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 services are a keystone of information diffusion information. Treatment CFs receive a direct, centralized in agriculture. This paper exploits a large randomized con- training. Control communities are exposed to the classic trolled trial to track diffusion of a new technique in the T&V model. Information diffusion was tracked through classic Training and Visit (T&V) extension model, relative to two nodes: from agents to CFs, and from CFs to others. a more direct training model. In both control and treatment Directly training CFs leads to large gains in information communities, contact farmers (CFs) serve as points-of-con- diffusion and adoption, and CFs learn by doing. Diffu- tacts between agents and other farmers. The intervention sion to others is limited: other males adopt the technique (Treatment) aims to address two pitfalls of the T&V model: perceived as labor saving, with an effect size of 75 percent. i) infrequent extension agent visits, and ii) poor quality 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 Siyao Zhu Development Research Group The World Bank August 04, 2014 JEL Classications : O12, O13, O30, D80. Keywords : information failures; technology diusion; agriculture; Africa. ∗ Corresponding authors' emails are: fkondylis@worldbank.org; v.mueller@cgiar.org. Research discussed in this publication has been funded by the International Initiative for Impact Evaluation, Inc. (3ie) through the Global Development Network (GDN), the Mozambique oce of the United States Agency for In- ternational 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, Insti- tutions, and Markets (PIM) led by the International Food Policy Research Institute (IFPRI) and nanced by the CGIAR Fund Donors. The authors beneted from comments provided by Jenny Aker, Madhur Gautam, Markus Goldstein, and David Spielman and during presentations at the CSAE (Oxford), the Mid-Western Economic Devel- opment Conference, the Development Impact Evaluation Seminar Series at the World Bank, and the IFPRI Seminar Series. The views expressed in this article do not reect those of the World Bank, 3ie or its 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 impressive research assistance. 1 Introduction Agricultural innovation is a necessary condition to accelerate productivity and achieve food security in Africa (Hazell, 2013). Recent eorts focus on designing mechanisms to overcome constraints on farmers' adoption, such as underdeveloped input delivery systems (Shiferaw, Kebede, and You, 2008), high acquisition costs (Suri, 2009), and time inconsistency (Duo, Kremer, and Robinson, 2011). A growing literature recognizes the role of information failures in the agricultural technolog- ical diusion process, focusing on conditions for eective communication between peers (Munshi, 2004; Bandiera and Rasul, 2006; Conley and Udry, 2010; Magnan et al., 2012, McNiven and Gilligan, 2012; BenYishay and Mobarak, 2013). Despite its formative eect on diusion (Feder, Just, and Zilberman, 1985; Davis, 2008), evidence on the ecacy of extension to help farmers overcome infor- mation failures is mixed (Bindlish and Evenson, 1997; Purcell and Anderson, 1997; Gautam, 2000; Anderson and Feder, 2007; Benin et al., 2007; Davis et al., 2010; Waddington and White, 2014). Recent experiments show potential in improving learning and adoption through participatory ap- proaches in extension, e.g., eld trials, farmer eld schools, and innovation platforms (Agyei-Holmes et al., 2011; Duo, Kremer, and Robinson, 2011; Pamuk, Bulte, and Adekunle, 2013; Duo, Kenis- ton, and Suri, 2014). We focus on the eectiveness of the Training and Visit (T&V) system, applied in several developing countries, to encourage the adoption of sustainable land management (SLM) practices (Gautam, 2000; BenYishay and Mobarak, 2013). Garden variety T&V models try to expand the geographic coverage of extension by engaging extension agents with a village point-of-contact (contact farmer). The model relies on extension agents (EAs) to transmit information about new technologies to contact farmers (CFs). In central Mozambique, the government randomly assigned 50 communities to the classic T&V model and 150 communities to a modied T&V model to disseminate information on SLM. The modied T&V model oers CFs a direct, centralized training on SLM practices, with similar content and breadth to the EA training. By centralizing and standardizing the training, the intervention addresses two pitfalls in the classic T&V system: i) low transmission of information due to infrequent EA visits, and ii) receipt of poor quality information, e.g., through EA lters on information. There were no other changes in the infrastructure or composition of extension services across treatment and control groups. Comparing CFs' SLM knowledge and adoption across treatment arms provides a 2 direct test of the assumption that a T&V model is conducive to EA-to-CF knowledge transfers, as measured against a direct, central CF training. We next measure how exogenous variations in CF training quality aect other farmers' SLM 1 learning and adoption. Both models ignore constraints on CF outreach and demand-side constraints on adoption in the context of social learning. 2 Given these limitations, we anticipate farmers to be aected by the revised T&V model in two ways. First, farmers may be exposed to informa- tion on additional SLM practices with enhanced frequencies of exposure, if trained CFs alter their demonstration activities and investment portfolio. Second, the direct training may both enlighten the CF and improve the quality of information available. Under those circumstances, risk-averse farmers may be inclined to adopt given reduced uncertainty of the benets of SLM (Rosenzweig and Binswanger, 1993; Rosenzweig and Wolpin, 1993; Ghadim, Pannell, and Burton, 2005; Dercon and Christaensen, 2011). The conservation agriculture technologies examined in this study are not akin to the input and crop adoption trials most commonly depicted in the literature (Munshi, 2004; Bandeira and Rasul, 2006; Conley and Udry, 2010; McNiven and Gilligan, 2012). The time and labor allocation required to implement the techniques often do not yield benets in terms of labor savings from weeding or improvements in soil quality until two or more years post implementation (FAO, 2008). This may be less desirable to subsistence farmers facing high discount rates. Yet, some techniques are considered benecial to farmers facing tough agronomic conditions. For example, water saving technologies, such as contour farming and micro-catchments, render farmers on sloped or arid land more resilient to losses from erosion and drought. Given high levels of risk aversion with regard to their cultivation decision, the intervention may change the behavior of some farmers by reducing the uncertainty of a technology's benets in the short term. Furthermore, by introducing multiple SLM techniques 1 First, the original criticism of the T&V model of a lack of accountability and incentive structures to encourage outreach remains in both models (Gautam, 2000). Second, transaction costs associated with communicating to households in remote areas or outside of their social network may continue to limit the diusion process. 2 Common demand-side constraints may inhibit the intervention's success. First, farmers may be inclined to free ride on the learning of others and delay the adoption until protable (Foster and Rosenzweig, 1995). Second, within a community, farmers may be heterogeneous along dimensions of crop choice, soil conditions, and management style, which can aect the diusion process (Munshi, 2004; Conley and Udry, 2010). Third, the characteristics of the primary adopter (CF) are likely to aect how other farmers internalize the information. For instance, farmers may be more inclined to learn from others' with similar characteristics (Feder and Savastano, 2006; Bandiera and Rasul, 2006; BenYishay and Mobarak, 2013), while CFs may be quite dissimilar than their communal peers. Social inuence may be more relevant to a community of passive learners, requiring interventions that improve the ties and visibility of previous adopters (Hogset and Barrett, 2010). 3 (mulching, strip tillage, micro-catchments, contour farming, crop rotation, improved fallowing, and row planting), we are likely increase the likelihood of exposure to at least one relevant technique to a group of heterogeneous farmers. Demonstration activities as well as CFs' private adoption were much higher in communities where the CFs had been centrally trained 15 months after the training. This increase in adoption augmented knowledge of the techniques adopted 27 months after the initial training. Taken together, our results suggest that this wedge in CF knowledge likely corresponds to learning-by-doing in this treatment arm, as the actual benets of the techniques are exposed through increased practice. Exploiting this exogenous information shock at the village level to measure the impact of boost- ing CFs' demonstration, knowledge, and adoption activities on practices within the community provides mixed results. While CFs' activities and knowledge in the revised T&V model successfully increased others' adoption of one of the techniques adopted by the CFs, this is not the case for all demonstrated techniques and all farmers. Only men increased their micro-catchments' adoption by 3 percentage points (an eect size of 75%) 15 months after the intervention. Of the additional techniques adopted by the trained CFs (strip tillage, micro-catchments, and contour farming) in response to the intervention, the adoption of micro-catchments is more likely to achieve positive net short-term benets as it does not require major investments in tools and labor in its implementa- tion (Liniger et al., 2011). In our study, other farmers exposed to the intervention only perceived micro-catchments as a labor-saving technique and CFs spent less time on agricultural tasks. The indirect evidence suggests that farmers are likely to act on the information they received, when the technology requires little up-front investment and short-term cost savings are expected. In what follows, we summarize the limitations of the agricultural extension network in Mozam- bique and improvements provided by the intervention (Section 2). We then describe the evaluation design and empirical strategies used to identify the impact of the modalities used to deliver infor- mation to the contact and other farmers (Section 3). Section 4 presents estimates of the impact of the intervention on the contact and other farmers' knowledge and adoption of SLM practices. Section 5 discusses the implications of this study for future adoption studies. 