How Digital Literacy Shapes Farmers’ Adoption of Green Agricultural Technologies: Evidence from Specialty Crop Producers in Jilin Province

Abstract

In the context of accelerating digital transformation and increasing pressure for sustainable agriculture, understanding how digital capabilities shape farmers’ green production behavior has become a critical issue. This study examines how digital literacy influences the adoption of green agricultural technologies as well as the mechanisms driving this relationship, based on survey data collected from 422 ginseng farmers in Jilin Province, China. An ordered Probit model is employed, complemented by an instrumental variable approach and mediation analysis to ensure robust causal inference. The findings indicate that digital literacy plays a significant role in encouraging farmers to adopt green production technologies, and this effect remains robust after addressing potential endogeneity issues. Mechanism analysis indicates that digital literacy mainly functions by strengthening farmers’ multi-dimensional technological cognition, including technological attribute cognition, technological ecological cognition, technological economic cognition, and technological risk cognition, while the information cost channel is not supported. Interestingly, heterogeneity analysis further shows that the impact of digital literacy remains relatively consistent across different generations, farm sizes, and levels of cooperative participation. This finding suggests that digital literacy functions as a “foundational and universally effective capability”, particularly in highly specialized and regionally concentrated agricultural systems such as ginseng production. This study contributes to the existing literature by shifting the focus from information access to cognitive empowerment in explaining green technology adoption, and by refining the understanding of the boundary conditions under which digital literacy exerts uniform effects. The findings provide important implications for digital rural development policies aimed at promoting sustainable agricultural transformation.

1. Introduction

Digital transformation is fundamentally reshaping economic and social systems and has profound implications for both farming practices and the transition toward sustainable development. Against this backdrop, adopting green production technologies serves not only as an effective approach to environmental governance but also a critical component of sustainable agricultural transformation [1,2]. Promoting green agriculture requires technological innovation as a driving force, and scholars generally emphasize both the foundational role of green technologies in supporting sustainable agricultural transitions [3] while highlighting the importance of strengthening farmers’ motivation to adopt such technologies [4]. Despite these efforts, farmers’ adoption of green technologies remains limited due to high costs, uncertain risks, and information constraints [5]. Amid the rapid expansion of the digital economy, digital technologies have become more prevalent in farmers’ production and daily life, transforming cognitive patterns and behavioral habits. Digital literacy, as an important form of human capital in the digital age, reflects individuals’ ability to access, process, and use digital information, and acts as a key link between digital technologies and behavioral outcomes. Farmers, as the primary micro-level agents in agricultural production systems, rely on digital literacy to mitigate information asymmetries and insufficiencies encountered in factor markets. Understanding the mechanisms and pathways through which digital literacy stimulates green technology adoption is therefore essential for advancing sustainable agricultural transformation and high-quality development.

Existing studies have examined determinants of farmers’ green technology adoption through multiple perspectives. Broadly, these factors are categorized into intrinsic attributes, including aging [6], cognition [7,8], and risk perception [9,10], and external environmental factors, such as social capital [11], government regulation [12], ecological compensation [13], and cooperative participation [14]. Together, these dimensions provide a core analytical framework for examining adoption behaviors. With the proliferation of digital technologies in agriculture, scholars have increasingly explored the linkages between digital elements and green transformation. Research has highlighted the roles of the digital economy [15], digital technologies [16], and digital tools [17] in promoting sustainable production practices. Evidence suggests that digitalization improves information flow, optimizes resource allocation, and encourages environmentally friendly behaviors among farmers. The concept of digital literacy has evolved alongside technological advances and rising public digital competence. While no universally accepted definition exists, digital literacy is generally recognized as a composite of abilities and attitudes enabling individuals to effectively utilize digital tools in learning, work, and daily life [18,19]. Empirical studies demonstrate that digital literacy facilitates behavior changes, such as reducing fertilizer usage [20], optimizing cropping structures [21], and adopting environmentally conscious practices [22], with credit accessibility [23], social capital [24], and cognition [25,26,27] serving as mediating mechanisms.

This study offers significant innovations and theoretical contributions to the field of digital literacy and its impact on the adoption of green agricultural technologies. First, while a wealth of literature has examined the relationship between digital literacy, technological cognition, and the adoption of green agricultural technologies, most of the existing research has focused on general agricultural technologies or traditional crops. There is a noticeable gap in studies specifically addressing the context of specialty crops, such as ginseng farming. Ginseng, as a distinctive agricultural product in Jilin Province, involves high technical barriers and ecological requirements, and farmers’ adoption decisions are often influenced by information asymmetry and risk uncertainty. By focusing on ginseng farmers in Jilin Province, this study fills the gap in the literature on green agricultural technology adoption, particularly the role of digital literacy in specialty crop contexts, which has not been sufficiently explored. Second, from the perspective of cognitive empowerment, this study introduces and validates the pathway through which digital literacy enhances farmers’ multidimensional technological cognition—including technological attribute cognition, ecological cognition, economic cognition, and risk cognition—to facilitate the adoption of green production technologies. This theoretical framework innovatively emphasizes the profound influence of digital literacy on the cognitive processes of farmers, surpassing the traditional view of digital literacy as merely an information acquisition tool. It offers a new perspective on the promotion of green agricultural technologies. Moreover, by developing a three-dimensional digital literacy measurement system (comprising basic digital literacy, digital learning literacy, and digital business literacy), this study provides an operational model for quantifying digital literacy, advancing its application within the human capital theory. Finally, this study conducts a heterogeneity analysis of ginseng farmers in Jilin Province, further verifying the universal effects of digital literacy on technology adoption across different generational cohorts, farm scales, and cooperative participation. This finding challenges the emphasis on differences between farmer groups in many existing studies, offering important evidence of the broad applicability of digital literacy as a foundational capability for agricultural green transformation.

In conclusion, this study not only deepens the theoretical understanding of the relationship between digital literacy, technological cognition, and green technology adoption but also provides practical empirical insights for agricultural policy-making in Jilin Province and similar regions. The findings offer significant policy implications, especially in the context of digital rural development and sustainable agricultural transformation, contributing to the long-term sustainability of agricultural practices.

2. Theoretical Framework and Hypothesis Development

2.1. Direct Impact of Digital Literacy on Green Technology Adoption

Shifting from traditional farming practices to environmentally sustainable production essentially entails the continuous introduction, diffusion, and application of agricultural technologies. Under traditional rural development models, farmers’ limited knowledge, cognitive capacity, and access to information often result in pronounced risk-averse behaviors and path dependency, manifesting as insufficient awareness of new technologies and low adoption willingness, thereby creating a “low-equilibrium” state that constrains technological progress and green transformation [28]. Schultz (1964), in Transforming Traditional Agriculture, emphasizes that human capital is a critical driver of agricultural modernization, with its level directly determining farmers’ ability to understand and absorb new technologies [29]. Extending this perspective, cognitive human capital refers to individuals’ capabilities to acquire, process, and apply knowledge to solve problems and make rational decisions in complex environments. Unlike traditional human capital, which primarily emphasizes physical and technical skills, cognitive human capital highlights information-processing, analytical reasoning, and decision-making capacities, and exhibits stronger knowledge spillovers and marginal returns [30,31]. Within the evolving digital economy, digital literacy can be understood as an important dimension of cognitive human capital, enabling farmers to efficiently access, interpret, and utilize digital information, thereby influencing their technology adoption decisions. In this study, digital literacy is conceptualized as cognitive human capital, linking digital capabilities with behavioral outcomes and providing a theoretical foundation for understanding the micro-level mechanisms underlying green technology adoption.

