Policy Research Working Paper 10041 Incentivizing Social Learning for the Diffusion of Climate-Smart Agricultural Techniques Guigonan Serge Adjognon Tung Nguyen Huy Jonas Guthoff Daan van Soest Development Economics Development Impact Evaluation Group May 2022 Policy Research Working Paper 10041 Abstract Unsustainable land use is a key threat to both economic both the transfer of information from the trained to the development and environmental conservation in developing peer farmers, as well as the peer farmers’ sustainable land countries. This study implemented a randomized controlled management practices adoption rates. Offering financial trial in arid Burkina Faso to test the effectiveness of financial incentives thus mitigates two of the most important barriers incentives in stimulating the adoption of sustainable land to the adoption of sustainable land management practices management practices (SLMPs). It did so in the context —the (perceived) lack of private benefits and insufficient of a so-called cascade training program, in which some diffusion of the technical implementation information farmers were trained in the implementation of sustainable from the trained farmers to their peers. Finally, the study land management practices, who were then asked to dis- documents that adoption of sustainable land management seminate their newly acquired knowledge and expertise practices generates substantial increases in crop productivity to other farmers in their social networks. The study finds and agricultural income already after one agricultural cycle. that offering payments conditional on adoption improves This paper is a product of the Development Impact Evaluation Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at gadjognon@worldbank.org and jguthoff@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Incentivizing Social Learning for the Diffusion of Climate-Smart Agricultural Techniques∗ Guigonan Serge Adjognon1 , Tung Nguyen Huy2 , Jonas Guthoff1 , and Daan van Soest2 1 Development Impact Evaluation Department (DIME), The World Bank Group, Washington D.C., USA 2 Dep. of Economics and Tilburg Sustainability Center, Tilburg University, The Netherlands JEL Classifications: O13, Q15, Q16, Q24. ∗ We thank the Climate Investment Fund for their financial support, and Burkina Faso’s Forest Investment Program (FIP) for their generous collaboration in the implementation of this research. We obtained IRB ap- proval from Tilburg University (2019-008), and we registered the project in the AEA’s registry for randomized controlled trials under registry ID AEARCTR-0006461. We thank participants of the 2021 EAERE and 2021 Bioecon conferences as well as seminar participants at Tilburg University for their constructive comments and suggestions. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank Group, or the governments they represent. All remaining errors are our own. Please send all correspondence to d.p.vansoest@tilburguniversity.edu. 1 Introduction Sustainable land management practices are thought to be both socially desirable as well as productivity enhancing (World Bank, 2008; Liniger et al., 2011b). An example of such practice is the construction of bunds which are ridges that prevent the erosion of soil from heavy rains by slowing down the runoff water. Vancampenhout et al. (2006) estimates in an agronomic experiment in dry Ethiopia that such bunds can increase crop productivity by 16%, but take-up of such technologies in Ethiopia is low (Teklewold et al., 2013). Adoption of these practices is of paramount interest especially in arid Sub-Saharan Africa where soil erosion and soil depletion threaten the long-run viability of agriculture on existing arable lands as well as necessitate the continued conversion of forested areas to create new arable land (World Bank, 2008; FAO, 2019). Despite their promise of not just being environmentally desirable but also privately op- timal, take-up of these sustainable land management practices is typically low. Lack of benefits, actual or perceived, has been documented to be an important barrier to sustainable land management practices (SLMP) adoption (Jack, 2011; Beaman and Dillon, 2018), and so is the lack of knowledge of how to implement them (Conley and Udry, 2010; Jack, 2011). These two barriers may, in fact, be related, because of the cost-benefit considerations of infor- mation acquisition. If the technologies are deemed to be not very beneficial (or even costly) to implement, farmers are also less likely to be receptive of SLMP information provision – let alone to actively start searching for the information themselves. In this paper we test whether a financial incentive for the adoption of sustainable land management practices increases take-up, and to what extent it can help overcome the above two key barriers – limited (or even negative) perceived private benefits, and the lack of implementation know-how. We cooperated with the government of Burkina Faso and im- plemented a randomized controlled trial (RCT) to estimate the impact of offering farmers financial compensation conditional on the adoption of up to nine different SLMPs. The stan- dard program is a so-called cascade training scheme (Banerjee et al., 2013; BenYishay and Mobarak, 2018; Kondylis et al., 2017; Behaghel et al., 2020). A selected set of farmers were invited to participate in a four-day training program offered by Burkina Faso’s agricultural extension services. The program consisted of providing information on the benefits and costs of each of the nine SLMPs, as well as training in how to implement them. Upon completion 2 of this training, the trained farmers (henceforth referred to as “contact farmers”) were asked to actively disseminate their newly acquired knowledge and expertise among other farmers in their existing social network (henceforth “peer farmers”). Our main treatment consisted of offering cash transfers to the farmers in our treatment group, where the amount to be paid depended on the number of technologies present on our peer farmers’ lands at endline. In the control group, we implemented a cascade training scheme without offering financial incentives on the adoption of practices. We find that in our cascade training program, conditional payments increase SLMP adoption among peer farmers. Compared to the control group, where farmers were not of- fered payments, offering financial compensation conditional on technology adoption increases SLMP adoption by 0.38 standard deviation. Offering financial payments obviously improves the cost-benefit ratio of SLMP adoption, but we also find that it is especially the less widely used (as measured at baseline) that experience the largest increase in uptake. Importantly, we also document that offering financial incentives increases the peer farmers’ demand for knowledge and expertise. More specifically we find that offering financial compensation for downstream adoption resulted in more frequent meetings between peer and contact farmers, in peer farmers being more likely to reach out to their contact farmers to ask for advice, and in more effort by contact farmers to support their peer farmers’ adoption efforts. At endline, we also document that “a lack of information” is significantly less of an adoption barrier in the payment treatment than in the control group. We thus find that offering conditional adoption payments renders the cascade training scheme more effective. We also test whether the effect of a given amount of financial incen- tive on information exchange and SLMP uptake can be improved by allocating part of the incentive to contact farmers. Contact farmers then have a direct financial stake in the peer farmers’ SLMP decision, which provides incentives for them to disseminate SLMP informa- tion. We implement two sub-treatments within the conditional payment treatment. While in both sub-treatments the payment to be disbursed is conditional on the number of SLMPs adopted by the peer farmer at endline, we vary how the payment is allocated between the peer farmer and her contact farmer. In one sub-treatment, the peer payment treatment, the full amount of the incentives is disbursed to the peer farmer; in the other sub-treatment, the payment is split, 80-20, between the peer and contact farmer. Interestingly, we find no 3 statistically significant difference in either our output or input measures of communication between the peer payment and the split payment treatment arms, and also no difference in the number of SLMPs adopted. This is an important outcome as it implies that if the downstream demand for information is sufficiently strong, supply will follow – independent of whether the supplier of information is rewarded directly for supplying information, or not. This result is reminiscent of the Coase theorem (Coase, 1960) which, in this context, states that for a given surplus generated by transferring knowledge, the amount of information transferred is independent of the initial allocation of the surplus. This result is of key policy interest as it suggests that the standard cascade training scheme will become increasingly more effective, because of the following reasoning. Based on these results we conjecture (but cannot prove) that, even without adoption subsidies, SLMP adoption is productivity enhancing already in the short run. With continued SLMP diffusion farmers update their expectations about the surplus generated by SLMP adoption (for example via learning by watching), and the exchange of information is predicted to improve even without offering direct financial incentives for information dissemination. Our study contributes to three different strands in the literature. First, it speaks to the literature on the effectiveness of subsidizing sustainable land use practices. Despite the fact that the SLMPs are thought to increase agricultural production, subsidies are warranted because of the relatively low take-up rates using standard information dissemination prac- tices. Farmers may be reluctant to invest in technologies that require considerable up-front costs and/or yield uncertain private returns in the distant future. Society may, however, be willing to subsidize the adoption of these technologies. By definition the societal benefits of sustainable techniques exceed the private benefits (as the agent adopting them only reaps a small and oftentimes negligible share of the environmental gains these technologies give rise to), and also the societal discount rate is typically lower than the private one. As such, our intervention can be viewed as an example of a “Payments for Ecosystem Services” (PES) scheme – a policy that aims to stimulate the private provision of nature conservation by offering financial compensation conditional on actual environmental service delivery (Wun- der, 2007; Engel et al., 2008; Engel, 2016). PES schemes have been shown to be effective in inducing forest and water conservation (Jayachandran et al., 2017; Börner et al., 2017); our study contributes to this literature by documenting the effectiveness of PES on the diffusion 4 of sustainable land management practices. Second, our study speaks to the literature on the efficient dissemination of information on sustainable land practice implementation. Cascade training programs have been devel- oped as an alternative to the traditional model of government agricultural extension services for essentially two reasons (Krishnan and Patnam, 2014; Kondylis et al., 2017; BenYishay and Mobarak, 2018). One, large-scale diffusion of agricultural innovations is challenging in Sub-Saharan Africa, as the available resources are oftentimes insufficient for a nation-wide coverage of high-quality government extension services. And two, the standard extension services’ approach of top-down information provision (from an extension worker to a farmer) is not always effective in convincing the latter of the desirability of adopting the new tech- nology – oftentimes because of doubts whether the new technology is sufficiently well suited for the local agronomic circumstances. Cascade training systems may be able to alleviate (if not overcome) both issues. They hold the promise of being both more efficient as well as more effective than the traditional diffusion model. They may be more efficient as relatively few farmers need to be trained directly. And they may also be more effective, as information provided by a fellow farmer from the same region may be perceived as more reliable and bet- ter adapted to the local agronomic conditions than the information provided by (non-local) government extension workers. The evidence on the effectiveness of cascade training programs is, however, mixed at best. Takahashi et al. (2019) experiment with a cascade training program aimed at disseminating rice management practices in Côte d’Ivoire and find that knowledgeable contact farmers are able to successfully disseminate the management practices among their peers. Kondylis et al. (2017) document for the case of Mozambique that information dissemination is still limited, however, because of the contact farmers’ low willingness to actively share their knowledge. Shikuku et al. (2019) explore whether social recognition and private in-kind rewards are effective in inducing contact farmers to better disseminate their newly acquired knowledge and expertise. They find that although both interventions increase dissemination effort, neither results in increased SLMP uptake among peer farmers. A series of recent studies suggest that financial incentives may be more promising. In their seminal study implemented in Malawi, BenYishay and Mobarak (2018) offer financial compensation to contact farmers dependent on their peer farmers’ rate of adoption of specific agricultural practices, and find 5 that this increases both information dissemination and increased uptake among peer farmers. Sseruyange and Bulte (2018) make use of a financial literacy cascade training program among farmers in Uganda, and find that offering contact farmers money as a function of their peers’ knowledge acquisition substantially improves peer farmers’ financial literacy test scores. As is evidenced by Berg et al. (2017), financial payments are also effective in raising the overall effectiveness of cascade training schemes in domains other than agricultural extension work. Berg et al. focus on the uptake of health insurance in India, and find that such incentives are able to increase insurance uptake even among peers that are not socially close.1 All these studies have in common that offering financial compensation to contact farmers does not only provide incentives for knowledge sharing – the payments also increase the total surplus associated with knowledge dissemination and technology uptake. By implementing both the peer and split payment schemes (in addition to the control group), we shed light on the question whether, next to the size of the total surplus, there is an additional impact of the way in which the surplus is distributed. Third, the insights provided in this paper are not limited to providing a proof of concept of PES in inducing SLMP uptake – it also allows us to actually estimate the farmers’ short-run benefits (positive, or negative) of SLMP adoption. While the sustainable land use literature claims that SLMPs are not only socially desirable but also productivity enhancing, reliable productivity estimates are still scant (Liniger et al., 2011a; Pretty et al., 2011; Pittelkow et al., 2015; Pretty et al., 2018). Productivity impacts are typically estimated using either matching designs (Abdulai and Huffman, 2014; Kassie et al., 2015; Manda et al., 2016) or by exploiting time variation in the SLMP adoption decision (Arslan et al., 2015; Khonje et al., 2018; Kassie et al., 2018; Tesfaye et al., 2021). Estimates may, however, be biased because neither method is able to fully control for the role of unobservable characteristics in determining the (timing of the) adoption decision. Partly due to this, productivity estimates differ substantially between studies. For example, estimates of the productivity impacts of intercropping range between no effect (Arslan et al., 2015) and 136% (Thierfelder and Wall, 2010), and also anything in between (see for instance Manda et al. (2016); Tesfaye 1 Indeed, social distance between farmers can be a barrier to knowledge dissemination in a cascade training scheme. For example, Kondylis et al. (2016) study the role of farmers’ gender within a cascade training program in Mozambique and find that the training does not increase female peer farmers’ awareness and knowledge about agricultural technologies if their contact farmers are male. BenYishay et al. (2020) show that male peer farmers are reluctant to request information from female contact farmers. Berg et al. (2017) show that financial incentives can help mitigate these social barriers to information dissemination. 6 et al. (2021)). As do BenYishay and Mobarak (2018) and Takahashi et al. (2019), our study complements this literature by exploiting the treatment-induced exogenous variation in SLMP adoption to estimate the average short-run impact on agricultural revenues. We find that, in the year of implementation, the impacts of SLMP adoption on agricultural productivity and income are positive and sizeable. As such, our study is among the few that provide causal evidence on the short-run impacts of SLMP adoption on farmer welfare. The remainder of this paper is organized as follows. We explain the details of the cascade training program, the determination of the payments for adoption, and the experimental design in Section 2. Section 3 describes our empirical framework and the characteristics of the farmers in our experiment. We present our treatment effect estimates on the adoption of practices and on agricultural production in Section 4, and those on knowledge dissemination in Section 5. Section 6 concludes the paper. 