4 2 Agricultural Extension Constraints in Mozambique and Intervention 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) promoted the development of extension networks (Eicher, 2002). This expansion is set to continue moving forward (Gêmo, Eicher, and Teclemariam, 2005 ). Extension agents (EAs) are employed by the District Services for Economic Activities (Serviços Distritais de Actividades Económicas) and oper- ated at the sub-district level to disseminate information and new techniques. The system assumes that information ows are a linear process: agricultural innovations are created by researchers, then distributed by extension workers, and lastly adopted by producers (Pamuk, Bulte, and Adekunle, 2013). Country-wide, 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 villages every year based on their achievements and potentials (Coughlin, 2006). Only 15 percent of farmers report receiving extension services (Cunguara and Moder, 2011). To address these supply-side bottlenecks, the World Bank promoted the T&V model of extension. In practice, a communal representative, the CF, is designated to receive information on improved techniques from the EA and disseminate it to his community through village-level demonstration activities. Under this model, increased frequencies in training and visits would be made by the EAs to a select group of CFs (Feder and Anderson, 2004). The present National Plan for Agricultural Extension (PRONEA 2007-2014) and Extension Master Plan (2007-2016) aim to develop the decentralization of services at the district level and expand the T&V model, increase participation of targeted groups (women and marginal farmers), and enhance partnerships with other actors, such as the private sector and NGOs (Callina and Childiamassamba, 2010). Despite the importance placed on extension services and, particularly, the T&V model by the national government, there are no rigorous studies that validate this policy action. The literature that evaluates the T&V system is conducted using non-experimental methods (Gautam, 2000). Recent work attempts to correct for the non-random assignment of extension services (Cunguara and Moder, 2011) and nds a positive impact of extension on farm income in 5 Mozambique. In what follows, we describe the details of the extension network and the T&V model present at baseline. 2.1 Extension Network in Mozambique's Zambezi Valley Our experiment is set in ve districts of central Mozambique, Mutarara (Tete province), Maríngue and Chemba (Sofala province), Mopeia and Morrumbala (Zambézia province). This area receives support from a large, World Bank-Government of Mozambique (GoM) investment that aims to support the development of the extension network. The project provides three levels of agricultural technical assistance: each district has a facilitator, an environmental specialist, and eight EAs. A district is sub-divided into four administrative posts ( posto administrativo ) that include about 8-10 communities ( aldeia ). Each community has a designated CF who receives direct assistance from the two EAs placed in his administrative post, 3 who in turn receive direct assistance from the district- level technical sta. CFs are expected to provide advice to their peers within the community through demonstration activities, as well as being responsive to farmers' demands for technical assistance. We examine whether the T&V model is as eective in getting CFs to demonstrate and learn new technologies as a direct training program. The classic T&V model was adapted from the historical comprehensive model, since monthly EA trainings were dicult to sustain nancially (Gautam, 2000). Instead, EAs periodically received an extensive training (2010 and 2012). Moreover, in the classic T&V model, the majority of CFs receive visits from EAs monthly rather than twice a month (as planned under the historical, comprehensive T&V model) (Anderson and Feder, 2004). The conditions underlying most extension networks raise concerns about the ecacy of the classic T&V model. If EAs are challenged to reach communities in the rst place, designating CFs in these communities may not adequately address the supply issue of extension services. Moreover, information may get diluted from the central level to the CFs. EAs may not suciently train the CFs to ensure know-how is transmitted. There is also no guarantee that EAs transmit a clear to-do list to CFs to conduct dissemination activities in the communities. By comparing the T&V model to 3 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 were on average in their position 3 years with a standard deviation of 3. This indicates the majority of CFs were already commissioned by the system prior to the intervention. 6 a central training of the CFs, keeping the extension system and infrastructure constant, we provide a direct test of the rst modality of the T&V model, which assumes EAs will successfully train all CFs (with no information lters). 2.2 External Validity As mentioned above, our study area is limited to ve districts of Mozambique's Zambezi valley. While this is a large-scale program, it is not immediately clear that our results would hold in other contexts. Our study likely provides an upper-bound estimate of the T&V model relative to the impact of central training. Even though the ratio of EAs per administrative post is on par with the national average of 1.89 (Gêmo and Chilonda, 2013), 4 our study districts are receiving enhanced support from the local Services for Economic Activities oce. Hence, it is unclear that EAs in our sample face a smaller set of demands on their time, leading them to be more available to train CFs. A competing assumption is that as EAs receive more attention from the district services in our study area, they may be exposed to more trainings and, therefore, too busy to provide assistance to their communities. We provisionally rule this out, as we did not hear of additional trainings to EAs over our study period. 3 Experimental Design, Data and Identication At baseline, CFs and EAs were operating in all communities of our ve study districts. From these districts, we randomly selected 200 communities (with 200 CFs) in 16 administrative posts, to which 30 EAs are assigned. All EAs received SLM training. We randomly assigned CFs in 150 (Treatment) communities to a similar, centrally-administered training on SLM, which we describe in more detail in the next sub-section. CFs in the remaining 50 (Control) communities were supposed to receive SLM training from their EAs, which corresponds to the  status quo  T&V 4 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. 7 modality. 5 The randomization was stratied at the district level. To test for eective knowledge diusion in the T&V model and isolate the additional eect of a central training of CFs, we hold constant all other typical T&V interventions across treatment and control communities. Specically, and in line with the status quo modality, all CFs in the experimental sample are supposed to receive assistance from their EAs as well as a toolkit to create a demonstration plot within the community ( aldeia ). These demonstration plots are then used by (1) EAs to teach and assist each CF in implementing at least one of the agricultural practices of the CF's choice, and (2) the CF to demonstrate the elected new techniques to farmers in their community. It is important to note that all CFs in treatment and control communities are encouraged to maintain demonstration plots within the project areas. 6 Usage is quite high and not statistically dierent across treatment and control communities: 82 (83) percent of the CFs in treated (control) communities maintained a demonstration plot in 2012. By 2013, the use of demonstration plots increased to 90 (93) percent. Our experiment allows us to compare information diusion across two modalities: direct, centrally administered training, vs. second-hand, EA-led training. The extent of information diusion across the two modalities can then be measured through observed variations in the technique-mix learned and adopted across treatment status, not at the extensive margin of CFs' demonstration activities. Our design implies that each EA will work with both Treatment and Control CFs in his administrative post. 7 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's time away from the other treatment status. For instance, Treatment CFs may be more engaged with 5 The full design consists of multiple treatment arms. A second treatment arm was overlaid to 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 1 CF trained on SLM in the community on farmer outcomes. A third treatment arm was overlaid to the rst two that attempted to provide dierent performance-based incentives for the CFs to reach farmers. Although we do not measure this eect explicitly, the third treatment arm is controlled for in the regression analysis. 6 There is no instruction, however, as to what type of plot should be used for demonstration. CFs can elect to use their own, private plot or use a communal land. Hence, we present the adoption results for any plot (own or demonstration). 7 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). We further include EA-team indicators in our regression analysis (noting that each CF is assisted by two EAs (not one EA) within the administrative post). 8 the techniques they have learned and request more follow-up visits from their EAs. Reassuringly, we nd that Control and Treatment CFs received equal amounts of attention from their EAs in the year after the training, both at the extensive and intensive margins. 8 We account for unobserved heterogeneity across EAs using EA xed eects. 3.1 SLM Trainings Our training intervention encompasses seven 9 SLM techniques: Mulching, Crop Rotation, Strip Tillage, Micro-catchments, Contour Farming, Row Planting, and Improved Fallowing. Mulching covers the soil with organic residues to maintain soil humidity, suppress weeds, reduce erosion, and enriches the quality of the soil cover. Crop rotation rotates crops on a given plot to improve soil fertility 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. Micro-catchments (approximately 15-cm deep permanent holes) are constructed 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 can improve productivity by improving access to sunlight and facilitates weeding and other cultivation practices (e.g., mulching and intercropping) by providing space between rows. Improved fallowing reduces the productivity losses from fallowing land by targeted planting of species that enrich soil in a shorter time frame than traditional fallowing. Table 1 summarizes the functionality of each of these techniques (Liniger et al., 2011). 10 In 8 While access to EAs was also surveyed at endline, the statistics are contaminated by the fact that EAs visited Treatment CFs to advertise the second SLM training. Hence, reported access to EAs mechanically goes up in the treatment at endline. Thus, Table A.2 focuses on midline observations only to prove balance across extension agent characteristics. It is also important to note the fact that there were no EA dierences across groups at midline has implications on the interpretation of the intervention's impacts. One interpretation is that EAs were diligent about visiting all CFs irrespective of the intervention which would imply that the main advantage of the treatment is improving the quality of the information delivered. An alternative interpretation is that EAs might be irresponsive to the increase in the demand for services that a centralized training might have produced. This would motivate the centralized training at increasing frequency to address the low transmission of information due to infrequent EA visits. 9 Intercropping was included in the curriculum, which allows for the cultivation of several crops at once. We exclude this technique 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 (Tables A.3-A.4). 10 As row planting is often used to reinforce some of the above practices (e.g., mulching and strip tillage), the independent number of advantages of row planting are undocumented. 