Mechanistically, digital literacy may influence adoption through multiple pathways, including the alleviation of information constraints and the enhancement of cognitive evaluation. Among these, cognitive evaluation is expected to play a more central role in shaping farmers’ decision-making processes. Farmers with higher digital literacy possess stronger information identification and processing capabilities, mitigating information asymmetries and reducing adverse selection and moral hazard, which lowers transaction costs [32,33]. Additionally, digital literacy broadens information channels, enhancing both the breadth and depth of information access. Diverse sources improve risk recognition and response capacity and strengthen awareness of the benefits of green technologies, thereby increasing perceived utility [34,35]. Improved information-processing ability enables more accurate cost–benefit evaluations of green technologies. Under rational economic assumptions, when expected benefits increase and uncertainty decreases, farmers are better positioned to engage in the adoption of these technologies. Finally, as a form of human capital, digital literacy lowers technical understanding and application barriers, facilitating the spread and practical use of green technologies [36]. Given the importance of adoption costs, digital literacy can reduce search and evaluation costs, enhancing decision efficiency [37,38,39].

Based on the above analysis, the following hypotheses are proposed:

H1: 

Digital literacy is positively associated with farmers’ adoption of green technologies.

H1a: 

Basic digital literacy is positively related to adoption behavior.

H1b: 

Digital learning literacy is positively related to adoption behavior.

H1c: 

Higher levels of digital business literacy are linked to increased technology adoption.

2.2. Mechanisms of Digital Literacy on Technology Adoption

To clarify the theoretical transmission mechanism, this study integrates cognitive behavioral theory and the theory of planned behavior into a unified analytical framework. Cognitive behavioral theory emphasizes that external stimuli influence individual behavior through internal cognitive processing structures, while the theory of planned behavior further specifies that behavioral intention is shaped by individuals’ evaluations of expected outcomes.

In this study, digital literacy is conceptualized as a form of cognitive human capital that acts as an external capability, influencing farmers’ internal cognitive evaluations of green technologies. These cognitive evaluations—reflected in perceptions of technology attributes, ecological benefits, economic returns, and risks—correspond to the “attitude toward behavior” component in the theory of planned behavior, which represents the most directly relevant cognitive channel in this study, while subjective norms and perceived behavioral control are implicitly captured through farmers’ resource endowments and social contexts. Thereby shaping farmers’ adoption decisions. Accordingly, digital literacy affects adoption behavior indirectly through cognitive evaluation, forming a structured pathway from capability to cognition to action.

With cognitive biases often underlying non-rational actions [40]. The theory of planned behavior further posits that behavioral intentions depend on evaluations of anticipated outcomes [41]. In agriculture, farmers’ adoption of green technologies depends on their cognition of technology attributes, which is strongly influenced by digital literacy [42]. Digital literacy enhances information acquisition and processing, improving farmers’ understanding and evaluation of green technologies, thus shaping subjective value perception—the integrated assessment of benefits, costs, and risks—which is a key antecedent of adoption [43,44]. Farmers with low digital literacy struggle to form comprehensive cognitive assessments, leading to uncertainty and reduced adoption likelihood; by contrast, high digital literacy enables systematic evaluation through digital tools, resulting in more rational decision-making and stronger adoption intentions.

This forms a theoretically grounded pathway of “digital literacy (cognitive capability) → technological cognition (attitudinal evaluation) → adoption behavior,” which explicitly links cognitive behavioral theory with the theory of planned behavior. In this framework, digital literacy is treated as the independent variable, technological cognition as the mediating variable, and green technology adoption as the dependent variable. Accordingly, we propose

H2: 

The effect of digital literacy on technology adoption operates through farmers’ cognitive evaluations.

We operationalize technological cognition along four dimensions: technology attributes, ecological benefits, economic benefits, and risk perception [45]. Digital literacy affects these dimensions by enabling deeper understanding of technical features, enhancing ecological awareness via digital channels, and supporting comprehensive economic and risk evaluations, thereby promoting adoption. Farmers with lower digital literacy face cognitive limitations, restricting adoption. The following specific hypotheses are proposed:

H2a: 

Cognition of technological attributes acts as a pathway linking digital literacy to adoption behavior.

H2b: 

Ecological cognition serves as a mediating channel in this relationship.

H2c: 

Economic cognition functions as an intermediary mechanism between digital literacy and adoption behavior.

H2d: 

Risk awareness plays a mediating role in shaping adoption decisions.

Additionally, based on Akerlof’s “market for lemons” theory, information asymmetry is a key source of market failure, arising from high costs of information acquisition and processing [46]. In agricultural production, farmers face information constraints in accessing technologies, purchasing inputs, and marketing outputs, which increase adoption thresholds. Farmers with higher levels of digital literacy are better able to access, assess, and integrate information via digital tools, thereby lowering information search costs and reducing decision-making risks. From the perspective of information economics, information costs represent an additional mechanism linking digital literacy to behavioral outcomes, complementing the cognition-based pathway. Hence, digital literacy also affects adoption via the “information cost reduction → decision efficiency → adoption” pathway:

H3: 

The influence of digital literacy on adoption behavior is partly transmitted through reductions in information-related costs.

H3a: 

Reduction in search costs mediates this effect.

H3b: 

Reduction in acquisition costs mediates this effect.

2.3. Heterogeneity in Digital Literacy Effects

The influence of digital literacy on technology adoption could be heterogeneous across different farmer groups. Drawing on life-course theory, the technology acceptance model, and resource-based perspectives, this study examines heterogeneity across generations, farm scale, and organizational embedding. Younger farmers, having grown up in a digitally rich environment, possess greater information access and technological adaptability, leading to stronger effects of digital literacy [47]. In terms of farm scale, larger-scale operations benefit from resource endowments and economies of scale, amplifying the positive effects of digital literacy [48,49]. Regarding cooperative participation, social network and organizational embedding theories suggest that cooperatives provide information, training, and risk-sharing mechanisms, which may substitute for or complement digital literacy. Accordingly,

H4: 

The positive association between digital literacy and green technology adoption varies across farmer groups. Specifically, the effect is stronger among younger farmers, those operating at larger scales, and those participating in cooperatives.

Based on the preceding theoretical analysis and hypothesis development, the conceptual framework of this study is presented in Figure 1.

3. Methodology

3.1. Data Collection

Data for this study were obtained through a field survey conducted by the authors from July to November 2024, targeting ginseng farmers in Jilin Province. Jilin represents the core cultivation region of Changbai Mountain ginseng, accounting for approximately 60% of global ginseng production, with the most complete ginseng industry chain in China. To ensure representativeness, the survey focused on the main ginseng-producing areas of Tonghua City, Baishan City, and Yanbian Prefecture along the Changbai Mountain range. Within these regions, 10–15 townships engaged in ginseng production were randomly selected as the initial survey areas. Subsequently, based on the actual cultivation scale, individual villages and their ginseng farmers were randomly sampled. Of the 450 questionnaires distributed, 439 were returned, yielding a response rate of 98%. To comprehensively capture the necessary information, a dual approach combining standardized questionnaires and semi-structured interviews within participatory rural appraisal was employed. After systematically checking for completeness and logical consistency, the final dataset consisted of 422 valid questionnaires, yielding an effective rate of 96%.

3.2. Variable Selection

3.2.1. Dependent Variable: Adoption of Green Ginseng Production Technologies

This study defines green ginseng production practices as an integrated production system guided by the principles of ecological sustainability, resource conservation, and environmental friendliness. Its core objectives are to mitigate soil degradation, reduce pesticide and heavy metal residues, and improve both the quality of ginseng and the content of its bioactive compounds. Based on existing studies [50,51], green production practices can be divided into two stages: the pre-production stage and the production stage. The pre-production stage includes site selection, soil management, and fertilization based on soil testing. The production stage involves the application of organic fertilizers and green manures, as well as integrated pest and weed management.

