How agricultural cooperatives facilitate sustainable technology adoption in China: resource access or peer influence?

The trade-off between improving human welfare and protecting the environment is a significant challenge to sustainable rural development (Kanter et al., 2018; Tudi et al., 2021). For decades, farming communities in developing economies have experienced notable environmental degradation due to agricultural practices (Gilbert, 2012; Mhazo et al., 2016; Rani et al., 2021). Addressing these issues is complicated by the uneven distribution of environmental benefits and costs, and agent-level innovations are needed to internalize social costs (Buchanan and Faith, 1981). Consequently, there has been a strong emphasis on the diffusion of sustainable agricultural technologies (Lee, 2005), which can be defined as innovations and practices increasing productivity while minimizing environmental harm, conserving natural resources, strengthening ecological resilience, and supporting long-term socio-economic viability of farming system (Arhin et al., 2023). Recent studies systematically exploring potential drivers and synthesizing relevant policy information (Takahashi et al., 2020; Rosário et al., 2022).

Agricultural cooperatives, or the groups where individual farmers pool resources to collectively achieve their production or marketing goals, are increasingly found to play a prominent role in sustainable agricultural technology adoption (see Candemir et al., 2021, for a recent literature review). Existing literature generally indicates a positive impact on the adoption of agricultural chemicals (Abebaw and Haile, 2013), improved seeds (Manda et al., 2020), integrated pest management practices (Ma and Abdulai, 2019) and sustainable practices (Liang et al., 2023). Moreover, cooperative membership is found to be associated with higher agricultural income resulting from technology adoption (Yang et al., 2021; Khan et al., 2022). It may also complement access to agricultural extension services and improve household welfare in multiple dimensions (Wossen et al., 2017).

The diffusion of innovations is a complicated process that involves heterogeneous farmers, information asymmetries, learning costs and social influences. Here, agricultural cooperatives may play a pivotal role by reducing material and informational barriers that often constrain adoption (Rogers, 2003). Through mechanisms such as input support (Blekking et al., 2021), credit provision (Nowfal et al., 2025), farmer field schools (Davis et al., 2012), collective training programs (Zhang et al., 2020), peer-to-peer learning (Munshi, 2004), cooperatives facilitate experiential learning and adoption of new technologies. These functions position agricultural cooperatives as critical institutional platforms that not only disseminate innovations but also support the adoption of sustainable agricultural technologies.

Recent studies have identified overall effects of agricultural cooperatives on farmers’ adoption behaviour (Luo et al., 2022; Wei et al., 2022; Chen et al., 2025). Certain pathways are further explored. Nwankwo et al. (2009) suggest that cooperatives play a crucial role in disseminating technological information, and members generally manifest higher willingness to adopt, given the trust they place in the platform. Agricultural cooperatives may also stimulate technology adoption through the provision of productive resources as well as technical support or training (Naziri et al., 2014; Zhang et al., 2023). While these studies offer valuable insights, a more systematic understanding of the mechanisms is needed. This necessitates a consolidated framework and comprehensive data that could jointly facilitate rigorous empirical testing. Moreover, the ongoing rural digitalization in China could substantially affect these mechanisms. For instance, information flow (technical support, training, etc.) is eased by alternative pathways (mobile apps, social media and online platforms). Likewise, by enabling online ordering, real-time price comparisons and coordinated distribution through cooperatives, rural digitalization enhances farmers’ access to agricultural inputs while reinforcing peer learning and leadership influence. These profound changes deserve scholarly investigation. Findings on possible mediating mechanisms could not only enhance the external validity of existing literature but also inform policy making, promoting sustainable agricultural technology adoption through cooperative initiatives.