9 Sub-Saharan Africa, mulching and crop rotation oer the greatest number of advantages in terms of water eciency, soil fertility, and improving plant material. They were also the most common techniques applied by farmers in the Zambezi Valley at baseline, though adoption rates were far from universal (Tables A.5-A.6). Use of strip tillage, micro-catchments, and contour farming is also deemed eective at improving water eciency and soil fertility. For each SLM technique, we asked the CFs in our study whether they believed the practice aected productivity, the land preparation eort, the planting seed eort, and harvesting eort in our midline survey. The greatest percentages of CFs (in the control group) report mulching and crop rotation techniques increase productivity (60% and 36%) and reduce land preparation eorts (45% and 26%). As mentioned, these techniques were also the most popular at baseline. The percentage of CFs (in the control group) that perceive strip tillage and micro-catchments to be productive only slightly lags behind the percentage reported for crop rotation (29% and 24%, respectively). This suggests that some (but not all) SLM technologies pose as reasonable instruments to test knowledge diusion under the T&V model in the Zambezi valley. We worked with technical sta from the Ministry of Agriculture (MINAG) to develop an edu- cational agenda for the EAs and CFs on these SLM practices. The EAs were given two three-day training courses in SLM techniques in October 2010 and November 2012 (prior to the main planting season), delivered by their district technical sta with support from the central MINAG project team. Half of the training sessions were devoted to in-class lectures, and the other half consisted of hands-on plot demonstrations. The syllabus included a thorough review of the advantages of each technique over less-environmentally desirable ones. The weeks that followed those two train- ings, Treatment CFs were invited to attend the same course, delivered by the same district-level technical sta 11 with support from MINAG sta12 . After the rst training, all CFs (Control and Treatment) received a new toolkit 13 (bicycle, tools to plow the land, and smaller articles) and the mandate to disseminate the techniques most 11 In some districts, district sta relied on their EAs to help during the hands-on sessions. This could contaminate our results by lowering the amount of on-farm attention Treatment CFs subsequently received from their EAs. If anything, this implies that we will underestimate information ow in the centrally-ran training arm, and overestimate it in the T&V model. 12 Given the low literacy of farmers, a lm covering all techniques substituted the initial lecture format in the second training of the CFs. 13 The toolkit distribution was planned, regardless of our intervention, by the project sta, as the previous distri- bution had been done in 2007 and the items were deemed too old to function in 2010. 10 pertinent to their local area on their demonstration plots. 14 A second toolkit with similar items (including a bicycle) was provided to all CFs again in July 2012 before the second training. The only dierence between our Treatmentand Control CFs is the modality through which they received training on the selected seven SLM techniques. 15 EAs were told to transfer their knowledge to Control CFs and assist both Treatment and Control CFs in setting up demonstration activities. Inviting Treatment CFs to the district-level trainings was left to each EA team, at the admin- istrative post level. EAs were given the list of randomly chosen Treatment CFs, and the district sta explained the physical impossibility of training all CFs at once and that a lottery had been used to select the participating CFs. An attendance sheet was taken at training by the district sta. In 2010, only four Treatment CFs did not attend the training, and all are in the Mopeia district. 16 Since district stas may have an incentive to misreport attendance, we performed independent au- dits. 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. 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 back-checks indicate that attendance was genuine. Similar checks were performed on the 2012 training. While the attendance list fully lines up with our back-checks, participation was not universal, and contamination was quite substantial. Sixty-three (sixteen) percent of the treated (control) communities had at least one CF attend the training. 17 As these gures signal signicant exposure of Control CFs to the treatment in 2012, they foreshadow our weakened ability to statistically dierentiate the two training models in the 2013 (second follow-up) survey round. 14 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. 15 It is possible that in attending the centralized training, CFs might feel special. Our treatment eect may incorporate how this feeling may empower CFs to implement their lessons in practice. Although this was beyond the scope of the evaluation, we did ask CFs to report their state of happiness. We found 91 percent of the CFs (in the control group) and 85 percent of the CFs (in the treatment group) were happy and there is not statistically signicant dierence in those proportions (Table A.1). 16 These CFs were trained by the EA on an individual basis, and the follow-up training was veried. 17 The contamination likely was caused by a combination of EA and self-selection. CFs in the control group could have easily learned about the trainings from peers located elsewhere. Clearly, the EAs were involved in organizing the training. A politically connected CF might have been admitted by the EA or district ocer as the project fostering the evaluation was ending. 11 3.2 Data We conducted two follow-up surveys, a 2012 (midline) round, and a 2013 (endline), which form a panel of households and CFs in the study area. 18 A baseline census survey was administered to all CFs in August 2010, before the district-level randomization. Figure 1 illustrates the timing of the surveys and CF trainings over the course of four years. Midline and endline surveys collected household demographics, individual and plot-level SLM adoption, and household production information for approximately 4,000 non-CF households in 200 communities ( aldeias, that mostly overlap with Mozambique's enumeration areas) (Figure 2). A listing of households residing in each community was performed, from which we drew a random sample of 18 non CF-households per community. Our eld work included ve survey instruments: a household questionnaire; a household agricultural production questionnaire; a CF questionnaire; an extension agent questionnaire; and a community questionnaire. The household survey was also administered to CF households, in addition to the specic CF survey. The present analysis exploits the information from the household and CF surveys. Both midline and endline surveys were conducted during the primary planting season. In each survey round, households were visited twice: pre- and post-harvest. This is because SLM practices are most visible just after planting (pre-harvest, from February to April), while production data can only be obtained after harvest (May-June). Hence, all household surveys were administered during FebruaryApril, with the exception of the agricultural production module. The agricultural pro- duction (and CF, community, and extension agent) surveys were administered post-harvest during May and June in 2012 and June through August in 2013. 3.3 Balance We use data from the baseline CF survey and retrospective information collected in the 2012 house- hold survey to check for balance across treatments. Table 2 indicates minor dierences between CFs in the treatment and control communities. Treatment CFs spent almost four more hours a 18 Following McKenzie (2012), we optimize our probability of detecting a signicant impact under a budget con- straint by conducting two follow-up data collections rather than a baseline and a follow-up. 12 week working as a CF (pre-intervention) with slightly more recent training when we condition on being formally trained. Control CFs were exposed to a greater number of techniques prior to the intervention. 19 In spite of these dierences, (recalled) pre-intervention adoption rates among CFs in control and treated communities are similar. 20 Farmers' (recalled) baseline SLM learning and adoption rates are also similar across treatments (Table 3). 21 Taken together, these results suggest that we will provide a conservative measure of the relative impact of directly training CFs. Our estimates might understate the impact of direct training, and overestimate the impact of T&V model. In addition, the fact that CFs are more knowledgeable in SLM than the average farmer at baseline further suggests that the impact estimates of the training program are likely not generalizable to the average farmer. 3.4 Measuring Information Diusion and Behavioral Change Central to identifying variations in information diusion is measuring changes in agricultural prac- tices. Our study rests on the reliability of our markers of individual SLM knowledge and adoption outcomes. We focus on three outcomes: a knowledge score (see Kondylis, Mueller, and Zhu (2014) for details of the exam), the number of techniques the respondent identied by name, and the num- ber of techniques the respondent claims to adopt on any plot. 22 Objective adoption measures were also collected for two plots per household and largely corroborate the self-reported outcomes (see Kondylis, Mueller, and Zhu (2014) for a detailed comparison). 23 19 Given 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 CF and other farmers' knowledge and adoption of individual SLM techniques at baseline are reported in Tables A.5-A.6. Because these values are based on recall data, the tests should be interpreted with caution. 21 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. 22 For CFs, the majority of the analysis rests on their adoption of techniques on any plot which includes their own and demonstration plots. There are slight dierences between adoption measures which include and exclude the demonstration plot for a minority of CFs who's demonstration plots rest on communal land (29%). We verify the results are not driven by communal propriety of the demonstration plot and report those robustness checks in some but not all of our CF tables of results. 23 Our decision to focus on the knowledge score and self-reported adoption outcomes is motivated by the conclusions of Kondylis, Mueller, and Zhu (2014). Using the Smallholders' midline survey data, we nd that learning outcomes based on knowledge exams provide more precision when compared to know-by-name questions, 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. In our triangulation of the self-reported vs. observed adoption, we nd that false reporting is negligible. Since objective measures of adoption are only collected for a subset of plots in the sample (one per respondent), we 13 Since the CFs were encouraged to choose the most relevant techniques to their local conditions, we focus on aggregate measures of knowledge and adoption for our main results. However, restricting the analysis to aggregate measures of knowledge and adoption may lead us to overlook patterns of substitution across techniques attributable to the intervention. For example, we may underestimate the impact of the intervention if CFs substitute away from already-disseminated technologies to the benet of some newer techniques within the proposed package. We therefore also present how knowledge and adoption of specic techniques indicators are aected by the intervention. Technique- specic knowledge is captured by whether the respondent answers correctly at least 1 of 3 knowledge questions pertaining to the practice. Knowledge, adoption and perception of the SLM techniques were collected at the individual level from the household questionnaire. Two respondents were interviewed: the household head and his/her partner or spouse. If a polygamous household was encountered, the main spouse was interviewed. 24 Thus, 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. Specically, 179 and 172 villages were interviewed for the contact farmer survey in 2012 and 2013, respectively; 2,536 male and 3,716 female non-CFs were surveyed in 2012, and 3,115 female and 2,175 male non-CFs in 2013. 25 Selective sample attrition is of denite concern, and we detail below how selective attrition is addressed in our specications. 3.5 Empirical Strategy We measure information diusion through a direct training model relative to the traditional, T&V extension network: from EA to CF, and CF to others. A particularly attractive feature of our design is that we track information diusion through an existing network. We rst measure to what extent instead focus on a more inclusive measure of adoption provided by self-reports of men and women surveyed in the sample. 