WPS7000 Policy Research Working Paper 7000 Seeing Is Believing? Evidence from an Extension Network Experiment Florence Kondylis Valerie Mueller Siyao Zhu Development Research Group Impact Evaluation Team August 2014 Policy Research Working Paper 7000 Abstract Extension is designed to enable lab-to-farm technology dif- CFs and access the same extension network. In treatment fusion. Decentralized models assume that information flows villages, CFs additionally receive a three-day, central training from researchers to extension workers, and from extension on the new technology. They track information transmission agents to contact farmers (CFs). CFs should then train through two nodes of the extension network: from exten- other farmers in their communities. Such a modality may sion agents to CFs, and from CFs to other farmers. Directly fail to address informational inefficiencies and accountabil- training CFs leads to a large, statistically significant increase ity issues. The authors run a field experiment to measure the in adoption among CFs. However, higher levels of CF adop- impact of augmenting the CF model with a direct CF train- tion have limited impact on the behavior of other farmers. ing on the diffusion of a new technology. All villages have This paper is a product of the Impact Evaluation Team, Development Research Group. 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://econ.worldbank.org. The authors may be contacted at fkondylis@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 Seeing Is Believing? Evidence from an Extension Network Experiment * Florence Kondylis Development Research Group The World Bank Valerie Mueller * Development Strategy and Governance Division International Food Policy Research Institute Jessica Zhu Agricultural and Applied Economics University of Wisconsin, Madison JEL Classifications: O1, O3, Q1, D8 Keywords: information failure; technology diffusion; agriculture; Africa. * Corresponding authors' emails are fkondylis@worldbank.org; v.mueller@cgiar.org. This research was funded by the International Initiative for Impact Evaluation, Inc. (3ie) through the Global De- velopment Network (GDN); the Mozambique oce of the United States Agency for International Development; the Trust Fund for Environmentally and Socially Sustainable Development; the Belgian Poverty Reduction Partnership and the Gender Action Plan; and the CGIAR Research Program on Policies, Institutions, and Markets (PIM) led by the International Food Policy Research Institute (IFPRI) and nanced by the CGIAR Fund donors. The au- thors beneted from comments provided by Jenny Aker, Luc Behagle, Madhur Gautam, Markus Goldstein, Maria Jones, Rashid Lajaaj, Mark Lundell, Mushq Mobarak, Tewodaj Mogues, Glenn Sheri, David Spielman; and semi- nar participants at the CSAE (Oxford), the Mid-Western Economic Development Conference, the NEUDC at Boston University, The Ohio State University, the Paris School of Economics, the University of Georgia, the World Bank, and IFPRI. The views expressed in this article do not reect those of the World Bank, 3ie, or their members. The authors would like to thank Pedro Arlindo, Jose Caravela, Destino Chiar, Isabel Cossa, Beatriz Massuanganhe, and Patrick Verissimo for their collaboration and support throughout the project. John Bunge, Ricardo da Costa, and Cheney Wells provided excellent eld coordination; Siobhan Murray and João Rodrigues impressive research assistance. 1 1 Introduction Agricultural innovation is necessary to accelerate growth and achieve food security in Africa (Hazell, 2013). Despite availability of yield-enhancing technologies, adoption rates in Sub-Saharan agricul- ture remain low (Gollin et al., 2005). A growing literature identies information failures as an impediment to the technological diusion process in agriculture (Bandiera and Rasul, 2006; Conley and Udry, 2010; Munshi, 2004). Less documented are the modalities through which information can best diuse and boost adoption of productive farming practices. Agricultural extension services are designed to facilitate the diusion of innovations from lab to farm. In developing countries, they account for large shares of government expenditures on agriculture (Akroyd and Smith, 2007). These substantive investments are seldom supported by causal evidence regarding their eectiveness as a whole, or of a particular modality (Anderson and Feder, 2004). Contact farmers (CFs), who serve as points of contact between extension agents (EAs) and other farmers, are ubiquitously used as messengers of information in developing countries. Ecacy of the CF modality rests on two key assumptions. First, EAs will eectively train CFs to adopt and demonstrate new technologies to peers. Frequent EA visits are supposed to elicit a process of experiential learning among CFs. Second, other farmers' exposure to CFs will encourage wider adoption in the community, through a peer learning process. Despite some evidence of implementation and accountability constraints, and perhaps for lack of a viable policy alternative, the CF model persists across Africa (Gautam, 2000). Formally documenting returns to additional low-cost, scalable interventions to help leverage these large investments in agricultural extension services could signicantly aect the path of technology diusion. We exploit a large-scale, government-run randomized controlled trial (RCT) to measure the impact of augmenting the CF model with a direct training on the diusion of a new technology in central Mozambique. Our treatment consists of adding a direct CF training to an existing CF model, holding everything else constant. In practice, CFs in treatment villages receive two three-day trainings (one in 2010, one in 2012) on a yield-enhancing technology at district headquarters, by the same experts and using the same curriculum as provided to EAs. All EAs were trained on sustainable land management (SLM) and were expected to train their local CFs to demonstrate the technology to other farmers in 200 villages. All CFs were provided demonstration kits to encourage 2 adoption and diusion of information to other farmers. We augmented the CF model by centrally training CFs on SLM in 150 randomly selected (treatment) communities. The training format was part lectures, part hands-on, with similar content and breadth as the EA training. The central training is the only dierence between treatment and control, and all 200 villages adhere to the status quo CF model. We use two rounds of follow-up survey data on 200 CFs and a random sample of over 5,000 other farmers to examine the impact of adding a central training on knowledge and adoption of the technology, as well as agricultural production. The CF model enables a process of experiential learning among CFs through the use of demon- stration activities and regular on-site feedback from EAs. This practice is similar to on-the-job, learning-by-doing processes in other labor markets. Neoclassical growth theories suggest learning- by-doing may be of equal importance to formal training in explaining human capital formation as a production input (Lucas, 1988). While learning-by-doing theories are supported in the context of rm or plant-level studies (Levitt et al., 2012; Thompson, 2010), empirical evidence of the sig- nicance of learning through extension programs on agricultural growth is mixed (Bindlish and Evenson, 1997; Purcell and Anderson, 1997; Gautam, 2000; Anderson and Feder, 2007; Benin et al., 2007; Davis et al., 2012; Waddington et al., 2014). We contribute to this research agenda by formally documenting the impact of augmenting an existing, decentralized extension model with a relatively low-cost centralized training modality. Adding a direct training may aect technology adoption among CFs through three broad cate- gories of mechanisms: increased quantity of information, enhanced learning experience, and channels other than knowledge. First, the curriculum in a direct training may increase the quantity of infor- mation transferred (e.g., number of techniques taught). Central trainings are oered in an enclosed setting under the supervision of project sta, and extension agents present the material from the course manual. This plausibly increases the chance that the intended curriculum is covered. Second, the centralized format of the training may enhance the learning experience. For instance, the formal setting may add credibility to the information. While the use of course materials hardly aects learning indicators in other settings (Tan et al., 1999; Glewwe et al., 2004, 2009), use of computing technology as a complementary input to a standard curriculum has been shown to have a positive eect on learning (Banerjee et al., 2007; Linden, 2008). For these reasons, the information set shared during a training held at district headquarters may be (perceived to be) of higher quality 3 than what is given during eld visits from the EAs. Peer learning will also likely be more pronounced during a centralized training, as CFs with similar characteristics get to share information and jointly interact with the material. Third, a direct training could increase CFs' adoption through other channels than knowledge. For instance, the training may improve EA-to-CF accountability. Directly trained CFs may demand more information from EAs (Björkman and Svensson, 2009; Banerjee et al., 2010). Being formally trained could also build empowerment, reinforcing the identity of the CF as community messenger and their propensity to lead village-level demonstration activities. Similarly, attending a training in the district town for a few days may make CFs feel special relative to control CFs. Alternatively, a centralized training could create a momentum among peers to adopt the new practices, akin to herd behavior (Banerjee, 1992; Karlan et al., 2014). We nd a statistically signicant increase in CF adoption of SLM when CFs have access to a direct training in addition to the status quo extension modality. Private returns in the form of labor savings and yield benets in dry years accompany increases in adoption. However, CFs' knowledge scores on the SLM curriculum are unaected by the direct training. Increasing demonstration of SLM practices could reinforce the perceived benets of SLM among peers by increasing knowledge and reducing the uncertainty of SLM benets (Foster and Rosenzweig, 1995). Yet, boosting CF demonstration and adoption through a direct training does not aect other farmers' practices within the community in our context. Patterns of CF-farmer interactions suggest that the direct training did not additionally stimulate CFs to fulll their role as village messengers. Interestingly, variations in treatment eects by CF characteristics and similarity in cropping patterns indicate that relevance of expected cost-benet margins aect the diusion process. For example, pit planting adoption rates increase when a CF's crop portfolio matches the farmers'. Hence, our results corroborate the idea that the proximity of the source of information may be the primary constraint on changes in farmer behavior (Munshi, 2004; Feder and Savastano, 2006; Bandiera and Rasul, 2006; Conley and Udry, 2010) and, therefore, that the process of CF selection may aect the pace of diusion (Beaman et al., 2014; BenYishay and Mobarak, 2014). Overall, our ndings suggest augmenting decentralized extension programs with a direct training modality can improve their eectiveness in getting CFs to demonstrate new technologies. Our cost- benet analysis shows net private returns of up to USD 76 per CF. Yet, a direct CF training leads 4 to modest diusion to others in the community. Taken together, these results imply that adding a direct training modality on its own may not be enough to reform the speed of technology diusion. Further study is needed to build up the evidence base, using larger samples, improved measurement techniques, and testing complementary policy actions to make extension services work for farmers. In what follows, we detail the Mozambique extension policy and network at baseline (Section 2). We then describe the evaluation design and empirical strategies used to identify the impact of adding a direct CF training on technology diusion (Section 3). Section 4 presents estimates of impact on CF knowledge, adoption, and productivity, other farmers' adoption and knowledge, as well as measures of cost-eectiveness. Section 5 discusses implications of this study for policy and future research. 