In this study, the dependent variable is constructed based on the following five clearly defined green production practices: site selection, soil management, soil testing–based fertilization, organic fertilizer and green manure application, and integrated pest and weed management. In this study, farmers’ adoption of green production technology is measured by the number of green production practices actually implemented at the farm level. Specifically, we adopt adoption breadth as the core indicator, a measurement approach that has been widely used in the literature on technology adoption [25,35,49]. It is worth noting that all five green production practices examined in this study are key components specified in the Technical Standards for High-Quality Ginseng Cultivation. From an institutional perspective, they collectively constitute the fundamental requirements of green production; the omission of any single practice may result in a deviation from the conceptual scope of green production. Furthermore, based on a preliminary survey of 20 households and interviews with agricultural extension officers, these practices are generally comparable in terms of economic input requirements and technical difficulty.

Given these institutional constraints and empirical characteristics, this study employs an equal-weight aggregation approach as the baseline measure of farmers’ green production technology adoption. This measure is designed to capture the overall extent of farmers’ engagement in green production rather than the marginal contribution of individual practices. Therefore, it is conceptually consistent and methodologically appropriate within the analytical framework of this study.

3.2.2. Independent Variable: Digital Literacy

Digital literacy is conceptualized as a multidimensional competency that empowers individuals to appropriately utilize digital tools, access up-to-date information and knowledge, participate in social interactions through platforms such as WeChat, and carry out commercial transactions via digital channels in contemporary digital contexts. Grounded in Martin and Grudziecki’s three-layer digital literacy framework, the European Commission’s DigComp 2.2, and the work of Su et al. [52], this study establishes a context-specific digital literacy measurement system for ginseng farmers. The system encompasses three dimensions—basic literacy, learning literacy, and commercial literacy—with a total of 14 items. These three dimensions correspond to distinct but complementary capability domains: basic digital literacy reflects access to and operation of digital tools, digital learning literacy captures the ability to acquire and process information, and digital business literacy represents the capacity to engage in digitally enabled market interactions. This classification ensures conceptual clarity and aligns with the functional differentiation emphasized in existing digital capability frameworks. All items are measured using a five-point Likert scale (1 = strongly disagree; 5 = strongly agree). The entropy weight method, an objective weighting technique, is adopted to calculate the composite digital literacy score, and the corresponding indicator weights are reported in

Table 1. Measurement of Indicator Weights for Ginseng Farmers’ Digital Literacy.

First-Level Indicator	Second-Level Indicator	Third-Level Indicator	Weight
Digital Literacy	Basic Digital Literacy	The family owns at least one smart Internet-enabled device	0.0524
		Perceives mastery of Internet skills as important	0.0538
		Familiarity with social networking tools such as QQ, WeChat, and Douyin.	0.0711
		Able to shop online (on platforms such as Taobao and JD.com)	0.0709
		Able to make payments and settle fees via mobile phone	0.0625
		Maintains a cautious attitude toward operations such as entering ID numbers and fund transfers	0.0661
	Digital Learning Literacy	Able to learn living and production skills via smart devices	0.0734
		Able to search for needed information online	0.0818
		Able to distinguish valid from invalid information among massive online content	0.0741
		Has participated in ginseng cultivation-related training online	0.0770
	Digital Business Literacy	Participated in online activities for purchasing fresh ginseng	0.0660
		Obtain information and negotiate fresh ginseng prices via the Internet.	0.0749
		Sell ginseng products through digital video and live streaming platforms.	0.0711
		Complete transactions of ginseng products via e-commerce platforms	0.1050

.

The reliability and validity of the scale are presented in Appendix

Table A1. Reliability and validity of the digital literacy scale (original dimension structure).

Dimension	No. of Items	Cronbach’s α	KMO	Bartlett’s Test of Sphericity	Cumulative Variance Explained (%)
Basic Digital Literacy	6	0.858
Digital learning literacy	4	0.870
Digital business literacy	4	0.867
Total scale	14	0.891	0.890	χ2 = 3575.885, df = 91, p < 0.001	71.29%

Note: Exploratory factor analysis (principal component method with varimax rotation) extracted three factors, cumulatively explaining 71.29% of the total variance. All items loaded above 0.63 (unrotated) or 0.76 (rotated) on their theoretically assigned dimensions, indicating adequate construct validity. For item e6 (“cautious about ID and fund transfers”), the rotated factor loading was slightly higher on the learning literacy dimension than on the basic literacy dimension. Based on the theoretical framework of DigComp 2.2, which classifies security-related competencies under foundational digital literacy, we retain the original classification to maintain theoretical consistency and comparability with subsequent analyses. All subscale Cronbach’s α values exceed the recommended threshold of 0.70. Total scale Cronbach’s α = 0.891, KMO = 0.890, Bartlett’s test of sphericity: χ²(91) = 3575.885, p < 0.001.

. Overall, the Cronbach’s α for the full scale is 0.891, and the α values for the three subscales range from 0.858 to 0.870. Given the exploratory nature of the construct and the adaptation to a rural context, we employ EFA as an initial validation approach. Exploratory factor analysis yields a KMO value of 0.890 and a statistically significant Bartlett’s test of sphericity (p < 0.001). The three extracted factors collectively account for 71.29% of the total variance, and all item factor loadings exceed 0.63. These results collectively demonstrate satisfactory reliability and construct validity of the scale.

3.2.3. Mediating Variables

In this study, we consider two types of mediating variables: technology cognition and information cost.

Technology cognition is assessed across four dimensions: technological attribute cognition, ecological cognition, economic cognition, and risk cognition. Each dimension is measured using two items. Reliability tests show that the Cronbach’s α coefficient for each dimension exceeds the commonly accepted threshold of 0.7 (see

Table 2. Measurement items and reliability of key variables.

Variable	Category	Measurement Item	Cronbach’s α
Technology Cognition	Technological attribute cognition	The operation of green ginseng production technology is not complex and is easy to learn	0.8063
		The adoption of green ginseng production technology does not increase labor and time input
	Technological Ecological Cognition	The adoption of this technology can prevent soil erosion and reduce environmental pollution	0.7785
		The adoption of this technology can improve soil fertility
	Technological Economic Cognition	The adoption of this technology can improve ginseng quality and enable better market prices	0.8682
		The adoption of this technology can reduce input costs (e.g., fertilizers and pesticides) and increase net income
	Technological Risk Cognition	I am concerned that adopting this technology may reduce ginseng yield and affect income	0.7867
		I am concerned that ginseng produced using this technology may not obtain a good market price
Information Cost	Information Search Cost	Accuracy of information search on ginseng green standardized production technologies	—
	Information Acquisition Cost	Difficulty in obtaining information on ginseng green standardized production technologies	—

), confirming satisfactory internal consistency. Each dimension score is calculated as the equal-weighted average of its two corresponding items. To examine discriminant validity among the four dimensions, we computed Pearson correlation coefficients between the dimension scores. As shown in Appendix

Table A2. Discriminant validity of technology cognition dimensions (Pearson correlations).

Dimension	(1)	(2)	(3)	(4)
(1) Technological Attribute cognition	1.000
(2) Technological Ecological Cognition	0.279 ***	1.000
(3) Technological Economic Cognition	0.634 ***	0.488 ***	1.000
(4) Technological Risk Cognition	0.316 ***	0.465 ***	0.481 ***	1.000

Note: *** p < 0.001. All correlation coefficients are below 0.70, indicating no multicollinearity concerns. The highest correlation is between Technological Character Cognition and Technological Economic Cognition (r = 0.634), which is expected as they belong to the same higher-order construct of technology cognition. The lowest correlation is between Technological Attribute cognition and Technological Ecological Cognition (r = 0.279), supporting discriminant validity. These moderate correlations justify treating the four dimensions as separate mediators in the mechanism analysis (Section 4.3).

, the coefficients range from 0.279 to 0.634 and are all statistically significant at the p < 0.001 level. These moderate correlations indicate that the four dimensions are conceptually related but sufficiently distinct, thus providing support for discriminant validity.