Agricultural cooperatives in China have received governmental endorsement since the 1990s, with a State Council directive first issued to support voluntary farmers’ organizations and a piloting program of agricultural cooperatives launched in 2002 (Huang and Liang, 2018). Government-led and entrepreneur-led cooperatives are the leading type of agriculture cooperatives, where the former focus more on agricultural production (e.g. input and information provision) while the latter facilitate processing and marketing of agricultural products. In government-led cooperatives, ownership, control, and benefit rights are more equally distributed among members, while in entrepreneur-led cooperatives, core members dominate capital contribution, control rights and dividends distribution (Liu et al., 2024). Agricultural cooperatives work extensively on different sectors of the value chain, including production, processing and marketing (Qu et al., 2023). In China, they are best understood as hybrid organizations. While the Farmers’ Professional Cooperatives Law formally emphasizes voluntariness, democratic governance and member control (National Farmer Professional Cooperative Law, Standing Committee of the National People’s Congress, 2007), in practice, many cooperatives are initiated, incentivized or guided by local governments through policy targets, subsidies and technical support. Regardless of the organization structure, governments at various levels have proactively channelled significant resources to farmer cooperatives, providing staff training, resource access, marketing assistance and financial underpinning through legislation. Chinese cooperatives differ significantly from those in other Asian economies, such as the Japan Agricultural Cooperatives and South Korea’s National Agricultural Cooperative Federation. Both of these are highly centralized, nationwide, multi-tier cooperative network that integrates production support, agricultural extension, input supply, credit, insurance and marketing. In contrast, Chinese agricultural cooperatives are highly localized. Over 2.1 million agricultural cooperatives have been officially registered as of 2018 (Xinhua Net, 2019). The rapid growth of these cooperatives amid ongoing agricultural transformation therefore makes China an ideal site for study.

As the world’s largest producer, importer and consumer of food, China has experienced rather uneven regional development. For several decades following the economic reform and market opening, significant discrepancies have emerged between coastal and inland regions, with real per capita GDP and its growth rate being lowest in Western China (Fleisher et al., 2010). Agricultural output constitutes a much higher share of GDP in this region, while it faces challenges in aspects like agricultural productivity (Xu, 2017), agricultural export (Zheng et al., 2021), modernization (Wang and Kuang, 2023) and sustainability (Zhou and Wen, 2023). These disparities collectively underscore the importance of focusing on Western China.

Using a large rural household survey, this study aims to identify whether agricultural cooperatives stimulate sustainable agricultural technology adoption in Western China’s horticultural sector, and to explore the possible mechanisms that make this happen. Our contribution is threefold. First, we develop a conceptual model that consolidates key mediating mechanisms linking cooperative membership to sustainable technology adoption, which provides an adjustable analytical framework for similar investigations. Second, we extend existing literature that emphasizes the roles of agricultural cooperatives in resource and information provision to further incorporate two complementary dimensions of peer effects, thereby presenting a more comprehensive picture of mediation processes. Third, our empirical identification of related mechanisms directly reveals their respective effectiveness and potential interconnections, enabling us to offer more precise policy recommendations.

Olson’s (1965) collective action theory emphasizes the difficulty of cooperation in large groups without selective incentives, while Ostrom (1990) shows that communities can overcome these problems through locally crafted rules, monitoring, and trust-based governance. Agricultural cooperatives provide a governance context that allows members to coordinate to achieve shared goals despite incentives to free ride. They encompass various autonomous associations involved in input provision, production assistance, postharvest processing, wholesaling and retailing. Given the diverse activities along the chain, cooperatives can assume very different roles. Existing studies have identified several benefits of agricultural cooperatives in China, including better access to productive resources and credit (Zhang et al., 2023), increased cropland utilization (Ma and Zhu, 2020), encouragement of safe production practices (Li et al., 2021), reduced chemical overuse (Liu and Wu, 2022) and provision of technical support and training (Liu et al., 2022).