2 Agricultural Extension Constraints in Mozambique 2.1 National Extension Coverage Mozambique's agricultural extension network was created in 1987 and began to operate in 1992 after the peace agreement. During the past two decades, the Ministry of Agriculture (MINAG) has promoted and expanded extension networks (Eicher, 2002; Gemo et al., 2005). EAs are employed by the District Services for Economic Activities (Serviços Distritais de Actividades Económicas) and operate at the subdistrict level to disseminate information and new techniques. The system assumes that information ows linearly: agricultural innovations are created by researchers, then distributed by extension workers, and nally adopted by producers (Pamuk et al., 2014). Countrywide, coverage is as low as 1.3 EAs per 10,000 rural people (Coughlin, 2006). Given this shortage, EAs are inclined to visit the same set of villages every year based on their achievements and potentials (Coughlin, 2006). Only 15 percent of farmers report receiving extension services (Cunguara and Moder, 2011). At the time our study was designed, the present National Plan for Agricultural Extension and Extension Master Plan aimed to develop the decentralization of services at the district level; in- crease participation of targeted groups (women and marginal farmers); and enhance partnerships with other actors, such as the private sector and nongovernmental organizations (Gallina and Chidia- 5 massamba, 2010). Given the importance the government places on decentralized extension services and the lack of rigorous evidence to date, formally documenting the impact of this policy action seems warranted (Gautam, 2000). 1 In what follows, we describe the details of the status quo exten- sion model operating at baseline in our study area. 2.2 Study Area We worked in ve districts of central Mozambique: Mutarara (Tete Province), Maríngue and Chemba (Sofala Province), and Mopeia and Morrumbala (Zambézia Province; Figure 1). This area receives nancing from a large World BankGovernment of Mozambique investment to sup- port the development of the extension network (Smallholder project ). The project provides three levels of agricultural technical assistance: each district has a facilitator, an environmental specialist, and eight EAs. A district is subdivided into four administrative posts (posto administrativo ) that include about 810 communities (aldeia ). EAs periodically receive training from the district spe- cialists. 2 Each community has a designated contact farmer (CF) who receives direct assistance from the two EAs placed in his administrative post. 3,4 CFs receive visits from EAs monthly. They were instated to respond to other farmers' demands for technical assistance and provide advice through demonstration activities. A CF model of extension may not foster learning and adoption among CFs. EAs are typically challenged to reach the communities they serve. Designating CFs may therefore not adequately ad- dress the supply-side constraints of extension services. Another concern is that information may get diluted from the central level to CFs. For instance, EAs may not cover all techniques, suciently train the CFs, nor adhere to the expected format. Since CFs do not know what curriculum their EA should follow, accountability may be low. Finally, periodic visits from the EA may not be sucient 1 Recent work has employed a quasi-experimental design to evaluate the impact of extension and found a positive impact of extension on farm income in Mozambique (Cunguara and Moder, 2011) 2 In October 2010 and November 2012, these trainings were dedicated to SLM. 3 The ratio of EAs per administrative post in our study area is on par with the 2013 national average of 1.89 (Gêmo and Chilonda, 2013). This ratio is calculated using the 2010 gures from the Direçåo Nacional de Extenså Agraria (DNEA), available at the following URL: http://www.worldwide-extension.org/africa/mozambique/s-mozambique. 4 EAs can choose which CFs to work with, and do not necessarily split responsibilities. Hence, a given CF may interact with both EAs in his administrative post. CFs are typically chosen by the community. In 2010, CFs had been in their position for three years on average, with a standard deviation of 3. This indicates the majority of CFs were already commissioned by the project prior to our intervention. 6 in getting CFs motivated to demonstrate to others in their community. The underlying assumption of the CF model is that, through peer learning, a change in CF demonstration eort should aect the process of diusion to other farmers in the community. By exogenously aecting CFs' adoption of a new technology, our experiment directly tests whether the CF model is suited to promoting technology adoption on a large scale. Allowing the ITT estimates to vary by CF characteristics provides qualitative evidence of existing barriers to knowledge transfer. 3 Experiment and Data We run a large eld experiment to test for eective knowledge diusion under the CF model of extension, and isolate the additional impact of directly training CFs. A new technology, SLM, was disseminated through the extension network for the rst time in 2010. Our study started in October 2010 and ended after the main 2013 cropping season, thus spanning three main agricultural seasons (Figure 2). We collected three rounds of data: a rapid CF baseline and two CF and household-level follow-ups, respectively, 15 and 27 months after the rst SLM demonstration season. By baseline, we refer to September 2010 and earlier. The initial demonstration season in our study was 2011. Our surveys captured the 2012 and 2013 adoption seasons. This section details the experiment and data sources. 3.1 Sustainable Land Management Sustainable land management (SLM, or conservation farming) is a yield-enhancing farming tech- nology that consists of a bundle of techniques adapted to local crops and agro-ecological condi- tions (Haggblade and Tembo, 2003; Thierfelder et al., 2015). 5 In the Zambezi valley, the recom- mended SLM technology package encompasses seven SLM techniques: Mulching, Crop Rotation, 5 A direct implication is that, while positive yield eects of SLM are relatively well documented for Southern Africa (Haggblade and Tembo, 2003; Thierfelder et al., 2015), there is little evidence on the returns of individual SLM techniques. 7 Strip Tillage, Pit Planting, Contour Farming, Row Planting, and Improved Fallowing. 6 Mulching covers the soil with organic residues to maintain soil humidity, suppress weeds, reduce erosion, and enrich the quality of the soil cover. Crop rotation rotates crops on a given plot to improve soil fertil- ity and reduce the proliferation of plagues. Strip-tillage prevents opening the soil, such as through plowing, harrowing, or digging on land surrounding the seed row. Pit planting consists of construct- ing permanent holes 15 cm deep around the base of a plant, such as maize, to aid water and nutrient accumulation. Contour farming is the use of crop rows along contour lines fortied by stones (or vegetation) to reduce water loss and erosion on sloped land. Row planting improves productivity by improving access to sunlight and facilitating weeding and other cultivation practices (for instance, mulching and intercropping) by providing space between rows. Improved fallowing reduces tempo- rary productivity losses from fallowing through targeted planting of species that recharge the soil in a shorter time frame. There are important complementarities across these techniques, which are expected to generate savings in labor time during the main season. For instance, combining strip-tillage with pit planting will ensure that pits do not need to be excavated every year. This should save labor at the seeding stage over the traditional methods of tillage and planting in ridges. Similarly, strip-tillage and contour farming combined will save time, since the terraces will not have to be prepared every year for seeding. Combining mulching and pit planting can also help maximize the nutrient retention of the soil around the maize crop and minimize the need for weeding. We asked CFs to recall their familiarity with these techniques at baseline (Table A.1). SLM exposure varied widely across techniques and farmer types. Twenty-one percent of CFs had heard of improved fallowing relative to 10 percent of other farmers. In contrast, 76 percent of CFs knew of mulching, compared to 34 percent of other farmers. This suggests that some, if not all, SLM technologies taught in the CF training and disseminated by EAs pose as reasonable instruments to track knowledge diusion in the Zambezi valley. 7 6 Intercropping was included in the curriculum, but is excluded from the analysis as it was already widely adopted at the time of the intervention by CFs (98 percent) and other farmers (76 and 81 percent of women and men, respectively). Including the technique bears little consequence on our point estimates (not reported). 7 The project had started to disseminate mulching, strip tillage, row planting, and crop rotation as early as 2008. However, the formal practice was sparse at the time of the intervention and most EAs and CFs had not received a formal training on SLM techniques, or been instructed to transfer their knowledge to their peers. 8 3.2 Training We now describe the trainings delivered to EAs and CFs in the context of our study. First, all EAs serving administrative posts within our study area received two three-day training courses on SLM techniques in October 2010 and November 2012 (prior to the main planting season). Technical sta from the Ministry of Agriculture (MINAG) developed the educational agenda on SLM practices, and the training was delivered by MINAG's district technical sta with support from one sta from the central project team. Half of the training sessions were devoted to in-class lectures, and the other half consisted of hands-on demonstrations. The syllabus included a thorough review of the advantages of each SLM technique over less-environmentally desirable ones. 8 The EA training also highlighted good practices in fostering interactions between EAs and CFs. The centralized CF trainings were held a few weeks after those of the EAs. The content of the CF training was similar to that received by EAs, and was delivered by the same district-level and central MINAG sta. 9 , 10 The cost of a direct training per CF per year was 74 USD. Over three agricultural seasons, this represents a modest 12.8 percent increase in the total salary and training cost of the extension network. 11 After these trainings were completed, all EAs worked with their CFs to disseminate the SLM techniques most pertinent to their local area on their (own or communal) demonstration plots, regardless of their CFs' treatment status. All CFs received a new toolkit (a bicycle, tools to plow the land, and smaller articles). 12 A second toolkit with similar items (including a bicycle) was provided to all CFs again in July 2012. The only dierence between treatment and control CFs is 8 The main charts from the class can be found here: http://siteresources.worldbank.org/INTDEVIMPEVAINI/Resources/Flipchart_deAC_anonymized.pdf. The general curriculum used by the MINAGRI sta is provided on this site: http://siteresources.worldbank.org/INTDEVIMPEVAINI/Resources/Manual_AC_FINAL.pdf. The hands-on component of the training was not recorded but followed closely the techniques discussed in class. 9 In some districts, district sta relied on their EAs to help during the hands-on sessions. This could contami- nate our results by lowering the amount of on-farm attention treatment CFs subsequently received from their EAs. This may lead us to underestimate information ow in the central training arm, and overestimate it in pure CF model. Reassuringly, as mentioned above, we do not nd that EAs devoted more time to visiting CFs in treatment communities, relative to control. 10 Given the low literacy of farmers, a lm covering all techniques substituted the initial lecture format in the second training of the CFs in 2012. 11 The monthly EA salary costs were at USD 210 per EA from data provided by the DNPDR and the Smallholders project team, and EA training costs ran at USD 370 per training. Each EA supervised on average 5 treated CFs over the course of 36 months. There are obviously other, non-wage costs to running an extension network. To the extent that we do not account for these additional costs, we overestimate the relative cost of adding a direct CF training. 12 The toolkit distribution was planned, independently of our intervention, by the project sta. The previous distribution had been done in 2007 and, by 2010, the items were deemed too old to function. 9 that treated CFs received an additional direct training on SLM. 3.3 Experimental design At baseline, CFs and EAs in our ve study districts operated under the CF model of extension in all communities. From these districts, we randomly selected 200 communities (with 200 CFs) in 16 administrative posts, to which 30 EAs were assigned. All EAs received SLM training. We randomly assigned CFs in 150 treatment communities to the augmented version of the CF model (treatment), stratifying the assignment at the district level. Control (50) and treatment (150) CFs received SLM training during visits from their EAsthe status quo CF extension modality. Treatment CFs additionally received the direct CF training described above. 