Information cost is captured by two dimensions: information search cost and information acquisition cost. Both are measured using single-item indicators reflecting the accuracy of searching for green production technology information and the difficulty of obtaining such information, respectively. Higher search accuracy and lower acquisition difficulty correspond to lower information costs. Consistent with existing literature [53,54], single-item measurement is appropriate here because these variables represent concrete and factual assessments.

3.2.4. Control Variables

The analysis controlled for variables including: (i) farmer characteristics: gender, age, education, health status, village leadership role, and off-farm employment; and (ii) production and management factors: labor availability, farm scale, number of plots, cooperative membership, and participation in technical training. Regional-level fixed effects were also included to control for unobservable heterogeneity in resources and economic development across municipalities.

3.2.5. Instrumental Variables

To mitigate potential endogeneity caused by unobserved confounders or reverse causality between digital literacy and green technology adoption, this study employs village-level digital governance capacity as an instrumental variable. The instrument satisfies the standard criteria of relevance to the endogenous regressor and exogeneity to the error term. All variables defined in Section 3.2.1, Section 3.2.2, Section 3.2.3, Section 3.2.4 and Section 3.2.5 are reported in

Table 3. Descriptive statistics.

Variable Category	Variable Name	Variable Definition & Assignment	Mean	Std. Dev.
Dependent Variable	Green Production Technology	Total adoption score of five green practices	3.7772	1.3948
Independent Variable	Digital Literacy	Calculated by employing the entropy method	0.6216	0.1906
	Basic Digital Literacy		0.2427	0.0835
	Digital Learning Literacy		0.1894	0.0794
	Digital Business Literacy		0.1895	0.0801
Mediating Variables	Technological attribute cognition	Calculated by the equal-weight averaging method	0.5873	0.2391
	Technological Ecological Cognition		0.6582	0.2339
	Technological Economic Cognition		1.2575	0.3686
	Technological Risk Cognition		0.5400	0.2233
	Information Search Cost	Accuracy of information search (1 = very inaccurate, 5 = very accurate)	3.5170	0.8250
	Information Acquisition Cost	Convenience of obtaining information (1 = very difficult, 5 = very easy)	3.5146	0.7979
Control Variables	Gender	1 = male, 0 = female	0.8886	0.3149
	Age	Age (years)	52.7369	8.7341
	Education Level	1 = primary or below, 2 = junior high, 3 = senior high/technical secondary, 4 = junior college or above	2.0308	0.6834
	Health Status	1 = average, 2 = good, 3 = very good	2.2204	0.6901
	Village Cadre Identity	1 = yes, 0 = no	0.0758	0.2650
	Off-farm Employment Status	1 = engaged in part-time activities, 0 = no	0.4834	0.5003
	Number of Laborers	Number of family laborers engaged in ginseng cultivation	2.0379	0.6707
	Operation Scale	Area of household ginseng land (mu)	31.7495	60.3296
	Number of Plots	Number of household ginseng plots	2.1967	1.4348
	Cooperative Membership	1 = joined a cooperative, 0 = no	0.2109	0.4084
	Participation in Technical Training	1 = participated in ginseng green production training, 0 = no	0.7844	0.4118
	Regional Dummy Variables	1 = Tonghua City, 0 = others	0.4621	0.4992
		1 = Baishan City, 0 = others	0.2441	0.4300
		1 = Yanbian Korean Autonomous Prefecture, 0 = others	0.2938	0.4560
Instrumental Variable	Digital Governance Capacity of Towns and Villages	Total number of digital services provided by the town/village: digital government services, smart Party building services, digital village affairs, digital cultural publicity, digital medical services	4.1800	0.7234

(Descriptive Statistics).

3.3. Empirical Model

3.3.1. Ordered Probit Model

Given that the dependent variable—farmers’ green production behavior—is an ordinal measure, an ordered probit (OProbit) model was employed to evaluate the effect of digital literacy on technology adoption. The model specification is

$$At{n}_{ij}=\alpha +\beta D{L}_{iJ}+\gamma x_{ij}+{\mu }_j+{\epsilon }_{ij}$$

(1)

In this specification, subscripts $i$ and $j$ denote individual farmers and their corresponding prefectures; $At{n}_{ij}$ represents the green technology adoption behavior; $D{L}_{ij}$ denotes the core independent variable of digital literacy; $X_{ij}$ denotes the control variables included in the model; $\alpha$ represents the intercept; $\beta$ and $\gamma$ are parameters to be estimated; ${\mu }_j$ accounts for regional fixed effects; and ${\epsilon }_{ij}$ denotes the random error term. By including municipality- and prefecture-level fixed effects, the model controls for unobserved regional characteristics that could simultaneously influence adoption and digital literacy, mitigating potential bias.

3.3.2. Mediation Effect Model

Building on the theoretical framework, the effect of digital literacy on green technology adoption can be decomposed into two channels: enhancing technology cognition and reducing information costs. Following the causal stepwise regression procedure and mediation analysis methods outlined by Jiang [55], the following models were specified. First, the baseline model, which estimates the total effect of digital literacy on adoption, is given by Equation (1) in Section 3.3.1.

Next, the mediating variable $Me{d}_{ij}$ (technology cognition or information costs) is regressed on digital literacy and control variables:

$$Me{d}_{ij}={\alpha }_1+{\beta }_{2}D{L}_{ij}+{\gamma }_2{X}_{ij}+{\mu }_j+{\epsilon }_{ij}^{\prime }$$

(2)

Finally, the adoption variable is regressed on both digital literacy and the mediator:

$$At{n}_{ij}={\alpha }_2+{\beta }_{3}D{L}_{ij}+{\beta }_{4}Me{d}_{ij}+{\gamma }_3{X}_{ij}+{\mu }_j+{\epsilon }_{ij}^{″}$$

(3)

In these models, ${\alpha }_1$ and ${\alpha }_2$ are constant terms; ${\beta }_2$, ${\beta }_3$, ${\beta }_4$, ${\gamma }_2$, ${\gamma }_3$ are coefficients to be estimated; and ${\epsilon }_{ij}^{\prime }$ and ${\epsilon }_{ij}^{″}$ denote random error terms.

4. Empirical Results and Discussion

Building on the baseline regression results, this study further examines the impact of digital literacy on farmers’ adoption of green production technologies from four perspectives: endogeneity treatment, robustness checks, mechanism analysis, and heterogeneity analysis. Specifically, an instrumental variable (IV) approach is employed to address potential endogeneity concerns. A set of robustness checks is performed to verify the reliability of the empirical results. Furthermore, the underlying mechanisms are explored from the perspectives of technological cognition and information costs, while heterogeneity analysis is performed across different groups of farmers with distinct characteristics. Through these complementary analyses, this section aims not only to strengthen causal inference but also to reveal the underlying mechanisms by which digital literacy shapes farmers’ adoption of green production practices, as well as to clarify the conditions under which these effects are most pronounced.

4.1. Main Effects Analysis

4.1.1. Baseline Regression Results

The baseline regression results show that, after controlling for individual characteristics, production and management factors, and regional fixed effects, digital literacy has a statistically significant positive effect on farmers’ adoption of green production technologies at the 1% significance level (

Table 4. Ordered Probit Estimates of the Influence of Digital Literacy on Farmers’ Adoption of Green Production Practices.