These benefits generally fall into two broad categories: improved access to productive resource access and enhanced access to information. Both are tangible benefits that can promote agricultural technology adoption (Zhang et al., 2023). However, there are additional benefits that are often overlooked due to their intangibility. Rogers (2003) highlights the importance of social networks, opinion leaders and institutions in shaping information flows and reducing uncertainty in technology diffusion. In realistic ways, agricultural cooperatives may foster peer effects among members, which are common in various types of networks (Bramoullé et al., 2020). Social learning as a cognitive process allows individuals to acquire knowledge and skills by observing others and the outcomes of their actions, which may stimulate later behavioural change (Bandura, 1977). In cooperatives, peer learning occurs because farmers are embedded in networks of repeated interactions, where network structure and position shape access to information and influence (Granovetter, 1973; Wasserman, 1994). Frequent interactions through the ties among cooperative members facilitate observational learning and imitation, while trusted ties enhance the credibility of peers’ experiences with new technologies. Peer learning about technologies and practices is frequently recognized as a driver of adoption (Conley and Udry, 2010; Zhou et al., 2020; Xu et al., 2022), though it is rarely explored in the context of agricultural cooperatives.

In rural China, people typically strive to maintain a positive self-image within their networks and often engage in face-saving behaviours among peers (Kwong and Cheung, 2003). Consequently, many farmers would be sensitive to how they look like to others, especially when they are cooperative members and thus perceive themselves as role models. In this way, they may be more willing to adopt sustainable technologies. We refer to this phenomenon as self-perceived leadership. It motivates them to genuinely align their behaviour with socially valued norms, such as adopting sustainable technologies to maintain a positive self-image and influence peers. According to social identity theory (Tajfel and Turner, 1979), individuals derive part of their self-concept from their social group membership and are motivated to act in ways that maintain a positive and distinctive group identity. Here, farmers who strongly identify with the cooperative may internalize its norms, such as sustainability and innovation as part of “who they are.” When individuals further see themselves as representative or exemplary members of the group, they perceive a leadership role, even in the absence of formal authority. Self-perceived leadership differs from social desirability bias, which describes a tendency to misreport one’s real behaviour to please others (Nederhof, 1985; Krumpal, 2013). The former shapes actual adoption choices rather than merely affecting how behaviours are reported, and it operates through self-identity and role expectations rather than external pressure alone. Importantly, self-perceived leadership may coevolve with peer learning, emerging from successful learning experiences and, in turn, motivating further learning. Nevertheless, the distinct characteristics of these processes justify treating them as separate mechanisms.

As the current analysis focuses on sustainable technology adoption, primarily occurring at the front end of the value chain, relevant cooperatives tend to emphasize agricultural supply procurement and production rather than marketing agricultural products. It is therefore hypothesized that agricultural corporate membership can stimulate sustainable technology adoption in four different ways, two through improved resource access, including both physical productive resources and information, while the other two capture peer influence, namely peer learning and self-perceived leadership. Peer influence can occur bidirectionally: peer learning captures influence from others, while self-perceived leadership measures expected influence on others.

These mediation relationships are visually presented in Figure 1. It is important to note that the mechanisms may vary depending on the nature of agricultural cooperatives and with digital technology penetration, necessitating case-by-case investigation. Specifically, the traditional centralized intermediary roles of agricultural cooperatives, such as pooling resources and disseminating information, maybe partly substituted by digital technologies. These advancements may enable farmers to access inputs, credit, technical advice and market information directly, often in real time and at lower transaction costs. This reduces farmers’ dependence on cooperatives as gatekeepers of information and resources. Also, while it is intuitively hypothesized that cooperative membership affects peer effects, reverse causality is possible as actively learning farmers and/or those who perceive themselves as leaders could be more likely to join cooperatives. Hence, our results should be cautiously treated as suggestive rather than definitive, and the identified effects as correlational rather than causal.

These hypothesized relationships are empirically tested using regression modelling. To first test the overall effect of cooperative membership on technology adoption, a baseline model is estimated:

where adoption decision is predicted by cooperative membership (⁠$CoopMember$⁠) and a series of conventional controls with coefficient vectors $α1$ and $α2$ capturing their respective effects. $α0$ is the constant and $ϵ$ is the error term. Estimate of $α1$ is of our primary interest, which reveals the overall effect hypothesized.

