input use (e.g., reducing excess fertilizer) requires both knowledge diffusion and technology adoption. By lowering the cost of precision tools, the OSCF directly addresses allocative inefficiency—the gap between actual and optimal input use [62].

Thirdly, small-scale farms and the management practices of smallholders are significant barriers to mitigating the excessive chemical fertilizer use in China [1]. In contrast, large-scale farming has emerged as a promising avenue for ensuring food security while promoting sustainable agricultural development [63,64]. Facing the liquidity constraints, farmers operating small-scale farms often have to overuse chemical fertilizer to compensate for the lack of investment in labor and advanced technologies [16,65]. The OSCF policy primarily targets large farms, family farms, agricultural cooperatives, and agricultural enterprises, facilitating an enhancement in the scale and specialization of agricultural production. Additionally, the OSCF policy emphasizes the establishment of large-scale demonstration areas, contributing to the expansion of agricultural production scales in the pilot areas. Large-scale farms benefit from economies of scale, leading to more efficient resource allocation and reducing chemical fertilizer use [32]. Moreover, large-scale farmers often possess advanced management knowledge and skills, enabling them to improve their cultivation methods and achieve sustainable development and improved agricultural productivity, further reducing chemical fertilizer use [66–68]. In summary, the OSCF policy can reduce chemical fertilizer use by prompting farmers to achieve economies of scale, thereby improving the utilization efficiency of agricultural inputs and agricultural productivity. This reflects economies of scale theory [69], where larger operational scales reduce per-unit input costs (e.g., fertilizer) through mechanization and bulk purchasing.

Based on the above analysis, the hypothesises of this study are put forward:

1. Hypothesis 1. (H1): The OSCF policy can reduce chemical fertilizer use.
2. Hypothesis 2. (H2): The OSCF policy reduces chemical fertilizer use by improving agricultural technical efficiency.
3. Hypothesis 3. (H3): The OSCF policy reduces chemical fertilizer use by improving agricultural scale efficiency.

3.1. Variable setting

3.1.5. Data sources.

This study evaluates the macro-level impact of the OSCF policy on chemical fertilizer use in China. The data are mainly sourced from China County Statistical Yearbook and local statistical yearbooks of each province and prefecture-city. Specifically, chemical fertilizer use, rural power consumption, expenditures of rural residents, monthly minimum wage, the net income of rural residents, and the employed population in agricultural industry are sourced from local statistical yearbooks from each prefecture-level city. Annual average precipitation is derived from the ERA5-Land dataset. Cultivation area and yield data for various crops are collected from county government websites. Missing data are provided through the China Economic and Social Big Data Research Platform on the official website (https://data.cnki.net/) as much as possible. The list of pilot counties selected in 2017 is obtained from the Ministry of Agriculture and Rural Affairs of China. However, subsequent batches of pilot counties are non-public. Therefore, data for these counties is collected manually through formal requests for public disclosure submitted to the agricultural departments of each province, municipality, and autonomous region.

The data spans from 2012 to 2022, which is based on the following reasons. Firstly, the 18th National Congress of the Communist Party of China in 2012 officially advocated for ecological civilization, emphasizing protection and harmony with nature and initiating significant environmental policy changes. Secondly, starting the analysis in 2012 ensures symmetry between the pre-treatment (2012–2017) and post-treatment (2017–2022) periods, facilitating a more straightforward comparison between the two periods and deepens our comprehension of the policy’s impact [82].