Variable	Adoption of Green Production Technologies
	Model (1)	Model (2)	Model (3)	Model (4)
Digital Literacy	2.3559 ***
	(0.3233)
Basic Digital Literacy		4.2265 ***
		(0.7334)
Digital Learning Literacy			4.3779 ***
			(0.8053)
Digital Business Literacy				3.6436 ***
				(0.7043)
Gender	−0.2776	−0.2968	−0.2468	−0.2195
	(0.2130)	(0.2043)	(0.2091)	(0.2171)
Age	−0.0052	−0.0068	−0.0062	−0.0071
	(0.0070)	(0.0068)	(0.0071)	(0.0069)
Education Level	−0.0372	−0.0161	−0.0244	−0.0299
	(0.0973)	(0.0988)	(0.0994)	(0.1008)
Health Status	0.0322	0.0621	0.0332	0.0483
	(0.0863)	(0.0855)	(0.0851)	(0.0845)
Village Cadre Identity	0.4131	0.3839	0.4358	0.4323
	(0.3003)	(0.3019)	(0.2826)	(0.3218)
Off-farm Employment Status	−0.2970 **	−0.2750 **	−0.2889 **	−0.2979 **
	(0.1244)	(0.1250)	(0.1238)	(0.1266)
Number of Laborers	0.0361	−0.0028	0.0616	0.0170
	(0.0818)	(0.0808)	(0.0816)	(0.0790)
Operation Scale	0.0018 *	0.0016	0.0018 *	0.0015
	(0.0010)	(0.0010)	(0.0010)	(0.0009)
Number of Plots	0.0337	0.0380	0.0384	0.0606
	(0.0429)	(0.0432)	(0.0437)	(0.0428)
Cooperative Membership	0.0278	0.0825	0.0127	0.0947
	(0.1519)	(0.1534)	(0.1553)	(0.1516)
Participation in Technical Training	0.5859 ***	0.6229 ***	0.6829 ***	0.6326 ***
	(0.1533)	(0.1540)	(0.1559)	(0.1510)
Regional Dummy Variables	Yes	Yes	Yes	Yes
N	422	422	422	422
Prob > chi2	0.0000	0.0000	0.0000	0.0000
Pseudo R2	0.1154	0.0982	0.0996	0.0922

Note: The significance levels indicated by asterisks in this table are: * p < 0.10, ** p < 0.05, *** p < 0.01. Robust standard errors are reported in parentheses.

). Higher levels of farmers’ digital literacy are associated with the adoption of a broader range of green production technologies, providing preliminary support for the hypothesis that digital literacy promotes agricultural green transformation.

Further analysis of the multidimensional structure of digital literacy (Models 2–4) examines the effects of basic digital literacy, digital learning literacy, and digital business literacy separately. All three dimensions are statistically significant at the 1% level, with digital learning literacy showing the largest coefficient (4.3779), followed by digital foundational literacy (4.2265) and digital business literacy (3.6436). These results suggest that while all dimensions contribute to facilitating farmers’ green production decisions, their relative effects differ. This effect may stem from digital literacy’s ability to enhance farmers’ access to information and understanding of technologies, thereby lowering barriers to adoption.

Regarding control variables, off-farm employment consistently exhibits a negative and statistically significant effect on green technology adoption (coefficient ≈ −0.28, significant at the 5% level), indicating that time and labor constraints faced by part-time farmers may limit their adoption behavior. Conversely, participation in technical training shows a robust, positive, and statistically significant effect across all model specifications (coefficients ranging from 0.58 to 0.68 at the 1% significance level), highlighting that training enhances farmers’ technical awareness and their capacity to apply technologies, thereby promoting adoption. Farm size shows a modest positive effect, reaching significance at the 10% level in some specifications (e.g., Models 1 and 3). However, some control variables, including gender, age, education, health status, and cooperative participation, do not exhibit statistically significant effects. While these factors are typically considered important in adoption studies, their limited influence in this context may be due to the homogeneity of our sample, where farmers’ demographic characteristics are relatively uniform. Moreover, the significant role of digital literacy and technical training may have overshadowed the effects of these factors.

Regarding model fit, while the pseudo R² values range from 0.09 to 0.12, which is typical of cross-sectional survey data, we acknowledge that unobserved factors, such as household income or government subsidies, may influence adoption decisions. These factors were not included in the model due to data limitations, but future research could incorporate them for a more comprehensive understanding of adoption behavior. Despite the insignificance of some control variables, the overall model remains robust. The pseudo R² values, along with the statistically significant effects of digital literacy, technical training, off-farm employment, and farm operation scale, suggest that the key determinants of green technology adoption are well captured.

Overall, the models are jointly significant (Prob > chi² = 0.0000), indicating moderate explanatory power. Taken together, the findings demonstrate that digital literacy and its sub-components exert a consistently positive influence on farmers’ adoption of green production technologies.

4.1.2. Addressing Endogeneity

Given the potential bidirectional causality between digital literacy and green technology adoption, as well as potential omitted variable bias, baseline estimates may suffer from endogeneity concerns. To mitigate this issue, an instrumental variable Probit (IV-Probit) model is applied, using township digital governance capacity as an instrument for digital literacy. This variable captures the total number of digital public services provided at the township level, including e-government services, smart Party-building platforms, digital village administration, digital cultural services, and telemedicine systems [35]. The validity of this instrument is grounded in the institutional features of rural governance in China, where digital governance primarily facilitates farmers’ exposure to digital environments through enhanced infrastructure and service accessibility. The diffusion of green agricultural technologies, however, depends predominantly on agricultural extension systems and market-based mechanisms rather than digital governance itself. Consequently, we expect the influence of township digital governance on green technology adoption to be mediated primarily through digital literacy, rather than through a direct channel.

To further reduce potential confounding bias, we control for participation in agricultural technical training and cooperative membership. Under this setting, the effect of township digital governance capacity on green technology adoption is reasonably interpreted as operating through digital literacy, rather than reflecting a direct institutional effect. In estimation, the reference group (Tonghua City) exhibits no variation in the dependent variable, leading to perfect prediction for the regional dummy (Baishan City). This is a mechanical result in the IV-Probit model, and we exclude these observations to avoid separation. Therefore, the final sample consists of 227 farm households from Baishan City and Yanbian Korean Autonomous Prefecture, which represent key ginseng-producing regions in Jilin Province. Importantly, this subsample shows no systematic differences from the full sample in observable characteristics, such as farm size, household head age, and education level, ensuring that it remains broadly representative of ginseng-producing households in the region. However, we acknowledge that the IV estimates identify a Local Average Treatment Effect (LATE) for these core production regions.

The first-stage results (

Table 5. IV-Probit Estimation Results of the Effect of Digital Literacy on Green Production Technology Adoption and Weak-Instrument Robustness Tests.

Variable	Green Production Technology Adoption
	Model (1)	Model (2)	Model (3)	Model (4)
Digital Literacy	7.3009 ***
	(1.8309)
Basic Digital Literacy		11.2267 **
		(5.1424)
Digital Learning Literacy			16.2440 ***
			(3.2890)
Digital Business Literacy				14.5256 ***
				(1.2853)
Control Variables	Yes	Yes	Yes	Yes
Regional Dummy Variables	Yes	Yes	Yes	Yes
N	227	227	227	227
Wald chi2	32.22	25.41	43.94	152.06
Prob > chi2	0.0022	0.0204	0.0000	0.0000
First-stage F-statistic	13.79	22.20	9.34	6.18
AR test chi2(1)			24.80	24.80
p-value of AR test			0.0000	0.0000
AR 95% confidence set			[15.17, 52.08]	[17.53, 76.05]

Notes: (1) Robust standard errors (heteroskedasticity-consistent) are reported in parentheses; (2) ** p < 0.05, *** p < 0.01; (3) The Anderson–Rubin (AR) test and AR confidence sets are based on weak-instrument-robust inference in a linear IV framework; the Wald test is not robust under weak instruments and is therefore not used as the primary basis for inference; (4) The first-stage F-statistics for Models (1)–(2) exceed the conventional rule-of-thumb threshold of 10, indicating no weak-instrument concern in these specifications; hence, AR test results are not separately reported for these models.