2 Program Description and Experimental Design 2.1 The Cascade Training Program Our RCT is embedded in a large-scale environmental conservation project, the Forest In- vestment Program (FIP), implemented by the government of Burkina Faso with financial support from the World Bank, the African Development Bank, and the Climate Investment Fund. One of the FIP’s key objectives is to reduce the dependence of rural communities on (the unsustainable) exploitation of nearby forest areas – especially of those with protected forest status. Burkina Faso’s protected forests are threatened by an increased demand for agricultural land, caused by rapid population growth as well as by dwindling productivity on existing agricultural lands (Pouliot et al., 2012; Goldstein and Udry, 2008; Etongo et al., 2015). The decline in agricultural productivity can be mitigated (and even reversed) by implementing so-called Sustainable Land Management Practices (SLMPs) – techniques and measures aimed at reducing soil depletion and erosion, as well as the use of sustainable inputs like organic fertilizers (Liniger et al., 2011a; Pretty et al., 2011). In cooperation with the FIP we identified nine SLMPs that were deemed to be most promising in arid Burkina Faso; see Table 1. The nine SLMPs span three agricultural do- mains: agronomy, agro-sylvo-pastoralism, and agro-forestry. The agronomy-oriented SLMPs 7 focus on maintaining land productivity by conserving soil nutrients and retaining rainwa- ter on farmers’ plots. They include planting seeds in purposely prepared pits, constructing bunds (made from earth or stone) on plot perimeters, and building adequate structures for composting crop residue. The SLMPs in the agro-sylvo-pastoral domain (sometimes also referred to as integrated crop and livestock management) consist of producing and storing fodder from residues of agricultural production and from direct cultivation of forage crops, as well of (re-)using agricultural and forest by-products. These practices enhance agricultural productivity and reduce the grazing pressure on nearby lands. Finally, practices in agro- forestry aim to improve soil and water management by conserving tree and shrub cover on agricultural plots, to improve nutrient recycling and to reduce soils’ exposure to direct sun- light. These nine SLMPs were selected because they were expected to improve short-term growing conditions as well as agricultural resilience to climate change, and are especially well-suited for low-input agriculture in arid countries (Liniger et al., 2011b). Table 1: Overview of our project’s nine focal SLMPs. Group Practice Agronomy Pit planting (Zaï) Stone and earth bunds Heap and pit composting Agro-sylvo-pastoral Mowing and conservation of natural fodder Forage crop cultivation Use of agricultural and wood by-products Agroforestry Controlled clearing Assisted natural regeneration Living hedges The FIP aimed to foster adoption of these nine SLMPs in 32 communes (municipali- ties, each consisting of several villages or hamlets) across five different regions: Boucle du Mouhoun, Centre Sud, Centre Ouest, Est and Sud-Ouest (see Figure 1). These regions were selected because they were among the FIP’s target areas. The essence of the FIP project was to stimulate the adoption of SLMPs via dissemination of SLMP knowledge and expertise by means of a cascade training program; the timeline of the RCT is presented in Figure 3. In April 2019, the FIP recruited 320 farmers, 10 in each of the 32 communes, to participate in a four-day training on the nine key SLMPs described above. These farmers were selected because they had participated in one or more earlier activities organized by the FIP, and 8 because they were thought to be effective entry points for the diffusion of the practices in their respective communities. During the recruitment process, the farmers were informed of the general set-up of the intervention: (i) that they themselves would receive training in the implementation of nine climate-change resilient SLMPs that are considered effective in raising long-run agricultural yields; and (ii) that they would be expected to actively transfer the acquired knowledge and expertise to fellow farmers in their village. All farmers agreed to participate. Upon having accepted the invitation, each of these 320 farmers were asked to provide the names of five fellow farmers in their village whom they would expect to be interested in the adoption and usage of (some of) the SLMPs, and whom they would be willing to disseminate their newly acquired knowledge and expertise to. Importantly, the contact farmers in the treatment group were not informed of any possible payments until after they had completed the training and had returned to their villages. Therefore, the con- tact farmers’ treatment status cannot have affected either their choice of whom to nominate as peer farmers nor the extent to which they paid attention during the SLMP training. Our experimental sample thus consists of, in total 1920 farmers – the 320 so-called contact farmers who received the training, and the 1600 so-called peer farmers to whom the contact farmers may or may not have transferred their newly acquired SLMP knowledge and exper- tise. Comparing contact and peer farmers, the former are, on average, older, more educated, wealthier, have more land, and also have more experience with SLMP usage at baseline than their peers; see Table A4 in Appendix A. This suggests that the contact farmers are indeed well-positioned to understand the benefits of these techniques and to disseminate them. This is in line with the FIP’s goal of selecting contact farmers who are more likely to be good transmitters of knowledge given their education and experience, and they are similar to the lead farmers in BenYishay and Mobarak (2018) and also to the contact farmers in Kondylis et al. (2017). The four-day training program for the contact farmers was developed by experts from the Ministry of Agriculture. The training itself was implemented at the commune level, in May and June 2019, by specially-trained government extension workers. On the first day of the training, the contact farmers received information on the longer-run consequences of soil erosion, on the theoretical benefits of each of the nine SLMPs in terms of reducing soil erosion and maintaining land productivity, as well as under what circumstances each SLMP 9 Figure 1: Geographic location of the 32 communes involved in the RCT. Boucle du Mouhoun Est Centre-Ouest Centre-Sud Sud-Ouest Legend Not part of intervention Peer payment Split payment Control (no payment) was likely to be particularly useful or effective. The remaining three days were dedicated to practical training on the actual implementation of each of the nine SLMPs on demonstration plots. To evaluate the effectiveness of the training and quantify the learning outcomes, we administered a test at the beginning and the end of the training on the content that was taught during the training. Figure 2 shows the distribution of farmers’ test scores before and after training. It shows that farmers’ knowledge improved substantially: the score of the median farmer increased by 30%, from 47% correct answers to 61%. Upon completion of the training, all contact farmers were provided with a knowledge dissemination kit that included cheat sheets summarizing the key information on the ben- efits and implementation processes of each of the nine SLMPs. In addition, each contact farmer received an SLMP implementation kit containing agricultural equipment (including a wheelbarrow, a pickaxe, a shovel, a fork, and a bundler) as well as inputs (seeds and plants). Contact farmers were told that they were free to use these tools and inputs to facilitate the 10 Figure 2: Distribution of the contact farmers’ scores on the SLMP knowledge test before and after the training. Note: The knowledge test consisted of multiple choice questions covering all nine SLMPs to be diffused via the cascade training scheme. The scores presented in this figure are adjusted for the expected score obtained −0.41 by pure guessing, which is 41%. Denoting the unadjusted score by x, the adjusted score is: xadj = x 1−0.41 . implementation of the various SLMPs on their own lands, but that these tools and inputs were also meant to be made available to their peer farmers upon request. At the end of the four-day training session all contact farmers went back to their villages, and each of them was asked to actively disseminate the newly acquired information to all other farmers in their village including the five peer farmers they had previously selected to be included in the study. They were also explicitly told that the project team would visit them as well as their five peers, at the end of the agricultural season, to evaluate the outcomes of the dissemination and SLMP adoption processes. 11 Figure 3: Timeline of the study. Selection Endline data of contact Training collection and peer of contact (adoption farmers farmers verification) April May June ... December ... July 2019 2020 Collection Payment of baseline Treatment/Control disburse- data info (certifi- ment cates) relayed to the peer and contact farmers 2.2 Experimental Design Our main intervention consisted of offering financial compensation depending on the number of SLMPs present, at endline, on each peer farmer’s land. In May 2019 our survey team georeferenced the perimeter of a maximum of five plots managed by each participant – those that the farmer had planned to cultivate in the 2019 agricultural season, and also those that were planned to be left fallow.2 In the treatment group the amount to be disbursed thus depended on the number of SLMPs present, at endline, on each peer farmer’s georeferenced plots. The payment scheme is presented in Table 2. If, in December 2019, between one and three SLMPs were present on a peer farmer’s land, 30,000 FCFA would be disbursed; if the number of SLMPs present was between seven and nine, 50,000 FCFA would be paid out. Table 2: The total amount of money to be disbursed as a function of the number of SLMPs present on a peer farmer’s land. # of SLMPs present Payment 0 0 FCFA 1-3 30,000 FCFA (≈ $ 50) 4-6 40,000 FCFA (≈ $ 68) 7-9 50,000 FCFA (≈ $ 85) Four aspects of the payment scheme require additional discussion. First, payments were not conditional on the number of SLMPs adopted during the 2019 agricultural season, but rather on the number of SLMPs present on a peer farmer’s land at endline. This was 2 For practical reasons, we implemented the rule that if the number of plots managed by a participant was larger than five, she was asked to point out the five cultivation plots most suited for SLMP implementation. In our sample, farmers had control over an average of 1.66 plots, and the maximum of five plots was only binding for three farmers. 12 because the FIP considered that imposing strict additionality was unfair vis-à-vis those farmers who had already adopted SLMPs prior to the start of the intervention. Second, the payment scheme was such that the average payment per SLMP decreased with the cumulative number of SLMPs adopted. We implemented this because the adoption process is likely to be characterized by non-negligible set-up costs, such as the time and effort spent on acquiring information on the range of available technologies (Liniger et al., 2011a; Giger et al., 2018). Note that these two design choices affect the incentives to adopt, but not the internal validity of our RCT. Third, the amounts of money to be disbursed were sizeable. With an average farmer household’s annual agricultural production of about $960 (WB, 2016, p. 52), the payments offered amounted to between 5% and 9% of an average farmer’s annual revenues, or between 40% and 70% of the country’s average per-capita food consumption of about $120 (WB, 2016, p. 29). Fourth, and importantly, the disbursement of the payments only started in July 2020 – well after the end of the 2019 agricultural season; see Figure 3.3 While informing treated farmers that they would be eligible to receive compensation based on the number of SLMPs adopted is expected to have increased SLMP adoption – with possibly subsequent conse- quences for input use, crop choice, quantities harvested and agricultural revenues earned – it is unlikely to have affected farmers’ budget constraint during the 2019 agricultural season itself. The prospect of receiving payments, almost a year after the start of the intervention, can only affect investment decisions via mechanisms other than SLMP adoption if the farmer is able to borrow against such future payments. We will come back to this in Section 4.2.1. The schedule presented in Table 2 determined the total amount of payments to be dis- bursed in the treatment group; farmers in the control treatment were not offered any fi- nancial compensation. The conditional payments treatment consisted, however, of two sub- treatments that only differed in how the payment was disbursed. In one sub-treatment, the full amount was disbursed to the peer farmer; in the other, the peer farmer would receive 80% of the payment, and her contact farmer would receive the other 20%. We will refer to these two sub-treatments as the peer and split payment treatments, respectively. All farmers in 3 A delay was foreseen because of the administrative processes needed to clear the payments for each individual farmer, conditional on the independent verification of the end-line number of SLMPs present on each treatment farmer’s land. At the start of the intervention farmers were aware that they would receive the payments only after a delay because of administrative reasons; the actual length of the delay was longer due to the impact of the global pandemic. The bulk of the payments took place in July and August 2020, and the last payments were made in November 2020. 13 the treatment group were offered personalized certificates detailing the relevant conditional payment scheme.4 As stated before, all farmers in the two sub-treatments were informed of the details of the relevant payment scheme only after the contact farmers had returned from the training. Both the selection of the peer farmers and the contact farmers’ effort to gain knowledge in their training are thus independent of whether financial incentives were offered, and also of how the transfers were to be divided between the peer and contact farmers. Regarding the RCT’s implementation, assignment to each of the three (sub-) treatment groups was randomized at the commune level, stratified by region. Of the 32 communes, 12 were assigned to the control group, and ten to each of the two sub-treatment groups. That means that there were, in total, 720 farmers in the control group and 600 in each of the two sub-treatment groups. We motivate these design choices as follows. First, we randomized treatments at the commune level to minimize the chances of imperfect treatment implemen- tation, conflict and biased impact estimates because of spillovers. Government extension services are organized at the commune level, and hence commune level randomization avoids the risk of incorrect treatment implementation that might occur when randomizing at the level of the village or even the contact farmer. Randomization at the commune level also mitigates concerns regarding both treatment spillovers and possible conflict. Spillovers can occur if contact farmers in the treatment group directly communicate with farmers in the control group, or if control group farmers observe the increased uptake of the technologies in the treatment groups and they decide to also adopt more SLMPs themselves. Given the communes’ size and the distances between them, spillovers are less likely to be a concern when using cluster-randomized treatment assignment at the commune level than if we would we have randomized at the village or even the individual contact farmer level.5 Second, our decision to assign 10 communes to each of the two payment sub-treatment groups and 12 to the control group was driven by statistical power considerations. We wanted to have a high-powered test of the overall impact of offering financial incentives, but we also wanted to be able to have a good chance of detecting the disbursement mechanisms’ differential impact if there is one. Assuming both to be equally important and using z to denote the 4 Farmers in the control group also received personalized certificates that show the name of the farmer, the village, commune, and region name, the name of the georeferenced plots, and the name of the contact farmer. In the payment group, the certificate also detailed the condition and the structure of the payments. See Figure A1a-A1b in the Appendix. 5 The average Euclidean distance between farmers in the control group and those in the closest treatment commune is 18 kilometers. Distances via the road network are likely to be substantially larger. 14 number of (sub-) treatment groups (in addition to the control group), statistical power is √ maximized when assigning a share of 1/(z + z ) of the available treatment units to each √ √ of the z treatment groups, and hence a share of z/(z + z ) of the available units to the control group (List et al., 2011). Intuitively, because the control group is used as a reference outcome for the test of the effectiveness of either treatment, the joint statistical power of the two sub-treatment impact estimates is maximized by (slightly) oversampling the con- trol group. Because of indivisibility, we approximate this optimal allocation by assigning 10 treatment units to each of the two sub-treatment groups, and 12 to the control group. With this treatment allocation, assuming an intra-cluster correlation of 0.1 and baseline covariates being able to explain 30% of the variation, we have an 80% chance of detecting an adoption impact of 0.4 standard deviation, or better. 2.3 Main Effects Identified in the Experiment Our experimental design allows us to estimate three effects. First, the impact of offering financial compensation on SLMP adoption, and then especially whether offering compensa- tion indeed stimulates the adoption of especially the lesser known (or lesser used) SLMPs. In the longer run, the usage of SLMPs is expected to be beneficial for the farmer adopting them, but their adoption may be less than perfect because of incomplete knowledge about the practice or because of the (perceived) riskiness of implementing them (Teklewold et al., 2013; Karlan et al., 2014). Offering farmers payments conditional on their adoption is thus expected to increase take-up, as the conditional payments increase the cost-effectiveness of the SLMPs (Foster and Rosenzweig, 1995; Koundouri et al., 2006). Conditioning payments on the number of adopted practices out of nine (rather than just on the adoption of one specific practice) gives farmers a choice to consider which ones to adopt. Therefore the esti- mated impacts on take-up rate also reflects farmers’ perception about which practices they expect to be beneficial given their own constraints and the SLMPs already in place on their plots (Teklewold et al., 2013; Kassie et al., 2015; Kpadonou et al., 2017). Second, in case of a significant impact on the uptake of SLMPs, this analysis will also yield insight into the farmers’ short-run benefits (positive, or negative) of SLMP adoption – by comparing agricultural productivity and income between the treatment and control groups. Because the payments were not disbursed until at least seven months after the end of the 15 agricultural season, any difference in endline outcomes between the treatment and control groups is driven by the prospect of receiving money. Note however, that the estimate is unbiased if and only if farmers cannot borrow against (uncertain) future income (see below). Third, offering peer farmers financial incentives to adopt SLMPs increases the benefits of adoption, and hence peer farmers’ demand for knowledge and expertise from the contact farmer (Conley and Udry, 2010; Dupas, 2014). Offering adoption payments to the peer farmers presumably increases their willingness to pay for more detailed information on how to implement the SLMP technologies. In turn, the contact farmer may also have incentives to actively provide the required knowledge and expertise – think of side payments from the peer to the contact farmer, or the peer farmer reciprocating to the contact farmer in ways other than via a direct financial transfer. With perfect markets for information, the contact farmer’s incentives to share information are independent of the allocation of the payments (100% for the peer farmer, or 80% to the peer farmer and 20% for her contact farmer). If this is indeed the case, the initial payment allocation will neither affect the SLMP adoption rates by the peer farmers, nor the knowledge transfer from the contact farmer to their peer farmer. The alternative hypothesis is that the market frictions affect the efficient transfer of information, and hence providing the contact farmer with a direct stake in their peer farmers’ SLMP adoption decisions is expected to result in higher actual adoption rates. 