Then, mediation analysis is performed to test each of the four hypothesized mechanisms how cooperative membership affects sustainable agricultural technology adoption. The regression is further specified in a way similar to Cutler and Lleras-Muney (2010), where individual mediators are incorporated:

where $β3$ captures the mediation effect; $β1$ and $β2$ are respectively coefficients of cooperative membership and conventional controls; $β0$ is the constant and $u$ is the error term. The mediators are functions of cooperative membership, and thus the impacts of cooperative membership on sustainable agricultural technology adoption can be obtained as: $∂Adoption/∂CoopMember=β1+β3(∂Mediator/∂CoopMember)$⁠. Here, $β1$ is the direct effect while $β3(∂Mediator/∂CoopMember)$ captures the indirect/mediation effect.

To further evaluate the role of possible mechanisms, $Mediator$ is specified as a linear function:

where $δ0$⁠, $δ1$ and $δ2$ are coefficients and $v$ is random error that is unrelated to cooperative membership. The total impacts of cooperative membership on sustainable agricultural technology adoption can then be rewritten as $β1+β3δ1$⁠. By estimating the reduced form Eq. (1), it can be shown that $α1ˆ=β1ˆ+β3ˆδ1ˆ$⁠, where the mediation effects $β3ˆδ1ˆ=α1ˆ−β1ˆ$⁠. The explanatory power of mediation effects in proportion terms is therefore $(α1ˆ−β1ˆ)/α1ˆ=1−β1ˆ/α1ˆ$⁠.

We use a cross-sectional rural household survey from Western China in 2023. Farmers cultivating and generating more than half of their household income from annual horticultural crops are the focus of the survey. Horticultural crops have the second-highest cultivation area (after cereals) and contribute nearly 80% of the value added in crop production in China (China Academy of Agricultural Sciences, 2021). It is therefore policy meaningful to investigate this substantial sector.

The survey encompasses three provinces, including Shaanxi, Sichuan and Yunnan, which are selected purposefully to represent the mountainous landscape in Western China (49.8% of Shaanxi, 79.5% of Sichuan and 88.6% of Yunnan are mountains). Within each province, stratified random sampling was implemented, with three counties selected and five villages chosen from each county. In each village, face-to-face interviews were attempted with 30 randomly selected farmers whose primary income source was horticultural cultivation. Among the 1,350 total attempts, 1,247 provided complete information (effective response rate 92.4%).

Four sustainable technologies are considered, including conservation tillage (including no tillage, shallow tillage, contour tillage and basin tillage), crop rotation, organic fertilizer and drip irrigation. Conservation tillage, crop rotation and organic fertilizer have been widely adopted in China for decades (He et al., 2010; Niu and Ju, 2017; Jia et al., 2019; Zhao et al., 2020), which are labour-intensive technologies. Drip irrigation was also introduced in China decades ago, yet its adoption rate remains low given its capital-intensive nature (Yang et al., 2023). The adoption of each technology is recorded dichotomously (1 indicating adoption and 0 for non-adoption).

Below, covariates include measures of labour and capital availability, measured by the number of agricultural labourers in the household and annual consumption expenditure per capita. Additionally, whether the farmer had off-farm employment in the past year is controlled for, which could substitute for agricultural production and affect technology adoption (Huang et al., 2020). Various socio-economic factors are incorporated based on literature findings, including cultivate area (Hu et al., 2022), land tenure (Deininger and Jin, 2006) and village cadre status (Zhang et al., 2020). Furthermore, household financial status is controlled for using a binary indicator reflecting whether the household had unmet credit needs in agricultural production in the past year (Zeng et al., 2018). Terrain (hilly or plain) and transportation conditions (whether the village is located on a main road) are also included. Conventional demographic measures are incorporated.