) demonstrate that the instrumental variable is strongly correlated with digital literacy in all specifications, satisfying the relevance condition. The first-stage F-statistics in Models (1) and (2) are 13.79 and 22.20, respectively, exceeding the conventional threshold of 10 (Stock and Yogo [56]), which suggests that weak-instrument concerns are minimized in these specifications. In Models (3) and (4), the F-statistics are lower (9.34 and 6.18), indicating potential concerns with weak instruments. To address this, we perform the Anderson–Rubin (AR) test, which is robust to weak instruments and does not rely on instrument strength assumptions. The AR test results consistently show that the causal effect of digital literacy on green technology adoption remains statistically significant across all specifications, with confidence intervals excluding zero.

While some specifications exhibit weaker instruments, the AR tests indicate that the main findings are robust under weak-instrument-robust inference. However, the relatively low F-statistics in Model (4) suggest that the estimated effects for digital business literacy in this specification may be less precisely estimated. We acknowledge that stronger instruments would improve the precision of these estimates and suggest that future research may benefit from exploring alternative instruments to strengthen causal identification.

To further evaluate the plausibility of the exclusion restriction, we conduct a sensitivity analysis following Conley et al. (2012) [57], which allows for the possibility that township digital governance may directly affect green technology adoption. We vary the direct effect (δ) within the range of [0, 0.481], constructing adjusted 2SLS estimates and 95% confidence intervals. The results show that the estimated effects of both digital literacy and digital business literacy remain positive and statistically significant unless δ exceeds approximately 0.10, which corresponds to about 20% of the reduced-form effect. Given that the primary role of township digital governance is to enhance public service delivery and digital infrastructure, while green technology diffusion depends largely on agricultural extension and market incentives, such a large direct effect of township governance is unlikely under the prevailing institutional conditions. Thus, the exclusion restriction remains plausible in this context, and the instrument’s exogeneity is considered credible. For further details, see Figure 2 and Appendix

Table A3. Conley sensitivity analysis: union of confidence intervals.

Model	Endogenous Variable	$\widehat{\widehat{\mathbf{\beta }}}$_IV (δ = 0)	UCI [Lower, Upper]	Robust Region
(3)	Digital learning literacy	25.68	[−13.99, 39.73]	δ < 0.10
(4)	Digital business literacy	31.48	[−13.24, 53.76]	δ < 0.10

Note: $\widehat{\hat{\mathrm{\beta }}}$_IV represents the 2SLS coefficient when the instrument is assumed to be strictly exogenous (δ = 0). Important: The coefficients reported here are from a linear 2SLS model and are not directly comparable in magnitude to the IV-Probit coefficients reported in Table 5 of the main text, due to differences in model specification (linear vs. nonlinear) and scaling. However, both sets of estimates consistently support a positive causal effect. UCI (union of confidence intervals) is computed for δ ranging from 0 to 0.481 (the reduced-form total effect of the instrument on green technology adoption). The robust region indicates the range of δ for which the UCI excludes zero. For both digital learning literacy and digital business literacy, the causal effect remains statistically significant when the assumed direct effect δ does not exceed approximately 20% of the reduced-form total effect (δ < 0.10), supporting the plausibility of the exclusion restriction under reasonable assumptions.

, which present the union of confidence intervals from the sensitivity analysis.

Overall, the IV-Probit results provide consistent evidence of a positive causal relationship between digital literacy and green technology adoption. Although some specifications show weak-instrument concerns, the findings remain robust under weak-instrument-robust inference and sensitivity analysis. These results highlight the critical role of digital literacy in facilitating agricultural green transformation in rural China. Therefore, the exclusion restriction is plausible under the institutional conditions in rural Jilin, and the instrument is considered exogenous.

4.2. Robustness Checks

To further evaluate the robustness of the baseline findings, this research performs a series of tests from three perspectives: variable measurement, model specification, and data treatment. First, regarding variable measurement, an alternative objective weighting method—the coefficient of variation (CV) method—is used to reconstruct the digital literacy index, followed by re-estimation of the regression models. Second, in terms of model specification, an OLS model is used instead of the ordered Probit model to test the sensitivity of the results. Third, continuous variables are winsorized at the 1st and 99th percentiles to mitigate the influence of outliers. The results consistently show that, regardless of the measurement method, model specification, or data treatment, digital literacy remains positively and significantly associated with farmers’ adoption of green production technologies (p < 0.01). Both the direction and significance of the estimates remain largely unchanged, supporting the robustness of the empirical results (

Table 6. Results of Robustness Checks.

Variable	Adoption of Green Production Technologies
	Digital Literacy Measured by Coefficient of Variation	OLS Test	Winsorize Test
Digital Literacy	0.1944 ***	2.5738 ***	2.3559 ***
	(0.0268)	(0.3286)	(0.3233)
Constant		2.0436 ***
		(0.5743)
Control Variables	Yes	Yes	Yes
Regional Dummy Variables	Yes	Yes	Yes
p-value	0.0000	0.0000	0.0000
R2/Pseudo R2	0.1152	0.3284	0.1154

Note: In this table, the significance levels indicated by asterisks are as follows: *** p < 0.01; robust standard errors are reported in parentheses.

).

4.3. Mechanisms

To shed light on the underlying mechanisms linking digital literacy to farmers’ adoption of green production technologies, this research examines two potential channels: the cognitive pathway and the information cost pathway. Methodologically, a stepwise regression approach is first employed by sequentially introducing six potential mediators—attribute cognition, ecological cognition, economic cognition, risk cognition, search cost, and acquisition cost—to examine changes in the coefficient of the key explanatory variable (As shown in

Table 7. Incremental Regression Results.

Variable	Adoption of Green Production Technologies
	Model (1)	Model (2)	Model (3)	Model (4)	Model (5)	Model (6)
Digital Literacy	2.2745 ***	2.2421 ***	2.2784 ***	2.2293 ***	2.5107 ***	2.4851 ***
	(0.3253)	(0.3233)	(0.3266)	(0.3270)	(0.3360)	(0.3346)
Attribute cognition	0.1573 ***
	(0.0536)
Ecological cognition		0.2900 ***
		(0.0825)
Economic cognition			0.2076 **
			(0.0814)
Risk cognition				0.2745 ***
				(0.0760)
Search cost					0.1763 **
					(0.0765)
Acquisition cost						0.1882 **
						(0.0893)
Control Variables	Yes	Yes	Yes	Yes	Yes	Yes
Regional Dummy Variables	Yes	Yes	Yes	Yes	Yes	Yes
N	422	422	422	422	422	422
Pseudo R2	0.1211	0.1250	0.1212	0.1261	0.1411	0.1410

Note: The significance levels indicated by asterisks in this table are: ** p < 0.05, *** p < 0.01. Robust standard errors are reported in parentheses.

). Subsequently, a Bootstrap procedure with 5000 replications is conducted to estimate bias-corrected 95% confidence intervals for the indirect effects, offering a more robust test of the mediation effects (

Table 8. Bootstrap Mediation Effect Results.

Mediator	Indirect Effect	Bias-Corrected 95% CI	Sobel Z	Share of Mediation Effect
Attribute cognition	0.049 *	[0.010, 0.093]	2.452 **	22.3%
Ecological Cognition	0.071 **	[0.021, 0.127]	2.537 **	13.8%
Economic Cognition	0.059 *	[0.008, 0.116]	2.270 **	18.6%
Risk Cognition	0.069 ***	[0.019, 0.122]	2.694 ***	15.6%
Search Cost	−0.046	[−0.112, 0.009]	−1.542	—
Acquisition Cost	−0.021	[−0.084, 0.038]	−0.707	—

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Bootstrap replications are set to 5000. The proportion of the mediating effect (indirect effect/total effect × 100%) is reported only when the Bootstrap test indicates statistical significance. Sobel test results are provided as a supplementary robustness check.

presents the results).