3 Empirical Framework 3.1 Identification Strategy We implement two types of modelling approaches. First, we use a simple linear model to capture the intention-to-treat impact of offering conditional payments on key outcome variables capturing SLMP uptake, agricultural production, and knowledge dissemination: 1 yicr 0 = α + βyicr + τ Tcr + γ ′ Xicr + δ ′ Wcr + Rr + ϵicr . (1) t In equation (1), yicr denotes the outcome variable of interest observed at either baseline (t = 0) or endline (t = 1), for farmer i living in commune c in region r. When avail- able, controlling for the baseline value of the outcome variable improves the precision of 16 the treatment estimates (see McKenzie, 2012).6 Next, Tcr captures the treatment status of commune c in region r where the farmer is located. We run two types of analyses, so that P ooled , T Split }. In one, we pool the peer and split payment groups to estimate the Tcr = {Tcr cr average impact of offering conditional payments on our outcome variables of interest. Treat- P ooled which takes on value 1 if the peer and contact farmers ment status is then denoted by Tcr in commune c in region r have been assigned to either the peer payment treatment or the split payment treatment, and zero otherwise. In the other, we test how the initial allocation of the conditional payments affects outcomes. In that case, the analysis is restricted to just Split the peer and split payment groups. Treatment status is then captured by Tcr which takes on value 1 if the peer and contact farmers in commune c in region r were assigned to the split payment sub-treatment, and 0 if they were assigned to the peer payment subtreatment. That means that in equation (1), τ is the estimate of either the average impact of offering P ooled ), or the differential impact, for a given surplus, of the conditional payments (if Tcr = Tcr Split way in which the conditional payments are distributed (if Tcr = Tcr ). Regarding the other covariates in equation (1), vector (Xicr ) contains a series of baseline characteristics of the farmer (age, gender, and level of education), her household (family size and composition, and asset index) and farm (total agricultural land area, a land quality indi- cator, employment of family and hired labor, and agricultural and non-agricultural household income). In case farmer i is a peer farmer, vector Xicr also contains the knowledge score ob- tained by the contact farmer she was matched with, as an indicator of the (potential) quality of SLMP adoption information farmer i (may have) had access to. This is important because with a standard deviation of about 14%, there is substantial variation in contact farmers’ test scores obtained; see Figure 2 and Table A1 in the Appendix. The vector of baseline commune level controls, Wcr , includes the share of farmers in the commune who have one or more SLMPs in place at baseline, as well as its quadratic term. These covariates are intended to control for possible social aspects associated with technology adoption (including learning by watching spillovers, or the strategic consideration to postpone adoption to wait and see; Bandiera and Rasul, 2006). Finally, Rr is a vector of strata (region) fixed effects, and ϵicr is 6 This modeling approach is typically referred to as the ANCOVA specification. The key difference with standard difference-in-difference models is that in equation (1) β can be estimated freely, while in the difference-in-difference (DID) models β is restricted to be equal to 1. The extent to which ANCOVA outperforms DID thus depends on the extent to which the value of the coefficient on the lagged dependent variable differs from 1 (McKenzie, 2012). 17 the idiosyncratic error term which is clustered at the level of randomization – the commune level; see Abadie et al. (2017). For ease of interpretation, in the main body of this paper we present equation (1)’s regression results using ordinary least squares. Results of robust- ness tests regarding estimation methods (e.g. negative binomial models for the analysis of SLMP adoption) and hypothesis testing (randomized inference, and also multiple hypothesis testing) are presented in Appendix A.5. We use equation (1) to estimate the impact of financial incentives on, among others, the number of SLMPs at endline, but we are also interested in uncovering which types of technologies experience the largest increase in take-up. One approach would be to separately estimate equation (1) for each of the nine SLMPs. This would, however, disregard the fact that the decisions of whether to adopt each of these practices, are not independent. Using s1 and s2 to index SLMPs, this implies that for s1 ̸= s2 we have cov(ϵicr,s1 , ϵicr,s2 ) = σs1 ,s2 ̸= 0. We take into account the possibility of unobserved factors simultaneously affecting the adop- tion decisions of multiple SLMPs by estimating our nine models using seemingly unrelated regression (SUR; see Wooldridge, 2010). The analyses of the SLMP-specific adoption deci- sions provides insight into the farmers’ ex-ante assessment of the relative profitability (or more generally, desirability) of adopting techniques in the various domains (agronomy, agro- forestry and agro-sylvo-pastoralism). Equation (1) provides us with the intention-to-treat impacts of financial incentives, but we are also interested in what is the effect of implementing the practices on agricultural pro- duction (interpreted as the treatment-on-the-treated effect). This can be done by estimating the following two-stage least-squares (2SLS) model: 1 #SLM Picr 0 = α + β #SLM Picr P ooled + τ Tcr + γ ′ Xicr + δ ′ Wcr + Rr + ϵicr , (2) 1 1 yicr = µ + θ #SLM P icr + ψ ′ Xicr + η ′ Wcr + Rr + νicr . (3) t to denote the number of SLMPs present at time t = {0, 1} on farmer i ’s Using #SLM Picr land, equation (2) estimates the exogenous increase in SLMP adoption as induced by the prospect of receiving conditional payments; equation (3) then estimates the marginal impact 1 , by regressing that variable on the of SLMP adoption on the key variable of interest, yicr 1 predicted number of SLMPs (#SLM P icr ) as derived from equation (2). If offering financial 18 incentives significantly and substantially increases SLMP uptake without any direct impact 1 itself (that is, if the exclusion restriction holds; see Section 4.2.1), the coefficient on on yicr these predicted values of the number of adopted practices (θ) captures the marginal impact of SLMP adoption on the relevant outcome variable of interest. 3.2 Data and Descriptive Statistics We collected two types of data; see also the timeline presented in Figure 3. First, we col- lected information on the types as well as number of SLMPs present at baseline and at endline (in May and December 2019, respectively) on each farmer’s land. At baseline we identified which plots were eligible for SLMP implementation, and georeferenced up to five plots that were controlled by the farmers.7 We also documented the type of SLMPs that were already in place, as well as the plot they were located on. Independent teams went back to each of the georeferenced plots to verify the presence and the types (and hence also the number) of SLMPs. Independent verification is important, especially at endline, because the endline number of SLMPs at a peer farmer’s land determined the amount of money to be disbursed. To ensure this, the teams were tasked to document the presence of the SLMPs by taking photographs of the identified SLMPs and by asking follow-up questions about the implementation of the practices. Identification of the SLMPs at endline was possible because all practices had visible traces on the field: bunds, living hedges, and solutions to assist tree growth (for assisted natural regeneration) stayed intact after the harvest; composting pits and storage of fodder were implemented after harvest, right before endline data collection; and even the holes of pit planting remained visible at endline (see Figure A2 in the Ap- pendix). The incentives of enumerators to misreport the presence of SLMPs in farmers’ favor were also small given that teams consisted of at least two survey enumerators and they did not return to the field to disburse the payments after the verification. Second, we implemented two surveys, one at baseline and one at endline. In the baseline survey we collected information on the socio-demographic characteristics of all participants as well as on their family’s size and composition, farm and non-farm activities, indicators of wealth and assets, and behavioral traits. We also collected more refined information at the plot-level on how farmers cultivate their land. The endline survey was administered 7 As stated before, the constraint of a maximum of five plots was binding for just 3 of the more than 1,900 farmers in our sample. 19 seven months after the baseline, in December 2019, at the end of the agricultural season. Overall, from the 1,914 farmers interviewed at baseline, 1,901 (99.3%) were surveyed again at endline (see Table A2 in the Appendix). The potential bias from systemic attrition is thus negligible. The baseline and endline surveys were similar, but not identical. That means we have baseline and endline values for many variables, but not all. Our main outcome variables from the endline survey capture agricultural production, farmer livelihood, and communication between contact and peer farmers. We analyze the impacts of the payments and the adoption of practices on agricultural production and liveli- hoods. On the input side of agricultural production we georeferenced the total area culti- vated, the share of plots which were manually sowed (instead of mechanically), the amount of fertilizers (chemical and organic) and pesticides applied on the plots, number of household members who worked on the plots, and the amount of money spent on hired labor. Agricul- tural production was captured by crop productivity, calculated as the total quantity of crop produced by farmers divided by the size of area (in hectares) on which the crop was produced. Livelihood outcomes were captured by the value of agricultural production as a measure of agricultural income.8 Other livelihood outcomes include total income from keeping livestock and the amount of income obtained from non-agricultural activities. We implement the survey one to two months after harvest therefore we do not expect measurement error in agricultural production and livelihood income due to the length of the recall period. Any measurement error in these outcomes is expected to be orthogonal to treatment assignment as payments were conditional on the number of SLMPs present on the georeferenced at end- line and were unrelated to agricultural production. The last set of outcomes measured the extent to which knowledge was disseminated along the cascade training scheme. We captured interactions between contact and peer farmers by asking how frequently they met to discuss the SLMPs, how frequently the contact farmers verified if SLMPs were correctly adopted on the plots of the peers, and how often peer farmers asked their contact farmers for advice on the practices. Table 3 presents the baseline characteristics of our peer farmers as well as the results 8 We construct the main agricultural production outcomes based on farmers’ responses on how much they harvested of each crop on each agricultural plot. We convert quantities to kilograms and sum the produced quantities at the crop-farmer level. We also sum the estimated value of harvest from the crop-plot level to the farmer level as a measure of total agricultural revenue. We ask farmers about agricultural production at the crop-plot level to elicit deliberation and more precise responses. 20 of the balance tests across the three (sub-)treatment arms. Overall, our baseline sample included 1,914 farmers, 319 contact and 1,595 peer farmers.9 On average, the peer farmers in our sample were 41 years old, about 17 percent of them were female, and about 29 percent had at least some primary education. Furthermore, almost 75 percent were household heads, and 86 percent were living in rudimentary dwellings10 and, on average, managed less than two plots with a total surface of about five hectares. Regarding the relationship between peer and contact farmers, 43% of the peer farmers are kins of the contact; the remaining peers are neighbors, friends of the contact farmer, or acquaintances from social groups. Finally, note that the baseline use of SLMPs was quite low; on average, farmers used less than 2.5 SLMPs on their lands.11 9 With 10 contact farmers per commune and with five peers for each contact farmer, we intended to have a sample size of 1,600 peer farmers (and 320 contact farmers). Due to security concerns we were unable to reach and survey one contact farmer and his corresponding 5 peer farmers in the East region. 10 This is a measure of housing quality that captures one aspect of household wealth. It equals to one if the floor, the wall or the roof is made of rudimentary materials and zero otherwise. 11 Comparing the control to the pooled payment group show smaller and less significant differences, but the characteristics are still jointly correlated with the payment indicator, see Table A3 in the Appendix. 21 Table 3: Baseline characteristics of peer farmers, and the results of the pair-wise balance tests. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Total Control group Peer payment group Split payment group T-test (P-value) Normalized difference Variable N/[Clusters] Mean/SD N/[Clusters] Mean/SD N/[Clusters] Mean/SD N/[Clusters] Mean/SD (2)-(3) (2)-(4) (3)-(4) (2)-(3) (2)-(4) (3)-(4) Age 1595 41.377 600 42.382 500 40.542 495 41.002 0.148 0.205 0.545 0.169 0.128 -0.042 [32] (19.035) [12] (23.968) [10] (18.716) [10] (8.829) Female respondent (0/1) 1595 0.173 600 0.140 500 0.176 495 0.210 0.301 0.053* 0.273 -0.099 -0.186 -0.086 [32] (0.734) [12] (0.870) [10] (0.557) [10] (0.683) Respondent is Head of Household (0/1) 1595 0.740 600 0.775 500 0.746 495 0.693 0.504 0.065* 0.121 0.068 0.187 0.118 [32] (0.776) [12] (1.004) [10] (0.508) [10] (0.614) Has some primary education (0/1) 1595 0.285 600 0.298 500 0.320 495 0.234 0.683 0.050** 0.031** -0.047 0.144 0.191 [32] (0.721) [12] (0.801) [10] (0.748) [10] (0.480) Adults in household 1595 11.695 600 12.217 500 11.084 495 11.679 0.333 0.651 0.436 0.170 0.077 -0.092 [32] (20.279) [12] (24.799) [10] (18.273) [10] (17.042) Deprived house (0/1) 1595 0.858 600 0.875 500 0.802 495 0.893 0.258 0.721 0.178 0.200 -0.056 -0.253 [32] (1.216) [12] (1.025) [10] (1.463) [10] (1.188) Asset index 1595 -0.136 600 -0.058 500 -0.174 495 -0.191 0.479 0.470 0.962 0.052 0.058 0.008 [32] (9.141) [12] (8.869) [10] (8.669) [10] (10.772) Association membership (0/1) 1595 0.657 600 0.635 500 0.614 495 0.727 0.654 0.159 0.064* 0.043 -0.197 -0.241 [32] (1.437) [12] (1.740) [10] (1.400) [10] (1.016) Hired labor in previous agri. season (0/1) 1595 0.544 600 0.595 500 0.550 495 0.475 0.509 0.163 0.316 0.091 0.241 0.150 22 [32] (2.003) [12] (2.081) [10] (2.067) [10] (1.955) Number of plots under the control of the farmer 1595 1.699 600 1.802 500 1.516 495 1.760 0.051* 0.844 0.119 0.366 0.049 -0.297 [32] (2.708) [12] (2.667) [10] (2.329) [10] (2.891) Number of eroded plots 1595 2.453 600 2.588 500 2.294 495 2.451 0.191 0.549 0.485 0.211 0.097 -0.110 [32] (3.989) [12] (3.677) [10] (4.465) [10] (3.964) Landholdings (ha) 1595 4.976 600 5.168 500 4.018 495 5.712 0.014** 0.554 0.030** 0.269 -0.107 -0.396 [32] (14.935) [12] (13.233) [10] (5.677) [10] (20.956) Number of SLMPs adopted at baseline 1595 2.349 600 2.367 500 2.234 495 2.444 0.488 0.918 0.347 0.096 -0.056 -0.150 [32] (5.782) [12] (5.891) [10] (6.301) [10] (5.633) Contact farmer is family member (0/1) 1594 0.432 599 0.474 500 0.400 495 0.412 0.156 0.289 0.878 0.149 0.125 -0.025 [32] (1.090) [12] (0.727) [10] (1.035) [10] (1.479) Income from agricultural production (IHS transformed) 1595 12.815 600 13.086 500 12.894 495 12.409 0.269 0.347 0.553 0.092 0.227 0.158 [32] (11.334) [12] (3.806) [10] (3.909) [10] (19.936) Household has income from non-agricultural activities (0/1) 1595 0.518 600 0.538 500 0.494 495 0.519 0.229 0.739 0.733 0.089 0.038 -0.050 [32] (1.090) [12] (0.881) [10] (0.912) [10] (1.505) Notes : Average values of the characteristics for the total sample of peer farmers as well as for each of the three sub-samples thereof, are presented in columns (1)-(4); standard deviations are presented in parentheses. Columns (5)-(7) present the p -values for the treatment group indicators from regressing the characteristic on the treatment indicators and region fixed-effects. Standard errors are clustered at the commune level. Columns (8)-(10) present the normalized differences between each of the sub-treatment and the control group. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Comparing the average values of all the baseline characteristics across the three treat- ment arms, differences are generally small.12 The results of the pairwise t -tests are presented in columns (5)-(8) in Table 3. We find that peer farmers’ characteristics are reasonably well-balanced across the treatment groups. Only eight of the, in total, 48 differences are sta- tistically significant at the 10% level, including gender, age, farmers’ education level and the size of their land holdings, and the differences themselves are also relatively small. Balance is thus decent, and this conclusion is reinforced when assessing the size of the normalized differ- ences for each of the characteristics; see columns (8)-(10) of Table 3. Normalized differences are generally preferred to t-tests because they provide a scale-free comparison, and imbal- ances are typically identified as problematic if normalized differences exceed 0.25 (Imbens and Rubin, 2015; Abadie and Imbens, 2011). Only five out of the 48 normalized differences are about or larger than 0.25 standard deviation. Still, we follow Bruhn and McKenzie (2009) and mitigate the consequences of possible imbalances on our impact estimates by including all variables with significant differences as covariates in our regression analyses. While Table 3 presents the overall picture of the (differences in) the characteristics of our sample, it is of particular interest to also check balance for the various types of SLMPs present at baseline. As shown in Table 4 farmers used on average 2.35 practices at baseline, and about 90% of them were using at least one SLMP. That is, farmers are familiar with at least one practice, but the overall take-up rate of all the practices are still low compared to set of all practices. Regarding the presence of each of the nine SLMPs, three make up the bulk of the practices already in place, with usage rates of 37% and higher: heap and pit composting, use of agricultural and woody by-products, and controlled land clearing. At a usage level of 27.5%, stone and earth bunds were also quite widespread. Pit planting (“Zaï”), forage crop cultivation, and living hedges were practiced by only very few farmers. The low baseline usage rates of especially these techniques suggest that incentivizing SLMPs adoption can substantially improve the spread of SLMPs among farmers. 12 Balance tests for the subsample of contact farmers are presented in Table A5 of the Appendix. We find no major imbalances for this subsample either. 