24 A polygamous household is characterized by the household head having more than one spouse or partner. Only 2.7 percent of the households in our sample are polygamous. 25 For analysis, we restrict the sample to farmers with complete information on household and individual charac- teristics. We have 179 and 168 CFs in 2012 and 2013, respectively; 2,475 male and 3,592 female non-CFs in 2012, and 3,098 female and 2,141 male non-CFs in 2013. 14 directly training CFs aects CFs' and others' knowledge and adoption. 26 We causally estimate the intent-to-treat eects (ITT) of a community being assigned to a central CF training (relative to a status quo T&V information diusion modality) on the SLM knowledge and adoption of CFs 27 and others in the community, Y, using a simple reduced-form specication: Yi,h,j = β0 + β1 Tj + β2 Xi,h,j + i,h,j (1) T takes the value 1 for each community j i with a centrally trained CF. Individual , household h, and administrative post indicator variables are included in the vector X to improve the precision of the estimated coecients. 28 Since CFs were exposed to a team of two EAs within each administrative post, we identify the ITT from within-EA-team variations, by including administrative post xed eects in all regressions . 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. Gender-dierentiated eects are presented throughout to allow for dierent functional form, as women cultivate their own plots separate from their husbands' and may face varying constraints on their time, input use, crop choice, and plot characteristics (Table A.7). We separate specications by round for a few reasons. First, nine percent of households attrited between 2012 and 2013. We nd evidence of selective attrition across household survey rounds, as individuals present in both midline and endline rounds appear statistically dierent than individuals only present during the midline and endline surveys (Table A.8). 29 Second, in spite of attrition being uncorrelated with the treatment (Table A.10), evolution in the realities of the program on the ground 26 CFs have the exibility to decide which SLM techniques to adopt on the demonstration plot. As the marginal value of adopting a technique will vary with the predominant crops grown, soil quality, topography, and other local conditions, demonstrated technique-mix is unlikely to be uniform across communities. 27 CF-level regressions are run using community-level CF outcomes and characteristics. In those communities where we (randomly) assigned a second woman CF, we measure increased village-level exposure by regressing the maximum (mean) value of binary (continuous) outcomes of CFs within the village on the maximum (mean) value of binary (continuous) covariates. 28 We address omitted variable bias by including variables that reect CF (or other farmers') demographic char- acteristics, the number of hours worked by the CF at baseline, administrative post indicators, and indicators for treatment arms not analyzed in the present study. 29 Attrition rates at the household and CF level are not statistically dierent nor correlated across treatment groups (Tables A.9-A.10). The characteristics of CFs do not vary on average over time (Table A.11). Household attrition rates appear consistent with other studies in the same region (de Brauw, 2014). A probit regression in which reects the probability of the household attriting indicates the percent of household members that were away in 2012 increases the probability of moving out of the sample (Table A.10). Age and the number of children of the household head reduces the probability of moving out of the sample. 15 compels us to split the sample by survey year. For instance, as mentioned above, contamination was quite large in 2012, while absent in 2010. Hence, results from the 2013 survey will likely underestimate the impact of the intervention, and our inferences draw heavily on the estimates provided by the 2012 survey. We perform two robustness checks to examine the sensitivity of our results to attrition. The rst diagnostic estimates (1) using the balanced panel. We show that the inclusion of individuals only present in one round 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). Upper (lower) bounds of the treatment eect are produced non-parametrically by trimming the tail of the distribution of the outcome variable in the treatment group below quantile p (and above quantile 1-p ), where p is the dierence in the proportions of non-missing observations between the treatment and control groups divided by the total number of observations in the treatment group. The technologies we disseminate are somewhat novel in the sense that baseline adoption is low. However, awareness of the techniques is quite high (Tables A.5-A.6). While there are large potential gains in knowledge and adoption as a result of the SLM training, farmers' responses are less likely to be driven by the freshness of the material. A caveat to the low novelty content of SLM training is that, should adoption prove low both at the CF and farmer levels, we will not be able to rule out demand-side from supply-side constraints without further investigation. We rst provide information on the potential costs savings associated with others' technological adoption (both in terms of changes in perceptions and realized labor savings). We additionally assess whether CF prole and CF similarity in production habits with others' inuenced the impact of the intervention, dierentiating ITT estimates by variants in CF and farmer characteristics. 3.6 Summary Statistics To understand the socioeconomic conditions in the project area, we briey describe the character- istics of the average farmer in our sample (drawing from statistics in Table A.12). The majority of individuals are women, due to the high prevalence of female headship (approximately 30 percent) 16 in the region (TIA, 2008) . The average plot owner was 38 years old with only 2 years of schooling. Most plot owners were married with 3 children, living in a single-room house made of mud and sticks, with palm or bamboo roofs (not reported). They possess 2 hectares of land on average with a standard deviation of 1.8. CFs are more knowledgeable (Tables 2 and 3), educated and wealthier (Table 4) than the average farmer. Communicator prole has been shown to aect the diusion process in ambiguous ways (Munshi, 2004; Bandiera Rasul, 2006; Feder and Savastano, 2006; Conley and Udry, 2010; BenY- ishay and Mobarak, 2013). While CFs are positively selected in attributes, they are also well-known in their communities: 81 and 90 percent of male and female farmers in the control group declare knowing them personally. However, only 79 and 67 percent of males and females report knowing that these individuals assume a role as CF in their community. Thus, barriers to knowledge diusion may stem from a lack of transparency in the roles of CFs rather than their dissimilarities with those they intend to serve. We explore the latter possibility explicitly by estimating heterogeneous eects of the treatment. 4 Results 4.1 CF Adoption and Learning-by-doing Table 5 provides the ITT estimates of aggregate measures of knowledge and adoption. Despite the variety of techniques adopted among control CFs at midline, we detect that CFs adopt an additional technique in response to the training. These eects are not driven by dierential access to extension agents (row 1, Table 5). We further disaggregate ITT estimates by the adoption of specic techniques (Table 6). Directly trained CFs are more likely to adopt techniques that were most uncommon at baseline, and perceived productive as outlined is Section 3.1. Eect sizes range from 24 to 41 percent. Adoption signicantly trended downward in both treatment and control villages. In fact, at endline, the impact of the intervention on adoption is insignicant for all techniques. 30 30 The maintenance costs oer one explanation for the disadoption in the subsequent year of the study. For example, 17 We next examine changes in CF knowledge scores to test whether direct training corrected for any loss in EA-to-CF information diusion associated with a pure T&V approach. By additionally comparing the adoption and knowledge gains, we can observe whether increased training helped lift a genuine information constraint to adoption, or whether the gains are purely achieved through in- creased salience of the techniques. Figure 3 graphs the eect sizes of the treatment on the knowledge and adoption of each SLM technique. 31 The left panels of Figure 3 indicate that directly training CFs did little to increase CFs' knowledge scores on the techniques relative to a pure T&V regimen at midline. Hence, the gains in adoption observed under the direct training modality at midline are attributable to increased salience of information rather than an actual learning eect. This is not surprising since CFs were suciently aware of the techniques from the outset. Tracking adoption and knowledge scores across years allows us to document CFs' learning-by- doing. In contrast to the adoption decay in control and treatment communities, CF knowledge of SLM signicantly expands in treatment areas (see Table A.13 for knowledge score point estimates). Treated CFs' knowledge scores associated with the adopted techniques go up one year after we detect a signicant increase in adoption. Moreover, the order of magnitude of these gains is remarkably similar to those achieved on adoption: the largest gains are experienced for contour farming, with slightly lower gains in strip tillage and micro-catchments. Taken together, the prole of this one-year lag from demonstration to learning suggests that CFs acquired knowledge through a learning-by- doing process. We show direct training has the potential to increase adoption of innovative practices both at the intensive and extensive margins. Although formal training on its own did not appear to lift any knowledge constraint among relatively skilled CFs, it increased adoption through added salience. This intimates a weakness of the T&V model, where EAs are not as eective in getting farmers to devote time to adopting new activities as a direct training. Liniger et al. (2011) indicate for micro-catchments the long-term benets appear less positive than the short-term benets. 31 The probability of detecting statistical signicance due to the intervention increases with the number of outcomes used. Our results are not robust according to the ’idàk (p-value=0.015) and Bonferonni (p-value=0.014) adjustments (Abdi, 2007). However, micro-catchment adoption at midline and contour farming knowledge at endline are robust to the adjusted p-values in a regression model that substitutes district indicators for the administrative post indicators. 18 4.2 CF Substitution of Techniques We next try to formalize whether the training caused CFs to modify their practices towards newer techniques brought to their attention by the direct training. To gauge the potential substitution eects, we exploit the (recall) baseline adoption measures to estimate the following regression, suppressing all subscripts in (1) but those that reect time: 32 Yt = β0 + β1 T + β2 Yt−1 T + β3 Yt−1 + β4 X + (2) where t signies the midline and t-1 baseline. The results in Table 7 are suggestive that the training might have encouraged but not substituted techniques (e.g., improved fallowing). Although the signs of the parameter of the variables interacting the treatment and previous adoption (β3 ) are negative for mulching and crop rotation, their magnitudes are similar to the ITT estimate. We err on the cautious side in the interpretation of these results, as clearly there appears to be a greater, though statistically insignicant, recall bias among treatment farmers for some techniques. 