13 This design allows us to isolate the additional eect of a direct training, implicitly testing for eective knowledge diusion under the CF model. For this purpose, we held constant all other extension interventions across treatment and control communities. Specically, in line with the status quo modality, all CFs in the study area receive assistance from their EAs and a tool kit to set up and maintain a demonstration plot within the community. These demonstration plots are used by (1) EAs to teach and assist CFs in implementing at least one of the agricultural practices of the CF's choice, and (2) CFs to demonstrate the new techniques to other farmers in their community. In practice, the CF-level random assignment was implemented as follows. Each EA team at the administrative post level was in charge of inviting treatment CFs to the central SLM trainings. During the EA trainings on SLM, district sta explained the physical impossibility of training all CFs at once and that a lottery had been used to select the participating CFs. EAs were then given the list of randomly chosen treatment CFs. An attendance sheet was taken at CF training by the district sta. In October 2010, only four treatment CFs did not attend the training (all in the Mopeia district), and there was no contamination to control CFs. 14 Since district sta may have 13 The full design consists of multiple treatment arms. A second treatment arm was overlaid on our central training that randomly assigned 75 of the 150 treated communities to have an additional trained female. This second treatment is the subject of a separate study. In the present study, we pool the two treatments together, to examine the impact of having at least one CF in the community trained on SLM on farmers' outcomes. A third randomized treatment arm was overlaid on the rst two that attempted to provide dierent performance-based incentives for the CFs to reach farmers in both villages that were assigned to the direct training and control communities. These incentives were never announced to the CFs, and we show that they did not have any statistically signicant eect on our outcomes of interest (not reported). Nonetheless, we control for this third treatment arm in the regression analysis. 14 These CFs were trained by the EA on an individual basis, and the follow-up training was veried. 10 an incentive to misreport attendance, we performed independent audits. First, we veried that the attendance list reected the (randomly assigned) eligibility, and found no contamination of the control group. Second, we showed up unannounced at the trainings in all ve districts and veried attendance verbally. Finally, attendance lists were back-checked: a random set of listed participants were visited in November and December of 2010 and asked whether they attended the SLM training. Our results from these audits indicate that attendance was genuine. Similar checks were performed on the 2012 training. While the attendance list was equally validated, participation was not universal and contamination was quite substantial. Of the treated communities, 63 percent had at least one CF attend the training, and 16 percent of control com- munities had a CF attend. 15 These gures signal statistically signicant exposure of control CFs to the treatment in 2012. While this may hamper our ability to statistically dierentiate the two training models on CF behavior in the 2013 (second follow-up) survey round, our results on other farmers at endline are arguably robust to this contamination. Increased demonstration by control CFs in the 2013 growing season is unlikely to have aected farmers' adoption in that same season. There are two important limitations to our identication strategy: one concerns the intensive margin of EA support to treatment CFs relative to control CFs, while the other relates to the extensive margin of EA attention. First, direct training and EA support are likely complementary inputs. Therefore our estimates capture the overall eect of augmenting the CF model with a direct training. We cannot disentangle the impact of learning during the central training from that of improved learning during regular EA-to-CF tutorials as a result of the direct training. Second, our design implies that each EA will work with both treatment and control CFs in his administrative post. 16 A threat to our identication stems from the fact that CFs may request dierent levels of attention from their EAs across treatment assignments, displacing EA time away from the other treatment statusthe extensive margin of EA attention across treatments. For instance, treatment CFs may request more follow-up visits from their EAs, cutting into the time 15 The contamination likely was caused by a combination of self-selection and EA oversight. CFs in the control group could have easily learned about the trainings from peers in other communities. Since EAs and district sta were involved in organizing the training, it is easy to see how a well-connected CF might have been invited in. 16 A limitation of working with an existing extension network is that we could not withhold information from a random group of CFs by shutting down their interactions with their assigned EAs. Given the small number of extension workers (30), reasonable levels of statistical power cannot be reached by assigning the intervention at the EA level. We do verify that extension agent characteristics are balanced across treatment and control communities at midline (Table A.2). 11 EAs devote to control CFs. Reassuringly, we nd that control and treatment CFs received equal amounts and types of attention from their EAs in the year after the training (Table A.3). 3.4 Data We conducted two follow-up surveys after the rst training and demonstration season (October 2010 to April 2011). A 2012 (midline) round and a 2013 (endline) round form a panel of households and CFs in the study area. 17 We randomly sampled 18 non-CF households in each community from a full listing performed by our enumerators ahead of the survey. At both midline and endline, households were visited twice: pre- and postharvest. This allows us to observe SLM practices when they are most visible, just after planting (preharvest, from February to April), and to record production data after harvest (from mid-May on). Hence, our eldwork ran from February to April and May to August in 2012 and 2013. Midline and endline surveys gathered longitudinal CF and household information on the two main agricultural seasons that followed the rst demonstration season. Our eldwork included ve survey instruments: a household questionnaire, a household agricultural production questionnaire, a CF questionnaire, an EA questionnaire, and a community questionnaire. The household survey was also administered to all 200 CF households, in addition to the specic CF survey. These surveys provide household demographics, SLM knowledge for the two main agricultural producers in the household, individual and plot-level SLM adoption, and production information for approximately 3,600 non-CF households in 200 communities. Since the plot roster identies the adult in charge of making agricultural decisions for each cultivated plot, we obtain individual measures of knowledge and adoption for a sample of 5,884 and 5,071 individuals at midline and endline, respectively. A rapid baseline survey was administered to all CFs in August 2010, before the district-level randomization. This provided data to perform balance tests on the success of the randomization, using the pre-intervention characteristics of CFs by treatment status. Figure 2 illustrates the timing of the surveys and CF trainings over the course of the four-year study. 17 Operational constraints precluded us from conducting a household survey at baseline. 12 3.5 Descriptive Statistics We briey describe the average characteristics of farming individuals in our sample (Table A.4). More than half of the individuals are women, and the prevalence of female headship is consistently high (approximately 30 percent) for the region (TIA, 2008). The average farmer is 38 years old with two years of schooling. Most plot owners are married with three children, and live in a single-room house made of mud and sticks with a palm or bamboo roof (not reported). Farmers possess 2.2 hectares of land on average, with a standard deviation of 2.1. CFs are more knowledgeable (Tables 1 and 2), more educated, and wealthier (Table 3) than other farmers. While CFs are positively selected in attributes, they are also well known in their communities: 84 percent of farmers in the control group declared knowing them personally. However, only 72 percent of the same group of farmers reported knowing that these individuals assumed a role as CF in their community. We also note that usage of demonstration plots was quite high and not statistically dier- ent across treatment and control communities (Table A.3). Of the CFs in treatment and control communities, 85 percent maintained a demonstration plot. 18 Thus, changes in patterns of CF- to-other-farmers information diusion across the two modalities can be interpreted as resulting from variations at the intensive margin of CFs' activities (e.g., number of techniques demonstrated, quality of the demonstration). 3.6 Balance We use data from the baseline CF survey as well as time-invariant and retrospective information collected in the 2012 household survey to check for balance across treatments. Table 1 indicates minor dierences between CFs in the treatment and control communities. Treatment CFs spent almost four more hours a week working as a CF (pre-intervention) and had slightly more recent training when we condition on being formally trained. Control CFs were exposed to a greater number 18 There was no instruction, however, as to what type of plot should be used for demonstration activities. CFs could choose to use their own, private plot or communal land. Hence, we present the demonstration results for any plot (own or not). 13 of techniques prior to the intervention. 19 In spite of these dierences, (recalled) pre-intervention adoption rates among CFs in control and treated communities were similar, as were other farmers' (recalled) baseline SLM learning and adoption rates (Table 2). 20 3.7 Measuring Information Diusion and Behavioral Change Central to identifying variations in information diusion is measuring changes in learning and agri- cultural practices. Our study rests on the reliability of our markers of individual SLM knowledge and adoption. We focus on three outcomes: a knowledge score, the number of techniques the re- spondent identied by name, and the number of techniques the respondent reported having adopted on any plot. 21 The knowledge score is a continuous measure based on the number of correct re- sponses provided in the 23-question exam, covering all SLM techniques. For CFs, the majority of the analysis rests on their self-reported adoption of techniques on any plot (demonstration or not). 22 Since the CFs were encouraged to choose the techniques most relevant to their local conditions, our main results focus on unweighted aggregate measures of knowledge and adoption. However, we create a second set of weighted knowledge and adoption outcomes as a robustness check. Prior to aggregation, we multiply the technique by a weight based on its relative importance to maize revenue. This is done as follows. First, we compute a vector of weights, based on a regression of maize revenue on adoption of the seven individual practices. Second, we compute adoption and knowledge 19 Given that CFs in treatment villages spend more hours a week working as a CF at baseline, we will include the variable as a control in the regression analysis. 20 Balance tests for the CFs' and other farmers' knowledge and adoption of individual SLM techniques at baseline are reported in Tables A.1 and A.5. Because these values are based on recalled data, the tests should be interpreted with caution. Even though mean comparisons indicate there are no statistically signicant dierences, recall bias may be present. We therefore do not exploit the recalled information beyond balance checks. 21 Our decision to focus on the knowledge score and self-reported adoption outcomes is motivated by the conclusions in Kondylis et al. (2015). Using the midline survey data, they nd learning outcomes based on knowledge exams provide more precision than know-by-name questions, inasmuch as they reveal the true knowledge of those individuals less familiar with the name of the technique yet more familiar with its purpose and usage. Objective adoption measures were also collected for two plots per household and largely corroborate the self-reported outcomes. In our triangulation of the self-reported versus observed adoption, we nd that false reporting is negligible. Since objective measures of adoption are collected for only a subset of plots (one per respondent) at midline and a subset of the sample at endline, we instead focus on a more inclusive measure of adoption provided by self-reports of interviewed men and women. 