The results indicate that digital literacy exerts a positive and statistically significant influence on the adoption of green technologies, primarily operating through the cognitive pathway. After introducing the various cognitive variables, the estimated coefficient of digital literacy decreases to varying magnitudes, and the Bootstrap results confirm that the corresponding indirect effects are statistically significant. This suggests that attribute cognition, ecological cognition, economic cognition, and risk cognition all serve as significant partial mediators. In other words, digital literacy appears to promote green technology adoption by enhancing farmers’ understanding of technological attributes, ecological benefits, economic returns, and risk controllability.

In contrast, although information cost variables are significantly associated with technology adoption in the stepwise regression results, the Bootstrap analysis shows that the bias-corrected 95% confidence intervals for both search cost and acquisition cost include zero, suggesting that their indirect effects are not statistically significant. This implies that relying solely on coefficient changes in stepwise regression may lead to misleading conclusions regarding mediation effects, whereas the Bootstrap approach provides more robust statistical inference [58]. Considering the relatively well-developed information infrastructure in Jilin Province, where the internet broadband penetration rate is 84.3% [59], it is likely that traditional information channels, such as government promotional services, agricultural extension services, and local media, already meet farmers’ needs. This widespread access to information likely reduces the marginal role of digital literacy in lowering information costs. In comparison, the cognitive empowerment effect of digital literacy appears to be more critical.

Overall, the evidence indicates that digital literacy appears to facilitate farmers’ adoption of green production technologies primarily via a cognitive enhancement mechanism, rather than through information cost reduction. However, due to the cross-sectional nature of the data and reliance on self-reported survey responses, these findings should be interpreted with caution regarding causal inference.

4.4. Heterogeneity Analysis

To further investigate group-level heterogeneity in the relationship between digital literacy and farmers’ adoption of green production technologies, the present study conducts subgroup analysis along three dimensions: generational characteristics, farm size, and cooperative participation. The dependent variable, measured as the count of green technologies adopted (0–5), was originally estimated using an ordered Probit model. However, subgrouping reduces sample sizes and may lead to limited within-group variation in certain variables (e.g., cooperative participation becoming constant), potentially causing convergence issues or instability in maximum likelihood estimation. Following Angrist and Pischke (2009) and Wooldridge (2010) [60,61], a Linear Probability Model (LPM) is adopted for subgroup regressions, which allows coefficients to be directly interpreted as marginal effects, facilitates cross-group comparisons, and provides more stable estimates in smaller samples. When combined with Bootstrap standard errors and Fisher-type combined tests, it enables valid inference on coefficient differences across groups. Appendix A confirms that the coefficient signs from the LPM are consistent with the marginal effects from the ordered Probit model, supporting the robustness of this approach. The LPM results are reported in

Table 9. Heterogeneity Analysis: Group Differences in the Influence of Digital Literacy on Farmers’ Green Technology Adoption among Ginseng Farmers (Linear Probability Model).

Variable	Generational Difference		Operation Scale		Cooperative Participation
	Older Generation	New Generation	Large-Scale	Small-Scale	Member	Non-Member
Digital Literacy	2.552 ***	1.510 ***	2.448 ***	1.958 ***	2.138 ***	2.115 ***
	(0.414)	(0.561)	(0.427)	(0.493)	(0.713)	(0.378)
Control Variables	Yes	Yes	Yes	Yes	Yes	Yes
Regional Dummy Variables	Yes	Yes	Yes	Yes	Yes	Yes
N	276	146	218	204	89	333
R-squared	0.241	0.249	0.353	0.200	0.323	0.271
Intergroup Coefficient Difference Test	Older & New Generation		Large & Small Scale		Member & Non-member
Coefficient Difference	1.042		0.490		0.023
p-value of Fisher’s Test	[0.068]		[0.244]		[0.492]

Notes: *** indicates statistical significance at the 1% levels. Robust standard errors are presented in parentheses. Values in square brackets present Fisher test p-values for coefficient differences across subgroups.

.

The empirical evidence consistently shows a positive and statistically significant effect of digital literacy on green technology adoption across all subgroups. However, in terms of generational differences, the Fisher test for coefficient differences shows a p-value of 0.068, which indicates weak statistical significance (p < 0.10), implying that the evidence for generational differences is limited. Similarly, there is no significant difference in the impact of digital literacy between large-scale and small-scale farmers (p = 0.244), nor between cooperative participants and non-participants (p = 0.492).

This suggests that while generational characteristics, farm size, and cooperative participation exhibit some variation, these factors do not significantly moderate the relationship between digital literacy and green technology adoption. As such, Hypothesis 4 is not supported, indicating that the impact of digital literacy on green technology adoption does not vary systematically across different farmer groups.

Potential Explanations for the Absence of Heterogeneity:

• Regional Concentration: The sample is drawn from the core ginseng-producing areas of Jilin Province (Tonghua, Baishan, and Yanbian), where natural, economic, and institutional conditions are highly similar. These regions exhibit homogeneous ecological environments, standardized technical extension systems, and similar market structures for ginseng products. As a result, regional concentration reduces variability across samples and limits the moderating role of individual factors such as age, farm size, and organizational participation.
• Industry Specialization: Ginseng cultivation is highly specialized, requiring significant technical expertise, long production cycles, and substantial investment risks. Unlike staple crop farming, ginseng cultivation is more reliant on systematic knowledge and has less variability in decision-making due to the technical demands. As digital literacy improves, farmers’ decision-making becomes more aligned, resulting in reduced observable heterogeneity.
• Digital Inclusiveness: Digital literacy serves as a universal capability that improves access to information, cutting across age groups, farm sizes, and cooperative participation. Digital tools have narrowed the generational divide, enhanced access for both smallholders and large-scale farmers, and complemented cooperative information channels, leading to a general empowerment effect that weakens differences between groups.

In conclusion, the lack of significant heterogeneity in this study can be attributed to these factors, suggesting that digital literacy plays a broadly applicable role in promoting green technology adoption, regardless of individual characteristics or production conditions.

5. Conclusions and Implications

5.1. Main Findings

Based on survey data from 422 ginseng farmers in Jilin Province, this study systematically investigates the impact of digital literacy on farmers’ adoption of green production technologies and its underlying mechanisms, employing an ordered Probit model, instrumental variable (IV) approach, mediation analysis, and heterogeneity tests. The reliability and validity tests of the digital literacy scale (Appendix Table A1) indicate that the scale demonstrates strong reliability (Cronbach’s α = 0.891) and structural validity (cumulative variance explained = 71.29%), providing a solid foundation for the subsequent analysis. The key findings are as follows:

• First, digital literacy significantly encourages farmers to adopt green production technologies. After accounting for individual characteristics, production and management factors, and regional fixed effects, digital literacy has a significant positive effect on green technology adoption at the 1% significance level. The causal effect remains robust after addressing potential endogeneity using an instrumental variable approach (IV-Probit estimates are significant in most specifications; sensitivity analysis in Appendix Table A3 further supports the robustness of the causal inference). Examining the multidimensional structure of digital literacy, all three dimensions—digital learning literacy (coefficient = 4.38), Basic Digital Literacy (coefficient = 4.23), and digital business literacy (coefficient = 3.64)—show significant positive effects, with digital learning literacy having the strongest impact. This result highlights that digital literacy, as a form of cognitive human capital, plays a pivotal role in driving agricultural green transformation.
• Second, digital literacy primarily operates through the “cognitive empowerment” pathway rather than the “information cost reduction” pathway. Mediation analysis reveals that the four dimensions of technological cognition—technological attributes cognition, ecological cognition, economic cognition, and risk cognition—serve as significant mediators, with all Bootstrap 95% confidence intervals excluding zero (see Table 8 in the main text). Appendix Table A2 shows that the correlations between these four dimensions range from 0.279 to 0.634, indicating good discriminant validity and making them suitable as independent mediators. The share of the mediation effect varies between 13.8% and 22.3%. In contrast, the mediation effects of information search costs and information acquisition costs were not statistically significant (Bootstrap confidence intervals included zero). This finding suggests that merely lowering the information acquisition threshold is insufficient to drive farmers’ adoption of green technologies. More importantly, digital literacy empowers farmers by enhancing their understanding of technological principles, ecological benefits, economic returns, and risk management. Notably, traditional stepwise regression showed significant results for the information cost pathway, but Bootstrap testing did not confirm these findings, indicating that relying solely on coefficient changes may lead to erroneous conclusions about mediation effects.
• Third, the promoting effect of digital literacy does not exhibit significant heterogeneity across different farmer groups. Heterogeneity analysis based on generational differences, farm scale, and cooperative participation shows that digital literacy has a significant positive effect in all subgroups. However, the differences in coefficients across groups were not statistically significant (Fisher test p-values of 0.068, 0.244, and 0.492, respectively). Appendix