4.3 Others' Knowledge and Adoption We now turn to CFs' ability to diuse knowledge to others. We exploit the random, positive shock introduced by the intervention in CF activity to measure the extent of CF-to-others knowledge transmission. Table 8 provides the mean aggregate knowledge and adoption rates of other farmers in the control communities, as well as the ITT estimates of changes in knowledge and adoption outcomes attributable to our intervention. Even though the margin for gains from receiving the information was larger than that of CFs, other farmers' aggregate SLM knowledge and adoption remained the same. This is in spite of farmers' reporting learning techniques, such as contour farming, explicitly from CFs (Table A.15). The absence of adoption is robust to balancing the panel at midline and accounting for selective attrition at endline (Table 9). These qualitative results indicate that demand-side constraints may continue to hinder farmers' adoption. 32 Note that we focus on the adoption practices on CFs' own plots here, since information on the demonstration plots were not collected for baseline through recall. 19 4.4 Farmers' Perceptions of Cost Savings Our midline survey asked farmers whether they perceived each technique to require more labor eort, equivalent labor eort, or less labor eort than the use of traditional cultivation practices. Farmers in the control group perceive all techniques to be labor intensive, with a range of less than 1 percent to 18 percent of farmers declaring the techniques decrease the amount of labor required (Table A.16). We nd that increasing exposure to SLM information through the trained CF aected farmers' perceptions of the adoption costs for micro-catchments only. The intervention signicantly increased the proportion of farmers who perceive micro-catchments to be labor-saving by 2 percentage points for men, amounting to a 100-percent increase relative to the control for male farmers. The changes in farmers' perceptions are complementary to the ITT estimates for others' adoption by technique. We observe male farmers exposed to the intervention were more likely to adopt micro- catchments by 3 percentage points (an eect size of 75% according to Table A.14). 33 Women do not act on the information they receive. Though complementary, the above inferences are not indicative of a causal relationship between perceptions and adoption. We lastly explore whether farmers were motivated by the demonstrated, short-term cost savings of the technology. In particular, we examine whether farmers' adoption rates coincide with CF labor savings in terms of the number of hours spent last week devoted to dierent agricultural tasks and the total number of weeks devoted to farming in the last year. Noting that our measure of labor eorts is inclusive of all techniques adopted by the CF (not exclusive to micro-catchments), we provide ITT estimates of the CF labor eorts for various agricultural tasks (Table A.18). The results in the Appendix indicate trained, CFs spent four fewer hours preparing land last week, and saved a total of one week spent on farming in the last year at midline. Thus, male farmers may be particularly motivated to adopt micro-catchments by the immediate gains in labor savings. Although we cannot make claims denitively using cost data, benet- cost assessments of other SLM studies in Africa suggest micro-catchments oer an additional cost- advantage. Unlike strip tillage and contour farming, additional tools are not required to create 33 The results are not robust to adjustments in the familywise error rates. However, the micro-catchment adoption of male farmers at midline is robust to the Sidak and Bonferonni p-value adjustments in a regression that substitutes district for administrative post indicators. 20 micro-catchments (Liniger et al., 2011). 34 4.5 CF and Others' Heterogeneity We lastly explore whether CFs' characteristics provoke heterogeneous responses among farmers. Working with an existing network of CFs, we could not exogenously vary their education, age, wealth, or cropping patterns. The results that follow cannot be interpreted as causal linkages but as descriptive evidence. Specically, it must be noted that, as CFs are, on average, of higher status than other farmers we do not have a counterfactual where average farmers are placed in a communicator role. We simply interact the treatment variable with CF characteristics, while controlling for both CFs and farmers' characteristics, following (1). Table 10 displays the results from regressions using farmers' adoption of micro-catchments as outcomes, since the adoption for other techniques remained unchanged by the treatment (Table A.14). Estimates of the treatment eect as well as the combined eect of the treatment and the treatment interacted with whether the CF completed his secondary education, was older than the median age, had above median landholdings, or produced the same two primary crops as the farmer are provided. We assume having similar primary crops to the CF is an exogenous variable, i.e. cropping decisions are xed before adoption decisions and cropping decisions are independent of the treatment. 35 For male farmers, exposure to increased CF activity yields larger point estimates when CFs are older and more educated. Only when we distinguish the female sample by whether they have similar crop portfolios as their CF do we observe a positive association between their adoption and treatment. This is somewhat consistent with the idea that social learning is more prevalent in homogeneous farming conditions (Munshi, 2004). Delays in women's adoption may stem from gendered dierences in production technologies and an inability to extrapolate demonstrated activities to their own plot. 34 Farmers may adopt micro-catchments in order to increase their resilience. We however do not witness any signicant changes in the ITT estimates when stratifying the sample by below and above median 30-year cumulative rainfall averages. 35 Although not shown here, whether the farmer grew the same primary two crops as the CF is not aected by the treatment at midline. Male farmers exposed to the intervention are more likely to grow the same primary two crops as the CF at endline. The intervention has no eect on the crop decisions of women, however. 21 5 Discussion Our study aims to reveal which mechanisms implementable within an existing extension network can improve the limited knowledge and adoption of novel agricultural practices in Mozambican rural communities. We examine innovation diusion through two nodes of an existing extension network: EA-to-CF and CF-to-others interactions. We nd that both modalities come short of eectively propagating innovative SLM techniques. Directly training CFs on SLM dominates a pure T&V approach to extension, as it is conducive to more demonstration, private adoption and learning-by- doing among CFs. This demonstrates that SLM techniques were in fact valued by sophisticated farmers, and that in-depth knowledge, not awareness, of the techniques constituted a barrier to adoption among CFs. Although the point estimates on the learning and adoption gains from direct training are small, the eect sizes are large. Running small-scale, low-cost trainings of designated communicators can provide a more ecient solution to enhance agricultural knowledge and practices than relying on extension workers to provide ad hoc training. Training a few seed adopters in a community may not be enough to boost adoption of a new technique. Studying the impact of an exogenous increase in CFs' activities shows that demonstration is not sucient to create learning within a community and to get others to adopt on a large scale. Male farmers choose to adopt one of three SLM techniques that the trained, CFs adopted. Looking at the multiplier eect of CFs' demonstration activities, these results imply that a one percentage point increase in CF demonstration of micro-catchments induces other male farmers to increase their adoption by 0.1 percentage points. Farmers' perceived costs of SLM techniques pose one obvious demand-side constraint to adop- tion. Adoption of micro-catchments increased 3 percentage points among male farmers exposed to the intervention with no changes in adoption for the other two techniques covered by the trained CFs. The average male farmer exposed to the intervention in our study is more likely to perceive micro-catchments as labor saving. Their CFs realized labor savings in the form of a 4-hour reduction in land preparation the week prior to the interview and a 1-week savings of working on the farm over the last year (although these are inclusive of all techniques CFs adopted). The descriptive evidence is consistent with farmers' beliefs updating in response to the intervention following their adoption of micro-catchments or after observing the CF's demonstration activities. 22 The prole of the seed adopters inuences whether women act on the information they receive. When focusing on the women who have similar cropping patterns as the CF, we observe social learn- ing is more prevalent within this demographic. Women are more likely to act on the information they receive: their micro-catchments' adoption increases by 3 percentage points at midline and 9 percentage points at endline. Providing messengers with amenable farming conditions may improve the targeting of female farmers in the provision of extension services. References [1] Abdi, H., 2007. The Bonferonni and ’idák Corrections for Multiple Comparisons, in N. Salkind, eds., Encyclopedia of Measurement and Statistics. Sage, CA. [2] Agyei-Holmes, A., Ayerakwa, H.M., Osei, R.D., and Osei-Akoto, I., 2011. Training and Farmer Productivity: An Evaluation using RCT for MCA-Ghana Programme. University of Ghana, Working Paper. [3] Anderson, J. R., and Feder, G., 2007. Agricultural Extension, in R.E. Evenson and P. Pingali, eds., Handbook of Agricultural Economics, Vol. 3, Agricultural Development: Farmers, Farm production, and Farm Markets. Elsevier, Amsterdam, pp. 23432378. [4] Bandiera, O., and Rasul, I., 2006. Social Networks and Technology Adoption in Northern Mozambique. Economic Journal 116, pp. 869-902. [5] Benin, S., Nkonya, E., Okecho, G., Pender, J., Nahdy, S., Mugarura, S., and Kato, E., 2007. Assessing the Impact of the National Agricultural Advisory Services (NAADS) in the Uganda Rural Livelihoods. IFPRI Discussion Paper 00724. Washington, DC: International Food Policy Research Institute (IFPRI). [6] BenYishay A., and Mobarak, A. M., 2013. Communicating with Farmers through Social Net- works. Economic Growth Center, Yale University, Working Paper. [7] Bindlish, V., and Evenson, R. E., 1997. The Impact of T&V Extension in Africa: The Experi- ence of Kenya and Burkina Faso. The World Bank Research Observer 12(2), pp. 183-201. [8] Conley, T., and Udry, C., 2010. Learning about a New Technology: Pineapple in Ghana. American Economic Review 100 (1), pp. 35-69. [9] Coughlin, P., 2006. Agricultural Intensication in Mozambique: Infrastructure, Policy and Institutional FrameworkWhen Do Problems Signal Opportunities?. Report commissioned by the African Food Crisis Study. Department of Sociology, Lund University. 