Data processing involves excluding provinces with unavailable information of polit counties (For example, the responses of Fujian, Shandong and Sichuan Province indicate the data are not public and unavailable), removing samples lacking necessary indicators, and supplementing missing data of remaining samples with corresponding county data from statistical yearbooks or interpolation. Specifically, interpolation for missing data is addressed through a three-tiered imputation protocol: firstly, for administrative or stable variables (e.g., institutional data with low volatility), missing values are filled using forward/backward filling to leverage temporal consistency; secondly, for time-trend variables (e.g., economic indicators), missing values are predicted using spatial-temporal interpolation based on adjacent counties’ data and historical trends; thirdly, remaining gaps are addressed via adaptive K-nearest neighbors imputation (Euclidean distance), which weights neighbors by probability density to preserve data distribution [83]. Then to mitigate potential biases arising from unit differences and skewness issues, particularly to address heteroscedasticity [84,85], all variables except dummy variables are normalized using z-score normalization. Descriptive statistics for the main variables are presented in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics of main variables.

https://doi.org/10.1371/journal.pone.0334751.t001

3.2. Double Machine Learning model construction

This study employs the DML approach to precisely evaluate the causal impact of the agricultural subsidy on chemical fertilizer use. The OSCF policy serves as a case of an exogenous policy shock, providing a unique opportunity for causal inference. Traditional causal inference methods, such as DID, synthetic control methods (SCM), and regression discontinuity design (RDD), often face limitations in addressing high-dimensional data, non-linear relationships, and endogeneity caused by omitted variables [86,87]. For instance, DID requires strict parallel trends assumptions and struggles with heterogeneous treatment effects, while SCM is sensitive to extreme values in the treatment group [88]. In contrast, DML excels in handling complex datasets, overcoming the limitations of traditional multiple linear regression models in dealing with non-linear relationships and the ‘curse of dimensionality’ [89,90]. Its robust framework effectively addresses challenges such as multicollinearity and high-dimensional variables, enhancing estimation accuracy [91,92].

Chemical fertilizer use is influenced by a multitude of factors, including economic, political, and social factors [93]. To optimize policy effect estimation, it is critical to control for high-dimensional confounders. Traditional regression models face the “curse of dimensionality” when incorporating numerous control variables, leading to biased estimates due to multicollinearity or oversimplified linear assumptions [94]. DML mitigates these issues by employing regularization algorithms to automatically filter and select relevant variables from high-dimensional data, thereby reducing bias and improving predictive accuracy [87,95]. For example, DML decomposes the estimation into two stages: first, predicting residuals of the outcome and treatment variables using machine learning models (e.g., random forests or Lasso), and second, estimating the causal effect via orthogonalized residuals [95]. This approach avoids regularization bias and model misspecification, which are common pitfalls in traditional methods [47,96].

Furthermore, the relationship between agricultural subsidies and fertilizer use is inherently complex and potentially non-linear. Linear regression models impose restrictive parametric assumptions, failing to capture interactions or threshold effects [97]. DML’s flexibility allows it to model both linear and non-linear relationships through machine learning algorithms like random forests or neural networks, which adaptively learn data patterns without pre-specified functional forms [98–100]. For instance, the OSCF policy’s impact may vary with regional economic development or initial fertilizer use levels-heterogeneities that DML can uncover through conditional average treatment effect estimation [96].

Endogeneity arising from unobserved confounders or reverse causality is another critical concern. Traditional instrumental variable approaches require valid instruments, which are often unavailable or weakly identified. DML addresses this by leveraging high-dimensional controls to approximate the confounding structure, thereby reducing omitted variable bias [87,95]. For example, our model incorporates 30 control variables across economic, structural, and technological dimensions, which traditional methods struggle to accommodate without overfitting.

In summary, DML offers three key advantages for this study: firstly, Robust handling of high-dimensional data through regularization and variable selection, mitigating multicollinearity and dimensionality challenges. Secondly, flexibility in modeling non-linear relationships via machine learning algorithms, avoiding restrictive parametric assumptions. Thirdly, reduced endogeneity bias by orthogonalizing treatment effects against confounders, enhancing causal identification. These strengths align with the complexities of evaluating the OSCF policy, making DML a superior choice over traditional methods for this research.