Table 1 presents descriptive statistics of the key variables. Four binary mediating variables are considered according to the hypothesized mechanisms outlined in Figure 1. These are constructed by asking respondents whether they have access to productive resources (e.g. physical inputs, loans) and technological information (e.g. training, contacts), as well as whether they believe it is possible to learn about technologies from adopters (e.g. relatives, friends, neighbours) and whether they think others would pay attention to their adoption behaviour and possibly follow suit. Adoption rates are higher among cooperative members (Table 1). Also, cooperative members and non-members differ systematically across most variables, with the former group being generally advantageous. This is intuitive given the likely self-selection into cooperative membership, which is possibly driven by various socioeconomic and demographic features of the household.

The baseline model in Eq. (1) is first estimated. Despite the binary outcome, a linear probability model estimated with ordinary least squares (OLS) is suitable as it does not rely on distributional assumption, and is more robust against model misspecification (Angrist, 2001; Lewbel et al., 2012). Empirically, the difference in terms of marginal effects is usually indistinguishable between the standard logit/probit model and the linear probability model (Angrist and Pischke, 2009).

Results are presented in Table 2. All four models are appropriately identified according to their F-statistics. Cooperative membership is positively correlated with adoption in all cases. Coefficient magnitudes range from 0.006 to 0.040, which are statistically significant. Notably, the coefficients of conservation tillage and drip irrigation are greater than those of crop rotation and organic fertilizer. This is interesting because conservation tillage and drip irrigation are more complex technologies that require greater practical learning. Hence, such a difference may stem from the increased learning demands.

Among the covariates, perceived land tenure is positively correlated with adoption. This is consistent with literature findings (e.g. Deininger and Jin, 2006). Financial constraints, as captured by a binary indicator of unmet credit need, hinder technology adoption, again reinforcing prior knowledge (Xu and Findlay, 2019). Moreover, education of household head is positively correlated with adoption, yet the coefficient magnitude is generally too small and the statistical significance is sometimes marginal. Apart from these, being a village cadre is positively associated with the adoption of conservation tillage and drip irrigation. Household wealth status is positively associated with capital-intensive drip irrigation.

While these results are generally intuitive, they demand further validation because cooperative membership is a farmer’s choice, which might introduce selection bias and generate inconsistency. We therefore turn to falsification tests to minimize the probability of type I error. Instrumental variable regression is employed to serve this purpose. Specifically, two instruments are adopted from the survey: the distance from the farmer’s house to the nearest agricultural cooperative office (in kilometres), and whether the farmer is aware of cooperative establishment in the area in the past five years. In rural China, cooperative establishment is largely driven by policy implementation in a top-down manner rather than individual farmer initiative and is usually promoted by local policies, campaigns, villagers’ forums and/or more recently digital platforms. Also, given the densely populated landscape (usually a few thousand people in a village), most villagers are extensively exposed to neighbours and peer villagers. Therefore, the role of farmers’ social network differences or variance in motivation should be minimal. For the same reason, extension activities and input promotion are primarily organized through cooperative channels rather than independently delivered in costly ways. Hence, variation in extension service intensity should be minimal. Both instruments should therefore be correlated with cooperative membership, while they are unlikely to affect the farmer’s technology adoption decision except through change in cooperative membership. Therefore, although the exclusion restriction of the instruments can never be empirically tested, it should be intuitively satisfied.

Two-stage least squares (2SLS) estimation is implemented for all four technologies, applying one instrument at a time. All models are properly identified, given the significance of the first-stage and overall F-statistics. Key coefficient estimates are reported in Table 3. All impact estimates are significant. In terms of the magnitudes, the instrumental variable estimates, while reasonably similar to baseline OLS coefficients, are slightly larger in most cases. While these may not strongly suggest negative selection, they collectively suggest that our baseline results are rather conservative.