22 There are slight dierences between adoption measures which include and exclude the demonstration plot. This is due to the fact that some CFs demonstrate on communal land (29%). We verify the results are not driven by communal propriety of the demonstration plot (not reported). Since 71% of demonstration activities are carried out on CFs' own plots, we choose to use pooled adoption on and o demonstration plots as our main marker of adoption. This improves our precision but does not aect our conclusions. 14 indices of practices weighted by these correlations between adoption and maize production. We additionally explore patterns of knowledge and adoption specic to individual techniques. We use responses from the same knowledge exam to quantify farmer knowledge of individual SLM techniques, categorizing questions by technique. The knowledge of a specic technique is a [0,1] continuous variable that depicts the share of questions pertaining to the practice that the respondent has answered accurately. The adoption of a technique is captured by a binary variable that indicates whether the farmer adopted the technique on at least one of his plots. Knowledge, adoption, and perception of the SLM techniques were collected at the individual level from the household questionnaire. Two respondents were interviewed: typically, the household head and the head's partner or spouse. 23 Our sample of CFs and other farmers consists of those who reported their personal information, participated in an agricultural knowledge exam with questions related to each specic SLM practice, and self-reported their SLM adoption rates. Our nal regres- sions samples consist of 347 CF-year observations and 10,955 person-year observations. 24 Selective sample attrition is of concern, and we address it in the next section. 3.8 Empirical Strategy We causally estimate the intent-to-treat (ITT) eects of a community being assigned to a direct CF training (relative to a status quo CF modality) on the SLM knowledge and adoption of CFs 25 and other farmers in the community, Y, using a simple reduced-form specication: Yihjt = β0 + β1 Tj + β2 Xi,h,j + νt + i,h,j . (1) 23 In the case of polygamous households, the main spouse was interviewed. Only 2.7 percent of our sampled households are polygamous. 24 The number of villages that were administered the CF survey were 179 and 172 in 2012 and 2013, respectively. The number of farmers interviewed in 2012 were 6,252, and 5,290 in 2013. Sample sizes vary in descriptive statistic tables and some regression tables, due to the addition of variables excluded from the main analysis. 25 CF-level regressions control for community-level CF outcomes and characteristics. In those communities where we (randomly) assigned an additional woman farmer to be trained, we measure increased village-level exposure by regressing the maximum value of CF outcomes within the village on the maximum (mean) value of binary (continuous) covariates. Switching to mean of outcomes, and controlling for mean and max of all covariates does not aect our conclusions. 15 T takes the value 1 for each community j with a trained CF. Individual i, household h, and commu- nity characteristics are included in the vector X to improve the precision of the estimated coecients. An indicator for the second follow-up survey,νt , is also included to capture the eect of time-specic events on behavior. 26 We estimate all main regression models on the pooled sample, controlling for survey-year xed eects. We also use the Huber-White heteroskedasticity-robust estimator to calculate the standard errors when using the sample of CFs. For the other farmer regressions, we cluster the standard errors at the community level to allow for arbitrary correlation of treatment eects within the community. 27,28 4 Results 4.1 CF Learning and Adoption We rst examine the ITT estimates of a direct training on unweighted aggregate measures of CFs' knowledge and adoption (Panel A, Table 4). While control CFs adopted on average 3.74 tech- niques, we detect that CFs adopt on average a 0.73 additional technique in response to the training (statistically signicant at the 5 percent level). The associated eect size is large at 0.39 standard deviations in the control group, or a 19.6% increase relative to the mean in the control. Next, we run similar specications with the weighted versions of the outcome. 29 Control CFs adopt on average 26 We include variables that reect CF (or other farmer) demographic characteristics: age, primary school com- pletion, whether the individual is single (and a separate widow dummy for the other farmer sample), number of children, total landholdings, the number of rooms in the house, the number of hours worked by the CF at baseline, an indicator for a missing response for the baseline CF variable, district indicators, and indicators for treatment arms not analyzed in the present study. Our results are robust to specications that omit the demographic characteristics (Tables A.6 and A.7) or replace district with administrative post xed eects (Tables A.8 and A.9). 27 Attrition rates at the household and CF level are not statistically dierent (Table A.10) nor correlated across treatment groups (Table A.11). Household attrition rates appear consistent with those of other studies in the same region (De Brauw, 2014). A probit regression reveals that the greater the percentage of household members away in 2012 and the incidence of being single produces a greater probability of the household moving out of the sample (Table A.11). Age, the number of children of the household head, and exposure to a precipitation shock reduce the probability of moving out of the sample. 28 We perform two additional robustness checks to examine the sensitivity of our results to attrition (not reported). The rst diagnostic estimates (1) using the balanced panel. We show that the inclusion of individuals present in both rounds aects the precision of our point estimates rather than their magnitude and sign. The second check bounds the treatment eect for selective attrition using a method proposed by Lee (2009). This check conrms that selective attrition is unlikely to aect our conclusions. 29 The linear model of maize revenue controls for adoption of each SLM technique, as well as household demographics, as in Table 4, and production inputs. The regression estimates are presented in Table A.12. The vector of weights 16 60.1% of the practices, and directly trained CFs increased adoption by about 10.6 percentage points (statistically signicant at the 5% level; Panel B, Table 4). This eect is similar in magnitude to that obtained on our unweighted index, with a similar eect size of 0.40 standard deviation in the control group, or a 17.6% increase relative to the mean in the control. Weighting our knowledge index conrms that directly training CFs did not aect their knowledge scores. Overall, unweighted and weighted results suggest that a direct training was eective in raising CFs' adoption of SLM practices, with little eect on knowledge scores. To shed light on changes in the technique mix, we disaggregate the ITT estimates of adoption by technique (Table 5). Despite positive point estimates for all practices, statistical signicance is achieved for only strip-tillage, pit planting and contour farming (statistically signicant at the 10%, 1%, and 10% levels, respectively). 30 The magnitude of these eects is substantial, ranging from 28.3% to 65% increases relative to the control mean. To account for multiple hypothesis testing, we adjust our inferences for familywise error rates (’idák, p-value = 0.015; Bonferroni, p-value = 0.014), following Abdi (2007). The eect on pit planting adoption is robust to multiple hypotheses testing (Figure 4), with a 28 percent increase in adoption relative to the control mean. 31 Finally, we examine changes in technique-specic knowledge as a result of the direct training. In line with our aggregate measures of knowledge, Table 6 indicates that adding a direct training to the CF model did little to increase CFs' knowledge scores. consists of the estimated regression parameter for each technique divided by the sum of the parameters over all seven techniques. Since some regression coecients are negative, we use improved fallowing as the reference weight. Thus, mulching takes value 0.227, strip tillage, 0.177, pit planting, 0.157, contour farming, 0.201, crop rotation, 0.193, and row planting, 0.045. 30 We note that adoption trended downward in both treatment and control villages (not reported). However, these changes are fully attributable to a fall in demonstration of SLM from midline to endline, while adoption on non- demonstration plots actually increases from midline to endline (not reported). We additionally use rainfall data to pro- vide partial evidence that this trend cannot be explained by climatic conditions (NASA 1/2x 1/21981/2013 precipita- tion data available at http://power.larc.nasa.gov/cgi-bin/cgiwrap/solar/hirestimeser.cgi?email=daily@larc.nasa.gov). Indeed, a dry shock during the rainy season prior to our midline survey (2011) and endline (2012) surveys could aect adoption. Figure 3 displays yearly standardized cumulative rainfall in the rainy season over the 1981/2012 period, and their 95% condence intervals. We observe that rainfall in the study years (2010/2012) are within normal range. Nonetheless, we test whether adoption decisions vary by exposure to a dry spell over the rainy season (where dry is dened as whether the cumulative rainfall during the growing season was below the 31-year 25th percentile). We nd that rainfall anomalies do not explain variations in adoption (Table A.13). 31 We also follow Anderson (2008) and address multiple inference in two additional ways (not reported): (1) using a free step-down resampling method to our p-values for familywise error rate, and (2) employing the false discovery rate control methodology proposed by Benjamini and Hochberg (1995). All yield the same results. 17 4.2 Private Returns on SLM Farmers will adopt a technique only if it demonstrates (public or private) positive returns. Recent observational and experimental evidence documents positive maize yield eects of SLM techniques in southern Africa, as well as substantial labor savings (Beaman et al., 2014; BenYishay and Mobarak, 2014; Haggblade and Tembo, 2003; Thierfelder et al., 2015). We examine private returns to SLM to explain the observed increase in adoption among CFs, beyond their willingness to comply with the training. In practice, we modify (1) to estimate the ITT eects on maize yields and revenue, input use, and on-farm labor allocation. Table 7 (Panel A) presents maize yields (revenue per hectare) and total revenue accrued from a direct SLM training. Given our low level of statistical power, we present results on the full sample and winsorizing yields at 1% to account for outliers. 32 Results show positive though imprecise point estimates, indicating eect sizes on the order of 0.13-0.24 standard deviations. Since most SLM practices disseminated have water-conserving properties (Liniger et al., 2011), we further account for the possibility that rainfall patterns in the main growing season may mediate the impacts of the intervention. In practice, we add a control variable to distinguish eects by whether the community experienced a dry spell during the survey round: Dry Year, which takes value one if the cumulative precipitation in a given location fell below the 30-year rst quartile, zero otherwise. 33 In line with recent experimental evidence on pit planting (Beaman et al., 2014), we nd that training CFs on SLM has positive and large, if noisy, eects on maize yields and revenues during drier spells (Table 7, Panel B). The magnitude of the eects, on the order of 0.35 standard deviations, or 37% increase in both yield and total revenue, are in line with ndings from the literature that claim increases of 50 to 100% (Haggblade and Tembo, 2003). The measured yield eects from receiving a SLM training corroborate the notion that farmer adoption of SLM technologies is motivated by short-term yield advantages. In the absence of any statistically signicant dierences in input use as a result of the direct training (Table 8), these yield eects in dry conditions are credibly attributable to SLM adoption. 32 Winsorizing yields at 1% does not aect the control group mean as the entire top 1% of the distribution is in the treatment group. 33 For these computations, we use NASA 1/2x 1/21981/2013 Precipitation data available at http://power.larc.nasa.gov/cgi-bin/cgiwrap/solar/hirestimeser.cgi?email=daily@larc.nasa.gov. This measure of weather event is used by others in the literature, see for instance Jayachandran (2006). 