Therefore, the study employs DML to assess the policy’s effect. The impact of the OSCF policy is evaluated through a two-stage process. In the first stage, the conditional estimations of residuals are decomposed into two separate prediction tasks: one for the outcome and another for the treatment. In the second stage, the average treatment effect is estimated using a final predictive model, incorporating the residuals from the outcome and treatment models [95]. Additionally, to account for the potentially complex and non-linear relationship between the OSCF policy and chemical fertilizer use, as well as to address heterogeneity in model estimation and mitigate the subjectivity inherent in the widely used partial linear models [86], this study uses a more generalized interactive model:

(1)

(2)

In this model, i represents the county and t denotes the year. Y is the explanatory variable related to the intensity of chemical fertilizer use. The dummy variable OSCF denotes the treatment variable associated with the OSCF policy, taking a value of 1 after the pilot program implementation and 0 otherwise. X is a high-dimensional vector of control variables. And U, V represent the error terms of the primary and auxiliary regressions, respectively.

4.1. Impact of OSCF policy on chemical fertilizer use

This study employs the DML approach to assess the impact of the OSCF policy on chemical fertilizer use. The sample is divided using a 1:4 ratio, and random forest models are employed for primary and auxiliary regression predictions, and Lasso and neural network algorithms are used in the robustness tests. Our choice of machine learning algorithms for the DML framework is guided by their complementary strengths in addressing the specific challenges of our dataset and research objectives. Alternatives like support vector regression may be computationally prohibitive for high-dimensional data and requires intensive kernel tuning, limiting reproducibility [101]. And the sequential tree-building of gradient boosting machines increases overfitting risks and computational costs without proportional gains in causal inference accuracy [102]. Given our focus on causal inference rather than pure prediction, random forest’s parallelized training and inherent regularization offered a more balanced tradeoff between accuracy and efficiency. Moreover, random forest excels at capturing complex, non-linear relationships and interactions among variables without requiring explicit specification [103]. This aligns with our need to model intricate socio-economic and agricultural dynamics in a high-dimensional setting. Besides, by aggregating predictions from multiple decorrelated trees, random forest mitigates overfitting while maintaining computational efficiency, making it suitable for large-scale county-level data. Additionally, random forest provides intuitive measures of variable importance, aiding in the interpretation of key drivers (e.g., economic development, initial fertilizer use) influencing policy effects [90,95,101]. On the other hand, the Lasso can efficiently select relevant covariates via L1 regularization, reducing multicollinearity and model complexity [91], and neural networks excels at modeling hierarchical and complex patterns in large datasets, ensuring scalability for 21,659 observations [98].

To ensure the robustness of our machine learning models, we rigorously optimize hyperparameters such as the number of trees, tree depth, and minimum leaf size in the random forest algorithm. Based on a preliminary analysis, we employ a stringent cross-validation to select hyperparameters that minimized out-of-sample prediction error while maintaining computational efficiency. For example, the number of trees was set to 100 to balance prediction stability and computational cost. This aligns with best practices in machine learning literature [103] and are validated through sensitivity analyses, confirming that results remained consistent across alternative hyperparameter configurations. The results are illustrated in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Benchmark regression results.

https://doi.org/10.1371/journal.pone.0334751.t002

Column (1) of Table 2 reveals a significantly negative impact of the OSCF policy on chemical fertilizer use at the 1% significance level. When controlling for county and time fixed effects in column (2), the results consistently show a negative and significant regression coefficient, with negligible changes in the coefficient value, thereby confirming the robustness of the findings. The model is further extended in columns (3) and (4) by incorporating primary and quadratic terms for control variables, respectively. In column (5), the model accounts for both county-fixed effects and time-fixed effects, as well as primary and quadratic terms of control variables, showing an increase in the regression coefficients for chemical fertilizer use associated with the OSCF policy. Nonetheless, the policy continues to have a significant negative effect on chemical fertilizer use, thereby supporting Hypothesis H1.