To further substantiate these results, we adopt a heteroskedasticity-based instrument and re-estimate the regression model across all four sustainable agricultural technologies (Lewbel, 2012). Lewbel’s instrument provides a robustness and identification-enhancing strategy. It is constructed using existing regressors, which is useful either when no external instrument is available or for testing the validity of external instruments (Baum and Lewbel, 2019). This approach constructs internal instruments from exogenous covariates under the assumptions that these covariates are uncorrelated with both the structural error term and the first-stage error, that the first-stage error exhibits heteroskedasticity conditional on the exogenous variables. Identification relies on systematic variation in the variance of the first-stage error with respect to the exogenous covariates, which ensures the relevance of the generated instruments. These identification assumptions can be satisfied in the case of a binary endogenous regressor (Lewbel, 2018), making it appropriate in our setting.

These estimates are reported in the third row of Table 3. The constructed instrument likewise has passed the weak identification test in all four technology-specific regressions. Regarding the impact estimates per se, they are again very close to our baseline results as well as the previous 2SLS estimates, lending further credence to the above findings.

Two more procedures are further exercised to test the robustness of these results. First, we remove certain variables that could be potentially endogenous and re-estimate the model with OLS. These variables include village cadre, land tenure and off-farm employment. The impact estimates of agricultural cooperative membership on sustainable technology adoption should stay the same were there little omitted variable bias. Second, we exercise logit estimation of the baseline model and obtain the marginal effects. These results are reported in the fourth and fifth rows of Table 3. The estimates are again reasonably close to our above OLS and 2SLS estimates. The high consistency between OLS and logit marginal effects is earlier observed by Angrist and Pischke (2009). Therefore, it is safe to conclude that related bias should not compromise the validity of our baseline results.

These exercises affirm the overall impact of cooperative membership on adoption. Although a few estimates lose some statistical significance across these tests, they are still consistently significant at 10% level and backed up by highly significant estimates obtained through alternative procedures.

In Figure 1, there are four hypothesized mediating mechanisms. Therefore, it is necessary to simultaneously incorporate these mediators into the mediation model and explore their roles. Empirical approaches to estimate mediation model with multiple mediators were first discussed in Preacher and Hayes (2008). In our setting, Eqs. (2) and (3) are first estimated as seemingly unrelated regressions, where the error terms are allowed to be arbitrarily correlated (Zellner, 1962). This exercise yields estimates of coefficients $β1$⁠, $β3$ and $δ1$⁠. Linear models are again used given its advantages (Angrist, 2001; Angrist and Pischke, 2009), and for the easiness of statistical inferences. Then, the indirect effects (⁠$β3δ1$⁠) are computed for each mediator. Finally, we bootstrap the standard errors of the indirect effects using 5,000 replications.

The key estimates of mediation models are reported in Table 4. Results of Eq. (2) show that, with mediators incorporated, the direct effects of cooperative membership are statistically significant, though generally smaller, for three of the four sustainable agricultural technologies. The effect on organic fertilizer adoption is gone (now insignificant). Among the four mediators, only peer learning and self-perceived leadership are generally significant as hypothesized, while productive resource access and information access are not. These findings highlight the importance of peer influence. Specifically, this could result from the ongoing rural digitalization in China, which facilitates easier input and information access for farmers. Therefore, input intermediation through agricultural cooperatives becomes less necessary, and information asymmetries decline with alternative digital channels. This is in accordance Zhang et al. (2020), which suggests that cooperatives act as repositories of productive knowledge, providing a platform that enables learning and capability formation among smallholders.

The only exception is the role of peer learning in crop rotation adoption, which is insignificant. Possibly, since crop rotation is a relatively easy practice, the role of learning is limited. Also, access to productive resources or information is generally insignificant across the four technologies. This is not as hypothesized, yet from Table 1 it is seen that the majority of farmers have reasonably good access to these resources. Moreover, with the increase of digital literacy, farmers have better access to those resources online, thereby weakening the role of these two factors.

Estimates of Eq. (3) are reported in the columns next to those of Eq. (2). Agricultural cooperative membership is significantly correlated with better information access, higher possibilities of peer learning and higher chances of self-perceived leadership. The role of cooperative membership in productive resource access is however insignificant, which may again be due to the lack of variation in the measure itself.