23 [10] Cunguara, B., and Moder, K., 2011. Is Agricultural Extension Helping the Poor? Evidence from Rural Mozambique. Journal of African Economics 20(4), pp. 562-595. [11] Davis, K., Nkonya, E., Kato, E., Mekonnen, D. A., Odendo, M., Miiro, R., and Nkuba, J., 2010. Impact of Farmer Field Schools on Agricultural Productivity and Poverty in East Africa. IFPRI Discussion Paper 00992. Washington, DC: International Food Policy Research Institute (IFPRI). [12] Davis, K., 2008. Extension in Sub-Saharan Africa: Overview and Assessment of Past and Current Models, and Future Prospects. Journal of International Agricultural and Extension Education 15 (3), pp. 15-28. [13] de Brauw, A. 2014. Gender, Control, and Crop Choice in Northern Mozambique. In- ternational Food Policy Research Institute Discussion Paper 1333. Available online at: http//www.ifpri.org/publication/gender-control-and-crop-choice-northern-mozambique. [14] Dercon, S., and Christiaensen, L., 2011. Consumption Risk, Technology Adoption and Poverty Traps: Evidence from Ethiopia. Journal of Development Economics 96(2), pp.159-173. [15] Duo, E., Keniston, D., and Suri, T., 2014. Diusion of Technologies within Social Networks: Evidence from a Coee Training Program in Rwanda. Available online at http://www.poverty- action.org/project/0462. [16] Duo, E., Kremer, M., and Robinson, J., 2011. Nudging Farmers to Use Fertilizer: Theory and Experimental Evidence from Kenya. American Economic Review 101(6), pp. 2350-90. [17] Eicher, C. K., 2002. Building African Models of Agricultural Extension: A Case Study of Mozambique. Washington, DC: The World Bank. [18] FAO, 2008. Conservation Agriculture. 2008-07-08 http://www.fao.org/ag/ca/index.html. [19] Feder, G., and Anderson, J., 2004. Agricultural Extension: Good Intentions and Hard Realities. World Bank Research Observer 19(1), pp. 41-60. [20] Feder, G., and Savastano, S., 2006. The Role of Opinion Leaders in the Diusion of New Knowledge: The Case of Integrated Pest Management. World Development 34(7), pp. 1287- 1300. [21] Feder, G., Just, R., and Zilberman, D., 1985. Adoption of Agricultural Innovations in Devel- oping Countries: A Survey. Economic Development and Cultural Change 33, pp. 255-298. [22] Foster, A., and Rosenzweig, M., 1995. Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture. Journal of Political Economy 103(6), pp. 1176- 1209. 24 [23] Gautam, M., 2000. Agricultural Extension: The Kenya experience, An Impact Evaluation. Washington, DC: The World Bank. [24] Gêmo, H. R., and Chilonda, P., 2013. Why did Mozambique's public extension halt the im- plementation of the National Agrarian Extension Program (PRONEA)?. Washington, DC: International Food Policy Research Institute (IFPRI). [25] Gêmo, H. R., Eicher, C. K., and Teclemariam, S., 2005. Mozambique's Experience in Building a National Extension System. Michigan State University Press, Michigan. [26] Ghadim, A., Pannell, D., Burton, M., 2005. Risk, Uncertainty, and Learning in Adoption of Crop innovation. Agricultural Economics 33, pp. 1-9. [27] Hazell, P., 2013. What Makes African Agriculture Grow?. 2012 Global Food Policy Report. Washington, DC: International Food Policy Research Institute (IFPRI). [28] Hogset, H., and Barrett, C., 2010. Social Learning, Social Inuence, and Projection Bias: A Caution on Inferences Based on Proxy Reporting of Peer Behavior. Economic Development and Cultural Change 58(3), pp. 563-589. [29] Kondylis, F., Mueller, V., and Zhu, S., 2014. Measuring Agricultural Knowledge and Adoption. Mimeo. [30] Lee, D., 2009. Training, Wages, and Sample Selection: Estimating Sharp Bounds on Treatment Eects. Review of Economic Studies 76, pp. 1071-1102. [31] Liniger, H., Studer, R. M., Hauert, C., and Gurtner M., 2011. Sustainable Land Management in Practice: Guidelines and Best Practices for Sub-Saharan Africa. Rome: Food and Agriculture Organization of the United Nations. [32] Magnan, N., Spielman, D., Lybbert, T., and Gulati, K., 2012. Leveling with Friends: Social Networks and Indian Farmers' Demand for Agricultural Custom Hire Services. Working Paper. [33] McKenzie, D., 2012. Beyond Baseline and Follow-Up: The Case for More T in Experiments. Journal of Development Economics 9, pp. 210-221.\ [34] McNiven, S., and Gilligan, D., 2012. Networks and Constraints on the Diusion of a Biofor- tied Agricultural Technology: Evidence from a Partial Population Experiment. University of California Davis, Working Paper. [35] Munshi, K., 2004. Social Learning in a Heterogeneous Population: Technology Diusion in the Indian Green Revolution. Journal of Development Economics 73, pp. 185-213. [36] National Agricultural Household Survey 2008 (TIA08). Michigan State University and The Ministry of Agriculture of Mozambique (MINAG). Available online at: http://fsg.afre.msu.edu/Mozambique/survey/index.htm. 25 [37] Pamuk H., Bulte, E., and Adekunle, A.A., 2013. Do Decentralized Innovation Systems Pro- mote Agricultural Technology Adoption? Experimental Evidence from Africa. Food Policy 44, pp.227-236. [38] Purcell, D.L., and Anderson, J.R., 1997.Agricultural Extension and Research - Achievements and Problems in National Systems. A World Bank Operations Evaluation Study. Washington, DC: The World Bank. [39] Rosenzweig, M., and Binswanger, H., 1993. Wealth, Weather risk and the Composition and Protability of Agricultural Investments. The Economic Journal 103(416), pp. 56-78. [40] Rosenzweig, M., and Wolpin, K., 1993. Credit Market Constraints, Consumption Smoothing, and the Accumulation of Durable Production Assets in Low-income Countries: Investments in Bullocks in India. Journal of Political Economy 101(21), pp.223-244. [41] Shiferaw, B., Kebede, T., and You, L., 2008. Technology Adoption under Seed Access Con- straints and the Economic Impacts of Improved Pigeonpea Varieties in Tanzania. Agricultural Economics 39, pp. 309-329. [42] Suri, T., 2009. Selection and Comparative Advantage in Technology Adoption. Econometrica 79(1), pp. 159-209. [43] Waddington, H., and White, H., and Anderson J., 2014. Farmer Field Schools: From Agricul- tural Extension to Adult Education. Systematic Review Summary 1. 3ie Synthetic Reviews. New Delhi: International Initiative for Impact Evaluation. 26 Figures and Tables Figure 1: Timeline of Trainings and Contact Farmer and Household Surveys 27 Figure 2: Geographical Distribution of (Non-CF) Households 28 Figure 3: Eect of SLM Training Intervention on Contact Farmers 29 Table 1: Advantages of SLM Techniques Technique Water Eciency Soil Fertility Improve Plant Material Improve Micro-Climate Mulching X X X X Strip tillage X X Micro-catchments X X Contour farming X X Crop rotation X X X Improved fallowing X Row planting - - - - Source: Sustainable Land Management in Practice, 2011. 30 Variables Treated Control Di. Mean SD N Mean SD N of Mean Baseline Survey CF age 38.858 9.348 148 40.160 10.559 50 -1.302 Ever being formally trained 0.350 0.479 140 0.447 0.503 47 -0.097 Number of years since formal training 2.157 2.239 51 3.409 3.202 22 -1.252* Experience as CF in years 2.243 2.401 144 2.653 2.570 49 -0.410 # of farmers assisted in last 7 days 18.034 16.095 147 19.100 14.333 50 -1.066 # of male farmers assisted in last 7 days 10.871 9.659 147 10.860 9.064 50 0.011 # of farmers assisted in last 30 days 37.060 28.320 133 38.370 26.441 46 -1.309 # 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 working as CF per week 16.322 12.498 143 12.960 12.034 50 3.362 Total acreage of cultivated land 3.184 1.619 144 3.070 1.542 50 0.114 # of households in the community 284.421 267.037 126 244.548 265.410 42 39.873 31 # of plots in the community 459.269 430.130 108 436.063 426.578 32 23.206 Midline Survey (recall) Number of techniques learned before 2010 2.839 2.362 137 3.286 2.255 42 -0.446 Number of techniques adopted before 2010 1.409 1.210 137 1.167 0.935 42 0.242 Number of observations 137 42 179 Source: Contact Farmer Baseline Survey, 2010; Household Survey, 2012. Note: ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. Table 2: Contact Farmers' Characteristics in Treated and Control Communities Table 3: Other Farmers' Characteristics in Treated and Control Communities Variables Treated Control Dierence Mean SD Mean SD of Mean Midline Survey 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 Landholdings (hectares) 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 (recall) Number of techniques learned before 2010 1.236 4.514 1.303 4.563 -0.066 Number of 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. Note: 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. 32 Table 4: Characteristics Comparison between Contact Farmers and Other Farmers Variables Midline Endline CFs Others Dierence CFs Others Dierence Mean Mean of Mean Mean Mean of Mean Household Characteristics Is the head of household 1.000 0.586 0.414*** 0.988 0.594 0.394*** Age 41.425 37.784 3.640** 43.341 38.775 4.566*** Years of schooling completed 5.469 2.003 3.467*** 5.494 2.114 3.380*** Single 0.017 0.062 -0.045 0.006 0.048 -0.042** Married 0.978 0.847 0.131*** 0.971 0.851 0.120*** Divorced, separated, or widowed 0.045 0.089 -0.044 0.07 0.101 -0.031 Number of children (ages < 15 years) 3.779 2.778 1.001*** 3.706 2.89 0.816*** Landholdings (hectares) 3.233 1.972 1.261*** 3.654 2.401 1.253*** Number of rooms in the house 1.777 1.432 0.345** 1.748 1.412 0.336** Housing walls made of brick 0.168 0.099 0.068 Housing roof made of tinplate 0.207 0.079 0.128** Production Grew maize 0.699 0.636 0.064 0.738 0.642 0.096 Grew sorghum 0.139 0.240 -0.101 0.157 0.274 -0.117 Grew cotton 0.202 0.097 0.105** 0.064 0.052 0.012 Grew sesame 0.243 0.161 0.082 0.337 0.148 0.189*** Grew cassava 0.069 0.171 -0.102 0.041 0.143 -0.102 Grew cowpea 0.225 0.354 -0.128 0.331 0.349 -0.018 Grew pigeon pea 0.202 0.189 0.013 0.186 0.216 -0.030 Farm Characteristics Plot size (hectares) 1.151 0.951 0.201* 1.301 1.166 0.135 Plot was at 0.807 0.639 0.167** 0.640 0.594 0.046 Plot was burnt 0.063 0.240 -0.178** 0.064 0.249 -0.185** Used herbicides/pesticides/fungicides 0.156 0.062 0.094** 0.087 0.021 0.067*** Used natural fertilizer 0.358 0.269 0.090 0.622 0.438 0.184 Used chemical fertilizer 0.127 0.008 0.119*** 0.052 0.005 0.047*** Number of observations 179 5,884 6,063 172 5,076 5,248 Sources: Household Survey, 2012, 2013. Note: 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. 33 Table 5: Eect of SLM Training Intervention on Contact Farmers Midline Endline Control ITT N Adjusted Control ITT N Adjusted Mean(SD) R2 Mean(SD) R2 Access to EAs EA visited CF 0.595 -0.079 179 -0.083 at least 1/month (0.093) EA visited CF 0.738 -0.088 179 -0.129 at least 1/half year (0.098) EA visited CF 0.786 0.021 179 -0.127 at least 1/year (0.093) Performance Knowledge Score 0.625 -0.002 179 -0.096 0.641 0.093*** 168 0.022 (0.201) (0.040) (0.142) (0.027) # of techniques 4.214 0.524 179 -0.079 4.048 1.320*** 168 0.040 known by name (1.601) (0.369) (1.667) (0.387) # of techniques 1.214 0.788*** 179 0.117 2.357 0.654** 168 0.032 adopted on own plot (1.001) (0.247) (1.340) (0.308) # of techniques 4.452 0.817* 179 -0.072 3.024 0.551 168 -0.029 adopted on any plot (1.928) (0.415) (1.569) (0.360) Source: Household Survey and Contact Farmer Survey, 2012, 2013. Note: Regressions include the following variables: a constant, age, completed at least primary school dummy, single dummy, number of children, total landholdings, the number of rooms in the household, baseline CF's number of years since formal training, missing dummy, posto indicators, and incentive treatment. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 34 Table 6: Eect of SLM Training Intervention on Contact Farmers' SLM Adoption Adoption on Midline Endline Any Plot Control ITT N Adjusted Control ITT N Adjusted Mean R2 Mean R2 Mulching 0.929 0.032 179 -0.130 0.929 0.031 168 -0.048 (0.051) (0.053) Strip Tillage 0.619 0.152† 179 -0.061 0.476 0.172 168 -0.060 (0.094) (0.114) Micro-catchments 0.643 0.216** 179 -0.087 0.476 0.057 168 -0.129 (0.099) (0.114) Contour Farming 0.405 0.171^ 179 -0.109 0.048 0.070 168 -0.133 (0.105) (0.071) Crop Rotation 0.905 0.050 179 -0.073 0.548 0.049 168 -0.037 (0.062) (0.100) Row Planting 0.524 0.097 179 -0.073 0.357 0.138 168 -0.091 (0.098) (0.103) Improved Fallowing 0.429 0.097 179 -0.044 0.190 0.034 168 -0.100 (0.103) (0.084) Source: Household Survey and Contact Farmer Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 5. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. †: P-value is 0.106. ^: P-value is 0.104. 