The findings are consistent with previous research that views agricultural subsidies as a potent policy instrument for promoting sustainable agricultural development [1,16]. As previously analyzed, the reduction in chemical fertilizer use can be attributed to three factors influenced by the OSCF policy. Firstly, the subsidy for organic fertilizer narrows the price gap with chemical fertilizers, encouraging farmers to adopt alternative organic fertilizers [4]. Secondly, the OSCF policy enhances farmers’ ability to absorb new knowledge and promotes the adoption of environmentally-friendly technologies [104], which in turn improves agricultural productivity and decreases overall chemical fertilizer consumption [60]. Thirdly, the policy’s linkage with large-scale farms incentivizes farmers to adopt large-scale and intensive cultivation modes, which typically operate on economies of scale, thereby improving agricultural productivity and reducing chemical fertilizer use [66]. The indirect effects of the OSCF policy on chemical fertilizer use through the improvement of agricultural productivity from technical and scale perspectives are further substantiated in the section dedicated to mediating effect analysis.

4.2. Robustness tests

This section aims to validate the robustness of the findings presented in the benchmark model. Several alternative methods are employed to assess the stability of the results across different analytical approaches and sample configurations. The robustness tests are conducted based on the benchmark model of column (5) in Table 2, and the results are presented in Tables 3 and 4. The robustness tests conducted using various methods consistently support the findings of the benchmark model. The OSCF policy is found to have a significant negative impact on chemical fertilizer use, and this effect is robust across different analytical approaches, sample configurations, and the presence of other concurrent policies.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. The results of the robustness test.

https://doi.org/10.1371/journal.pone.0334751.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Results after excluding the effect of other policies on chemical fertilizer use.

https://doi.org/10.1371/journal.pone.0334751.t004

4.2.6. Falsification test.

To further validate the robustness of our findings and address potential endogeneity, we conduct falsification tests in addition to the original robustness checks. These tests aim to verify whether the observed policy effects are spurious or driven by unobserved confounders. The results are presented in Table 5. we first construct a placebo OSCF policy by randomly assigning fictitious treatment years to counties in the control group. Using the same DML framework, we estimated the ‘effect’ of this placebo policy on chemical fertilizer use. Results show no statistically significant impact, confirming that the true OSCF policy’s effect is not attributable to temporal trends or coincidental shocks. Then, we replace the actual treatment group (OSCF pilot counties) with a randomly selected subset of control counties and re-ran the analysis. The estimated coefficient for the fictitious treatment group is negligible and statistically insignificant, reinforcing that the true treatment effect arises from the OSCF policy rather than unobserved county-level characteristics.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The results of the falsification test.

https://doi.org/10.1371/journal.pone.0334751.t005

4.2.7. Parallel trend test.

A fundamental prerequisite for ensuring the validity of policy evaluation is that the treatment and control groups should follow parallel trends in chemical fertilizer use prior to the implementation of the OSCF policy. To verify this critical assumption, we employ an event study methodology. As illustrated in Fig 2, the results demonstrate that during the pre-policy period (2012–2016), there are no statistically significant differences in chemical fertilizer application between the pilot counties and others. This indicates that the two groups exhibit comparable evolving patterns in chemical fertilizer use before the policy intervention, thereby satisfying the parallel trend assumption. In contrast, after the implementation of the OSCF policy, a notable and statistically significant divergence in the chemical fertilizer use emerges between the two groups, which aligns with the expected policy effects. These findings lay a solid foundation for the reliability of the subsequent DML estimates, confirming that the observed differences in chemical fertilizer use post-policy can be reasonably attributed to the OSCF policy rather than pre-existing systematic disparities between the groups.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Parallel trend test.