23 Table 4: Peer farmers’ usage of SLMPs at baseline (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Total Control group Peer payment group Split payment group T-test (P-value) Normalized difference Variable N/[Clusters] Mean/SD N/[Clusters] Mean/SD N/[Clusters] Mean/SD N/[Clusters] Mean/SD (2)-(3) (2)-(4) (3)-(4) (2)-(3) (2)-(4) (3)-(4) Number of SLMPs present at baseline 1595 2.349 600 2.367 500 2.234 495 2.444 0.488 0.918 0.347 0.096 -0.056 -0.150 [32] (5.782) [12] (5.891) [10] (6.301) [10] (5.633) At least one practice already present 1595 0.908 600 0.910 500 0.892 495 0.923 0.680 0.896 0.558 0.060 -0.048 -0.108 [32] (1.017) [12] (1.088) [10] (1.317) [10] (0.586) Zai 1595 0.045 600 0.045 500 0.030 495 0.059 0.774 0.646 0.535 0.078 -0.062 -0.139 [32] (0.852) [12] (1.015) [10] (0.530) [10] (0.968) Heap and pit composting 1595 0.470 600 0.480 500 0.456 495 0.471 0.462 0.570 0.852 0.048 0.019 -0.029 [32] (1.954) [12] (1.995) [10] (1.967) [10] (2.097) Stone and earth bounds 1595 0.275 600 0.278 500 0.254 495 0.293 0.717 0.865 0.605 0.055 -0.032 -0.087 [32] (1.319) [12] (1.399) [10] (1.218) [10] (1.441) Mowing and conservation of natural fodder 1595 0.173 600 0.210 500 0.150 495 0.152 0.341 0.284 0.975 0.155 0.151 -0.004 24 [32] (1.108) [12] (1.278) [10] (1.027) [10] (1.014) Forage crop cultivation 1595 0.059 600 0.053 500 0.052 495 0.073 0.991 0.727 0.670 0.006 -0.080 -0.086 [32] (0.828) [12] (0.659) [10] (0.844) [10] (1.053) Use of agricultural and wood by-products 1595 0.517 600 0.502 500 0.506 495 0.547 0.949 0.760 0.709 -0.009 -0.092 -0.083 [32] (2.410) [12] (2.448) [10] (2.676) [10] (2.335) Controlled clearing 1595 0.371 600 0.343 500 0.418 495 0.356 0.526 0.919 0.559 -0.154 -0.026 0.128 [32] (1.900) [12] (1.906) [10] (1.904) [10] (2.047) Assisted natural regeneration 1595 0.394 600 0.412 500 0.318 495 0.448 0.265 0.733 0.129 0.194 -0.074 -0.268 [32] (1.645) [12] (1.688) [10] (1.519) [10] (1.739) Living hedges 1595 0.046 600 0.043 500 0.050 495 0.046 0.679 0.835 0.808 -0.032 -0.015 0.016 [32] (0.353) [12] (0.318) [10] (0.252) [10] (0.492) Notes: : Average values, for the total sample of peer farmers as well as for each of the three sub-samples thereof, are presented in columns (1)-(4); standard deviations are presented in parentheses. Columns (5)-(7) present the p -values for the treatment group indicators from regressing the characteristic on the treatment indicators and region fixed-effects. Standard errors are clustered at the commune level. Columns (8)-(10) present the normalized differences between each of the sub-treatment and the control group. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Cascade training of farmers can ensure the accessibility of knowledge for farmers who only have experience with a small set of practices. This is especially evident if we look at the correlations from regressing the number of adopted practices at baseline on farmer characteristics. Table A6 shows that young, female, less educated, and poorer farmers tend to have fewer practices at baseline. For this group of farmers, formal sources of information tend to be less accessible, either due to a lack of a social network, financial means, or abilities to access and understand formal sources of information (Krishnan and Patnam, 2014; Vasilaky and Leonard, 2018; BenYishay and Mobarak, 2018). Incentivizing these groups to learn via existing social ties is potentially an effective method to disseminate information between these farmers. Below we will use these characteristics to test for possible heterogeneities in the impacts of SLMP adoption on agricultural productivity and livelihood outcomes. 4 The Impact of Conditional Payments on SLMP Adoption, and the Agricultural Consequences thereof 4.1 The Impact of Offering Conditional Payments on SLMP Adoption We start our analysis by addressing the first central question of this study: whether financial payments incentivize peer farmers to adopt SLMPs, and if so, what type of SLMPs are adopted – the already known ones, or also the ones that were used relatively little to date? We do so by pooling the peer payment and split payment groups, and then using equation (1) to estimate the impact of offering conditional payments on the number of SLMPs adopted. We then proceed by estimating the treatment effect for each of the nine SLMPs, to see what type of technologies saw the strongest increase in adoption. The results are presented in Table 5. As shown in column (1) of Table 5, offering conditional payments increased the number of SLMPs present at endline by about half a practice. This difference is statistically significant (with a p -value of 0.036 for the relevant t -test, and with a value of the F -test of 62.62) and it is also sizeable as it represents an increase of 0.381 standard deviation (or an 18% increase in SLMP usage). This result is robust to re-estimating the model using negative binomial regression that takes into account the count nature of the SLMP usage data; see Table A10 in the Appendix. 25 The average increase of half a practice adopted by peer farmers can mask substantial variation in uptake between the nine practices. Columns (2)-(10) of Table 5 present the treatment impact estimates of the payment incentives on the likelihood of adopting each of the nine SLMPs. All nine models are estimated simultaneously using seemingly unrelated regressions, because time constraints, land constraints and/ or technical complementarities and substitutabilities may result in the various SLMP adoption decisions being correlated. We find that the treatment increased adoption of almost all SLMPs. Four saw the largest increase in uptake – the establishment of stone and earth bunds, mowing and conservation of natural fodder, assisted natural regeneration, and living hedges; see Columns (4), (5), (9) and (10) of Table 5. The estimates range from 8 to 14 percentage point increases in the share of farmers adopting these practices, corresponding to effect sizes of between 0.20 and 0.41 standard deviation. The practices that saw the largest increases in take-up cover each of the three agricultural domains presented in Table 1; they were not concentrated in just one or two domains. We also observe positive and significant effects on the adoption of earth and stone bunds, which is typically thought of as labor intensive, and on the adoption of assisted natural regeneration and living hedges, which require investments in protecting existing trees and in planting shrubs (Liniger et al., 2011a). This suggests that offering conditional payments did not necessarily induce farmers to just adopt those SLMPs with the least costs. Finally, the practices with the highest percentage point increases in usage were among the technologies least frequently used in the region (as measured by their usage in the control group). This holds especially for mowing and conservation of natural fodder and living hedges (see Columns (5) and (10)).13 The conditional payments thus seem to have induced farmers to especially adopt the less known (or at least the lesser used) practices. 13 Zaï and forage crop production also had very low baseline utilization rates with, in relative terms, fairly large increases in uptake. These effects are not measured with sufficient precision for these impacts to be statistically significant; see Columns (2) and (6) in Table 5. 26 Table 5: Treatment effects on the average number of SLMPs in use as well as on the usage of each of the nine individual SLMPs. Agronomy SLMPs Agro-sylvo-pastoral SLMPs Agroforestry SLMPs # SLMPs Zai Heap and pit Stone and Mowing and cons. Forage crop Use of agr. and Controlled Assisted Living adopted composting earth bunds of nat. fodder cultivation wood by-products clearing natural regeneration hedges (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Payment treatment 0.507∗∗ 0.024 0.083 0.101∗ 0.131∗∗ 0.040 -0.057 -0.012 0.138∗ 0.084∗∗∗ (0.232) (0.032) (0.058) (0.055) (0.055) (0.064) (0.062) (0.104) (0.071) (0.021) F-Statistic 62.623 Observations 1574 1574 1574 1574 1574 1574 1574 1574 1574 Baseline outcome Yes No No No No No No No No No 27 Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Control mean 2.784 0.070 0.564 0.352 0.148 0.084 0.619 0.617 0.268 0.060 Control std.dev. 1.331 0.256 0.496 0.478 0.355 0.277 0.486 0.486 0.444 0.238 Effect size (in std.dev.) 0.381 0.095 0.167 0.210 0.370 0.144 -0.118 -0.024 0.311 0.354 Unit # share share share share share share share share share Notes: Column (1) is estimated using OLS regression. Columns (2)–(10) are estimated simultaneously using Seemingly Unrelated Regression. Standard errors, presented in parentheses, are clustered at the commune level. Covariates include age, gender, a dummy variable that captures whether the farmer finished primary education, household size, a household asset index, the farmer’s total agricultural land area, a binary variable that captures whether the farmer made use of hired labour, the number of plots threatened by erosion, the knowledge score of the peer farmer’s contact farmer, and a binary variable on whether household has non-agricultural sources of income. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. 4.2 The Short-run Productivity Impacts of SLMP Adoption and the Liveli- hood Consequences We thus document a sizeable increase in SLMP usage as induced by the prospect of receiving cash transfers conditional on the number of SLMPs used at endline. In this subsection we exploit the resulting exogenous variation in SLMP usage to estimate the short-run impact of SLMP adoption and usage on agricultural productivity. Whether our RCT provides reliable evidence on the short-run productivity impacts crucially depends on whether our treatment instrument, the prospect of receiving conditional payments well after the end of the current agricultural season, increased SLMP adoption without having affected any of the constraints faced by the farmers – especially their financial constraints. In other words, the key question is whether our instrument violated the 2SLS model’s exclusion restriction by affecting the key outcome variable of interest, agricultural revenues at endline, via mechanisms other than just via increased SLMP usage. In Section 4.2.1 we provide (suggestive) evidence that indeed offering future compensation conditional on current usage did not appreciably affect farmers’ budget constraints during the current agricultural season. In Section 4.2.2 we turn to estimating the short-run impacts on agricultural productivity, revenues, and livelihoods. 4.2.1 Future Payments and Current Budget Constraints Payments were not disbursed until at least seven months after the end of the 2019 agricultural season (see Figure 3). Still, the prospect of future payments may have affected endline agricultural revenues via mechanism other than just via increased SLMP take-up. For this to be the case, two conditions need to have been met. First, the prospect of future payments should have facilitated access to credit, allowing the farmers to borrow against these future cash transfers. Second, they should have invested these funds to optimize their production process – by acquiring more land, by hiring more labor, and/ or by purchasing more inputs such as chemical fertilizer and pesticides. We now explore the relevance of each of these two conditions in turn. We provide two pieces of (suggestive) evidence that farmers did not borrow against their future payments. First, access to credit is typically very poor among farmers in Burkina Faso. According to the 2017 Global Findex Database collected by the World Bank (Demirgüç- 28 Kunt et al., 2018), only 12% of the Burkinabé population borrowed from formal financial institutions, and only 6% of the Burkinabé borrowed money for agriculture. Adjognon et al. (2017) document that in Sub-Saharan Africa, contrary to conceived wisdom, traditional credit use, formal or informal, is extremely low (across credit type, country, crop and farm size categories), and also that farmers primarily finance modern input purchases with cash from nonfarm activities and crop sales. Our second piece of evidence on the implausibility of the access to credit channel is based on evaluating whether treatment households invested more in either alternative sources of income or in acquiring productive assets. In Table 6 we present the regression results aimed at detecting treatment differences in either non-agricultural sources of income, or in livestock revenues. Livestock is an asset farmers are especially likely to channel their funds to for two reasons: as a potentially productive asset (including the production of meat and milk; Wouterse and Taylor, 2008; Balboni et al., 2021), or to diversify risk (Fafchamps et al., 1998; Carter and Lybbert, 2012; Janzen and Carter, 2019). The results of this test are presented in Table 6. To mitigate the impact of outliers and to avoid having to drop those farmers from the analysis with no livestock or non-agricultural income, the dependent variables in this table are the inverse hyperbolic sine (IHS) transforma- tions of the revenues of livestock and non-agricultural sources of income. Using IHS implies that the treatment estimates are semi-elasticities; the coefficients presented can thus be in- terpreted as percentage changes (Burbidge et al., 1988; Bellemare and Wichman, 2020).14 As is clear from Table 6, we do not find any evidence that treatment farmers were more likely (or more able) to invest in these alternative sources of income than the farmers in the control group. Next, we assess whether the agricultural input mix used by treatment farmers is markedly different from that used by the farmers in the control group. We can test this by estimating to 14 Taking the natural logarithm is the standard way of reducing the impact of outliers on coefficient esti- mates, and the treatment coefficients can then be interpreted as percentage changes. However, the logarithmic transformation results in all observations being dropped√ from the analysis that have a zero value for the vari- able of interest. The IHS of variable z equals ln(z + z 2 + 1). Because essentially the IHS transformation boils down to shifting up the ln(z ) function by a constant (equal to ln(2)) for values of z that are not too close, the elasticities generated by the two functions are very similar as well (Bellemare and Wichman, 2020; Aihounton and Henningsen, 2021). We therefore apply the IHS transformation in all models in which either livestock income or non-agricultural income is the dependent variable; see Tables 8 and 9 as well as Figure 4. We also apply the IHS transformation to one other variable for which many observations have a zero value – hired labor; see Table 7. Because our explanatory variable of interest (T in equation (1), with coefficient τ ) is a dummy variable, the treatment’s percentage impact is approximated by eτ − 1 ≈ τ . 29 Table 6: Intention-to-treat effects of conditional payments on income from livestock and non-agricultural activities. IHS Income Livestock Non-agricultural (1) (2) Payment treatment 0.010 0.022 (0.110) (0.402) R2 0.142 0.105 Observations 1231 1549 Adjusted R2 0.129 0.094 Baseline outcome Yes Yes Covariates Yes Yes Region FE-s Yes Yes Control mean 9.915 5.313 Control std.dev. 4.283 5.931 Unit IHS(FCFA) IHS(FCFA) Notes: Both models are estimated using OLS. Standard errors, presented in parentheses, are clustered at the commune level. To control for outliers and because of the relatively high incidence of participants without livestock and/or non-agricultural sources of income, the dependent variables are transformed by applying the inverse hyperbolic sine (IHS) function. The treatment coefficients can thus be interpreted as percentage changes.For the vector of covariates, see Table 5. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. what extent treatment status affected the use of agricultural inputs including land, labor and fertilizer. While the adoption of SLMPs may affect the optimal input mix, changes therein are expected to be relatively minor unless the prospect of future payments substantially reduced the farmers’ current budget constraints. The results of this test are presented in Table 7. Overall, we find that the use of inputs is not markedly different between households in the treatment groups compared to those in the control group. The impacts are typically small and statistically insignificant. We do not find any changes in the area cultivated, the method of sowing, the number of household members who worked on the plots, the total cost of hired labor, and in the amount of chemical or organic fertilizers used in response to the payments for adoption; see Columns (1)-(6) of Table 7.15 While most of the changes in input use are relatively small and statistically insignificant, this does not hold for pesticides use. As shown in Column (7) of Table 7, having been 15 Although the coefficient on hired labor is not statistically significant, its point estimate implies that the prospect of receiving future payments results in a e−0.44 − 1 ≈ 35.3% decrease in the amount of money spent on hired labor (≈ USD$ 10). This seems like a large decrease, but it is inflated by the large share of farmers who are not using hired labor and the concave shape of the inverse hyperbolic function. These insights are reinforced by our analyses of the differences in the share of farmers employing hired labour and in the amount spent; see Columns (1) and (2) of Table A9. Neither of these shares differs between the payment treatment and the control groups. 30 Table 7: The impact of treatment status on agricultural input use. Cultivated # Manually Household IHS transf. of Use chem. Use org. Use Input area sowed plots labor hired labor cost fertizer fertilizer pesticides index (1) (2) (3) (4) (5) (6) (7) (8) Payment treatment 0.106 0.023 -0.702 -0.439 22.671 150.513 3.447∗∗∗ 0.047 (0.082) (0.050) (0.573) (0.399) (27.461) (201.122) (1.223) (0.037) Constant 0.465 1.015 -5.050 1.516 12.809 2386.687 25.741∗ -0.382 (0.886) (0.703) (5.161) (4.716) (333.630) (1999.779) (12.963) (0.402) Observations 1574 1574 1545 1574 1574 1574 1574 1574 Adjusted R2 0.925 0.762 0.144 0.242 0.294 0.178 0.207 0.522 Baseline outcome Yes Yes No No No No No No Covariates Yes Yes Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Yes Yes Control mean 4.874 1.641 3.634 6.089 161.010 1141.435 7.161 0.004 Control std.dev. 4.753 0.825 5.187 5.265 265.891 2173.358 10.952 0.550 Effect size (in std.dev.) 0.022 0.028 -0.135 -0.083 0.085 0.069 0.315 0.085 Unit Hectare Share # HH member IHS(FCFA) Kilogram Kilogram Liter Std.dev. Notes: All columns are estimated using OLS. Standard errors, presented in parentheses, are clustered at the commune level. As not all farmers use hired labor, expenditures on hired labor are transformed using the inverse hyperbolic sine function (IHS). The treatment estimates can thus be interpreted as percentage changes. For the vector of covariates, see Table 5. The input index is calculated as the unweighted average of the normalized values of the inputs in Columns (1)-(7) of this table (Kling et al., 2007). ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. offered conditional payments increased the usage of pesticides by almost 50 percent (from, on average, 7.2 liters in the control group to 10.6 liters in the treatment group). Based on a national average price of $7 per liter of pesticides (or 4000 FCFA; USDA, 2017), this would imply an increase in average expenditures of about $25. This increase in usage may be the result of some of the actors in the farmer’s supply chain to be willing to extend credit, because of the prospect of the farmer receiving payments in the due time. If this is the mechanism, then it is surprising to see that the supply chain’s willingness to offer credit was just limited to purchasing pesticides and not, for example, chemical fertilizers. An alternative explanation may be that the change in practices applied increased the marginal productivity of some inputs – in casu pesticides – resulting in an increase in the farmers’ willingness to purchase those inputs. Overall, the results presented in Tables 6 and 7 suggest that the prospect of receiving conditional payments in the future did not result in a large and instantaneous relaxation of the treatment farmers’ budget constraints during the current agricultural season. Thus it is unlikely that the treatment affected agricultural productivity via mechanisms other than increased SLMP usage. This conclusion is reinforced by the results presented in Column (8) of Table 7, where we estimate the impact of treatment status on an index of input use, constructed by averaging the normalized usages of each of our seven input variables.16 We 16 The index is created by standardizing each of the seven inputs using the means and standard deviations 31 do not find any impact of having been offered conditional payments on this index: the point estimates indicates a change in the index of just 0.047 standard deviations. 