18 An additional economic benet of applying SLM is in the form of large labor savings that follow from rening tillage operations and herbicide applications (Mazvimavi et al., 2011). These gains are expected to materialize from the second adoption season onward, since the rst year requires at least equal amount of land preparation as traditional practices (Haggblade and Tembo, 2003). Building on this literature and recent large-scale experimental evidence (BenYishay and Mobarak, 2014), our ndings support a delayed contribution of SLM to labor savings. We witness a substantial reduction in the number of hours spent seeding over the week preceding the interview and the total weeks spent farming at endline (Table 9). In particular, CFs spent 6.6 fewer hours seeding in the last week (a relative eect size of 0.63 standard deviations) and 7.1 fewer weeks farming over the last year (a relative eect size of 0.37 standard deviations). 34 While large, these point estimates are in line with the magnitude of eects mentioned in the literature (30 days per year reported in Haggblade and Tembo 2003). Since use of herbicides remained constant across treatment arms (Table 8), labor savings at endline are plausibly attributable to increased SLM adoption and complementary usage of the tools provided in the kit to minimize tillage operations. 4.3 Others Farmers' Knowledge and Adoption We now turn to CFs' ability to spread knowledge and adoption among other farmers in the commu- nity. We exploit the exogenous, positive shock in CFs' demonstration of SLM induced by the direct training to measure the extent of CF-to-others knowledge transmission. Since we cannot exclude the possibility that our treatment aected other farmers' adoption of SLM through channels other than demonstration, we adapt (1) and estimate the ITT of directly training CFs on adoption and knowledge on a random sample of other farmers in the community. Table 10 reports the ITT estimates of directly training CFs on CF-farmer interactions and other farmers' aggregate knowledge and adoption of SLM. First, we note that the direct training did not increase farmers' access to CFs. Second, other farmers' SLM knowledge and adoption are unaected by their exposure to a directly trained CF, despite the margin of gains in SLM awareness being larger than for CFs. These zero eects are robust to balancing the panel at midline, accounting for 34 These eects are robust to 1% top and bottom winsorizing. 19 selective attrition at endline (not reported), and cannot be explained by anomalous precipitations over the study period (Figure 3; FEWS 2012, 2013). Looking at ITT eects by technique conrms this general pattern (Table A.14). Recall the direct training led to a 15.9 percentage point increase in CF adoption of pit planting. Other farmers in treated communities were more likely to adopt pit planting by 2.8 percentage points (albeit a non-signicant eect). Placing a 95% condence interval around this point estimate allows us to rule out adoption rates higher than (0.028 + 1.96 × 0.018 =)6.3% for pit planting among other farmers. A back-of-the-napkin calculation on this (weak) pit planting result rules out a propagation rate higher than (6.3/15.9 =)39.6%. Adoption of pit planting by a CF would, at best, inspire less than half of a farmer in the community to adopt pit planting. This implies low and slow CF-to-other-farmers technology diusion in the augmented CF model: increasing demonstration of SLM has little eect on other farmers' behavior. Learning about other farmers' perceptions of labor savings associated with each SLM practice may shed light on the mechanisms underlying adoptionor, in our context, a lack thereof. We asked farmers whether they perceived each technique to require more labor eort, equivalent labor eort, or less labor eort than traditional cultivation practices. Farmers in the control group perceived all techniques to be labor intensive, with a range of less than 1 percent to 16 percent of farmers declaring the techniques to decrease the amount of labor required (not reported). Exposure to a trained CF does not favorably aect farmers' perceptions of adoption costs (not reported). These measures of communication and perceptions indicate that a direct training did not contribute to increasing CF-other farmers interactions, and that CFs' increased demonstration and use of SLM had little impact on other farmers' perceptions of these techniques. Lastly, we explore whether CFs' characteristics provoke heterogeneous responses among farmers. We focus on four CF indicators: above median educational attainment, above median age, above median landholdings, and production of the same two crops as the farmer. 35 Each regression separately adds an interaction of the treatment variable with one of the four indicators and the interacted indicator on its own. Working with an existing network of CFs, we could not exogenously 35 Our specication implies that having similar primary crops as the CF is an exogenous decision. Specically, we assume cropping decisions are made before adoption decisions, and cropping decisions are independent of the treatment. While we cannot verify the order of the respective planting decisions, we nd that other farmers' propensity to grow the same primary two crops as the CF is not aected by the treatment (not reported). 20 vary their education, age, wealth, or cropping patterns. Thus, the results that follow cannot be interpreted causally, but as descriptive evidence. In addition, CFs are, on average, of higher status than other farmers, which reduces the number of variations we have access to in establishing a counterfactual. Finally, to yield interpretable results, we need this exercise to focus on a single technique rather than on an aggregate measure of adoption. We focus on the adoption of pit planting among other farmers as the outcome, since it is the only single practice which was statistically signicantly adopted by CFs, when adjusting our inference for multiple hypothesis testing. Table 11 displays the results from interacted regression models accounting for farmer heterogene- ity in the treatment eects, with the additive eect of the treatment and its interaction with the CF characteristic reported in the last row. Overall, we nd CFs with above median total landholdings were 4.4 percentage points more likely to convince other farmers to adopt pit planting, a 64.7% increase relative to the control (signicant at the 10% level). Credibility in the source of informa- tion appears to inuence all farmers, CFs with larger farms perhaps commanding more trust and respect within the community. More interestingly, similarities in crop portfolios between CFs and other farmers appear to inuence adoption rates. Other farmers' adoption grew an additional 5.2 percentage points when they had access to a directly trained CF who grew similar crops to theirs. This is consistent with the idea that homogeneous farming conditions are conducive to social learn- ing (Munshi, 2004). Delays in adoption may stem from dierences in production technologies and an inability to extrapolate demonstrated activities to their own plot. 4.4 Cost-Benet Analysis To provide perspective on the cost-eectiveness of the program, we compare the average annual costs of directly training each CF to the average annual benets realized by the CF. We consider three scenarios where the private returns to a direct CF training are in the form of maize revenue, labor earnings, and both. 36 To compute the value of labor savings, we multiply the estimate of the intervention's impact on farm labor savings by the shadow value of labor. 37 Benets to other farmers 36 This approach ignores the social benets produced by the technologies which cannot be quantied over a short- term horizon, such as soil and water quality. 37 We price the shadow value of labor at the minimum agricultural wage oered in Mozambique. Minimum wage rates are provided by the U.S. State Department: http://www.state.gov/e/eb/rls/othr/ics/2013/204700.htm. Agri- culture is the lowest wage rate at 74 USD per month. 21 are excluded from all three scenarios given the non-signicant (negative) ITT point estimates on adoption. Calculations for the three scenarios are presented in Table 12. Positive net benets over one adoption season exist when we account for the labor benets only, and when we account for both labor and yield benets. The costs of training represent 12.8 percent of the total annual costs of running the extension system, 38 with benets ranging from USD -55 to USD 76 per CF per season. Although we fail to measure diusion to other farmers in our context, the predicted returns from CF's labor savings justify scaling a program of this nature. This motivates further research in extension modalities to improve the delivery of information to other farmers to augment the pool of program beneciaries. For instance, providing performance-based incentives to CFs and tempering the selection of CFs does appear to achieve greater rates of technological adoption within communities (Beaman et al., 2014; BenYishay and Mobarak, 2014; BenYishay et al., 2015). 5 Discussion Decentralized extension modalities continue to garner support in Africa despite criticism, often anecdotal, of their inecacy in reaching most farmers and providing relevant information. We designed an experiment in Mozambique to examine whether adding an in-depth centralized training on a new technology improves the knowledge and adoption of innovative agricultural practices. We show that adding a direct training to an existing CF model increases adoption of SLM among CFs. Net private returns come mostly in the form of labor savings associated with the new practices. Despite these gains in adoption, adding a direct training to the CF modality had little impact on CF knowledge scores. This could of course be the result of poor quality testing and measurement error (Laajaj and Macours, 2015). Alternatively, relative to the status quo extension modality, the direct training may not have changed adoption by increasing CFs' knowledge. Knowledge is not a necessary condition for adoption of a new technology, as manifestations of herd behavior indicate (Banerjee, 1992; Karlan et al., 2014). Instead, the direct training intervention may have gotten more 38 We focus on the costs specic to the SLM intervention, which include the annual per community cost of an extension agent and per community cost of the SLM training. 22 CFs to adopt SLM by strengthening their sense of identity as communicators in their community. The absence of dierences in the use of a demonstration plot, interactions with other farmers, and subjective happiness (not reported) across treatment arms however suggests CFs' dedication and esteem were unaected. Another possibility is that adding a centralized CF training may have heightened the quality and credibility of the information, beyond the scope of our knowledge test. The participatory nature of the training may have helped CFs convert the information into productive behavior, and fostered higher peer learning. Although adding a direct CF training successfully encouraged adoption of a new technology at the village level relative to the status quo model of extension, it failed to encourage higher diusion to other farmers in the community. There are a number of reasons why this may be the case. First, increased demonstration may not eectively address other barriers to adoption. Having access to a demonstration plot may need to be complemented by other learning inputs, such as CF time or other farmers' time. Adding a direct CF training does not address the fact that CFs' opportunity costs of time may limit interactions with peers. For instance, adding a performance-based incentive payment for contact farmers is shown to positively aect their impact in Malawi (BenYishay and Mobarak, 2014). Similarly, increasing demonstration of a yield-enhancing practice may not address other demand-side ineciencies, such as the tendency to delay adoption until protable (Foster and Rosenzweig, 1995), heterogeneity in farming conditions (Conley and Udry, 2010; Munshi, 2004), and social distance between messengers and peers (Feder and Savastano, 2006; Beaman et al., 2014; BenYishay and Mobarak, 2014). An alternative explanation is that farmers may learn more from their own experience than from their peers (Foster and Rosenzweig, 1995; Bryan et al., 2014; Dupas, 2014). Failing to notice a gap between knowledge and actual practice, and not the information set itself, may also pose a key barrier to learning. Hanna et al. (2014) nd that seaweed farmers in Indonesia acted on the information received only when it included descriptions of the relationship between yield and pod size from their own plot. If the main