Given the estimates above, it is now possible to compute the indirect effects of the four mediators. These are presented in the next column for each technology. The indirect effects are only significant for peer influence measures, namely peer learning and self-perceived leadership (except for the role of peer learning in crop rotation). The magnitudes of the identified indirect effects are rather small. In the case of conservation tillage, the indirect effect through peer learning (0.004) and that through self-perceived leadership (0.002) only collectively explain 15% of the total effect (0.037, OLS estimate from Table 2). For the other technologies, the indirect effects are similarly small (7% for crop rotation, 19% for organic fertilizer and 23% for drip irrigation).

The smallness of mediation effects is not uncommon. Walters (2019) argues it is nearly always the case in empirical studies, which may have several reasons. However, regression-based mediation analysis cannot definitively determine the underlying reason. Specifically, measurement error in the mediator or outcome may attenuate estimates, omitted mediators could mean additional unobserved pathways exist and contextual features of the study setting may limit the magnitude of peer effects.

Although the mediation effects are modest, their practical implications should not be understated. Within agricultural cooperatives, even small peer effects can accumulate and scale through repeated interactions, dense social networks and ongoing collective activities. A marginal increase in adoption driven by peer learning or self-perceived leadership among a subset of members may trigger demonstration and imitation effects, gradually normalizing sustainable practices within the cooperative. Over time, such incremental changes can generate meaningful diffusion as early adopters influence others, particularly in contexts where cooperatives facilitate frequent information exchange and social comparison. From a policy perspective, this suggests that interventions targeting peer learning platforms or reinforcing members’ role-model identities may yield disproportionate long-term benefits.

To check the robustness of these results, we consider the possible endogeneity of agricultural cooperative membership and employ instrumental variable techniques in the mediation models. This is not easy because the endogeneity of membership can transmit into mediators, resulting in more than one endogenous variable in Eq. (2). Previous studies have considered employing multiple instruments (e.g. Frölich and Huber, 2017), while models with multiple endogenous variables are difficult to identify, and the interpretation of simultaneous causalities is challenging (Angrist and Pischke, 2009). Dippel et al. (2020) propose a method of using a single instrument to identify the direct and indirect effects in mediation models where both treatment and mediator can be endogenous, which is more feasible and interpretable. A key assumption is that the only source of endogeneity between treatment and outcome comes through confounders that jointly affect the treatment and the mediator. This is a slightly stronger assumption than standard IV alone, but weaker than needing instruments for both treatment and mediator.

We extend this method to multiple mediators, where agricultural cooperative membership is endogenous. Our empirical estimation follows a two-step procedure. First, we identify the effect of cooperative membership on each mediator (⁠$δ1$⁠) through 2SLS estimation, where the instrument used in the first stage. Second, we identify the effects of cooperative membership and mediators on adoption (⁠$β1$ and $β3$⁠) through 2SLS, where in the first stage the same instrument is used to predict mediator values. Empirically, the above two external instruments and Lewbel’s (2012) heteroskedasticity-based constructed instrument are used.

The key estimates of this procedure are reported in Table 5. While these 2SLS estimates somehow differ from the OLS coefficients, the key findings remain. That is, indirect effects mainly come from peer influence (Table 4). Additional indirect effects are significant in a few other places. However, these are never consistently identified through alternative 2SLS procedures, or the estimate is too small to be policy meaningful.

Despite those discrepancies, the computed indirect effects are similar to the OLS estimates. Specifically, the 2SLS-estimated indirect effects on the adoption of conservation tillage, crop rotation, organic fertilizer and drip irrigation, respectively, range from 8% to 15%, 5% to 26%, 22% to 80% and 17% to 32%. These intervals reasonably nest the above OLS estimates of 15%, 7%, 19% and 23%. These exercises collectively confirm the role of peer influence. While the indirect effects are small, they are consistently significant. Indeed, mediation could occur through a complex, multi-channel system and likely only part of the indirect is uncovered here. These estimates, including direct, indirect and total effects, are further visualized in Figure 2, which focuses on the mediation mechanisms through peer learning and self-perceived leadership, which are consistently found playing roles.