4.2.2 The Impact of SLMP Usage on Agricultural Productivity and Farmers’ Livelihood Outcomes Having established that it is not very likely that the prospect of receiving conditional pay- ments affected agricultural outcomes via mechanisms other than via the increase in SLMP take-up, we now employ our two stage least squares model (see equations (2) and (3) in Section 3.1) to estimate the short-run impact of SLMP adoption on farm income and on the agricultural productivity of some of Burkina Faso’s most important crops. Using the treat- ment status of farmers to instrument SLMPs adoption, the first stage regression analyses have already been presented in Column (1) of Table 5. We now focus our attention on the second stage results presented in Table 8.17 Table 8: The impact of SLMP usage on farmers’ agricultural productivity and revenues. IHS Transfromed Income Productivity Agriculture Maize Millet Sorghum Cowpea Index (1) (2) (3) (4) (5) (6) # SLMPs 0.402∗ 320.018 379.850∗∗ 678.756∗ 86.331 0.482∗∗ (0.241) (232.197) (169.921) (405.474) (64.283) (0.217) Observations 1532 994 721 897 371 1467 R2 0.192 -0.013 -0.190 -0.802 0.083 -0.390 Adjusted R2 0.182 -0.040 -0.235 -0.855 0.016 -0.414 Baseline outcome Yes No No No No No Covariates Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Control mean 13.162 1151.229 717.798 745.729 289.560 -0.074 Control std.dev. 1.122 1213.386 933.566 726.059 411.660 0.759 Effect size (in std. dev.) 0.358 0.264 0.407 0.935 0.210 0.635 Unit IHS(FCFA) kg/ha kg/ha kg/ha kg/ha Std.dev. Notes: All columns are estimated using two-stage least squares regression. Standard errors, presented in parentheses, are clustered at the commune level. The dependent variable in Column (1) is the inverse hyperbolic sine of agricultural income; the coefficient can thus be interpreted as a percentage change. For the vector of covariates, see Table 5. The productivity index is calculated as the unweighted average of the normalized productivities of the four crops in Columns (2)–(5) (Kling et al., 2007). ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Column (1) of Table 8 presents our estimate of the marginal impact of SLMP usage on from the control group, and subsequently calculating the unweighted average. Following Kling et al. (2007) we do so to reduce the number of statistical tests (addressing multiple hypothesis testing) and to test the overall effect of payments on input use. The other indices presented in the remainder of this paper are constructed in essentially the same fashion (see Tables 8, 10, 11, and 12); the only difference is that in some instances some of the constituent variables needed to be recoded to realign their ordinal interpretation with those of the other variables that are included the index. 17 For the intention-to-treat impact estimates, see Table A7. 32 farmers’ agricultural income. We find that having one additional SLMP in place increases the adopter’s agricultural revenues by almost 40%. Taking the average agricultural production value of FCFA 468,139 (or ≈ $USD 790) in the control group, this percentage point difference amounts to $USD 316 (or 0.35 standard deviation). Our estimates are in line with the results of earlier studies that estimate the production gains from specific practices. BenYishay and Mobarak (2018) find that in the second year of their study in Uganda, pit planting had increased agricultural productivity by 19%, and also that composting had increased productivity by 105%. Similarly, Takahashi et al. (2019) report that the adoption of improved rice management techniques in Côte d’Ivoire resulted in a 46% increase in rice yields in the first year since adoption. Our study confirms these studies’ insights that SLMPs are not just likely to be productivity enhancing in the longer run, but in the shorter run as well. In the remaining columns of Table 8 we present the marginal impact of SLMP adoption on productivity of Burkina Faso’s four key crops (see Columns (2)–(5)) as well as on an agricultural productivity index thereof (see Column (6)). We show the productivity impact of SLMPs on Burkina Faso’s four most important crops: maize, millet, sorghum, and cowpea (FAO, 2021). We find sizeable productivity impacts of SLMPS adoption on all crops, and also on the overall productivity index. Yield increases range between 86 kg/ha (for cowpea, 0.21 standard deviation) and 678 kg/ha (for sorghum, 0.94 standard deviations), although only the yield responses of millet and sorghum are measured with sufficient precision to be significant. Overall, we find that adopting an additional SLMP raises productivity by almost half a standard deviation; see Column (6). 4.3 Heterogeneous Effects on Agricultural Livelihood Having documented that adopting an additional sustainable land management practice in- creases agricultural productivity and income even in the short run, we explore the importance of impact heterogeneity. Are there any subgroups among the peer farmers who benefited more than proportionally from the adoption of SLMPS? Or possibly more important, are there subgroups for whom SLMP adoption resulted in a decrease in agricultural productivity? To answer these questions, we first compare the distribution of the (IHS transformed) agricul- tural income between the different groups in a series of quantile regressions (Koenker and Hallock, 2001; Abadie et al., 2002; Koenker, 2005). We plot the estimated effects and the 33 corresponding 95% confidence interval for each quantile in Figure 4. The plot shows that the effects are largest for farmers at the lower end of the distribution (starting from an impact of 0.4, roughly a 40% increase in agricultural revenues), and also that the impact becomes smaller when moving through the various quantiles until they become insignificant at the 75th quantile. Figure 4: Quantile treatment effects on agricultural income. Notes: The continuous line in this figure presents the intention-to-treat impact estimates on the inverse hyperbolic sine of agricultural income using quantile regressions; the vertical axis of this figure thus reflects percentage changes. The percentage change impacts are estimated at 5percentile intervals, from the 5th to the 95th percentile. The shaded area around the continuous line is the 95% confidence interval around the corresponding point estimate, which is calculated using robust standard errors clustered at the commune level. Having documented that farmers with lower productivity tend to benefit more, we now proceed to use equation (1)’s intention-to-treat approach to test for the existence of heteroge- neous treatment effects for some of the key farmer characteristics for which we have baseline data. Column (1) of Table 9 presents our baseline estimate of the average impact of the conditional payment treatment on agricultural revenues. As shown in Columns (2)–(6) of Table 9, we do not find any evidence for heterogeneous treatment effects. Here, non-binary variables (like plot size and asset index) were re-coded into a binary variables so they take on a value of one for above-median values, and zero otherwise. As show in Table 9, the benefits in terms of agricultural income are similar regardless of gender (Column (2)), peer farmers’ education (Column (3)), wealth (Column (4) and (5)), or their number of degraded plots 34 (Column (6)). Based on the results presented in Figure 4 and in Table 9, we thus conclude that adoption of SLMPs is beneficial even in the first year, and especially so for those farmers with the lowest levels of agricultural income. Table 9: Testing for heterogeneous treatment effects on agricultural incomes. Average effect Heterogenous treatment effects No interaction Female Education Asset Farmsize Eroded plots (1) (2) (3) (4) (5) (6) Payment treatment 0.225∗∗ 0.216∗∗ 0.205∗∗ 0.256∗∗ 0.244∗∗∗ 0.238∗∗ (0.089) (0.097) (0.083) (0.097) (0.087) (0.094) Payment treatment × Covariate 0.085 0.082 -0.053 -0.060 -0.030 (0.140) (0.107) (0.123) (0.130) (0.092) Observations 1542 1542 1542 1542 1542 1542 Adjusted R2 0.420 0.421 0.421 0.423 0.445 0.420 Wald-Test p( βP aym + βP aym×Covar. = 0) 0.031 0.045 0.097 0.143 0.066 Baseline outcome No No No No No No Covariates Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Control mean -0.048 -0.048 -0.048 -0.048 -0.048 -0.048 Control std.dev. 0.779 0.779 0.779 0.779 0.779 0.779 Effect size (in std.dev.) 0.289 0.386 0.369 0.260 0.237 0.267 Unit IHS(FCFA) IHS(FCFA) IHS(FCFA) IHS(FCFA) IHS(FCFA) IHS(FCFA) Notes: All columns are estimated using OLS regression. The outcome variable is inverse hyperbolic since transformed agricultural income, therefore coefficients can be interpreted as percentage changes. In columns (2)-(6), the coefficient of the interaction between the treatment indicator and the dimension of heterogeneity is presented. Standard errors clustered at the commune level are in parentheses. For the vector of covariates, see Table 5. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01e. 5 Conditional Payments and Knowledge Dissemination As shown in Section 4.1, we find that offering financial incentives resulted in increased SLMP adoption. One mechanism for this impact is that the prospect of receiving conditional pay- ments directly improves the expected cost-benefit ratio of SLMP adoption. A second possible mechanism is that the treatment-induced increase in the profitability of SLMP adoption may have affected the effectiveness of the cascade training program. The extent to which infor- mation dissemination occurs in such programs is determined by both peer farmers’ demand for and the contact farmers’ supply of information. Did the increased (perceived) benefits of SLMP adoption result in an increase in peer farmers’ demand for information on how to implement them? And if so, did this increased demand induce contact farmers to provide it? We answer these questions by assessing the payments’ impact on the knowledge transmis- sion process in the cascade training scheme. We hypothesize that information about SLMP implementation is more valuable the higher the stakes associated with SLMP adoption, and 35 hence that higher stakes increase both the demand for and supply of information. In the presence of conditional payments peer farmers’ willingness to pay for information is higher, and if the market for information is efficient, more information will be exchanged. As a corollary and reminiscent of the Coase theorem, note that if the market for information is efficient, only the size of the surplus matters – not the initial allocation of the payment. Whether this is true or whether providing contact farmers with a direct financial stake in peer farmer adoption increases SLMP uptake, is an open question. In this section we thus test (i) whether the cascade training scheme works better if the adoption stakes are higher, and (ii) whether or not the flow of information can be improved further – for the same sur- plus – when giving contact farmers a direct financial stake in their peer farmers’ adoption decisions. In the next two subsections we test each of these two hypotheses. 5.1 The Impact of Higher Adoption Stakes on Information Dissemination In this subsection we test whether higher adoption stakes improve the exchange of knowledge and information in the cascade training program. We do so by combining observed behavior and survey evidence; see Table 10.18 The amount of information exchanged is the outcome of both demand and supply. Column (1) of Table 10 documents that offering conditional payments resulted in a 12 percentage points increase in the share of peer farmers having asked their contact farmers for advice (from 35% in the control group to 47% in the pooled treatment group). As shown in Column (2), this higher demand indeed resulted in more intensive information sharing: the share of peer farmers indicating that they frequently met with the contact farmer to discuss SLMP implementation is 13 percentage points higher in the treatment group than in the control group (about 55% in the pooled treatment group compared to 41% in the control group). Next, Columns (3) and (4) of Table 10 indicate that the treatment-induced increase in the demand for information induced the contact farmers to increase their information supply, in two ways. Although this difference fails to be significant at conventional levels (with a p = 0.11), Column (3) shows that contact farmers in the payment group had higher SLMP adoption rates than those in the control group. Since payments were offered conditional on peer farmer adoption, the adoption of more SLMPs by 18 In the table, as measures of knowledge exchange, we use the observed number of SLMPs adopted by contact farmers and a series of indicator variables on contact-peer farmer interaction which are equal to one if the frequency of interaction was at least once per month and zero otherwise. 36 contact farmers in the payment communities plausibly reflects a stronger willingness to lead by example and/ or to use their own land as demonstration plots. And as shown in Column (4), contact farmers in the payment communities are also found to have a higher propensity to actively engage in on-site monitoring and verification of whether and how peer farmers adopted the practices. Consequently, as shown in Column (5), these treatment-induced increases in information supply and demand mitigated the importance of the lack of SLMP implementation know-how as a barrier to adoption. The share of peer farmers not having adopted SLMP because of a lack of knowledge is estimated to be about 11 percentage points lower in the pooled treatment group than in the control group. Finally, upon combining all indicators into a single index together, we again find a significant increase of 0.27 standard deviation; see Column (6). We thus find that larger adoption surpluses (associated with offering financial incentives) increases the effectiveness of the cascade training program. Table 10: Treatment effects on contact farmer’s adoption and communication between farm- ers about SLMPs. PF asked Farmers # SLMPs adopted CF monitor % of SLMPs not adopted Knowl. exch. for advice discuss SLMPs by CF PF adoption due to lack of knowledge index (1) (2) (3) (4) (5) (6) Payment treatment 0.119∗ 0.134∗∗ 0.398 0.112∗ -0.110∗∗∗ 0.270∗∗∗ (0.063) (0.061) (0.245) (0.063) (0.033) (0.095) Observations 1573 1573 315 1573 1571 1574 Adjusted R2 0.078 0.073 0.268 0.082 0.119 0.189 Baseline outcome Yes Yes Yes Yes No Yes Covariates Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Control mean 0.356 0.408 3.900 0.336 0.231 0.356 Control std.dev. 0.479 0.492 1.746 0.473 0.263 0.479 Effect size (in std.dev.) 0.249 0.272 0.228 0.237 -0.416 0.564 Unit share share # share share Std.dev. Notes: All columns are estimated using OLS regression. Standard errors clustered at the commune level are in parentheses. For the vector of covariates, see Table 5. Upon having recoded the knowledge barrier variable (such that higher values reflect barriers being less important), the knowledge exchange index is calculated as the unweighted average of the normalized values of the indicators presented in Columns (1)–(5) of this table (Kling et al., 2007). ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. 5.2 The Impact of the Payment Structure on Information Dissemination and SLMP Uptake We thus find that offering conditional payments improved exchange of SLMP adoption infor- mation between peer and contact farmers. Plausibly the underlying mechanism is that the conditional payments increased the (expected) benefits of SLMP adoption, and thereby peer farmers’ demand for information. In this subsection we analyze whether the effectiveness of the cascade training program can be further enhanced by explicitly incentivizing, for the 37 same surplus, not just the demand for information but also its supply – by giving contact farmers a direct financial stake in their peer farmers’ adoption decisions. We do so by testing whether the results regarding demand and supply, as documented in Table 10, differ between the peer and split payment treatments. Whether information dissemination can be furthered even more by explicitly incentivizing the supply side is of obvious importance for the design of cascade training programs. However, it is also important from a more general perspective – whether it is necessary to continue to offer conditional adoption payments if, over time, farmers learn increasingly more about the beneficial consequences of SLMP adoption. The argument is as follows. As stated in the introduction, the perceived lack of private benefits is one adoption barrier, the lack of SLMP implementation know-how is another. If more and more SLMPs are being adopted, farmers who have not adopted yet receive increasingly more signals about the productivity impacts of SLMP adoption in the short and in the longer run, and hence the perceived insufficient profitability barrier is likely to become less important over time. However, the information barrier may continue to remain important unless the expected profitability-induced increase in demand for information results in an increased willingness among contact farmers to share this information. Differently stated, if the market for information is efficient, it may still be important for the government to offer the cascade training program, but it would no longer be necessary to financially incentivize SLMP adoption, nor to provide explicit incentives for information dissemination by contact farmers. Table 11 provides insight into whether, for the same surplus, the exchange of information is better when providing contact farmers with a direct financial stake in their peer farmers’ adoption. This table repeats the analysis of Table 10 by comparing the knowledge demand and supply indicators between the two conditional cash sub-groups, the peer and the split payment ones. These two treatments only differ in how the payment is split between the peer and the contact farmer, so that the size of the surplus remains the same. Column (1) of Table 11 shows that giving contact farmers a direct financial stake in their peers’ adoption does not result in significantly more practices being adopted by the peer farmers. We also do not find that the demand for information (Column (2)), the frequency of interaction (Column (3)), or the contact farmers’ efforts to disseminate knowledge (Columns (4) and (5)) are significantly different between the peer and split payment group. 38 And in line with the lack of significant differences in Columns (1)-(5), we also find no difference in neither the share of peer farmers reporting lack of knowledge as a barrier of adoption across the two groups (in Column (6)), nor in the overall knowledge dissemination index (see Column (7)). Therefore adoption of the SLMPs and knowledge dissemination do not depend on how the financial payment for adoption is allocated between the peer and the contact farmer. Table 11: The impact of offering direct financial incentives for information dissemination to the contact farmers. # SLMPs PF asked Farmers # SLMPs adopted CF monitor % of SLMPs not adopted Knowl. exch. adopt. by PFs for advice discuss SLMPs by CF PF adoption due to lack of knowledge index (1) (2) (3) (4) (5) (6) (7) Split payment 0.048 0.065 0.031 0.501 0.074 -0.024 0.134 (0.274) (0.068) (0.073) (0.320) (0.093) (0.040) (0.101) Observations 978 978 978 195 978 976 978 Adjusted R2 0.342 0.091 0.124 0.345 0.117 0.166 0.255 Baseline outcome Yes Yes Yes Yes Yes No Yes Covariates Yes Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Yes Control mean 3.307 0.435 0.521 4.061 0.415 0.136 0.237 Control std.dev. 1.925 0.496 0.500 1.963 0.493 0.218 0.674 Effect size (in std.dev.) 0.025 0.131 0.062 0.255 0.150 -0.112 0.199 Unit # share share # share share Std.dev. Notes: All columns are estimated using OLS regression. Standard errors clustered at the commune level are in parentheses. For the vector of covariates, see Table 5. Upon having recoded the knowledge barrier variable (such that higher values reflect barriers being less important), the knowledge exchange index is calculated as the unweighted average of the normalized values of the indicators presented in Columns (2)–(6) of this table (Kling et al., 2007). ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. So we find that offering financial incentives was equally effective in the peer and split payment groups. Providing contact farmers with a direct financial stake in their peer farmers’ SLMP adoption decisions does not affect the amount of effort the contact farmer provides in disseminating information on and in assisting the implementation of the sustainable land management technologies, and also the outcomes in terms of number of SLMPs adopted and of knowledge exchanged, are independent of the payment’s initial allocation. We consider two explanations to better understand the irrelevance of the payment allocation between peer and contact farmers. One possible reason is what we will loosely refer to as “altruism” – the relationship between the contact and peer farmer is such that the contact farmer is indifferent who ends up receiving what share of the payment. This can indeed be altruism within or between families, but it can also be more mechanical, in case the contact and peer farmer have a shared budget – for example if they are members of the same family. Indeed, no fewer than 43% of the peer farmers are kindred to their contact farmer, and hence the lack of a difference between the peer and split payment schemes may be due to the strength of (extended-) family ties (see Table A8). 39 A second possible explanation is that contact farmers are only willing to put in effort if they themselves become better off too, but that markets for information are perfect in the Coasean sense – independent of how the payments are allocated, (unobservable) side payments ensure implementation of the efficient amount of information dissemination. If this is indeed the case, we expect peers and contact farmers who are not in the same family to behave similarly in the two payment groups. If the markets for information are indeed efficient, the adoption of the SLMPs and knowledge dissemination should again be indepen- dent of how the financial payment for adoption is allocated between the peer and the contact farmer. We test the relevance of both the altruism mechanism and the efficient market hypothesis by re-estimating Table 11 and allowing for heterogeneous treatment effects based on kinship. More specifically, the key covariates of interest in our regression model are the split payment group indicator, a kinship indicator (which is 1 if the peer and contact farmer in a dyad are family related, and 0 otherwise), as well as their interaction term. In this model, the omitted category is thus the group of peer and contact farmers in the peer payment treatment who are not kindred. The coefficients on each of the three dummies can then be interpreted as follows. First, the coefficient on kin captures whether the outcome variable of interest differs depending on whether the individuals in the peer-contact dyad are family, for those peer and contact farmers in the peer payment group. A positive coefficient would be suggestive evidence of the importance of altruism in the transfer of information. Second, the coefficient on the split payment dummy captures whether the initial payment allocation affects the outcome variable of interest if the peer and contact farmer in the dyad are not kin. A significant coefficient would provide evidence that the initial payment allocation matters, and that markets are not fully efficient. Third, the coefficient on the interaction term (split payment times kin) reflects whether the initial payment allocation still affects the information exchange even if both farmers in the dyad are kin. The results of this analysis are presented in Table 12. We find that kinship matters at least to some extent, albeit that the outcomes are ambiguous; see the first row of Table 12. If the peer and contact farmers in a dyad are kin, they meet more often to discuss SLMP implementation (9.5 percentage points, or 0.19 standard deviations; see Column (3)). How- ever, whether this is indeed indicative of an improved exchange of information is not obvious, 40 as we also find that lack of knowledge is a barrier for the adoption of practices for a larger share of these farmers (3.5 percentage point, or 0.13 standard deviation; see Column (6)). Next, in all six models the coefficients on both the split payment indicator as well as on the interaction term are statistically insignificant. The lack of significance of the coefficient on the split payment dummy suggests that indeed the markets for information are sufficiently efficient to not affect the final outcomes, and the lack of significance on the interaction term suggests that the initial payment allocation is by and large inconsequential for the flow of information among kindred dyads as well. Table 12: Treatment effects of split payments compared to peer payment along family ties. # SLMPs PF asked Farmers # SLMPs adopted CF monitor % of SLMPs not adopted Knowl. exch. adopt. by PFs for advice discuss SLMPs by CF PF adoption due to lack of knowledge index (1) (2) (3) (4) (5) (6) (7) Kin -0.052 0.016 0.094∗ 0.538 0.059 0.035∗ 0.072 (0.147) (0.048) (0.045) (0.584) (0.054) (0.018) (0.055) Split payment 0.022 0.031 0.026 0.289 0.041 -0.008 0.099 (0.326) (0.067) (0.075) (0.878) (0.104) (0.041) (0.099) Split payment × Kin 0.198 0.072 0.005 0.220 0.072 -0.040 0.073 (0.259) (0.074) (0.067) (0.848) (0.078) (0.032) (0.093) Observations 978 978 978 195 978 976 978 R2 0.317 0.110 0.141 0.439 0.135 0.191 0.280 Adjusted R2 0.301 0.089 0.122 0.349 0.115 0.166 0.258 Baseline Outcome Yes Yes Yes Yes Yes No No Covariates Included Yes Yes Yes Yes Yes Yes Yes Region fixed effects Yes Yes Yes Yes Yes Yes Yes T-tests (p-values) – β80/20 + β80/20×Kin =0 0.527 0.275 0.733 0.099 0.264 0.274 0.187 – β80/20 = βKin 0.813 0.822 0.397 0.694 0.851 0.317 0.804 – β80/20×Kin = 0 0.346 0.936 0.798 0.367 0.229 0.445 Control mean 2.784 0.356 0.408 3.900 0.336 0.257 0.257 Control std.dev. 1.670 0.479 0.492 1.746 0.473 Effect size (in std.dev.) – β80/20 0.013 0.064 0.052 0.166 0.086 -0.029 0.369 – β80/20×Kin 0.119 0.150 0.011 0.126 0.152 -0.151 0.273 Notes: All columns are estimated using OLS regression. In the regressions, the split payment treatment indicator is fully interacted with an indicator on whether the peer and contact farmers are kin. Standard errors clustered at the commune level are in parentheses. For the vector of covariates, see Table 5. Upon having recoded the knowledge barrier variable (such that higher values reflect barriers being less important), the knowledge exchange index is calculated as the unweighted average of the normalized values of the indicators presented in Columns (2)–(6) of this table (Kling et al., 2007). ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. 6 Conclusions Adoption of sustainable agricultural practices in arid Sub-Saharan Africa is hindered by limited knowledge about the practices as well as by the low (perceived) private benefits of adoption. Cascade training programs, in which some farmers – the so-called contact farmers – are trained by government extension workers about the benefits and usage of new agricultural techniques and who are subsequently asked to disseminate their newly acquired knowledge and expertise among fellow – or peer – farmers in their local social network, have 41 been developed to overcome this information barrier. In this paper we argue that in the context of sustainable land management practices, aimed at conserving soil and water to reduce the need for new land clearing, conditional adoption payments can help overcome both the perceived lack of private benefits barrier as well as the information barrier. Offering compensation conditional on downstream adoption is likely to improve not only the new technology’s perceived cost-benefit ratio, but also the transfer of the contact farmer’s newly acquired knowledge and expertise to her peer farmers. Offering payments for downstream SLMP adoption increases the demand for knowledge and expertise, but it is an open question whether this will also translate into improved information transfer from the contact to the peer farmer. We implemented a Randomized Controlled Trial in arid Burkina Faso to test to what extent offering cash transfers, to be paid out conditional on downstream SLMP adoption, are effective in inducing increased uptake, and also in improving knowledge dissemination. The contact farmers in the control group participated in a cascade training program; upon completion of the training they were asked to disseminate the newly acquired knowledge to peer farmers in their network. Our two treatments consisted of offering financial compen- sation based on SLMP adoption by the peer farmers in these treatment groups. The two treatments only differed in the initial allocation of the payment. In the one treatment arm the peer farmer received the full amount whereas in the other treatment the payment was split, 80-20, between the peer and the contact farmer. We find that peer farmers adopted significantly more practices when there is a financial incentive to do so. Contact farmers also put in more effort to disseminate their information, and they also adopted substantially (although this difference is significant only at p = 0.11) more practices themselves – presumably to lead by example, or because of field demonstration purposes. Interestingly we do not find that the size of these effects vary with how the payment is initially allocated between the peer and the contact farmer. In other words, we find evidence that larger adoption surpluses render the cascade training scheme more effective, but that there is no reason to also provide direct financial stakes to contact farmers to ensure improved information dissemination. Finally, our RCT also speaks to the short-run benefits and costs of SLMP adoption. We find that offering payments, conditional on SLMP adoption and to be disbursed after the end of the agricultural season, results in substantially higher productivity 42 even in the first year of SLMP implementation. Because the payments were to be disbursed in the future, our intervention is unlikely to have substantially affected farmers’ agricultural production constraints, and hence the difference in agricultural productivity plausibly reflects the impact of SLMP adoption on productivity and revenues. Together, our results provide interesting new insights with respect to fostering the adop- tion of SLMPs – but probably also, more broadly, to agricultural development via technology diffusion. Our use of financial incentives can be defended on two grounds. First, the adver- tized SLMPs are (perceived) not to be privately profitable at least in the first few years of usage, while the global community benefits from the positive externalities generated by sus- tainable land use. Second, even if technologies are acknowledged to be profitable, financial incentives may still be necessary to ensure the active dissemination of knowledge and ex- pertise by the contact farmers (see for example BenYishay and Mobarak (2018), Sseruyange and Bulte (2018), and Shikuku et al. (2019)). Our results suggest that conditional subsidies may still be indispensable in the short run, but not in the longer run. With continued SLMP diffusion, farmers receive more and more signals about the productivity impacts of SLMP adoption in the short and in the longer run. Hence, the perceived lack of profitability barrier is likely to become less important over time. However, we also expect the information barrier to become less important over time. This is because we find (i) that a larger surplus gener- ated by SLMP adoption increases the effectiveness of our cascade training program, and (ii) that this increase is equally large independent of whether or not contact farmers are offered direct financial stakes in their peer farmers’ adoption decisions. If, over time, the perceived profitability of SLMP adoption goes up, the perceived size of the surplus of SLMP adoption also increases, and contact farmers are predicted to increase their dissemination effort over time as well – even so in the absence of a direct financial stake. While it is an open question to what extent the dynamics described above result in a socially optimal rate of SLMP adop- tion and dissemination, it is reassuring that two of the currently largest hindrances for the effectiveness of cascade training program, lack of perceived profitability and insufficient effort by contact farmers to disseminate their expertise, are likely to decrease with the continued implementation of cascade training programs. 43 References Abadie, A., Angrist, J. and Imbens, G. (2002). Instrumental variables estimates of the effect of subsidized training on the quantiles of trainee earnings. Econometrica, 70 (1), 91–117. —, Athey, S., Imbens, G. W. and Wooldridge, J. (2017). When Should You Adjust Standard Errors for Clustering? Working Paper No. 24003, NBER. — and Imbens, G. W. (2011). Bias-corrected matching estimators for average treatment effects. Journal of Business & Economic Statistics, 29 (1), 1–11. Abdulai, A. and Huffman, W. (2014). The adoption and impact of soil and water con- servation technology: An endogenous switching regression application. Land Economics, 90 (1), 26–43. Adjognon, S. G., Liverpool-Tasie, L. S. O. and Reardon, T. A. (2017). Agricultural input credit in Sub-Saharan Africa: Telling myth from facts. Food Policy, 67, 93–105, agriculture in Africa – Telling Myths from Facts. Aihounton, G. B. and Henningsen, A. (2021). Units of measurement and the inverse hyperbolic sine transformation. The Econometrics Journal, 24 (2), 334–351. Anderson, M. L. (2008). Multiple inference and gender differences in the effects of early in- tervention: A reevaluation of the abecedaian, perry preschool, and early training projects. Journal of the American Statistical Association, 103, 1481–1495. Arslan, A., McCarthy, N., Lipper, L., Asfaw, S., Cattaneo, A. and Kokwe, M. (2015). Climate smart agriculture? assessing the adaptation implications in Zambia. Jour- nal of Agricultural Economics, 66 (3), 753–780. Athey, S. and Imbens, G. W. (2017). The econometrics of randomized experiments. In Handbook of Economic Field Experiments. Volume 1., North Holland. Balboni, C. A., Bandiera, O., Burgess, R., Ghatak, M. and Heil, A. (2021). Why do people stay poor? Working Paper 29340, National Bureau of Economic Research. 44 Bandiera, O. and Rasul, I. (2006). Social networks and technology adoption in northern Mozambique. The Economic Journal, 116 (514), 869–902. Banerjee, A., Chandrasekhar, A. G., Duflo, E. and Jackson, M. O. (2013). The diffusion of microfinance. Science, 341 (6144). —, Duflo, E., Goldberg, N., Karlan, D., Osei, R., Parienté, W., Shapiro, J., Thuysbaert, B. and Udry, C. (2015). A multifaceted program causes lasting progress for the very poor: Evidence from six countries. Science, 348 (6236), 1260799. Beaman, L. and Dillon, A. (2018). Diffusion of agricultural information within social networks: Evidence on gender inequalities from Mali. Journal of Development Economics, 133, 147–161. Behaghel, L., Gignoux, J. and Macours, K. (2020). Social Learning in Agricul- ture: Does Smallholder Heterogeneity Impede Technology Diffusion in Sub-Saharan Africa? CEPR Discussion Paper 15220, Centre for Economic Policy Research. Bellemare, M. F. and Wichman, C. J. (2020). Elasticities and the inverse hyperbolic sine transformation. Oxford Bulletin of Economics and Statistics, 82, 50–61. BenYishay, A., Jones, M., Kondylis, F. and Mobarak, A. M. (2020). Gender gaps in technology diffusion. Journal of Development Economics, 143, 102380. — and Mobarak, A. M. (2018). Social learning and incentives for experimentation and communication. The Review of Economic Studies, 86 (3), 976–1009. Berg, E., Ghatak, M., Manjula, R., Rajasekhar, D. and Roy, S. (2017). Motivating knowledge agents: Can incentive pay overcome social distance? The Economic Journal, pp. 1–31. Bruhn, M. and McKenzie, D. (2009). In pursuit of balance: Randomization in practice in development field experiments. American Economic Journal: Applied Economics, 1 (4), 200–232. Burbidge, J. B., Magee, L. and Robb, A. L. (1988). Alternative transformations to handle extreme values of the dependent variable. Journal of the American Statistical As- sociation, 83 (401), 123–127. 45 Börner, J., Baylis, K., Corbera, E., de Blas Jordi Honey-Rosés, D. E., Persson, U. M. and Wunder, S. (2017). The effectiveness of payments for environmental services. World Development, 96, 359–374. Carter, M. R. and Lybbert, T. J. (2012). Consumption versus asset smoothing: testing the implications of poverty trap theory in Burkina Faso. Journal of Development Eco- nomics, 99 (2), 255–264. Coase, R. H. (1960). The problem of social cost. Journal of Law and Economics, 3, 1–44. Conley, T. G. and Udry, C. R. (2010). Learning about a new technology: Pineapple in Ghana. American Economic Review, 100 (1), 35–69. Demirgüç-Kunt, A., Klapper, L., Singer, D., Ansar, S. and Hess, J. (2018). The Global Findex Database 2017: Measuring Financial Inclusion and the Fintech Revolution. Data set, World Bank, Washington, DC., accessed in August 2021 https://globalfindex.worldbank.org/sites/globalfindex/files/ databank/Global%20Findex%20Database.xlsx. Dupas, P. (2014). Short-run subsidies and long-run adoption of new health products: Evi- dence from a field experiment. Econometrica, 82, 197–228. Engel, S. (2016). The devil in the detail: A practical guide on designing payment for environmental services. International Review of Environmental and Resource Economics, 9, 131–177. —, Pagiola, S. and Wunder, S. (2008). Designing payments for environmental services in theory and practice: An overview of the issues. Ecological Economics, 65, 663–674. Etongo, D., Djenontin, I. N. S., Kannien, M., Fobiessie, K., Korhonen-Kurki, K. and Djoudi, H. (2015). Land tenure, asset heterogeneity and deforestation in southern Burkina Faso. Forest Policy and Economics, 61, 51–58. Fafchamps, M., Udry, C. and Czukas, K. (1998). Drought and saving in West Africa: are livestock a buffer stock? Journal of Development Economics, 55 (2), 273–305. 