constraint to adoption of a protable practice such as SLM is not a lack of exposure or knowledge, but a failure to notice its benets, then augmenting the CF model will have little eect on the pace of diusion within the community. While we cannot reject that adding a direct training to a decentralized extension model is a cost eective intervention, more work is needed to understand the potential of community-level demon- 23 stration activities on technology diusion. The prole of the seed adopters inuences whether farmers act on the information they receive. When focusing on farmers with similar cropping patterns as their CF, we observe modest (yet statistically signicant) technology diusion. Com- plementary interventions, such as assigning dierent types of seed adopters (Beaman et al., 2014; BenYishay and Mobarak, 2014; BenYishay et al., 2015) or encouraging experiential learning in the community (Jones et al., 2015), may increase the pace of technology diusion in the context of decentralized extension services. References Abdi, H., 2007. The bonferonni and ²idák corrections for multiple comparisons. Encyclopedia of measurement and statistics 3, 103107. Akroyd, S., Smith, L., 2007. Review of public spending to agriculture. A Joint Study by the De- partment for International Development and the World Bank 5. Anderson, J. R., Feder, G., 2004. Agricultural extension: Good intentions and hard realities. The World Bank Research Observer 19 (1), 4160. Anderson, J. R., Feder, G., 2007. Agricultural extension. Handbook of agricultural economics 3, 23432378. Anderson, M., 2008. Multiple inference and gender dierences in the eects of early intervention: A reevaluation of the abecedarian, perry preschool, and early training projects. Journal of the American Statistical Association 103 (484), 14811495. Bandiera, O., Rasul, I., 2006. Social networks and technology adoption in northern mozambique*. The Economic Journal 116 (514), 869902. Banerjee, A., 1992. A simple model of herd behavior. Quarterly Journal of Economics 107, 797817. Banerjee, A., Cole, S., Duo, E., Linden, L., 2007. Remedying education: Evidence from two randomized experiments in india. The Quarterly Journal of Economics 122 (3), 12351264. Banerjee, A. V., Banerji, R., Duo, E., Glennerster, R., Khemani, S., 2010. Pitfalls of participatory programs: Evidence from a randomized evaluation in education in india. American Economic Journal: Economic Policy, 130. Beaman, L., BenYishay, A., Magruder, J., Mobarak, A. M., 2014. Can network theory based tar- geting increase technology adoption. Benin, S., Nkonya, E., Okecho, G., Pender, J., Nahdy, S., Mugarura, S., Kayobyo, G., 2007. As- sessing the impact of the national agricultural advisory services (naads) in the uganda rural livelihoods. IFPRI Discussion Paper. 24 BenYishay, A., Jones, M., Kondylis, F., Mobarak, A. M., 2015. Are gender dierences in performance innate or socially mediated? BenYishay, A., Mobarak, A., 2014. Social learning and communication. NBER Working Paper 20139. Bindlish, V., Evenson, R. E., 1997. The impact of t&v extension in africa: The experience of kenya and burkina faso. The World Bank Research Observer, 183201. Björkman, M., Svensson, J., 2009. Power to the people: evidence from a randomized eld experiment of a community-based monitoring project in uganda. Quarterly Journal of Economics 124 (2), 735769. Bryan, G., Chowdhury, S., Mobarak, A., 2014. Under-inversment in a protable technology: The case of seasonal migration in bangladesh. Econometrica 82 (5), 16711748. Conley, T. G., Udry, C. R., 2010. Learning about a new technology: Pineapple in ghana. The American Economic Review, 3569. Coughlin, P. E., 2006. Agricultural intensication in mozambique infrastructure, policy and institu- tional framework-when do problems signal opportunities. Maputo, EconPolicy Research Group, sponsored by Sida. Cunguara, B., Moder, K., 2011. Is agricultural extension helping the poor? evidence from rural mozambique. Journal of African Economies, ejr015. Davis, K., Nkonya, E., Kato, E., Mekonnen, D. A., Odendo, M., Miiro, R., Nkuba, J., 2012. Impact of farmer eld schools on agricultural productivity and poverty in east africa. World Development 40 (2), 402413. De Brauw, A., 2014. Gender, control, and crop choice in northern mozambique. Tech. rep., Inter- national Food Policy Research Institute (IFPRI). Dupas, P., 2014. Short-run subsidies and long-run adoption of new health products: Evidence from a eld experiment. Econometrica 82 (1), 197228. Eicher, C. K., 2002. Building african models of agricultural extension: A case study of mozambique. In: Workshop Extension and Rural Development: A Convergence of Views on International Approaches. pp. 1215. Feder, G., Savastano, S., 2006. The role of opinion leaders in the diusion of new knowledge: The case of integrated pest management. World Development 34 (7), 12871300. FEWS, January 2012. Agromet update: 2011/2012 agricultural season. FEWS, March 2013. Agromet update 2012/2013 agricultural season. Foster, A., Rosenzweig, M., 1995. Learning by doing and learning from others: Human capital and technical change in agriculture. Journal of political Economy, 11761209. Gallina, A., Chidiamassamba, C., 2010. Gender aware approaches in agricultural programmes mozambique country report: A special study of the national agricultural development programme (proagri ii). Tech. rep., Swedish International Development Cooperation Agency. 25 Gautam, M., 2000. Agricultural extension: The Kenya experience: An impact evaluation. World Bank Publications. Gemo, H., Eicher, C. K., Teclemariam, S., 2005. Mozambique's experience in building a national extension system. Michigan State University Press. Glewwe, P., Kremer, M., Moulin, S., 2009. Many children left behind? textbooks and test scores in kenya. American Economic Journal: Applied Economics 1 (1), 112135. Glewwe, P., Kremer, M., Moulin, S., Zitzewitz, E., 2004. Retrospective vs. prospective analyses of school inputs: the case of ip charts in kenya. Journal of development Economics 74 (1), 251268. Gollin, D., Morris, M., Byerlee, D., 2005. Technology adoption in intensive post-green revolution systems. American Journal of Agricultural Economics 87 (5), 13101316. Haggblade, S., Tembo, G., 2003. Conservation farming in zambia. International Food Policy Re- search Institute EPTD Discussion Paper #108. Hanna, R., Mullainathan, S., Schwartzstein, J., 2014. Learning through noticing: Theory and evi- dence from a eld experiment. The Quarterly Journal of Economics 129 (3), 13111353. Hazell, P., 2013. What makes african agriculture grow?. 2012 global food policy report. Washington, DC: International Food Policy Research Institute (IFPRI). Jayachandran, S., 2006. Selling labor low: Wage rresponse to productivity shocks in developing countries. Journal of Political Economy 114 (3), 538575. Jones, M., Kondylis, F., Mobarak, A. M., Stein, D., 2015. Evaluating the integrated agriculture productivity project in bangladesh. Karlan, D., Ratan, A. L., Zinman, J., 2014. Savings by and for the poor: A research review and agenda. The Review of Income and Wealth 60 (1), 3678. Kondylis, F., Mueller, V., Zhu, S. J., 2015. Measuring agricultural knowledge and adoption. Agri- cultural Economics 46 (3). Laajaj, R., Macours, K., January 2015. The reliability and validity of skills measurement in rural household surveys. Lee, D. S., 2009. Training, wages, and sample selection: Estimating sharp bounds on treatment eects. The Review of Economic Studies 76 (3), 10711102. Levitt, S., List, J., Syverson, C., 2012. Toward an understanding of learning by doing: Evidence from an automobile assembly plant. Journal of Political Economy 121 (4), 643681. Linden, L. L., 2008. Complement Or Substitute?: The Eect of Technology on Student Achievement in India. InfoDev. Liniger, H., Studer, R. M., Hauert, C., Gurtner, M., 2011. Sustainable land management in practice guidelines and best practices for sub-saharan africa. TerrAfrica, World overview of conservation approaches and technologies (WOCAT) and food and agriculture organization of the United Nations (FAO). 26 Lucas, R., 1988. On the mechanics of economic development. Journal of Monetary Economics 22 (1), 342. Mazvimavi, K., Myathi, P., Murendo, C., 2011. Conservation agriculture practices and challenges in zimbabwe. In: 5th World Congress of Conservation Agriculture Incroporating 3rd Farming Systems Design Conference, September 2011 Brisbane, Australia. Munshi, K., 2004. Social learning in a heterogeneous population: technology diusion in the indian green revolution. Journal of Development Economics 73 (1), 185213. Pamuk, H., Bulte, E., Adekunle, A. A., 2014. Do decentralized innovation systems promote agricul- tural technology adoption? experimental evidence from africa. Food Policy 44, 227236. Purcell, D., Anderson, J. R., 1997. Agricultural extension and research: Achievements and problems in national systems. World Bank Publications. Tan, J.-P., Lane, J., Lassibille, G., 1999. Student outcomes in philippine elementary schools: An evaluation of four experiments. The World Bank Economic Review 13 (3), 493508. Thierfelder, C., Matemba-Mutasa, R., Rusinamhodzi, L., 2015. Yield response of maize (zea mays l.) to conservation agriculture cropping system in southern africa. Soil & Tillage Research 146, 230242. Thompson, P., 2010. Learning by doing. Handbook of the Economics of Innovation 1, 429476. TIA, 2008. National agricultural household survey. Michigan State University and The Ministry of Agriculture of Mozambique (MINAG). URL http://fsg.afre.msu.edu/Mozambique/survey/index.htm. Waddington, H., White, H., Anderson, J., 2014. Farmer eld schools: from agricultural extension to adult education. Systematic Review Summary 1. Figures and Tables 27 Figure 1: Study Area and spatial distribution of sampled households 28 Figure 2: Timeline of Training and Survey Household and CF Household and CF Midline Survey Endline Survey Feb.2012-Apr.2012 Feb.2013-Apr.2013 CF (preharvest) (preharvest) Baseline Survey May.2012-Aug.2012 May.2013-Aug.2013 Jun.2010 (postharvest) (postharvest) 2010 2011 2012 2013 2014 Demonstration season Adoption season Adoption season Rainy season Rainy season Rainy season Oct.2010-Apr.2011 Oct.2011-Apr.2012 Oct.2012-Apr.2013 EAs and CFs EAs and CFs First training Second training Oct.2010 Nov.2012 Notes: EA = extension agent. CF = contact farmer. 29 Figure 3: Precipitation anomalies over time Notes: Standardized cumulative rainfall in the 200 study communities over the rainy season period (October through February). The rainy season starting in October 2012 is labeled as 2012. Standardized values are computed as a ratio of the distance between cumulative rainfall and average historical cumulative rainfall to the historical standard deviation across all 200 communities. 95% condence intervals are presented around each yearly value. Our study period includes 2010 (demonstration season), and 2011 and 2012 (adoption seasons). Midline and endline surveys correspond to 2011 and 2012, respectively. 30 Figure 4: Eect of SLM training intervention on contact farmers, controlling for familywise error rate 31 Variable Treated Control Dierence Mean SD N Mean SD N in mean Baseline survey Age 38.858 9.348 148 40.160 10.559 50 -1.302 Formally trained 0.350 0.479 140 0.447 0.503 47 -0.097 Years since formal training 2.157 2.239 51 3.409 3.202 22 -1.252* Years of experience as CF 2.243 2.401 144 2.653 2.570 49 -0.410 Number of farmers assisted in last 7 days 18.034 16.095 147 19.100 14.333 50 -1.066 Number of male farmers assisted in last 7 days 10.871 9.659 147 10.860 9.064 50 0.011 Number of farmers assisted in last 30 days 37.060 28.320 133 38.370 26.441 46 -1.309 Number of male farmers assisted in last 30 days 22.480 15.145 148 22.240 17.203 50 0.240 Hours worked as CF in last 7 days 14.813 12.726 144 12.340 11.573 50 2.473 Hours normally worked as CF per week 16.322 12.498 143 12.960 12.034 50 3.362 Total hectares of cultivated land 1.289 0.655 144 1.242 0.624 50 0.047 Number of households in the community 284.421 267.037 126 244.548 265.410 42 39.873 Number of plots in the community 459.269 430.130 108 436.063 426.578 32 23.206 32 Midline survey: Recalled Number of SLM techniques learned before 2010 2.839 2.362 137 3.286 2.255 42 -0.446 Number of SLM techniques adopted before 2010 1.409 1.210 137 1.167 0.935 42 0.242 Sources: Contact farmer baseline survey, 2010; Household survey, 2012. Notes: ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. CF abbreviates contact farmer. Table 1: Contact farmers' characteristics by treatment status Table 2: Other farmers' characteristics by treatment status Variable Treated Control Dierence Mean SD Mean SD in mean Midline survey: 2012 Is the head of household 0.585 0.493 0.588 0.493 -0.003 Male 0.420 0.493 0.414 0.493 0.005 Age 37.764 19.980 37.843 20.093 -0.079 Years of schooling completed 2.057 4.866 1.844 4.905 0.213 Single 0.063 0.504 0.058 0.509 0.005 Married 0.844 0.546 0.855 0.550 -0.011 Divorced, separated, or widowed 0.091 0.366 0.085 0.368 0.006 Number of children (ages < 15 years) 2.756 3.406 2.843 3.432 -0.087 Total hectares of owned land 2.004 3.995 1.880 4.033 0.124 Number of rooms in the house 1.427 2.116 1.444 2.138 -0.017 Housing walls made of brick 0.100 0.777 0.096 0.785 0.004 Housing roof made of tinplate 0.079 0.718 0.079 0.725 0.000 Midline survey: Recalled Number of SLM techniques learned before 2010 1.236 4.514 1.303 4.563 -0.066 Number of SLM techniques adopted before 2010 0.509 2.024 0.554 2.045 -0.045 Number of observations 4,385 1,499 5,884 Source: Household survey, 2012. Notes: T-test inferences are based on standard errors clustered at the community level. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. SLM=sustainable land management. 33 Table 3: Socioeconomic and farming characteristics of contact farmers and other farmers Means Dierence in mean CFs Other farmers Household Characteristics: In the current year Is the head of household 0.994 0.590 0.405*** Age 42.364 38.243 4.121*** Years of schooling completed 5.481 2.054 3.427*** Single 0.011 0.056 -0.044* Married 0.974 0.849 0.126*** Divorced, separated, or widowed 0.057 0.095 -0.038 Number of children (ages < 15 years) 3.744 2.830 0.913*** Total hectares of owned land 3.439 2.171 1.269*** Number of rooms in the house 1.763 1.423 0.340** Housing walls made of brick„ 0.168 0.099 0.068 Housing roof made of tinplate„ 0.207 0.079 0.128** Production: In the current rainy season Grew maize 0.725 0.637 0.088 Grew sorghum 0.139 0.255 -0.116 Grew cotton 0.133 0.076 0.058 Grew sesame 0.270 0.156 0.113* Grew cassava 0.058 0.157 -0.099 Grew cow pea 0.278 0.347 -0.069 Grew pigeon pea 0.191 0.200 -0.009 Farm characteristics: In the current rainy season Plot size (hectares) 1.215 1.047 0.167 Plot was at 0.727 0.616 0.111* Plot was burnt 0.060 0.244 -0.184*** Used herbicides/pesticides/fungicides 0.133 0.042 0.091*** Used natural fertilizer 0.484 0.351 0.133 Used chemical fertilizer 0.099 0.006 0.092*** Number of observations 351 10,960 11,311 Sources: Household survey, 2012, 2013. Notes: T-test inferences are based on standard errors clustered at the community level. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. CF abbreviates contact farmer. „ This variable is only available in Midline. 34 Table 4: Eect of a direct SLM training on contact farmers' adoption and knowledge Ctrl.Mean ITT N R 2 [SD] Panel A: CFs' Knowledge and Adoption, unweighted Knowledge score 0.633 0.052 347 0.102 [0.173] [0.055] Number of techniques known by name 4.131 0.706 347 0.098 [1.626] [0.546] Number of techniques adopted on own plot 1.786 0.673** 347 0.224 [1.309] [0.225] Number of techniques adopted on any plot 3.738 0.733** 347 0.241 [1.889] [0.250] Panel B: CFs' Knowledge and Adoption, weighted Knowledge score 0.646 0.059 347 0.138 [0.189] [0.056] Number of techniques known by name 0.687 0.079 347 0.071 [0.236] [0.076] Number of techniques adopted on own plot 0.314 0.113** 347 0.248 [0.220] [0.031] Number of techniques adopted on any plot 0.601 0.106** 347 0.226 [0.265] [0.034] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey 2012, 2013. Notes: Regressions include the following variables: a constant, age, a completed primary school dummy, a single dummy, number of children, total landholdings, the number of rooms in the dwelling, baseline CF's number of years since formal training, a dummy for missing the baseline CF variable, district indicators, an incentive treatment dummy, and an endline dummy. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. CF=contact farmer; ITT=intent-to-treat eect. 35 Table 5: Eect of a direct SLM training on contact farmers' adoption of individual SLM techniques Adoption on any plot Ctrl. Mean ITT N R 2 Mulching 0.929 0.026 347 0.040 [0.046] Strip-tillage 0.548 0.159* 347 0.125 [0.072] Pit planting 0.560 0.159*** 347 0.093 [0.023] Contour farming 0.226 0.147* 347 0.241 [0.067] Crop rotation 0.726 0.066 347 0.166 [0.051] Row planting 0.440 0.096 347 0.081 [0.058] Improved fallowing 0.310 0.080 347 0.169 [0.070] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey 2012,2013. Notes: Regressions include the same explanatory variables as models in Table 4. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat; SLM=sustainable land management. 36 Table 6: Eect of a direct SLM training on contact farmers' knowledge of individual SLM techniques Knowledge score Ctrl. Mean ITT N R 2 Mulching 0.893 0.043* 347 0.135 [0.017] Strip-tillage 0.512 0.089 347 0.095 [0.057] Pit planting 0.798 0.050 347 0.075 [0.088] Contour farming 0.520 0.127 347 0.098 [0.116] Crop rotation 0.567 0.008 347 0.072 [0.046] Row planting 0.310 -0.026 347 0.149 [0.113] Improved fallowing 0.690 0.007 347 0.046 [0.040] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey 2012, 2013. Notes: Regressions include the same explanatory variables as models in Table 4. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat; SLM=sustainable land management. 37 (1) (2) (3) (4) (5) (6) (7) Control Mean T Dry Year TÖ N R 2 T+TÖ [SD] Dry Year Dry Year Panel A: Without controlling for precipitation . Revenue per Ha Original data 1441.540 326.399 347 0.089 [MZN/Ha] [1417.638] [319.157] Winsorize at 1% 1441.540 182.322 347 0.147 [1417.638] [121.077] Total revenue Original data 4408.667 1137.593 347 0.121 [MZN] [4687.363] [721.392] Winsorize at 1% 4408.667 722.089 347 0.152 [4687.363] [438.057] Panel B: Controlling for precipitation . Revenue per Ha Original data 1441.540 41.837 178.182 866.412* 347 0.097 908.249 [MZN/Ha] [1417.638] [399.865] [451.867] [354.536] [431.334] Winsorize at 1% 1441.540 9.163 133.582 528.280 347 0.157 537.443* [1417.638] [152.561] [440.298] [257.942] [237.038] 38 Total revenue Original data 4408.667 744.990 19.641 1185.839** 347 0.124 1930.829 [MZN] [4687.363] [632.045] [1014.573] [359.641] [977.493] Winsorize at 1% 4408.667 264.976 -213.874 1370.718** 347 0.156 1635.694* [4687.363] [310.816] [996.793] [453.943] [664.713] Sources: Contact farmer survey, 2010; Household survey 2012, 2013; NASA 1/2x 1/21981/2013 Precipitation data. Notes: Dry Year is a dummy indicating cumulative precipitation in the rainy season is in the rst quartile of the 1981-2013 historical average. All models include the same explanatory variables as models in Table 4. T+TÖDry Year (col 7) presents the total eect of the treatment T and its interaction with Dry Year on maize yield and revenue. The associated standard errors are in brackets. Signicance on the additive eect is determined by a Wald test. Ö=multiplied by; SLM=sustainable land management; MZN=Mozambican Metacais; Ha=hectare. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. Table 7: Eect of a direct SLM training on contact farmers' maize production Table 8: Eect of direct SLM training intervention on contact farmers' input use Ctrl.Mean ITT N„ R 2 [SD] Burnt farm plot 0.167 -0.015 347 0.037 [0.060] Used natural fertilizer 0.524 0.129 343 0.166 [0.067] Used chemical fertilizer 0.071 0.065 343 0.037 [0.042] Amount of chemical fertilizer used (kg) 2.681 35.227 341 0.027 [21.992] [19.514] Amount of chemical fertilizer used (l) 3.855 1.282 341 0.033 [27.973] [5.836] Use herbicides/pesticides/fungicides 0.250 0.022 343 0.082 [0.032] Amount of herbicides/pesticides/fungicides used (kg) 0.060 0.374 341 0.014 [0.361] [0.223] Amount of herbicides/pesticides/fungicides used (l) 4.367 -0.928 341 0.048 [11.714] [1.582] Sources: Contact farmer survey, 2010; Household survey, 2012, 2013. Notes: All models include the same explanatory variables as models in Table 4. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat; SLM=sustainable land management. „The main sample has 347 observations. Sample size varies across models due to missing values in the dependent variables. 39 Table 9: Eect of a direct SLM training intervention on contact farmers' labor allocation Pooled Sample Endline Control ITT N R 2 Control ITT N R 2 mean [SD] mean [SD] Hours spent on 6.095 -2.822 346 0.032 6.429 -1.272 168 0.073 preparation of land [14.550] [3.356] [15.353] [2.998] Hours spent on seeding 8.214 -3.558** 346 0.021 10.357 -6.567* 168 0.065 [15.996] [1.069] [17.862] [2.943] Hours spent on 2.607 -1.408 346 0.062 1.738 -0.917 168 0.087 transplantation [7.973] [0.957] [6.666] [0.736] Hours spent on irrigation 0.000 -0.038 346 0.028 [0.000] [0.044] Hours spent on sacha 10.583 0.454 346 0.198 5.833 -0.690 168 0.072 [15.576] [1.098] [14.252] [1.514] Hours spent on protection 0.000 0.969 346 0.042 0.000 0.513 168 0.122 [0.000] [0.922] [0.000] [0.499] Hours spent on harvesting 11.012 -2.288 346 0.122 15.810 -2.234 168 0.114 [18.260] [1.417] [19.573] [1.164] Total weeks spent on 28.262 -2.986 346 0.029 30.381 -7.084** 168 0.069 farming in last year [18.386] [2.210] [19.196] [2.508] Sources: Contact farmer survey, 2010; Household survey, 2012, 2013. Notes: All models include the same explanatory variables as models in Table 4. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat eect; SLM=sustainable land management. 40 Table 10: Eect of a direct SLM training intervention on other farmers' access to contact farmers, adoption, and knowledge Ctrl.Mean ITT N R 2 [SD] Access to CF Has access to any contact farmer in the last half year 0.170 0.032 10,955 0.046 [0.027] Other Farmer Knowledge and Adoption, unweighted Knowledge score 0.341 -0.004 10,955 0.055 [0.200] [0.012] Number of techniques known by name 1.654 0.000 10,955 0.022 [1.538] [0.120] Number of techniques adopted 0.845 -0.034 10,955 0.060 [0.891] [0.071] Sources: Contact farmer survey, 2010; Household survey, 2012, 2013. Notes: Regressions include the following variables: a constant, age, a completed primary school dummy, a dummy for male, a single dummy, a widow dummy, number of children, total landholdings, the number of rooms in the dwelling, baseline CF's number of years since formal training, a dummy for missing the baseline CF variable, district indicators, an incentive treatment dummy, and an endline dummy. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. CF=contact farmer; ITT=intent-to-treat eect; SLM=sustainable land management. 41 Table 11: Eect of a direct SLM training intervention on other farmers' adoption of pit planting, by CFs' characteristics CF characteristics Educ > Median Age ≥ Median Land ≥ Median Same Crop T 0.029 0.035* 0.018 0.025 [0.021] [0.020] [0.019] [0.019] CF characteristics 0.011 -0.010 -0.014 0.003 [0.020] [0.016] [0.016] [0.023] TÖCF characteristics 0.000 -0.011 0.025 0.026 [0.025] [0.023] [0.022] [0.027] N 9,836 9,836 9,836 9,968 R 2 0.010 0.011 0.010 0.010 Control mean 0.068 0.068 0.068 0.069 T+TÖCF characteristics 0.029 0.023 0.044* 0.052* [0.025] [0.023] [0.024] [0.030] Sources: Contact farmer survey, 2010; Household Survey, 2012, 2013. Notes: Regressions include the same explanatory variables as models in Table 10. The cuto values for CF characteristics correspond to the median values of education (7 years), age (41 years at ML, 43 years at EL), and landholdings (2.75 ha at ML, 3.5 at EL) in the sample of CFs. T+TÖCF characteristics (bottom row) presents the total eect of the treatment T and its interaction with the CF characteristic. The associated standard errors are in brackets. Signicance on the additive eect is determined by a Wald test. ***, **, and * indicate signicance at 1, 5, and 10 percent critical level. Ö=multiplied by; CF=contact farmer; SLM=sustainable land management. ML=midline; EL=endline. 42 Yield and Labor Yield Benets Labor Benets Benets Number of Beneciaries per community CFs 1 1 1 Average Costs per Beneciary Total cost of trainings 22,262 22,262 22,262 Annual cost of training 11,131 11,131 11,131 Annual training cost per farmer 74 74 74 Annual cost of extension agent per farmer 505 505 505 Average Benets per Beneciary Annual maize revenue 19 0 19 Weekly agricultural wage rate 19 19 19 Number of weeks in labor savings 0 7 7 43 Annual labor earnings 0 131 131 Net Average Benets per Beneciary Total net benets per CF -55 57 76 Notes: CF=Contact farmer. Figures presented in terms of 2012 USD, assuming exchange rate of 38 Metacais per 1 USD. Annual cost per extension agent is based on the monthly salary of the extension agent (211 USD) and assumes one extension agent services ve communities. Annual benets in maize revenue obtained from estimates of the ITT in the winsorized specication in Table 7. Annual benets in endline labor savings taken from estimates of the ITT on the number of weeks worked in Table 9. Table 12: Cost-Benet analysis of a direct SLM training intervention Appendix A: Additional Tables Table A.1: Pre-intervention SLM training across treatment status (recalled) Variables Treated mean Control mean Dierence in mean Contact Farmers: before 2010 Learned mulching 0.620 0.762 -0.141* Learned strip-tillage 0.321 0.429 -0.107 Learned pit planting 0.504 0.524 -0.020 Learned contour farming 0.307 0.381 -0.074 Learned crop rotation 0.591 0.690 -0.099 Learned row planting 0.285 0.238 0.047 Learned improved fallowing 0.212 0.262 -0.050 Number of observations 137 42 179 Other Farmers „: before 2010 Learned mulching 0.306 0.337 -0.031 Learned strip-tillage 0.182 0.227 -0.045 Learned pit planting 0.145 0.113 0.032 Learned contour farming 0.039 0.048 -0.009 Learned crop rotation 0.360 0.360 0.000 Learned row planting 0.104 0.114 -0.010 Learned improved fallowing 0.101 0.104 -0.003 Number of observations 4,385 1,499 5,884 Source: Household survey, 2012. Notes: „T-test inferences are based on standard errors clustered at the community level. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. SLM=sustainable land management. 44 Variables Treated Control Dierence Mean SD Mean SD in mean EA age 35.415 4.646 34.925 4.962 0.489 EA years of schooling completed 7.192 0.534 7.263 0.601 -0.071 Number of years worked as EA 6.388 5.919 5.355 4.329 1.033 Number of years worked in agricultural section, before became an EA 4.451 2.893 4.412 2.994 0.038 Number of training received over the past 5 years 9.624 5.265 9.645 5.563 -0.021 Received training from the Ministry of Agriculture 0.344 0.477 0.289 0.460 0.055 Received training from Smallholders' project 0.752 0.434 0.816 0.393 -0.064 Number of weeks in training during the last 12 months 1.244 0.601 1.276 0.601 -0.032 One of the main topics covered in the trainings was conservation agriculture 0.944 0.231 0.974 0.162 -0.030 Number of villages 125 38 163 Sources: Extension agent survey, 2012. Notes: EA=extension agent. We assign each village the EAs allegedly servicing their administrative post. Since there is more than one EA per administrative post, binary (continuous) outcomes reect the maximum (mean) value of EA responses for that post. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. 45 Table A.2: Extension agents' characteristics by treatment status Table A.3: Eect of a direct SLM training intervention on contact farmers' use of demonstration plots and access to extension agents Ctrl. Mean ITT N R 2 Used demonstration plot during the last year 0.845 -0.034 347 0.043 [0.066] EA visited CF at least once/month 0.512 0.001 347 0.033 [0.073] EA visited CF at least once/half year 0.631 0.055 347 0.047 [0.123] EA visited CF at least once/year 0.667 0.143 347 0.048 [0.123] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey, 2012, 2013. Notes: Regressions include the following variables: a constant, age, a completed primary school dummy, a single dummy, number of children, total landholdings, the number of rooms in the dwelling, baseline CF's number of years since formal training, a dummy for missing the baseline CF variable, district indicators, an incentive treatment dummy, and an endline dummy. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. CF=contact farmer; EA=extension agent; ITT=intent-to-treat eect. 46 Table A.4: Other farmers' characteristics Variables Mean SD Is the head of household 0.590 0.492 Age 38.243 14.430 Years of schooling completed 2.054 2.798 Single 0.056 0.229 Married 0.849 0.359 Divorced, separated, or widowed 0.095 0.293 Number of children [ages < 15 years] 2.830 2.041 Total hectares of owned land 2.171 2.064 Number of rooms in the house 1.423 0.724 Number of observations 10,960 Sources: Household survey, 2012, 2013. 47 Table A.5: Pre-intervention SLM adoption by treatment status (recalled) Variables Treated mean Control mean Dierence in mean Contact farmers: before 2010 Adopted mulching 0.489 0.405 0.084 Adopted strip-tillage 0.248 0.214 0.034 Adopted pit planting 0.190 0.167 0.023 Adopted contour farming 0.007 0.000 0.007 Adopted crop rotation 0.314 0.262 0.052 Adopted row planting 0.124 0.095 0.029 Adopted improved fallowing 0.036 0.024 0.013 Number of observations 137 42 179 Other Farmers: before 2010 „ Adopted mulching 0.181 0.203 -0.022 Adopted strip-tillage 0.087 0.118 -0.031 Adopted pit planting 0.059 0.036 0.023 Adopted contour farming 0.002 0.000 0.002 Adopted crop rotation 0.121 0.132 -0.011 Adopted row planting 0.055 0.059 -0.005 Adopted improved fallowing 0.005 0.005 0.000 Number of observations 4,385 1,499 5,884 Source: Household survey, 2012. Notes: „T-test inferences are based on standard errors clustered at the community level. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. SLM=sustainable land management. 48 Table A.6: Eect of a direct SLM training intervention on contact farmers' knowledge and adoption, basic specication Ctrl.Mean ITT N R 2 [SD] CFs' Knowledge and Adoption, unweighted Knowledge score 0.633 0.046 347 0.019 [0.173] [0.049] Number of techniques known by name 4.131 0.646 347 0.018 [1.626] [0.571] Number of techniques adopted on own plot 1.786 0.594* 347 0.023 [1.309] [0.232] Number of techniques adopted on any plot 3.738 0.752** 347 0.020 [1.889] [0.244] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey, 2012, 2013. Notes: Regressions include the following variables: a constant, treatment variables, an endline dummy, and district indicators. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. CF=contact farmer; ITT=intent-to-treat eect; SLM=sustainable land management. 49 Table A.7: Eect of a direct SLM training intervention on other farmers' knowledge and adoption, basic specication Ctrl.Mean ITT N R 2 [SD] Other Farmers' Knowledge and Adoption, unweighted Knowledge score 0.341 -0.003 10,955 0.049 [0.200] [0.012] Number of techniques known by name 1.654 0.013 10,955 0.012 [1.538] [0.119] Number of techniques adopted 0.845 -0.020 10,955 0.049 [0.891] [0.071] Sources: Contact farmer survey, 2010; Household survey, 2012, 2013. Notes: Regressions include the following variables: a constant, treatment variables, a male dummy, an endline dummy, and district indicators. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat eect; SLM=sustainable land management. 50 Table A.8: Eect of a direct SLM training intervention on contact farmers' adoption and knowledge, includes administrative post xed eects Ctrl.Mean ITT N R 2 [SD] CFs' Knowledge and Adoption, unweighted Knowledge score 0.633 0.046 347 0.103 [0.173] [0.036] Number of techniques known by name 4.131 0.859** 347 0.105 [1.626] [0.367] Number of techniques adopted on own plot 1.786 0.726*** 347 0.256 [1.309] [0.216] Number of techniques adopted on any plot 3.738 0.658** 347 0.244 [1.889] [0.258] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey, 2012, 2013. Notes: Regressions include the same explanatory variables as models in Table 4, except replacing district indicators with administrative post indicators. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. CF=contact farmer; ITT=intent-to-treat; SLM=sustainable land management. 51 Table A.9: Eect of a direct SLM training intervention on other farmers' access to contact farmers, adoption, and knowledge, includes administrative post xed eects Ctrl.Mean ITT N R 2 [SD] Other Farmers' Knowledge and Adoption, unweighted Knowledge score 0.341 -0.007 10,955 0.057 [0.200] [0.012] Number of techniques known by name 1.654 -0.001 10,955 0.020 [1.538] [0.116] Number of techniques adopted 0.845 0.003 10,955 0.053 [0.891] [0.066] Sources: Contact farmer survey, 2010; Household survey, 2012, 2013. Notes: Regressions include the same explanatory variables as models in Table 10, except replacing district indicators with administrative post indicators. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. CF=contact farmer; ITT=intent-to-treat; SLM=sustainable land management. 52 Table A.10: Attrition of contact farmers and other farmers Variables Treated Control Dierence Mean SD Mean SD in mean CFs attrited from Midline 0.109 0.313 0.048 0.216 0.062 Number of Observations 137 42 179 Household attrited from Midline„ 0.090 0.372 0.087 0.374 0.003 Number of Observations 2750 935 3685 Sources: Household survey, 2012, 2013; Contact farmer survey, 2012, 2013. Notes: CF=contact farmer. „T-test inferences are based on standard errors clustered at the community level. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. 53 Table A.11: Determinants of attrition (contact farmers and other farmers) CFs Other farming HH Treatment 1 0.053 Treatment 1 0.014 [0.066] [0.013] Treatment 3 0.002 Treatment 3 -0.013 [0.055] [0.012] Age -0.006** Age -0.001** [0.003] [0.000] Completed at least 0.056 HH head completed at -0.004 primary school [0.060] least primary school [0.012] Single -0.091 HH head Single 0.023 [0.190] [0.020] HH head divorced, 0.045*** widow, or separated [0.017] Total number -0.011 Total number -0.007** of children [0.013] of children [0.003] Total landholding 0.005 Total landholding -0.005 [hectares] [0.011] [hectares] [0.003] Total number of rooms -0.037 Total number of rooms -0.003 [0.032] [0.008] Number of years -0.034* Number of years 0.004 since formal training [0.020] since formal training [0.003] Missing above -0.145* Missing above 0.001 variable [0.076] variable [0.015] Household head -0.035 Household head -0.002 was female [0.088] was female [0.013] % of household -0.436 % of household 0.171*** members was away [0.350] members was away [0.065] HH has non-own 0.023 HH has non-own -0.013 farming work [0.078] farming work [0.012] HH has outside -0.015 HH has outside 0.003 employment [0.073] employment [0.017] 2012 precipitation -0.001 2012 precipitation -0.001* shock [0.002] shock [0.000] Constant 0.352 Constant 0.043 [0.308] [0.058] N 178 N 3662 R2 0.099 R2 0.014 Sources: Contact farmer survey, 2010, 2012, 2013; Household survey, 2012, 2013. Notes: Regressions include district xed eect. CF=contact farmer; HH=household. Household attrition measured by whether a household surveyed in 2012 could not be interviewed in 2013. The CF attrition outcome reects whether the village had at least one CF interviewed in 2012 but not in 2013. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. 54 Table A.12: Eects of contact farmers' adoption of SLM practices on maize revenue Total Maize Revenue Adopted mulching on own plot 1458.547 [1032.677] Adopted strip tillage on own plot 723.044 [572.914] Adopted pit planting on own plot 436.835 [1502.123] Adopted contour farming on own plot 1078.387 [3145.182] Adopted crop rotation on own plot 957.972 [960.541] Adopted row planting on own plot -1206.388 [1472.065] Adopted improved fallowing on own plot -1871.335 [968.155] N 342 R 2 0.187 Sources: Contact farmer survey, 2010, 2012, 2013; Household survey, 2012, 2013. Notes: Regressions include the same explanatory variables as models in Table 4. Additional controls include dummies for usage of all inputs displayed in Table 8, as well as labor allocated to maize production as displayed in Table 9. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. 55 Table A.13: Eect of a direct SLM training intervention on contact farmers' adoption of individual SLM techniques, controlling for lagged rainfall Adoption (1) (2) (3) (4) (5) (6) (7) any plot Control T Dry Year TÖ N R 2 T+TÖ Mean Dry Year Dry Year Mulching 0.929 0.045 0.034 -0.058 347 0.042 -0.013 [0.059] [0.077] [0.060] [0.035] Strip-tillage 0.548 0.174 0.101 -0.050 347 0.127 0.124 [0.099] [0.129] [0.119] [0.072] Pit planting 0.560 0.161*** 0.125 -0.017 347 0.097 0.144** [0.026] [0.104] [0.059] [0.046] Contour farming 0.226 0.146 0.052 -0.002 347 0.242 0.145** [0.097] [0.086] [0.099] [0.043] Crop rotation 0.726 0.059 -0.054 0.025 347 0.166 0.084 [0.047] [0.164] [0.098] [0.100] Row planting 0.440 0.040 0.198 0.137 347 0.117 0.178** [0.089] [0.111] [0.132] [0.050] Improved fallowing 0.310 0.101 0.166*** -0.074 347 0.175 0.027 [0.065] [0.035] [0.038] [0.078] Sources: Contact farmer survey, 2010, 2012, 2013; Household survey, 2012, 2013. Notes: Dry year indicates that the precipitation amount falls in the rst quartile of the historical distribution of cumulative rainfall during the rainy season (1981-2013). Regression model includes the same controls as in Table 4. T+TÖDry Year (col 7) presents the total eect of the treatment T and its interaction with Dry Year on adoption. The associated standard errors are in brackets. Signicance on the additive eect is determined by a Wald test. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat; SLM=sustainable land management. 56 Table A.14: Eect of a direct SLM training intervention on other farmers' SLM adoption Adoption Ctrl. Mean ITT N R 2 Mulching 0.336 -0.022 10,955 0.056 [0.033] Strip-tillage 0.181 -0.035 10,955 0.031 [0.028] Pit planting 0.071 0.028 10,955 0.016 [0.018] Contour farming 0.005 -0.003 10,955 0.006 [0.003] Crop rotation 0.153 0.017 10,955 0.005 [0.018] Row planting 0.079 -0.012 10,955 0.018 [0.015] Improved fallowing 0.020 -0.007 10,955 0.003 [0.006] Sources: Contact farmer survey, 2010; Household survey, 2012, 2013. Notes: Regressions include the same explanatory variables as models in Table 10. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level. ITT=intent-to-treat; SLM=sustainable land management. 57