35 Table 7: Eect of SLM Training Intervention on Contact Farmers' Adoption Controlling for Previous Adoption Adoption on Midline Own Plot Control ITT Adopted ITT * N Adjusted Treatment Mean Before Adopt Bef. R2 Eect (PV) Mulching 0.405 0.254** 0.537*** -0.235 179 0.141 0.103 (0.105) (0.131) (0.143) Strip Tillage 0.286 -0.035 0.769*** 0.159 179 0.656 0.165 (0.060) (0.104) (0.114) Micro-catchments 0.119 0.162** 0.662*** -0.061 179 0.346 0.674 (0.068) (0.131) (0.146) Contour Farming 0.000 0.022 0.981*** 0.000 179 0.428 . (0.017) (0.081) . Crop Rotation 0.262 0.196** 0.812*** -0.243 179 0.398 0.115 (0.089) (0.138) (0.153) Row Planting 0.119 0.077 0.810*** -0.157 179 0.488 0.242 (0.050) (0.121) (0.134) Improved Fallowing 0.024 -0.011 0.022 0.484** 179 0.161 0.033 (0.042) (0.205) (0.225) Source: Household Survey, 2012. Note: Regressions include the same explanatory variables as models in Table 5. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 36 Table 8: Eect of SLM Training Intervention on Other Farmers Midline Endline Control ITT N Adj. Control ITT N Adj. Mean(SD) R2 Mean(SD) R2 Access to CF Has access to any Female 0.080 0.028 3423 0.012 0.214 0.028 2951 0.004 contact farmer (0.025) (0.033) in the last half year Male 0.135 0.048 2461 0.016 0.301 0.001 2120 0.001 (0.037) (0.042) Performance Knowledge score Female 0.290 0.004 3423 0.006 0.369 -0.019 2951 0.009 (0.157) (0.016) (0.236) (0.026) Male 0.316 0.006 2461 0.016 0.416 -0.019 2120 0.013 (0.161) (0.016) (0.221) (0.024) # of techniques Female 1.457 0.079 3423 0.025 1.581 0.016 2951 0.007 known by name (1.485) (0.135) (1.457) (0.195) Male 1.709 0.018 2461 0.014 2.025 -0.171 2120 0.010 (1.588) (0.144) (1.610) (0.196) # of techniques Female 0.659 -0.055 3423 0.005 0.912 0.111 2951 0.003 adopted (0.765) (0.072) (0.932) (0.109) Male 0.749 -0.034 2461 0.011 1.175 -0.009 2120 -0.001 (0.820) (0.080) (1.002) (0.118) Source: Household Survey, 2012, 2013. Note: Regressions include the following variables: a constant, age, completed at least primary school dummy, single dummy, widow dummy, number of children, total landholdings, the number of rooms in the household, baseline CF's number of years since formal training, missing dummy, posto indicators, and incentive treatment. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 37 Midline (Balanced Sample)† Endline (Lee's Bounds) Control ITT N Adjusted Treatment Bound Con. Interval # of Sel. Obs. Trimming Mean(SD) R2 Lower Upper Lower Upper # of Obs. Proportion Access to CFs Has access to any Female 0.083 0.018 2706 0.010 0.027 0.042* -0.006 0.084 2706 0.015 contact farmer (0.026) (0.019) (0.024) 3423 in the last half year Male 0.135 0.047 1873 0.017 -0.013 0.007 -0.067 0.053 1874 0.019 (0.041) (0.031) (0.026) 2461 Performance Knowledge score Female 0.297 0.002 2706 0.005 -0.005 0.008 -0.025 0.031 2706 0.015 (0.160) (0.018) (0.012) (0.014) 3423 Male 0.314 0.006 1873 0.018 -0.020 -0.008 -0.046 0.018 1874 0.019 (0.162) (0.017) (0.015) (0.015) 2461 38 # of techniques Female 1.524 0.087 2706 0.025 -0.024 0.068 -0.148 0.255 2706 0.015 known by name (1.505) (0.145) (0.073) (0.110) 3423 Male 1.783 -0.052 1873 0.020 -0.268** -0.149 -0.470 0.019 1874 0.019 (1.609) (0.157) (0.120) (0.100) 2461 # of techniques Female 0.681 -0.036 2706 0.006 -0.020 0.029 -0.096 0.129 2706 0.015 adopted (0.776) (0.077) (0.045) (0.059) 3423 Male 0.785 -0.062 1873 0.017 -0.130* -0.058 -0.255 0.046 1874 0.019 (0.830) (0.084) (0.074) (0.062) 2461 Source: Household Survey, 2012, 2013. Note: †Regressions include the same explanatory variables as models in Table 8. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. Table 9: Selective Attrition, Balanced Sample & Lee's Bounds Table 10: Heterogeneity of ITT on Other Farmers' Adoption of Micro-Catchments Adoption of Midline Micro-catchments Ctrl. ITT T+T* T+T* T+T* T+T* N Adj. Mean Ed>7years Age≥41 Land≥2.75 Same Crop R2 Female 0.039 -0.012 0.024 3239 0.002 (0.014) (0.017) Male 0.039 0.017 0.056†† 2348 0.008 (0.020) (0.025) Female 0.039 0.007 -0.004 3239 0.000 (0.017) (0.014) Male 0.039 0.036 0.038† 2348 0.007 (0.023) (0.021) Female 0.039 0.000 0.006 3239 0.000 (0.016) (0.015) Male 0.039 0.039* 0.027 2348 0.006 (0.021) (0.024) Female 0.039 0.002 0.029† 3237 -0.001 (0.013) (0.017) Male 0.039 0.041** -0.001 2342 0.009 (0.018) (0.039) Adoption of Endline Micro-catchments Ctrl. ITT T+T* T+T* T+T* T+T* N Adj. Mean Ed>7years Age≥43 Land≥3.5 Same Crop R2 Female 0.082 0.019 0.004 2474 -0.002 (0.031) (0.033) Male 0.137 -0.016 -0.044 1775 -0.001 (0.038) (0.048) Female 0.082 0.047 -0.019 2474 0.003 (0.035) (0.028) Male 0.137 0.001 -0.053 1775 0.001 (0.048) (0.035) Female 0.082 -0.006 0.038 2474 0.000 (0.028) (0.036) Male 0.137 -0.029 -0.016 1775 0.001 (0.038) (0.045) Female 0.082 0.001 0.089† 2546 0.002 (0.025) (0.048) Male 0.137 -0.038 -0.007 1843 0.001 (0.033) (0.068) Source: Household Surveys, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 8. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. Interaction columns reect the combined values of the treatment and treatment interacted with the CF characteristics coecients. †††, ††, and † indicate values are signicant based on the treatment and the treatment interacted with the CF characteristics variable at the 1, 5, and 10 percent critical levels. 39 Appendix A: Additional Tables 40 Variables Midline Endline Treated Control Di. Treated Control Di. Mean Mean of Mean Mean Mean of Mean CF feeled happy 0.854 0.905 -0.051 0.868 0.884 -0.016 Being CF: to help community 0.905 0.810 0.096* 0.984 0.977 0.008 Being CF: to learn new techniques 0.562 0.452 0.110 0.736 0.767 -0.031 Being CF: prestige of the position 0.285 0.238 0.047 0.171 0.186 -0.016 Being CF: curiosity for farming techniques 0.190 0.167 0.023 0.302 0.302 0.000 CF used the demo plot in the past 12 months 0.796 0.857 -0.062 0.837 0.837 0.000 CF owned the demo plot 0.686 0.786 -0.100 0.713 0.721 -0.008 EAs gave CF agriculture inputs 0.540 0.524 0.016 0.566 0.512 0.054 EAs gave CF natural fertilizers 0.146 0.119 0.027 0.171 0.163 0.008 EAs gave CF chemical fertilizers 0.380 0.381 -0.001 0.357 0.302 0.054 EAs gave CF farming tools 0.175 0.119 0.056 0.372 0.419 -0.047 EAs gave CF any non-technical support 0.577 0.548 0.029 0.628 0.535 0.093 41 CF received conservation agriculture trainings 0.961 0.907 0.054 Number of CA trainings CF received 2.535 2.488 0.047 Training was easy to attend 0.775 0.721 0.054 Paid for transportation to attend the training 0.178 0.093 0.085 Number of observations 137 42 179 129 43 172 Source: Contact Farmer Survey, 2012, 2013. Note: ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. Table A.1: Advantages of Being a Contact Farmer in Treated and Control Communities Table A.2: Extension Agents' Characteristics in Treated and Control Communities at Midline Variables Treated Control Di. Mean SD Mean SD of 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 # of years worked as EA 6.388 5.919 5.355 4.329 1.033 # of years worked in agricultural section, before became an EA 4.451 2.893 4.412 2.994 0.038 # of training received over the past 5 years 9.624 5.265 9.645 5.563 -0.021 Received training from MINAG (government) 0.344 0.477 0.289 0.460 0.055 Received training from Smallholder project 0.752 0.434 0.816 0.393 -0.064 # of weeks in training during the last 12 months 1.244 0.601 1.276 0.601 -0.032 One of the main topic of trainings was conservation agriculture 0.944 0.231 0.974 0.162 -0.030 Number of observations 125 38 163 Source: Extension Agent Survey, 2012; Contact Farmer Survey, 2012. Note: ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 42 Table A.3: Eect of SLM Training Intervention on Contact Farmers (Includes Intercropping) Midline Endline Control ITT N Adj. Control ITT N Adj. Mean(SD) R2 Mean(SD) R2 Knowledge Score 0.642 -0.007 179 -0.111 0.661 0.067*** 168 -0.014 (0.152) (0.033) (0.121) (0.023) # of techniques 5.214 0.487 179 -0.086 5.024 1.343*** 168 0.042 known by name (1.601) (0.370) (1.689) (0.389) # of techniques 2.048 0.799*** 179 0.080 3.310 0.653** 168 0.038 adopted on own plot (1.125) (0.269) (1.352) (0.315) # of techniques 5.429 0.824* 179 -0.078 4.000 0.532 168 -0.021 adopted on any plot (1.965) (0.421) (1.562) (0.366) Source: Household Survey and Contact Farmer Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 5. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 43 Table A.4: Eect of SLM Training Intervention on Other Farmers (Includes Intercropping) Midline Endline Control ITT N Adj. Control ITT N Adj. Mean(SD) R2 Mean(SD) R2 Knowledge score Female 0.339 0.005 3423 0.010 0.410 -0.012 2951 0.012 (0.143) (0.013) (0.178) (0.018) Male 0.358 0.008 2461 0.022 0.449 -0.012 2120 0.013 (0.148) (0.014) (0.162) (0.016) # of techniques Female 2.377 0.083 3423 0.022 2.486 0.006 2951 0.003 known by name (1.525) (0.138) (1.533) (0.186) Male 2.652 0.025 2461 0.015 2.941 -0.159 2120 0.009 (1.622) (0.141) (1.666) (0.191) # of techniques Female 1.415 -0.045 3423 0.005 1.757 0.096 2951 0.002 adopted (0.879) (0.086) (1.052) (0.107) Male 1.560 -0.045 2461 0.012 2.044 -0.006 2120 -0.001 (0.924) (0.087) (1.100) (0.114) Source: Household Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 8. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 44 Table A.5: SLM Learning before 2010 in Treated and Control Communities (Recall) Variables Treated Mean Control Mean Dierence of Mean Contact Farmers Learned mulching 0.620 0.762 -0.141* Learned strip tillage 0.321 0.429 -0.107 Learned micro-catchments 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 † Learned mulching 0.306 0.337 -0.031 Learned strip tillage 0.182 0.227 -0.045 Learned micro-catchments 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 Sources: Household Survey, 2012. Note: †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. 45 Table A.6: SLM Adoption before 2010 in Treated and Control Communities (Recall) Variables Treated Mean Control Mean Dierence of Mean Contact Farmers Adopted mulching 0.489 0.405 0.084 Adopted strip tillage 0.248 0.214 0.034 Adopted micro-catchments 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 † Adopted mulching 0.181 0.203 -0.022 Adopted strip tillage 0.087 0.118 -0.031 Adopted micro-catchments 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 Sources: Household Survey, 2012. Note: †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. 46 Table A.7: Gender Barriers to Adoption (Mean Dierences within the Control Group) Variables Midline Endline Male Female Dierence Male Female Dierence Mean Mean of Mean Mean Mean of Mean Production Grew maize 0.650 0.601 0.048 0.753 0.583 0.170*** Grew sorghum 0.122 0.305 -0.183*** 0.179 0.332 -0.153** Grew cotton 0.188 0.024 0.164*** 0.096 0.014 0.082*** Grew sesame 0.248 0.092 0.156*** 0.181 0.093 0.088** Grew cassava 0.198 0.154 0.044 0.166 0.128 0.038 Grew cowpea 0.265 0.394 -0.129 0.351 0.347 0.004 Grew pigeon pea 0.204 0.167 0.036 0.233 0.180 0.053 Farm Characteristics Plot size (hectares) 1.015 0.821 0.194*** 1.292 0.983 0.309*** Plot was at 0.606 0.633 -0.027 0.570 0.551 0.019 Plot was burnt 0.243 0.268 -0.025 0.249 0.253 -0.003 Used herbicides/pesticides/fungicides 0.124 0.018 0.106*** 0.048 0.002 0.046*** Used natural fertilizer 0.278 0.296 -0.018 0.51 0.42 0.09 Used chemical fertilizer 0.009 0.001 0.007 0.01 0.00 0.01 Number of observations 565 675 1,240 481 657 1,138 Sources: Household Survey, 2012, 2013. Note: 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. 47 Table A.8: Other Farmers' Characteristics between Attrition Groups Variables Both Midline Endline Mean Di. Mean Di. Mean Di. Rounds Only Only B-ML B-EL ML-EL Is household head 0.609 0.506 0.460 0.102 *** 0.149 *** 0.046 * Age 38.321 35.898 36.287 2.423 *** 2.034 ** -0.389 Years of schooling completed 1.934 2.242 2.559 -0.308 ** -0.624 *** -0.316 Single 0.059 0.071 0.113 -0.012 -0.054 ** -0.042 ** Married 0.843 0.861 0.792 -0.019 0.050 * 0.069 *** Divorced, widow, or separated 0.097 0.060 0.095 0.038 *** 0.003 -0.035 ** Total number of children 2.776 2.788 2.784 -0.012 -0.009 0.003 Total number of rooms 1.414 1.493 1.395 -0.079 0.019 0.098 Total landholdings 1.999 1.879 2.292 0.120 -0.294 * -0.414 *** Number of observations 4580 1304 496 Sources: Household Survey, 2012, 2013. Note: 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. 