  Table A4. Heterogeneity Analysis: Group Differences in the Impact of Digital Literacy on Green Production Technology Adoption among Ginseng Farmers (Ordered Probit Model).

  Variable	Generational Difference		Operation Scale		Cooperative Participation
  	Older Generation	New Generation	Large-Scale	Small-Scale	Member	Non-Member
  Digital Literacy	2.4571 ***	2.0122 ***	2.7222 ***	2.0776 ***	2.6187 ***	2.1940 ***
  	(0.3945)	(0.5840)	(0.4936)	(0.4573)	(0.7310)	(0.3705)
  Control Variables	Yes	Yes	Yes	Yes	Yes	Yes
  N	276	146	218	204	89	333
  Pseudo R2	0.1195	0.1284	0.1729	0.0966	0.1593	0.1268

  Note: *** indicates statistical significance at the 1% levels. Robust standard errors are reported in parentheses. The coefficient signs and relative magnitudes across subgroups are consistent with those from the Linear Probability Model (LPM) reported in Table 9 of the main text (see Section 4.4 for detailed discussion). Pseudo R² values are lower in subgroup analyses due to reduced sample sizes, which is expected and does not compromise the validity of the estimates. For formal tests of coefficient differences across subgroups (Fisher’s combined test), refer to Table 9 in the main text; all p-values exceed conventional significance levels except for the generational difference (p = 0.068, marginally significant at α = 0.10).

  reports the ordered Probit model estimates for subgroup analysis, and the direction and magnitude of the coefficients are highly consistent with the linear probability model estimates in Table 9, confirming the robustness of the conclusion across model specifications. This “no heterogeneity” result can be explained from three perspectives: first, the sample is concentrated in the core ginseng-producing areas of Jilin Province (Tonghua, Baishan, and Yanbian), where natural, economic, and institutional conditions are highly homogeneous; second, ginseng cultivation is highly specialized, with a relatively uniform technological approach and limited decision-making space; third, digital literacy, as a universal capability, is inclusive, bridging generational gaps and empowering both smallholders and large-scale farmers. This finding suggests that the promoting effect of digital literacy is broadly applicable and does not vary significantly across different farmer characteristics or production conditions.

5.2. Limitations

This study has several limitations that should be considered when interpreting the findings:

• Cross-sectional Data: The study uses cross-sectional data, limiting the ability to draw causal inferences. While an instrumental variable approach was employed to address potential endogeneity, future research using longitudinal panel data could better capture the long-term effects of digital literacy on technology adoption.
• Self-reported Measures: The reliance on self-reported data introduces the potential for response bias. Future studies could incorporate objective measures (e.g., digital proficiency tests or third-party evaluations) to ensure more reliable data.
• Residual Endogeneity: Despite efforts to address endogeneity, unobserved factors, such as household income, may still influence both digital literacy and adoption. Additional confounding variables should be considered in future research.
• Limited Generalizability: This study focuses on ginseng farmers in Jilin Province, which limits its generalizability. Future research should examine diverse agricultural sectors or regions to assess the broader applicability of the findings.

These limitations should be kept in mind when interpreting the study’s conclusions. Further research is necessary to enhance the robustness and generalizability of the results.

5.3. Policy Implications

The results offer several important policy implications to promote sustainable agricultural practices, particularly in specialty crop sectors such as ginseng.

• Investing in Digital Literacy: Policymakers should prioritize the enhancement of farmers’ digital literacy as a critical pathway for adopting green production technologies. Strengthening rural digital infrastructure and offering targeted digital literacy training programs are essential to improve farmers’ skills in information gathering, online learning, and the practical application of digital tools. This would facilitate farmers’ access to green technologies and enhance their decision-making abilities.
• Focus on Cognitive Empowerment: Since digital literacy influences adoption through cognitive enhancement, policy programs should focus on improving farmers’ technological understanding. Training content should be application-oriented, with visualized and scenario-based materials that highlight the technical, ecological, and economic aspects of green technologies. This will help farmers make informed decisions about which technologies to adopt and how to evaluate their potential benefits.
• Inclusive Policies for All Farmers: Digital literacy shows a relatively uniform impact across age groups, farm sizes, and participation in cooperatives. Given this, digital literacy interventions should target the wider farming population while addressing potential gaps, especially for smallholders and older farmers. In particular, integrating digital tools with cooperative structures could amplify the effectiveness of the interventions, providing a support network for knowledge dissemination.
• Prioritize Cognitive Development Over Information Cost Reduction: In regions like Jilin, where information infrastructure is relatively well-developed, focusing on reducing information costs alone may not yield significant benefits. Policies should therefore emphasize enhancing farmers’ ability to critically evaluate digital information and make better technology adoption decisions. Cognitive empowerment—understanding the underlying principles, benefits, and risks—should be the central focus of intervention programs aimed at promoting sustainable agricultural practices.

These recommendations aim to leverage digital literacy as a tool for broadening access to green technologies and fostering agricultural transformation in rural China, contributing to long-term sustainability in agricultural practices.

5.4. Research Outlook

Enhancing farmers’ digital literacy is crucial for promoting green production technologies and advancing both agricultural development and rural ecological revitalization. This study provides valuable insights into the role of digital literacy in agricultural transformation. However, acknowledging the limitations discussed in Section 5.2, several avenues for future research could deepen our understanding of the mechanisms through which digital literacy influences technology adoption.

First, longitudinal studies using panel data would be essential for examining the long-term effects of digital literacy on technology adoption. The cross-sectional nature of this study limits our understanding of how digital literacy influences adoption behavior over time. Longitudinal data would allow for capturing dynamic changes and causal relationships across different stages of adoption.

Second, to address the self-reported measures in this study, future research should incorporate objective assessments of digital literacy and technology adoption (e.g., digital proficiency tests or third-party evaluations). This would mitigate response bias and provide more accurate insights into farmers’ actual digital capabilities and adoption behaviors.

Third, future studies should account for residual endogeneity by incorporating additional confounding variables that may influence both digital literacy and technology adoption, such as access to agricultural support services or household income. This would improve the precision of causal inferences.

Finally, generalizability could be enhanced by expanding the research to include a wider range of agricultural sectors and regions. This study focused on ginseng farmers in Jilin Province, which may limit the applicability of the findings. Exploring other agricultural contexts, both within China and globally, would help assess the broader relevance of the results. Additionally, integrating field-based performance evaluations (e.g., farm productivity, sustainability metrics) with digital literacy measures would offer a more comprehensive view of the adoption process [62].

By addressing these limitations, future research could provide deeper insights into how digital literacy drives the adoption and long-term impact of green technologies, supporting sustainable agricultural transformation and rural development across diverse contexts. Such studies could also inform targeted policies to enhance the adoption of green technologies in agriculture.