46 FAO (2019). Africa Regional Synthesis for The State of the World’s Biodiversity for Food and Agriculture. Tech. rep., Food and Agriculture Organization of the United Nations, Link. — (2021). GIEWS Country Brief: Burkina Faso. Tech. rep., Food and Agricultural Organi- zation of the United Nations, accessed in May 2021. Foster, A. D. and Rosenzweig, M. R. (1995). Learning by doing and learning from others: Human capital and technical change in agriculture. Journal of Political Economy, 103 (6), 1176–1209. Giger, M., Liniger, H., Sauter, C. and Schwilch, G. (2018). Economic benefits and costs of sustainable land management technologies: An analysis of wocat’s global data. Land Degradation & Development, 29 (4), 962–974. Goldstein, M. and Udry, C. (2008). The profits of power: Land rights and agricultural investment in Ghana. Journal of Political Economy, 116 (6), 981–1022. Imbens, G. W. and Rubin, D. B. (2015). Causal Inference for Statistics, Social, and Biomedical Sciences: An Introduction. Cambridge University Press. Jack, B. K. (2011). Market Inefficiencies and the Adoption of Agricultural Technologies in Developing Countries. Center for International Development at Harvard University and Agricultural Technology Adoption Initiative, J-PAL (MIT) and CEGA (UC Berkeley), unpublished manuscript. Janzen, S. A. and Carter, M. R. (2019). After the drought: The impact of microinsur- ance on consumption smoothing and asset protection. American Journal of Agricultural Economics, 101 (3), 651–671. Jayachandran, S., De Laat, J., Lambin, E. F., Stanton, C. Y., Audy, R. and Thomas, N. E. (2017). Cash for carbon: A randomized trial of payments for ecosystem services to reduce deforestation. Science, 357 (6348), 267–273. Karlan, D., Osei, R., Osei-Akoto, I. and Udry, C. (2014). Agricultural decisions after relaxing credit and risk constraints. The Quarterly Journal of Economics, 129, 597–652. 47 Kassie, M., Marenya, P., Tessema, Y., Jaleta, M., Zeng, D., Erenstein, O. and Rahut, D. (2018). Measuring farm and market level economic impacts of improved maize production technologies in Ethiopia: Evidence from panel data. Journal of Agricultural Economics, 69 (1), 76–95. —, Teklewold, H., Jaleta, M., Marenya, P. and Erenstein, O. (2015). Under- standing the adoption of a portfolio of sustainable intensification practices in eastern and southern Africa. Land Use Policy, 42, 400–411. Khonje, M. G., Manda, J., Mkandawire, P., Tufa, A. H. and Alene, A. D. (2018). Adoption and welfare impacts of multiple agricultural technologies: evidence from eastern Zambia. Agricultural Economics, 49 (5), 599–609. Kling, J. R., Liebman, J. B. and Katz, L. F. (2007). Experimental analaysis of neigh- borhood effects. Econometrica, 75, 83–119. Koenker, R. (2005). Quantile Regression. Econometric Society Monographs, Cambridge University Press. — and Hallock, K. F. (2001). Quantile regression. Journal of Economic Perspectives, 15 (4), 143–156. Kondylis, F., Mueller, V., Sheriff, G. and Zhu, S. (2016). Do female instructors reduce gender bias in diffusion of sustainable land management techniques? experimental evidence from Mozambique. World Development, 78, 436–449. —, — and Zhu, J. (2017). Seeing is believing? evidence from an extension network experi- ment. Journal of Development Economics, 125, 1–20. Koundouri, P., Nauges, C. and Tzouvelekas, V. (2006). Technology adoption under production uncertainty: Theory and application to irrigation technology. American Journal of Agricultural Economics, 88, 657–670. Kpadonou, R. A. B., Owiyo, T., Barbier, B., Denton, F., Rutabingwa, F. and Kiema, A. (2017). Advancing climate-smart-agriculture in developing drylands: Joint analysis of the adoption of multiple on-farm soil and water conservation technologies in west african sahel. Land Use Policy, 61, 196–207. 48 Krishnan, P. and Patnam, M. (2014). Neighbors and extension agents in Ethiopia: Who matters more for technology adoption? American Journal of Agricultural Economics, 96 (1), 308–327. Liniger, H., Studer, R. M., Hauert, C. and Gurtner, M. (2011a). Sustainable Land Management in Practice: Guidelines and Best Practices for Sub-Saharan Africa (Field Application). Tech. rep., TerrAfrica, World Overview of Conservation Approaches and Technologies (WOCAT) and Food and Agriculture Organization of the United Nations (FAO), retrieved from www.fao.org. —, —, — and — (2011b). Sustainable Land Management in Practice – Guidelines and best Practices for Sub-Saharan Africa. Tech. rep., TerrAfrica, World Overview of Conservation Approaches and Technologies (WOCAT) and Food and Agriculture Organization of the United Nations (FAO). List, J., Sadoff, S. and Wagner, M. (2011). So you want to run an experiment, now what? Some simple rules of thumb for optimal experimental design. Experimental Eco- nomics, 14, 439–457. Manda, J., Alene, A. D., Gardebroek, C., Kassie, M. and Tembo, G. (2016). Adop- tion and impacts of sustainable agricultural practices on maize yields and incomes: Evi- dence from rural Zambia. Journal of Agricultural Economics, 67 (1), 130–153. McKenzie, D. (2012). Beyond baseline and follow-up: The case for more T in experiments. Journal of Development Economics, 99 (2), 210 – 221. Pittelkow, C. M., Liang, X., Linquist, B. A., van Groenigen, K. J., Lee, J., Lundy, M. E., van Gestel, N., Six, J., Venterea, R. T. and van Kessel, C. (2015). Productivity limits and potentials of the principles of conservation agriculture. Nature, pp. 365–368. Pouliot, M., Treue, T., Obiri, B. D. and Ouedraogo, B. (2012). Deforestation and the limited contribution of forests to rural livelihoods in West Africa: Evidence from Burkina faso and Ghana. AMBIO, 41 (7), 738–750. Pretty, J., Benton, T. G., Bharucha, Z. P., Dicks, L. V., Flora, C. B., Godfray, H. C. J., Goulson, D., Hartley, S., Lampkin, N., Morris, C., Pierzynski, G., 49 Prasad, P. V. V., Reganold, J., Rockström, J., Smith, P., Thorne, P. and Wratten, S. (2018). Global assessment of agricultural system redesign for sustainable intensification. Nature Sustainability, pp. 441–446. —, Toulmin, C. and Williams, S. (2011). Sustainable intensification in African agricul- ture. International Journal of Agricultural Sustainability, 9 (1), 5–24. Shikuku, K. M., Pieters, J., Bulte, E. and Läderach, P. (2019). Incentives and the diffusion of agricultural knowledge: experimental evidence from northern Uganda. American Journal of Agricultural Economics, 101 (4), 1164–1180. Sseruyange, J. and Bulte, E. (2018). Do incentives matter for the diffusion of financial knwoeldge? experimental evidence from Uganda. Journal of African Economies, 27, 1–20. Takahashi, K., Mano, Y. and Otsuka, K. (2019). Learning from experts and peer farmers about rice production: Experimental evidence from Cote d’Ivoire. World Development, 122, 157 – 169. Teklewold, H., Kassie, M. and Shiferaw, B. (2013). Adoption of multiple sustainable agricultural practices in rural ethiopia. Journal of Agricultural Economics, 64 (3), 597–623. Tesfaye, W., Blalock, G. and Tirivayi, N. (2021). Climate-smart innovations and rural poverty in Ethiopia: Exploring impacts and pathways. American Journal of Agricultural Economics, 103 (3), 878–899. Thierfelder, C. and Wall, P. C. (2010). Rotation in conservation agriculture systems of Zambia: Effects on soil quality and water relations. Experimental Agriculture, 46 (3), 309–325. USDA (2017). Cotton and Products Update - 2017 West Africa Update. Tech. rep., USDA Foreign Agricultural Service. Vancampenhout, K., Nyssen, J., Gebremichael, D., Deckers, J., Poesen, J., Haile, M. and Moeyersons, J. (2006). Stone bunds for soil conservation in the northern ethiopian highlands: Impacts on soil fertility and crop yield. Soil and Tillage Research, 90 (1), 1–15. 50 Vasilaky, K. N. and Leonard, K. L. (2018). As good as the networks they keep? Im- proving outcomes through weak ties in rural uganda. Economic Development and Cultural Change, 66 (4), 755–792. WB (2016). Burkina Faso: Poverty, Vulnerability, and Income Source. Tech. rep., World Bank, Report No. 115122. Wooldridge, J. M. (2010). Economic Analysis of Cross Section and Panel Data. The MIT Press, second edition edn. World Bank (2008). World Development Report 2008: Agriculture for Development. Wash- ington DC: the World Bank. Wouterse, F. and Taylor, J. E. (2008). Migration and income diversification: Evidence from Burkina Faso. World Development, 36 (4), 625–640. Wunder, S. (2007). The efficiency of payments for environmental services in tropical con- servation. Conservatrion Biology, 21, 48–58. Young, A. (2018). Channeling Fisher: Randomization Tests and the Statistical Insignifi- cance of Seemingly Significant Experimental Results. The Quarterly Journal of Economics, 134 (2), 557–598. 51 A Appendix A.1 Certificates 52 Figure A1: Example of certificates handed out to peer farmers (a) In control groups (b) In split payment groups Notes : Certificates that peer farmers have received for their participation in the experiment. All certificates show the name of the farmer, the village, commune, and region of residence, the name of the georeferenced agricultural plots where SLMP adoption would be verified at endline, and the name of the contact farmer. In payment communes, the certificate also details the53 conditions, structure, and allocation of payments. The certificate in the peer payment group is similar to that in the split payment group except for the last remark. A.2 Verification of the SLMPs Figure A2: Comparison of pit planting at the beginning and at the end of the agricultural season (a) Implementation of pit planting (b) Holes of pit planting at endline Notes : The two figures show the how pit planting looks like at the beginning of the agricultural season and after harvest (at endline). The second figure shows that the holes are still visible around the crop remains at endline. Pit planting is highlighted separately from the other SLMPs because this is the only practices which is optimally implemented at the beginning of the agricultural season and which stays intact only for one agricultural season. 54 A.3 Descriptive statistics and balance 55 Table A1: Scores of the contact farmers on the SLMP knowledge tests. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Total Conrol group Peer payment Split payment Difference Normalized difference Variable N Mean/SE N Mean/SE N Mean/SE N Mean/SE (2)-(3) (2)-(4) (3)-(4) (2)-(3) (2)-(4) (3)-(4) Agronomy (in percent) 317 0.907 120 0.912 99 0.914 98 0.893 -0.002 0.019 0.021* -0.028 0.207 0.256 (0.005) (0.009) (0.008) (0.008) Integrated crop and livestock management (in percent) 317 0.734 120 0.748 99 0.718 98 0.732 0.030 0.016 -0.014 0.207 0.113 -0.097 (0.008) (0.013) (0.015) (0.015) Agroforestry (in percent) 317 0.662 120 0.655 99 0.668 98 0.665 -0.012 -0.009 0.003 -0.093 -0.072 0.021 (0.008) (0.010) (0.015) (0.015) Total score (in percent) 317 0.770 120 0.775 99 0.769 98 0.765 0.006 0.010 0.005 0.073 0.141 0.053 (0.005) (0.006) (0.009) (0.009) Notes : Average raw knowledge scores of contact farmers in the whole sample and in the three treatment groups are presented in Columns (1)-(4); standard deviations are presented in parentheses. Knowledge scores for each agronomic subsection are the share of correct answers within the specific section of the test. Columns (5)-(7) present the p -values for the treatment group indicators from regressing the characteristics on the treatment group indicators and region fixed-effects. Standard errors are clustered at the commune level. Columns (8)-(10) present the normalized differences for each of the test parts. ***, **, and * indicate significance at the 1, 5, and 10 percent critical 56 level. Table A2: Differential attrition rates between treatments. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Total Conrol group Peer payment Split payment Difference Normalized difference Variable N Mean/SE N Mean/SE N Mean/SE N Mean/SE (2)-(3) (2)-(4) (3)-(4) (2)-(3) (2)-(4) (3)-(4) Attrited (0/1) 1914 0.007 720 0.006 600 0.005 594 0.010 0.001 -0.005 -0.005 0.008 -0.052 -0.059 (0.002) (0.003) (0.003) (0.004) Notes : Columns (1)-(4) present the share of farmers in the whole sample as well as in each treatment group who were not surveyed at endline. Standard deviations are reported in parentheses. Columns (5)-(7) present the p -values for the treatment group indicators as obtained by regressing the characteristics on the treatment group indicators and region fixed-effects. Standard errors are clustered at the commune level. Columns (8)-(10) present the normalized differences. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 57 Table A3: Differences in baseline characteristics of peer farmers comparing the control and payment groups. (1) (2) (3) (4) (5) Total Control group Payment group T-test (P-value) Normalized difference Variable N/[Clusters] Mean/SD N/[Clusters] Mean/SD N/[Clusters] Mean/SD (2)-(3) (2)-(3) Age 1595 41.377 600 42.382 995 40.771 0.117 0.148 [32] (19.035) [12] (23.968) [20] (14.353) Female respondent (0/1) 1595 0.173 600 0.140 995 0.193 0.102 -0.140 [32] (0.734) [12] (0.870) [20] (0.620) Respondent is Head of Household (0/1) 1595 0.740 600 0.775 995 0.720 0.162 0.126 [32] (0.776) [12] (1.004) [20] (0.581) Has some primary education (0/1) 1595 0.285 600 0.298 995 0.277 0.493 0.046 [32] (0.721) [12] (0.801) [20] (0.685) Adults in household 1595 11.695 600 12.217 995 11.380 0.369 0.124 [32] (20.279) [12] (24.799) [20] (17.331) Deprived house (0/1) 1595 0.858 600 0.875 995 0.847 0.636 0.079 [32] (1.216) [12] (1.025) [20] (1.338) Asset index 1595 -0.136 600 -0.058 995 -0.183 0.409 0.055 58 [32] (9.141) [12] (8.869) [20] (9.512) Association membership (0/1) 1595 0.657 600 0.635 995 0.670 0.528 -0.074 [32] (1.437) [12] (1.740) [20] (1.260) Hired labor in previous agri. season (0/1) 1595 0.544 600 0.595 995 0.513 0.223 0.165 [32] (2.003) [12] (2.081) [20] (1.976) Number of plots under the control of the farmer 1595 1.699 600 1.802 995 1.637 0.226 0.200 [32] (2.708) [12] (2.667) [20] (2.707) Number of eroded plots 1595 2.453 600 2.588 995 2.372 0.224 0.153 [32] (3.989) [12] (3.677) [20] (4.152) Landholdings (ha) 1595 4.976 600 5.168 995 4.861 0.441 0.067 [32] (14.935) [12] (13.233) [20] (16.154) Number of SLMPs adopted at baseline 1595 2.349 600 2.367 995 2.339 0.688 0.020 [32] (5.782) [12] (5.891) [20] (5.869) Relation to CF - Kinship 1594 0.432 599 0.474 995 0.406 0.134 0.137 [32] (1.090) [12] (0.727) [20] (1.242) Income from agricultural production (IHS transformed) 1595 12.815 600 13.086 995 12.652 0.280 0.158 [32] (11.334) [12] (3.806) [20] (14.063) Household has income from non-agricultural activities (0/1) 1595 0.518 600 0.538 995 0.507 0.426 0.064 [32] (1.090) [12] (0.881) [20] (1.214) Notes : Average values, for the total sample as well as for each of the two sub-samples, are presented in Columns (1)-(3); standard deviations are presented in parentheses. Column (4) presents the p -values for the payment indicator from regressing the characteristics on the indicator and region fixed-effects. Standard errors are clustered at the commune level. Column (5) presents the differences normalized by the sample variance. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table A4: Differences in baseline characteristics peer and contact farmers. (1) (2) (3) (4) (5) Total Contact farmer Peer farmer T-test (P-value) Normalized difference Variable N/[Clusters] Mean/SE N/[Clusters] Mean/SE N/[Clusters] Mean/SE (2)-(3) (2)-(3) Age 1914 42.039 319 45.351 1595 41.377 0.000*** 0.366 [32] (0.440) [32] (0.654) [32] (0.477) Female respondent (0/1) 1914 0.177 319 0.194 1595 0.173 0.216 0.056 [32] (0.017) [32] (0.019) [32] (0.018) Has some primary education (0/1) 1914 0.313 319 0.451 1595 0.285 0.000*** 0.358 [32] (0.019) [32] (0.035) [32] (0.018) Adults in household 1914 11.935 319 13.135 1595 11.695 0.000*** 0.209 [32] (0.501) [32] (0.566) [32] (0.508) Deprived house (0/1) 1914 0.845 319 0.784 1595 0.858 0.002*** -0.205 [32] (0.032) [32] (0.045) [32] (0.030) Asset index 1914 -0.000 319 0.678 1595 -0.136 0.000*** 0.359 [32] (0.219) [32] (0.194) [32] (0.229) Association membership (0/1) 1914 0.688 319 0.840 1595 0.657 0.000*** 0.395 59 [32] (0.034) [32] (0.031) [32] (0.036) Hired labor in previous agri. season (0/1) 1914 0.560 319 0.643 1595 0.544 0.000*** 0.200 [32] (0.050) [32] (0.053) [32] (0.050) Number of plots under the control of the farmer 1914 1.742 319 1.956 1595 1.699 0.000*** 0.302 [32] (0.069) [32] (0.089) [32] (0.068) Number of eroded plots 1914 2.522 319 2.868 1595 2.453 0.000*** 0.281 [32] (0.106) [32] (0.155) [32] (0.100) Landholdings (ha) 1914 5.252 319 6.629 1595 4.976 0.000*** 0.343 [32] (0.370) [32] (0.465) [32] (0.374) Number of SLMPs adopted at baseline 1914 2.531 319 3.439 1595 2.349 0.000*** 0.738 [32] (0.147) [32] (0.177) [32] (0.145) Income from agricultural production (IHS transformed) 1914 12.895 319 13.294 1595 12.815 0.000*** 0.177 [32] (0.290) [32] (0.332) [32] (0.284) Household has income from non-agricultural activities (0/1) 1914 0.536 319 0.621 1595 0.518 0.002*** 0.205 [32] (0.027) [32] (0.036) [32] (0.027) Notes : Average values, for the total sample as well as for each of the two farmer types, are presented in Columns (1)-(3); standard deviations are presented in parentheses. Column (4) presents the p -values for the farmer type indicator derived from a region fixed-effects model. Standard errors are clustered at the commune level. Column (5) presents the differences normalized by the sample variance. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table A5: Baseline characteristics of the contact farmers in each of the three (sub-)treatment arms. (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Total Conrol group Peer payment Split payment T-test (P-value) Normalized difference Variable N/[Clusters] Mean/SE N/[Clusters] Mean/SE N/[Clusters] Mean/SE N/[Clusters] Mean/SE (2)-(3) (2)-(4) (3)-(4) (2)-(3) (2)-(4) (3)-(4) Age 319 45.351 120 45.158 100 44.540 99 46.404 0.654 0.415 0.186 0.062 -0.124 -0.184 [32] (0.654) [12] (1.071) [10] (0.893) [10] (1.435) Female respondent (0/1) 319 0.194 120 0.175 100 0.170 99 0.242 0.929 0.107 0.094* 0.013 -0.166 -0.179 [32] (0.019) [12] (0.028) [10] (0.037) [10] (0.032) Has some primary education (0/1) 319 0.451 120 0.450 100 0.540 99 0.364 0.149 0.332 0.022** -0.180 0.175 0.353 [32] (0.035) [12] (0.060) [10] (0.045) [10] (0.070) Adults in household 319 13.135 120 13.675 100 12.450 99 13.172 0.395 0.730 0.509 0.160 0.069 -0.096 [32] (0.566) [12] (1.038) [10] (0.933) [10] (0.989) Deprived house (0/1) 319 0.784 120 0.833 100 0.630 99 0.879 0.031** 0.545 0.017** 0.463 -0.128 -0.576 [32] (0.045) [12] (0.066) [10] (0.087) [10] (0.068) Asset index 319 0.678 120 0.788 100 0.725 99 0.496 0.672 0.313 0.477 0.032 0.140 0.108 [32] (0.194) [12] (0.289) [10] (0.394) [10] (0.361) Association membership (0/1) 319 0.840 120 0.800 100 0.870 99 0.859 0.330 0.405 0.863 -0.187 -0.154 0.033 [32] (0.031) [12] (0.055) [10] (0.056) [10] (0.053) Hired labor in previous agri. season (0/1) 319 0.643 120 0.675 100 0.670 99 0.576 0.998 0.330 0.308 0.011 0.205 0.194 [32] (0.053) [12] (0.096) [10] (0.090) [10] (0.092) Number of plots under the control of the farmer 319 1.956 120 2.050 100 1.810 99 1.990 0.215 0.844 0.337 0.247 0.063 -0.189 60 [32] (0.089) [12] (0.153) [10] (0.138) [10] (0.173) Number of eroded plots 319 2.868 120 3.050 100 2.640 99 2.879 0.272 0.700 0.389 0.230 0.102 -0.144 [32] (0.155) [12] (0.266) [10] (0.283) [10] (0.268) Landholdings (ha) 319 6.629 120 6.333 100 6.198 99 7.423 0.810 0.414 0.255 0.025 -0.196 -0.204 [32] (0.465) [12] (0.586) [10] (0.516) [10] (1.236) Number of SLMPs adopted at baseline 319 3.439 120 3.508 100 3.160 99 3.636 0.304 0.766 0.208 0.223 -0.087 -0.292 [32] (0.177) [12] (0.238) [10] (0.373) [10] (0.333) Income from agricultural production (IHS transformed) 319 13.294 120 13.640 100 13.483 99 12.684 0.436 0.285 0.406 0.107 0.322 0.270 [32] (0.332) [12] (0.188) [10] (0.150) [10] (1.045) Household has income from non-agricultural activities (0/1) 319 0.621 120 0.642 100 0.580 99 0.636 0.288 0.964 0.586 0.126 0.011 -0.115 [32] (0.036) [12] (0.047) [10] (0.053) [10] (0.091) Notes : Average values, for the total sample as well as for each of the three sub-samples, are presented in Columns (1)-(4); standard deviations are presented in parentheses. Columns (5)-(7) present the p -values obtained by means of region fixed-effects regressions. Standard errors are clustered at the commune level. Columns (8)-(10) present the normalized differences between each of the treatment and the control group. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table A6: Determinants of the number of SLMPS at baseline. Agronomy SLMPs Int. Crop & Livestock Manag. SLMPs Agroforestry SLMPs # SLMPs Zai Heap and pit Stone and Mowing and cons. Forage crop Use of agr. and Controlled Assisted Living adopted composting earth bounds of natu. fodder cultivation wood by-products clearing natural regeneration hedges (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Age 0.017∗∗∗ 0.002∗∗∗ 0.004∗∗∗ 0.003∗∗∗ 0.002∗∗ 0.001 0.000 -0.001 0.003∗∗∗ 0.002∗∗∗ (0.003) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) (0.001) Female respondent (0/1) -0.224∗ -0.003 0.037 -0.059∗ -0.022 0.005 -0.027 -0.031 -0.068∗∗∗ -0.056∗∗∗ (0.110) (0.016) (0.036) (0.034) (0.029) (0.012) (0.042) (0.039) (0.022) (0.019) Has some primary education (0/1) 0.155∗∗ 0.042∗∗ 0.026 0.049∗∗ 0.021 0.013 -0.013 -0.005 0.005 0.016 (0.068) (0.020) (0.025) (0.018) (0.024) (0.014) (0.025) (0.033) (0.025) (0.011) Adults in household -0.000 -0.001 0.002 -0.000 0.002 -0.001 0.004∗ -0.005 0.001 -0.002 (0.005) (0.001) (0.002) (0.001) (0.002) (0.001) (0.002) (0.003) (0.002) (0.001) Asset index 0.132∗∗∗ -0.004 0.019∗∗ 0.041∗∗∗ 0.026∗∗∗ 0.012∗∗∗ 0.019∗∗ 0.009 0.000 0.010∗∗∗ (0.017) (0.005) (0.008) (0.005) (0.007) (0.004) (0.007) (0.009) (0.007) (0.003) Hired labor in previous agri. season (0/1) 0.148∗ -0.011 0.050 0.021 0.014 -0.023 0.015 0.018 0.049 0.014 (0.075) (0.031) (0.036) (0.026) (0.024) (0.019) (0.039) (0.039) (0.034) (0.015) Number of plots under the control of the farmer 0.352∗∗∗ 0.048 0.189∗∗∗ 0.098∗∗∗ -0.035 -0.016 0.057 -0.082∗ 0.076∗∗ 0.017 61 (0.066) (0.030) (0.025) (0.030) (0.028) (0.025) (0.042) (0.044) (0.034) (0.011) Number of eroded plots -0.121∗∗ -0.005 -0.076∗∗∗ -0.039∗∗ 0.002 0.002 -0.013 0.047∗ -0.027 -0.013∗ (0.046) (0.007) (0.017) (0.018) (0.014) (0.012) (0.020) (0.026) (0.020) (0.007) Landholdings (ha) 0.025∗∗∗ -0.005∗ 0.000 0.006∗∗ 0.002 0.005∗∗ 0.005 0.003 0.007 0.002 (0.009) (0.002) (0.002) (0.002) (0.003) (0.002) (0.004) (0.006) (0.005) (0.001) Household has income from non-agricultural activities (0/1) 0.020 0.014 -0.015 0.022 0.034 -0.007 -0.075∗ 0.022 0.027 -0.002 (0.075) (0.014) (0.023) (0.025) (0.023) (0.019) (0.039) (0.031) (0.032) (0.013) Score total (en pct) - Apres 1.659∗∗ -0.096 0.441∗∗ 0.609∗∗ -0.366 0.247∗∗ 1.059∗∗∗ -0.567 0.378 -0.046 (0.720) (0.137) (0.173) (0.250) (0.249) (0.106) (0.312) (0.475) (0.428) (0.084) Number of adopters in commune -0.063 0.012 -0.005 -0.030∗ -0.037 0.014 -0.041 -0.000 0.009 0.014 (0.055) (0.012) (0.021) (0.017) (0.024) (0.017) (0.031) (0.043) (0.032) (0.009) Number of adopters in commune2 0.001∗ -0.000 0.000 0.000∗∗ 0.000∗ -0.000 0.001 0.000 0.000 -0.000 (0.001) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) Constant -0.562 -0.200 -0.621 0.232 0.834 -0.416 0.005 0.617 -0.737 -0.275 (1.221) (0.209) (0.497) (0.456) (0.500) (0.351) (0.702) (0.976) (0.729) (0.178) Observations 1889 1889 1889 1889 1889 1889 1889 1889 1889 1889 Observations 1889 1889 1889 1889 1889 1889 1889 1889 1889 1889 Adjusted R2 0.350 0.098 0.130 0.297 0.058 0.042 0.268 0.082 0.097 0.053 Mean 2.537 0.053 0.308 0.503 0.206 0.070 0.534 0.375 0.419 0.067 Std. dev. 1.477 0.224 0.462 0.500 0.405 0.256 0.499 0.484 0.493 0.251 Notes : All columns are estimated using OLS regression where the baseline number of SLMPs or the indicator of SLMP adoption at baseline is regressed on farmers’ baseline characteristics and region fixed effects. Standard errors, presented in parentheses, are clustered at the commune level. The regressions are run using the data of all farmers in the sample (i.e., both peer and contact farmers), with, in total, 1889 observations. The last two rows of the table present the mean and standard deviation of the outcome variables. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. A.4 Supplementary results Table A7: The impact of treatment status on agricultural revenues and individual crop yields. IHS Transfromed Income Productivity Agriculture Maize Millet Sorghum Cowpea Index (1) (2) (3) (4) (5) (6) Payment treatment 0.221∗∗ 181.379 197.541∗∗∗ 347.612∗∗∗ 67.253 0.255∗∗∗ (0.089) (136.579) (67.861) (82.401) (50.290) (0.071) Observations 1532 994 721 897 371 1467 Observations 1532 994 721 897 371 1467 Adjusted R2 0.420 0.020 0.032 0.014 0.058 0.015 Baseline Yes No No No No No Covariates Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Control mean 467494.812 1151.229 717.798 745.729 289.560 -0.074 Control std.dev. 540515.154 1213.386 933.566 726.059 411.660 0.759 Unit FCFA kg/ha kg/ha kg/ha kg/ha Std.dev. Notes : Column (1) reports the OLS estimate of the intention-to-treat effect on the inverse hyperbolic sine of agriculture income; the coefficient can thus be interpreted as a percentage change. Column (2)- (5) report similar intention-to-treat effect estimates on crop productivity calculated as the ratio of total quantity produced and of total area where the crop was cultivated. Column (6) reports the intention-to-treat effect estimate on the productivity index which is calculated as the unweighted average of the normalized productivities of the four crops (Kling et al., 2007). The samples in Columns (2)-(6) consist of peer farmers who produced the crops. Standard errors, clustered at the commune level, are presented in parentheses. For the vector of covariates, see Table 5. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 62 Table A8: Uncovering the types of ties between peer and contact farmers. Control group Peer payment group Split payment group Total Family 0.472 0.401 0.417 0.433 (0.500) (0.491) (0.494) (0.496) Neighbor 0.190 0.220 0.258 0.220 (0.393) (0.415) (0.438) (0.415) Friend 0.239 0.275 0.237 0.250 (0.427) (0.447) (0.426) (0.433) Village association/cooperative 0.0790 0.0842 0.0757 0.0796 (0.270) (0.278) (0.265) (0.271) Other 0.0202 0.0200 0.0123 0.0177 (0.141) (0.140) (0.110) (0.132) Observations 1583 Notes : The table presents the frequency of each type of ties between peer and contact farmers for each treatment group and for the whole sample. The numbers in the table represent the share of peer farmers within the group. Standard deviations are presented in parentheses. The total sample of peer farmers is 1583 in the table. 63 Table A9: Treatment effects on hired labor use and expenditure. Share of HH Conditional IHS. transf. of hiring labor hired labor expenditure (1) (2) Payment treatment -0.041 -0.027 (0.037) (0.097) Observations 1574 846 Observations 1574 846 Adjusted R2 0.242 0.165 Baseline Outcome No No Covariates Included Yes Yes Region fixed effects Yes Yes Control mean 0.577 10.549 Control std.dev. 0.494 0.950 Unit Share FCFA Notes : Column (1) presents the intention-to-treat estimates of payments on the share of peer farmers who hired labour during the current (2019) agricultural season via OLS regression. Column (2) presents the intention-to-treat effects on the inverse hyperbolic sine of expenditures on hired labour conditional of having hired labour in the current agricultural season; the coefficient can thus be interpreted as a percentage change. Standard errors, clustered at the commune level, are presented in parentheses. For the vector of covariates, see Table 5. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 64 A.5 Robustness checks We implement three types of robustness checks. We first test whether our results are robust to re-estimating our OLS regressions using probit and count models. Next, we adjust our statistical inferences of the main treatment effect estimates to take advantage of the ran- domized treatment assignment in the experiment. In the final step, we adjust our statistical inferences for multiple hypothesis testing. A.5.1 Non-linear models for count and binary outcomes We first evaluate the extent to which our intention-to-treat impact estimates are sensitive to relaxing the continuous outcome variable assumption. This assumption of the OLS regression does not hold when our outcome variable is the number of practices adopted by farmers, or when it is a binary variable. We use negative binomial regression for count variables and probit regression for binary outcome variables to re-estimate the effects presented in Tables 5, 7, 10 and 11. The results are presented in Tables A10–A13; coefficients presented in these tables are marginal effects. Overall, the results of the negative binomial and probit regression models are very similar to those obtained using OLS – in terms of signs and significance, and also quantitatively. The results in Tables A11 and A12 are qualitatively (and even quantitatively) identical to those in Tables 7 and 10, respectively. Comparing Tables 5 and A10, the only difference between the two is for the impact on stone and earth bund construction (see the fourth column in both tables); while the coefficient for this practice remains by and large unchanged in size, it just fails to be statistically significant in Table A10 (with p = 0.102). The only substantive difference caused by using probit or negative binomial estimation is with respect to the role of leading by example in the comparison between the two payment sub-treatments; compare Column (4) in Tables 11 and A13. While using negative binomial estimation reduces the size of the coefficient on the spilt payment indicator in Table A13, the difference does become statistically significant(p -value 0.06). In any case, this does not really affect our conclusion of the initial allocation not affecting outcomes, as all other results regarding the impact of changing the initial payment allocation remain unaffected. 65 Table A10: Robustness of treatment effects on SLMPs adoption using negative binomial and probit estimation. Agronomy SLMPs Agro-sylvo-pastoral SLMPs Agroforestry SLMPs # SLMPs Zai Heap and pit Stone and Mowing and cons. Forage crop Use of agr. and Controlled Assisted Living adopted composting earth bunds of natu. fodder cultivation wood by-products clearing natural regeneration hedges (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Payment treatment 0.421∗ 0.001 0.101 0.104 0.136∗∗∗ 0.012 -0.068 -0.011 0.136∗ 0.079∗∗∗ (0.223) (0.002) (0.067) (0.063) (0.053) (0.021) (0.069) (0.109) (0.071) (0.020) Observations 1574 1574 1574 1574 1574 1574 1574 1574 1574 1574 66 Baseline Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Covariates Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Control mean 2.784 0.070 0.564 0.352 0.148 0.084 0.619 0.617 0.268 0.060 Control std.dev. 1.670 0.256 0.496 0.478 0.355 0.277 0.486 0.486 0.444 0.238 Effect size (in std.dev.) 0.252 0.003 0.203 0.217 0.383 0.045 -0.139 -0.022 0.307 0.330 Unit # # # # # # # # # # Notes: This table replicates Column (1) of Table 5 using negative binomial regression, and Columns (2)–(10) of that table using probit. All coefficients are marginal effects. Standard errors, presented in parentheses, are clustered at the commune level. For the vector of covariates, see Table 5. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. Table A11: Robustness of treatment effects on input use using negative binomial and probit estimation. # Manually sowed plots Household labor (1) (2) Payment treatment 0.003 -0.715 (0.020) (0.492) Observations 1560 1555 Baseline Yes No Covariates Yes Yes Region FE-s Yes Yes Control mean 1.641 3.634 Control std.dev. 0.825 5.187 Effect size (in std.dev.) 0.004 -0.138 Unit Share # HH member Notes: This table re-estimates Columns (2) and (3) of Table 7 using, respectively, probit and negative binomial estimation. Coefficients are marginal effects. Standard errors, presented in parentheses, are clustered at the commune level. For the vector of covariates, see Table 5. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 67 Table A12: Robustness of treatment effects on knowledge exchange using negative binomial and probit estimation. PF asked Farmers # SLMPs adopted CF monitor % of SLMPs not adopted for advice discuss SLMPs by CF PF adoption due to lack of knowledge (1) (2) (3) (4) (5) Payment treatment 0.121∗ 0.133∗∗ 0.343 0.110∗ -0.254∗∗∗ (0.068) (0.065) (0.216) (0.067) (0.068) Observations 1573 1573 315 1573 1571 Baseline Yes Yes Yes Yes No Covariates Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Control mean 0.356 0.408 3.900 0.336 0.231 Control std.dev. 0.479 0.492 1.746 0.473 0.263 Effect size (in std.dev.) 0.253 0.271 0.196 0.232 -0.965 Unit share share # share share Notes: This table re-estimates Columns (1), (2), (4) and (5) of Table 10 using probit estimation, and Column (3) of that table using negative binomial estimation. Coefficients are marginal effects. Standard errors, presented in parentheses, are clustered at the commune level. For the vector of covariates, see Table 5. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 68 Table A13: Robustness of treatment effects of offering direct financial incentives for infor- mation dissemination to the contact farmers using negative binomial and probit estimation. # SLMPs PF asked Farmers # SLMPs adopted CF monitor % of SLMPs not adopted adopt. by PFs for advice discuss SLMPs by CF PF adoption due to lack of knowledge (1) (2) (3) (4) (5) (6) Split payment group 0.073 0.070 0.035 0.463∗ 0.075 -0.037 (0.224) (0.069) (0.081) (0.246) (0.094) (0.099) Observations 978 978 978 195 978 976 Baseline Yes Yes Yes Yes Yes No Covariates Yes Yes Yes Yes Yes Yes Region FE-s Yes Yes Yes Yes Yes Yes Control mean 3.307 0.435 0.521 4.061 0.415 0.136 Control std.dev. 1.925 0.496 0.500 1.963 0.493 0.218 Effect size (in std.dev.) 0.038 0.141 0.070 0.236 0.153 -0.167 Unit # share share # share share Notes: This table re-estimates Columns (1) and (4) of Table 11 using negative binomial estimation, and Columns (2), (3), (5) and (6) using probit estimation. All coefficients are marginal effects. Standard errors, presented in parentheses, are clustered at the commune level. For the vector of covariates, see Table 5. ***, **, and * indicate significance at the 1, 5, and 10 percent critical level. 69 A.5.2 Randomized inference So far, we have tested our treatment effects with conventional standard errors which are based on the asymptotic distribution of the treatment effects. However, one might question the use of asymptotically consistent standard errors given the relatively limited number of clusters in our experiment (32 communes) . To address this concern, we use Fisherian randomized inference to test the sharp null of no treatment effects following Athey and Imbens (2017) and Young (2018).19 We consider treatment assignment permutations that are stratified at the region level and clustered at the commune level. We calculate p -values based on randomized inference for four intention-to-treat effects of conditional payments (see equation (1)) that summarize our main findings: on the number of practices adopted by peer farmers (Column (1) of Table 5), on transformed agricultural incomes of peer farmers (Column (1) of Table 8), on peer farmers’ agricultural productivity index (Column (6) of Table 8), and on the knowledge dissemination index (Column (6) of Table 10). We present the results from randomized inference in Table A14. Columns (1) and (2) of Table A14 show the OLS point estimates and the corresponding conventional standard errors. The conventional p -values are presented in Column (3) and indicate that our treatment effect estimates are significant at the 5% level, or better. The p -values from randomized inference are presented in Column (4). Overall they are larger than those in Column (3), but the effects are still significant: those on crop productivity and on knowledge exchange are at the 5% level whereas the effects on agricultural income and peer farmer SLMPs uptake are at the 10% level. Our results are therefore robust to Fisherian exact tests. Table A14: Randomized inference tests of the main treatment effect estimates. Coefficients Std. Err. Conventional p-val RI p-val # SLMPs adopted 0.507 0.232 0.036 0.086 Agricultural income 0.221 0.089 0.018 0.052 Productivity index 0.255 0.071 0.001 0.009 Knowledge exch. 0.268 0.095 0.008 0.035 Notes: Results are from applying randomization inference on each regression using 10,000 permutations of treatment assignment. In each permutation, the treatment assignment was stratified on regions and was randomized at the commune (cluster) level. 19 Randomization inference generates the exact distribution of treatment effect estimates by taking different permutations of treatment allocation over the whole sample and re-estimating the point estimates for each permutation. 70 A.5.3 Multiple hypothesis testing Finally, we show the robustness of our results to adjustments for multiple hypothesis test- ing. We have partially addressed this issue by combining individual outcome variables into summary indices within outcome groups following Kling et al. (2007) and Anderson (2008) thereby reducing the number of tests. However, we still test the effect of payments over multiple outcome groups so we apply two types of adjustments. We apply multiple hy- pothesis testing adjustments for four of our key tests, similarly to the previous section (Ap- pendix A.5.2). First, we adjust for family-wise error rate (FWER) using the Bonferroni-Holm step-down procedure and the Westfall-Young free step-down method using re-randomization. Both adjust p -values upwards with the probability of making any kind of false rejections, but the Westfall-Young approach controls for potential correlation between these outcome groups via re-randomization (Anderson, 2008; Young, 2018). In Table A15, we compare the unadjusted p -values (Column (3)) to those adjusted with the Bonferroni-Holm procedure (Column (4)) and with the Westfall-Young procedure (Column (5)). The four estimated effects survive the Bonferroni-Holm adjustment process at the 5% significance level and the Westfall-Young process at the 10% level. Re-randomization in the Westfall-Young approach also allows us to jointly test the null of no overall effect of the experiment (bottom of Column (5)) which we can reject at the 5% level. Table A15: Adjustments for multiple hypothesis testing. (1) (2) (3) (4) (5) (6) FWER adjustments FDR adjustment Coefficients Std.err. Regular Bonferroni-Holm Westfall-Young Anderson’s p-value free-step down (re-randomization) sharpened q-s # SLMPs adopted 0.507 0.232 0.036 0.036 0.086 0.019 Agricultural income 0.221 0.089 0.018 0.036 0.086 0.018 Productivity index 0.255 0.071 0.001 0.004 0.025 0.005 Knowledge exch. 0.268 0.095 0.008 0.025 0.086 0.013 Joint . . . . 0.032 . Notes: The table presents adjusted p -values in Column (4)-(8). The Bonferroni-Holm adjustment does not use resampling nor treatment allocation permutations. We implement the Westfall-Young method with random permutations of treatment allocation (re-randomization, Monte Carlo simulations) instead of resam- pling (bootstrapped). Adjustment is based on 10000 random permutation. The joint test for all effects in the randomization based Westfall-Young algorithm is based on randomization-t statistics of Young (2018). To the extent that FWER adjustments are conservative because they control for any kind of false rejection, we also implement a false detection rate (FDR) adjustment. Instead 71 of controlling for the probability of any false detection, FDR adjusts for the expected share of false rejections (Anderson, 2008). The procedure calculates sharpened q -values which represent the share of false rejections if one were to reject the hypothesis at hand and all the other hypotheses with lower q -values. We present these sharpened q-values in Column (6) of Table A15. Controlling for FDR at q = .05 or q = .10 (conventional levels used by Anderson (2008) and Banerjee et al. (2015)), our treatment effects remain significant.20 The estimated effects of payments therefore survive both types of multiple hypothesis testing adjustments. 20 More precisely, controlling for FDR at 0.05 level, we can reject all null hypothesis and the expected false discovery rate will not be larger than 1.8% or 0.72 hypotheses. 72