48 Table A.9: Attrition of CFs and Households Variables Treated Control Di. Mean SD Mean SD of Mean CFs attrited from Midline 0.109 0.313 0.048 0.216 0.062 # of Obs. 137 42 179 Household attrited from Midline† 0.090 0.372 0.087 0.374 0.003 # of Obs. 2750 935 3685 Source: Household Survey and Contact Farmer Survey, 2012, 2013. Note: †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. 49 Table A.10: Probability of CF and Household Attrition CF Village Household Treatment 1 0.074 Treatment 1 0.013 (0.069) (0.014) Treatment 3 0.003 Treatment 3 -0.011 (0.056) (0.012) Age -0.006** Age -0.001** (0.003) (0.000) Completed at least 0.052 HH head completed at -0.004 primary school (0.064) least primary school (0.012) Single -0.097 HH head Single 0.023 (0.199) (0.021) HH head divorced, 0.043** widow, or separated (0.017) Total number -0.010 Total number -0.006** of children (0.013) of children (0.003) Total landholding 0.004 Total landholding -0.004 (hectares) (0.012) (hectares) (0.003) Total number of rooms -0.022 Total number of rooms -0.002 (0.035) (0.008) Number of years -0.040* Number of years 0.005 since formal training (0.024) since formal training (0.004) Missing dummy -0.138 Missing dummy 0.027 (0.112) (0.020) Household head -0.033 Household head 0.000 was female (0.096) was female (0.013) % of household -0.394 % of household 0.131** members was away (0.368) members was away (0.066) HH has non-own 0.033 HH has non-own -0.013 farming work (0.080) farming work (0.012) HH has outside -0.017 HH has outside 0.003 employment (0.078) employment (0.017) 2012 precipitation 0.000 2012 precipitation -0.001 shock (0.002) shock (0.000) Constant 0.352 Constant 0.030 (0.436) (0.081) N 178 N 3656 Adj. R-sq (0.078) Adj. R-sq 0.004 Source: Household Survey and Contact Farmer Survey, 2012, 2013. Note: Regressions include posto xed eect. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 50 Table A.11: Descriptive Statistics of Contact Farmers' Characteristics Variables Pooled Men Women Dierence Mean SD Mean Mean of Mean Midline Is the head of household 0.970 0.171 1.000 0.727 0.273*** Age 41.080 10.362 41.599 36.909 4.690** Years of schooling completed 5.487 2.872 5.672 4.000 1.672*** Single 0.015 0.122 0.011 0.045 -0.034 Married 0.945 0.229 0.972 0.727 0.244*** Divorced, separated, or widowed 0.040 0.197 0.017 0.227 -0.210*** Number of children (ages < 15 years) 3.759 2.151 3.847 3.045 0.802* Landholdings 3.218 2.284 3.248 2.984 0.264 Number of rooms in the house 1.769 0.919 1.780 1.682 0.098 Number of observations 199 177 22 Endline Is the head of household 0.938 0.242 1.000 0.690 0.310*** Age 43.311 10.703 43.515 42.500 1.015 Years of schooling completed 5.431 2.824 5.754 4.143 1.612*** Single 0.005 0.069 0.006 0.000 0.006 Married 0.938 0.242 0.982 0.762 0.220*** Divorced, separated, or widowed 0.057 0.233 0.012 0.238 -0.226*** Number of children (ages < 15 years) 3.646 2.177 3.772 3.143 0.630* Landholdings 3.680 2.296 3.772 3.314 0.459 Number of rooms in the house 1.760 0.935 1.744 1.857 -0.113 Number of observations 209 167 42 Source: Household Survey, 2012, 2013. Note: ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 51 Table A.12: Descriptive Statistics of Other Farmers' Characteristics Variables Pooled Men Women Dierence Mean SD Mean Mean of Mean Midline Is the head of household 0.586 0.493 0.944 0.329 0.615*** Age 37.784 14.507 40.312 35.967 4.345*** Years of schooling completed 2.003 2.802 3.351 1.033 2.319*** Single 0.062 0.241 0.074 0.054 0.020 Married 0.847 0.360 0.904 0.806 0.098*** Divorced, separated, or widowed 0.089 0.285 0.021 0.138 -0.117*** Number of children (ages < 15 years) 2.778 2.014 2.920 2.677 0.243*** Landholdings 1.972 1.772 2.110 1.873 0.236** Number of rooms in the house 1.432 0.730 1.491 1.389 0.102* Number of observations 5,884 2,461 3,423 Endline Is the head of household 0.594 0.491 0.919 0.360 0.559*** Age 38.775 14.322 41.193 37.038 4.155*** Years of schooling completed 2.114 2.793 3.608 1.041 2.567*** Single 0.048 0.214 0.056 0.042 0.014* Married 0.851 0.356 0.916 0.804 0.112*** Divorced, separated, or widowed 0.101 0.301 0.028 0.153 -0.125*** Number of children (ages < 15 years) 2.890 2.071 3.072 2.760 0.312*** Landholdings 2.401 2.336 2.602 2.257 0.346*** Number of rooms in the house 1.412 0.718 1.461 1.377 0.083 Number of observations 5,076 2,122 2,954 Source: Household Survey, 2012, 2013. Note: 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. 52 Table A.13: Eect of SLM Training Intervention on Contact Farmers' SLM Knowledge Knowledge Midline Endline Score Control ITT N Adjusted Control ITT N Adjusted Mean(SD) R2 Mean(SD) R2 Mulching 0.833 0.038 179 -0.099 0.952 0.030 168 -0.126 (0.258) (0.047) (0.139) (0.022) Strip Tillage 0.460 0.051 179 -0.102 0.563 0.122* 168 -0.076 (0.345) (0.072) (0.270) (0.066) Micro-catchments 0.798 -0.031 179 -0.068 0.798 0.111* 168 -0.001 (0.399) (0.082) (0.332) (0.057) Contour Farming 0.524 0.030 179 -0.094 0.516 0.181** 168 -0.039 (0.369) (0.073) (0.405) (0.077) Crop Rotation 0.540 -0.063 179 -0.102 0.595 0.067 168 -0.076 (0.329) (0.070) (0.271) (0.049) Row Planting 0.476 -0.155 179 -0.039 0.143 0.118 168 -0.102 (0.505) (0.103) (0.354) (0.088) Improved Fallowing 0.738 0.008 179 -0.124 0.643 0.024 168 -0.148 (0.276) (0.060) (0.229) (0.054) Source: Household Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 5. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 53 Table A.14: Eect of SLM Training Intervention on Other Farmers' SLM Adoption Adoption Midline Endline Ctrl. ITT N Adj. Ctrl. ITT N Adj. Mean R2 Mean R2 Mulching Female 0.248 -0.051 3423 0.009 0.390 0.056 2951 0.009 (0.044) (0.049) Male 0.248 -0.022 2461 0.004 0.512 0.005 2120 0.000 (0.041) (0.051) Strip Tillage Female 0.153 -0.008 3423 0.029 0.200 -0.004 2951 0.002 (0.039) (0.038) Male 0.161 -0.007 2461 0.023 0.225 -0.003 2120 0.006 (0.040) (0.044) Micro-catchments Female 0.041 0.000 3423 -0.001 0.084 0.022 2951 0.002 (0.013) (0.024) Male 0.039 0.033* 2461 0.006 0.139 -0.020 2120 -0.001 (0.017) (0.029) Contour Female 0.000 0.002 3423 0.000 0.007 -0.006 2951 0.010 Farming (0.001) (0.005) Male 0.000 0.002 2461 -0.001 0.017 -0.014 2120 0.016 (0.002) (0.010) Crop Rotation Female 0.131 0.005 3423 0.002 0.151 0.039 2951 0.003 (0.020) (0.028) Male 0.182 -0.013 2461 0.004 0.160 0.035 2120 -0.002 (0.024) (0.033) Row Planting Female 0.073 0.012 3423 0.010 0.061 0.002 2951 0.001 (0.020) (0.021) Male 0.093 -0.008 2461 0.011 0.095 -0.023 2120 0.003 (0.023) (0.026) Improved Female 0.014 -0.014* 3423 0.019 0.019 0.002 2951 0.005 Fallowing (0.008) (0.006) Male 0.026 -0.020* 2461 0.018 0.027 0.012 2120 0.002 (0.011) (0.013) Source: Household Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 8. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 54 Table A.15: From Whom Other Farmers Claim to Learn SLM Techniques Midline Endline Ctrl. Mean ITT N Adj. R2 Ctrl. Mean ITT N Adj. R2 Mulching extension 0.041 -0.007 5884 0.017 0.034 -0.012 5071 0.009 agent (0.011) (0.011) contact 0.091 0.008 5884 0.009 0.232 -0.011 5071 0.012 farmer (0.022) (0.034) other 0.296 -0.023 5884 0.021 0.301 -0.003 5071 0.003 farmer (0.041) (0.039) Strip extension 0.007 0.004 5884 0.005 0.003 0.005 5071 0.002 Tillage agent (0.006) (0.004) contact 0.023 0.003 5884 0.003 0.056 -0.010 5071 0.001 farmer (0.009) (0.019) other 0.166 -0.014 5884 0.036 0.162 0.024 5071 0.004 farmer (0.031) (0.036) Micro- extension 0.007 0.013** 5884 0.011 0.005 0.001 5071 -0.001 catchments agent (0.005) (0.003) contact 0.041 0.020 5884 0.006 0.089 -0.011 5071 0.012 farmer (0.018) (0.022) other 0.095 0.021 5884 0.004 0.078 0.012 5071 0.005 farmer (0.018) (0.023) Contour extension 0.002 0.002 5884 0.006 0.003 -0.002 5071 0.000 Farming agent (0.002) (0.002) contact 0.005 0.010** 5884 0.002 0.015 0.001 5071 0.007 farmer (0.005) (0.007) other 0.037 0.003 5884 0.013 0.013 -0.008 5071 0.004 farmer (0.011) (0.006) Crop extension 0.021 0.004 5884 0.011 0.012 -0.004 5071 0.002 Rotation agent (0.010) (0.005) contact 0.030 0.012 5884 0.006 0.115 -0.021 5071 0.011 farmer (0.011) (0.028) other 0.302 -0.004 5884 0.006 0.253 0.035 5071 0.002 farmer (0.025) (0.035) Row extension 0.004 0.002 5884 0.001 0.003 0.000 5071 0.000 Planting agent (0.003) (0.003) contact 0.010 0.000 5884 0.000 0.030 0.001 5071 0.000 farmer (0.004) (0.014) other 0.099 0.018 5884 0.007 0.048 0.003 5071 0.002 farmer (0.025) (0.018) Improved extension 0.007 0.001 5884 0.002 0.005 -0.004 5071 0.001 Fallowing agent (0.006) (0.003) contact 0.011 -0.002 5884 0.001 0.025 -0.003 5071 0.004 farmer (0.005) (0.009) other 0.073 -0.020 5884 0.007 0.083 -0.007 5071 0.004 farmer (0.019) (0.022) Source: Household Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 8. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 55 Table A.16: Other Farmers' Perceptions of Technique's Labor Savings Compared to Traditional Method Technique Saves Midline Labor Time Control Mean ITT N Adjusted R2 Mulching Female 0.133 -0.025 3423 0.008 (0.035) Male 0.142 -0.015 2461 0.018 (0.034) Strip Tillage Female 0.156 -0.003 3423 0.018 (0.036) Male 0.177 -0.030 2461 0.006 (0.040) Micro-catchments Female 0.008 0.006 3423 0.002 (0.005) Male 0.008 0.019** 2461 0.001 (0.009) Contour Farming Female 0.005 0.009 3423 0.005 (0.006) Male 0.008 -0.006 2461 0.002 (0.005) Crop Rotation Female 0.063 0.007 3423 0.000 (0.017) Male 0.090 0.011 2461 0.005 (0.021) Row Planting Female 0.041 0.010 3423 0.003 (0.014) Male 0.055 0.009 2461 0.009 (0.018) Improved Fallowing Female 0.034 -0.010 3421 0.009 (0.014) Male 0.043 -0.011 2461 0.008 (0.016) Source: Household Survey, 2012. Note: Regressions include the same explanatory variables as models in Table 8. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 56 Table A.17: Contact Farmers' Perceptions of SLM Techniques Variables Increase Reduce land Reduce Reduce N productivity preparation planting harvesting eort seed eort eort Mulching Treated 0.569 0.423 0.307 0.234 137 Control 0.595 0.452 0.310 0.238 42 Strip tillage Treated 0.314 0.270 0.321 0.139 137 Control 0.286 0.238 0.286 0.190 42 Micro-catchments Treated 0.212 0.088 0.124 0.073 137 Control 0.238 0.143 0.119 0.071 42 Contour farming Treated 0.095 0.022 0.044 0.036 137 Control 0.071 0.048 0.071 0.071 42 Crop rotation Treated 0.401 0.255 0.190 0.182 137 Control 0.357 0.262 0.214 0.190 42 Row planting Treated 0.182 0.117 0.168 0.073 137 Control 0.167 0.095 0.095 0.071 42 Improved fallowing Treated 0.124 0.036 0.036 0.044 137 Control 0.095 0.071 0.048 0.048 42 Source: Household Survey, 2012. Note: ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 57 Table A.18: Eect of SLM Training Intervention on Contact Farmers' Labor Time Midline Endline Control ITT N Adj. Control ITT N Adj. Mean(SD) R2 Mean(SD) R2 Hours spent on 5.762 -4.124** 178 -0.064 6.429 -2.167 168 -0.070 preparation of land (13.879) (1.977) (15.353) (2.857) Hours spent on seeding 6.071 -1.183 178 -0.036 10.357 -6.386** 168 -0.083 (13.767) (2.717) (17.862) (3.081) Hours spent on 3.476 -2.105 178 0.020 1.738 -0.751 168 -0.082 transplantation (9.094) (1.613) (6.666) (1.516) Hours spent on irrigation 0.000 -0.214 178 -0.130 (0.000) (0.380) Hours spent on sacha 15.333 0.258 178 -0.050 5.833 -0.093 168 -0.082 (15.550) (3.176) (14.252) (2.539) Hours spent on protection 0.000 1.061 178 -0.131 0.000 0.656 168 -0.027 (0.000) (0.805) (0.000) (0.658) Hours spent on harvesting 6.214 -0.988 178 -0.132 15.810 -3.314 168 -0.011 (15.645) (2.391) (19.573) (3.711) Total weeks spent on 26.143 -1.233 178 -0.114 30.381 -9.298** 168 -0.056 farming in last year (17.512) (3.615) (19.196) (3.720) Source: Household Survey, 2012, 2013. Note: Regressions include the same explanatory variables as models in Table 5. ***, **, and * indicate signicance at the 1, 5, and 10 percent critical level for t statistics. 58