https://doi.org/10.1371/journal.pone.0334751.g002

4.3. Heterogeneity analysis

4.3.1. Developed and underdeveloped counties.

The economic disparity among different regions of China results in varying impacts of the OSCF policy on chemical fertilizer use across counties. Many scholars argue that environmental policy instruments may be more effective in developed areas due to higher investment in pollution control and a greater priority on environmental protection over economic growth [106,108,109]. To explore this, we categorize the sample counties into developed and underdeveloped regions by the median of GDP per capita and conduct separate regression analyses for each region. The results presented in Table 6 reveal that the OSCF policy significantly reduces chemical fertilizer use in developed areas, whereas this effect is minimal in underdeveloped regions. This aligns with previous studies, which demonstrate that the environmental policy instruments may be more effective in developed regions in China [110,111]. This suggests that the OSCF policy has a more pronounced impact on reducing chemical fertilizer use in developed areas, implying the OSCF policy may be more effective in these areas.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Heterogeneity analysis results.

https://doi.org/10.1371/journal.pone.0334751.t006

This may be explained by the following reasons: Firstly, infrastructure and market access. Developed counties possess better rural infrastructure (e.g., organic fertilizer distribution networks, digital extension services), enabling efficient adoption of organic substitutes. In contrast, underdeveloped regions face logistical bottlenecks, raising transaction costs for organic inputs. [112]. Secondly, farmer capacity. Developed regions often have a higher economic base, which means that farmers can afford to invest in better agricultural inputs. Moreover, farmers in developed regions exhibit higher environmental awareness and technical literacy, facilitating the uptake of precision farming techniques promoted by the OSCF policy [113]. Thirdly, policy enforcement. Local governments in developed counties prioritize environmental goals over short-term economic gains, rigorously implementing OSCF targets [111].

To address these disparities, Tailored interventions are critical. For underdeveloped regions, complementary investments in rural infrastructure (e.g., organic fertilizer storage facilities) and farmer training programs are needed to amplify OSCF’s impact.

4.4. Mediation analysis

Mediation analysis is a valuable tool for understanding the pathways through which an independent variable influences a dependent variable [119]. The theoretical analysis suggests that the OSCF policy’s impact on reducing chemical fertilizer use is primarily mediated by improvements in technical and scale efficiency. This section explores these underlying mechanisms using the technical efficiency (TE) and scale efficiency (SE) as mediating variables. The causal mediation analysis of DML is used to analyze these mechanisms based on the Lasso algorithm [100].

The first row of Table 7 presents the results of the mediation analysis focusing on the role of TE. The analysis reveals a significant negative direct and indirect effect of the OSCF policy on chemical fertilizer use in both the treatment and control groups, supporting Hypothesis H2. The OSCF policy facilitates investment in advanced agricultural technologies and practices by addressing liquidity constraints, enabling more precise fertilizer application and reducing chemical fertilizer use. Furthermore, training programs and knowledge transfer initiatives enhance farmers’ understanding of efficient fertilizer management, leading to more informed decision-making and reduced chemical fertilizer use. Additionally, the policy encourages the development of new and improved fertilizer technologies, offering more efficient alternatives to traditional fertilizers.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Mechanism test results.

https://doi.org/10.1371/journal.pone.0334751.t007

The second row of Table 7 explores the mediating role of SE. The analysis demonstrates a significant negative indirect effect of the OSCF policy on chemical fertilizer use in both the treatment and control groups. After accounting for the SE pathway, the direct effects remain significantly negative, supporting Hypothesis H3. The OSCF policy promotes the consolidation of operations and adoption of larger-scale production methods. Larger-scale farms benefit from economies of scale, leading to more efficient resource allocation and enabling investment in better infrastructure and sophisticated management practices, including efficient fertilizer management. These factors contribute to reduced chemical fertilizer use in large-scale farming operations.

4.5. Feature importance

To enhance the interpretability of our machine learning-based analysis, we further estimated the importance of different features using random forest feature importance scores (based on mean decrease in Gini impurity). This allows us to clarify which pre-specified control variables most strongly shape the relationships between the OSCF policy and chemical fertilizer use.