Given these results, a natural further step is to explore the interrelationship between peer effects. Specifically, we would like to understand whether learning from others and showcasing leadership to others facilitate or replace one another. Two subsample analyses are therefore implemented. To investigate the role of peer learning alone, we break the respondents into those who believe it is possible to learn about agricultural technologies from others and those who do not think so, and re-estimate Eq. (2). Similarly, we construct two subsamples of respondents with and without self-perceived leadership, and again re-estimate Eq. (2).

Key estimates are presented in Table 6. It is evident that either type of peer influence is universally recognized (with at least 10% significance) if the other influence is present. Specifically, peer learning is predominantly observed among individuals who view themselves as leaders, while self-perceived leadership is notably evident among those who acknowledge opportunities for peer learning. In instances where only one type of peer influence is present, the statistical significance of the other type becomes inconsistent, preventing reliable inferences. In other words, peer learning and self-perceived leadership tend to complement rather than substitute one another.

The complementary nature of peer learning and self-perceived leadership aligns with social identity and network theories. Weak social ties among cooperative members may help spread non-redundant information, facilitating observations and the flow of information (Granovetter, 1973). Therefore, even modest individual-level effects may have meaningful aggregate impacts within cooperative contexts. In our case, peer learning may motivate individuals to try out these innovations, while self-perceived opinion leaders may better exercise their influence in learning networks. Hence, leaders socially recognized and engaged in peer learning networks can both model behaviour and diffuse knowledge, boosting adoption rates (Rogers, 2003). This finding reinforces the synergistic effects of the two types of peer influence, and highlights the importance of policies facilitating such influence.

We examine the role of agricultural cooperative membership in sustainable agricultural technology adoption using a rural household survey from Western China, focusing on four specific technologies. Econometric modelling confirms the overall positive relationship. We further explore four potential mechanisms underlying this impact in a consolidated framework, two capturing resource access, while the other two measuring peer influence. Mediation analysis provides consistent evidence in support of peer influence but not resource access, and further reveals synergetic effects between the two peer effects.

Our findings underscore the critical role of agricultural cooperatives in promoting sustainable technology adoption (e.g. Ma and Abdulai, 2019; Manda et al., 2020). However, we have identified alternative pathways to realize these effects, as traditional mechanisms emphasizing access to productive resources or information no longer effectively drive adoption, as many rural communities in China have grown out of these constraints. In contrast, peer influence are significant factors. These insights emphasize the need for cooperative training programs to facilitate peer communication. Practical measures may include the establishment and enhancement of formal or informal community learning platforms, collaborative projects, peer-to-peer networks/demonstration and participatory research initiatives. Additionally, leveraging modern digital technologies can facilitate rapid dissemination of information online and instant learning, particularly as farmers’ digital literacy improves. Moreover, policy support for knowledge exchange involving farmers; extension services and non-governmental organizations could foster a more integrated learning system. While our results do not directly evaluate or endorse these measures, they warrant careful consideration in related policy designs. In this process, certain unintended consequences also need to be minimized, if possible, such as peer conformity overtaking real learning, and exclusion of marginal farmers.

Despite our contribution, the observational nature of our study likely limits the external validity of findings, and the cross-sectional data may not capture the dynamic relationship between cooperative membership and technology adoption. Also, the current study is confined to dichotomous technology adoption measures and mediators, and horticultural producers in three western provinces, which may limit the external validity of findings. Future investigations may apply suitable study designs to available data, and/or experimental approaches to better establish causal relationships. They may also possibly cover complimentary sectors such as grain and livestock, consider cooperatives of different member sizes and look at cooperative heterogeneity in terms of size, governance structure and other features, to further test the mediating mechanisms.