The importance scores reveal a clear hierarchy aligned with agricultural production principles and policy implementation realities (Fig 3). Agricultural development level (ADL) emerges as the most influential factor, reflecting its foundational role in shaping resource allocation efficiency—higher ADL corresponds to more standardized production management and technological adoption, directly reducing fertilizer waste through improved technical efficiency [54]. Agricultural industry structure (AIS) ranks second, validating our heterogeneity findings that non-grain crop regions (focused on fruits, vegetables, and tea) exhibit stronger policy responsiveness due to the targeted design of OSCF measures [37].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Feature importance.

https://doi.org/10.1371/journal.pone.0334751.g003

Agricultural mechanization (AM) and agricultural marketization level (AML) highlight the importance of technical progress and market mechanisms: mechanization enhances precision in fertilizer application [16], while higher marketization facilitates coordinated promotion of organic fertilizer substitutes through supply chain networks [23]. Education level (EL) further supports the role of human capital in adopting scientific fertilization practices [113]. Land scale operation (LSO) aligns with our scale efficiency mechanism, with its relatively moderate score potentially reflecting uneven adoption of large-scale farming in the sample [32]. Collectively, these covariates reinforce that the OSCF policy operates through interconnected channels of technical advancement, structural adjustment, and regional development capacity.

4.6. Policy implications

Our findings offer several policy implications for promoting sustainable agricultural development. Firstly, the OSCF policy demonstrates the potential of agricultural subsidies to reduce chemical fertilizer use, crucial for long-term sustainability in China and globally. Policymakers should recognize and leverage the significant role of agricultural subsidies in fostering sustainability, potentially informing policy reforms in other regions.

Secondly, the technical effect of the OSCF policy highlights the importance of enhancing farmers’ access to advanced knowledge and environmentally-friendly agricultural techniques. Policymakers should prioritize measures to facilitate the adoption of new agricultural technologies through intensified agricultural technology dissemination, collaboration with enterprises and research institutions, and increased opportunities for farmers to participate in technical training and production experience sharing [26,113,120].

Thirdly, the scale effect of the OSCF policy emphasizes the benefits of larger-scale farming operations in achieving economies of scale and reducing fertilizer use. Policymakers should focus on strengthening land management and promoting rational land transfer to ensure efficient resource allocation and overcome the inefficiencies of small-scale farms. This includes providing incentives for the development of land rental markets and establishing more land transfer service platforms [1].

Finally, the heterogeneous analysis indicates that differentiated policy instruments are crucial for maximizing effectiveness of the OSCF policy across different regions. Policymakers should prioritize developed regions, cash crop production areas, and regions with high initial chemical fertilizer use to enhance the impact of the OSCF policy. For underdeveloped regions, grain-producing areas, and regions with low initial fertilizer use, additional support and complementary policy instruments should be considered to address specific challenges and promote sustainable agricultural practices.

4.7. Limitations and future works

This study offers a comprehensive understanding of the OSCF policy’s impact on chemical fertilizer use, providing valuable insights for policymakers and farm managers to promote sustainable agricultural development. However, certain limitations are acknowledged. Firstly, while many factors influence fertilizer use, this study focuses primarily on macro-level factors, potentially overlooking micro-level influences. Future research should integrate macro and micro data to consider all potential driving factors comprehensively. Secondly, although this study conducts heterogeneity analysis on policy effects and provide a brief illustration of key features, it has not fully examined the impact of all influencing factors on policy outcomes. Further investigation is needed to understand the contributions of different factors to fertilizer use reduction through subsidy policies. This will enable policymakers to implement more targeted measures and maximize policy efficiency. Besides, future research could integrate geospatial data to explore spatial autocorrelation or cluster effects. Such analyses would further elucidate how localized environmental and infrastructural factors moderate policy impacts. Lastly, this study only examines the policy impact within pilot areas, neglecting potential spillover effects on surrounding regions. Future research should consider the broader implications of agricultural subsidy policies to assess their full impact on sustainable agricultural development.