Policy Research Working Paper 9840 Knowledge and Adoption of Complex Agricultural Technologies Evidence from an Extension Experiment Denise Hörner Adrien Bouguen Markus Frölich Meike Wollni Development Economics Development Policy Team November 2021 Policy Research Working Paper 9840 Abstract In most of Sub-Saharan Africa, agricultural extension extension-only and extension combined with video increase models have become more decentralized and participa- adoption and knowledge of the package, especially of its tory and thus, rely on effective farmer-to-farmer learning, more complex components; although on average, there while increasingly including non-traditional forms of edu- is no additional effect of the video intervention on adop- cation. At the same time, agricultural technologies become tion. Knowledge and the number of adopted practices also more complex and are now often promoted as integrated increase among farmers not actively participating in exten- packages, likely to increase the complexity of the diffusion sion activities, suggesting information diffusion. For this process. Based on a randomized controlled trial, this study group, the additional video intervention has a reinforcing assesses the effects of ‘farmer-to-farmer’ extension and a effect, and particularly fosters adoption of the integrated video intervention on adoption of a complex technology package. package among 2,382 smallholders in Ethiopia. Both This paper is a product of the Development Policy Team, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at denise.hoerner@uni-goettingen.de. 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 Knowledge and adoption of complex agricultural technologies – Evidence from an extension experiment – Denise Hörner, Adrien Bouguen, Markus Frölich, and Meike Wollni JEL classification: D83, O13, O22, O33, Q16 Key words: Randomized controlled trial, integrated soil fertility management, technology dif- fusion, agricultural extension systems, rural development Denise Hörner (corresponding author) is a postdoctoral researcher and Meike Wollni is a professor at the Depart- ment of Agricultural Economics and Rural Development, University of Goettingen, Germany; their email ad- dresses are denise.hoerner@uni-goettingen.de and mwollni1@gwdg.de. Adrien Bouguen is assistant professor at the Department of Economics, Santa Clara University, United States; his email address is abouguen@scu.edu. Markus Frölich is a professor at the University of Mannheim and Center for Evaluation and Development (C4ED), Germany, and J-PAL Affiliate; his email address is froelich@uni-mannheim.de. This work was financially sup- ported by the German Research Foundation (DFG) [grant numbers RTG1666 GlobalFood and 390367541]; in particular for data collection, funding of the video intervention and data analysis. The German Agency for Inter- national Cooperation (GIZ) provided funding for baseline data collection. Further support for research analysis was provided by DFG through SFB 884 at University of Mannheim. The thank all members of the Integrated Soil Fertility Management Project (GIZ-ISFM+) for their support, and Ibrahim Worku for his excellent field work coordination as well as the whole team of research assistants and enumerators, and farmers who gave their time. The authors are grateful to Eric Edmonds and three anonymous reviewers for their constructive comments and valuable advice that have greatly helped to improve this manuscript. We thank participants at the NBER Transfor- mation of Rural Africa Conference, the Annual International Conference of the Research Group on Development Economics 2019, the Agricultural & Applied Economics Association Annual Meeting 2019, the PEGNet Confer- ence 2019 as well as seminar participants at C4ED in Mannheim and DARE in Göttingen for helpful comments. This study is registered under the AEA RCT ID: AEARCTR-0004406, a previous version has been published in the NBER Working Paper Series (Hörner et al. (2019)). A supplementary online appendix is available with this article at The World Bank Economic Review website. 1. Introduction The slow adoption of new agricultural technologies is an important factor in explaining persis- tent productivity deficits and poverty among the rural population in developing countries, es- pecially in Sub-Saharan Africa (SSA). Recent evidence shows that farmers delay in particular the uptake of integrated system technologies, i.e. packages of agricultural practices that should be jointly applied in order to deploy their full productivity-enhancing potential (Noltze, Schwarze, and Qaim 2012; Sheahan and Barrett 2017; Ward et al. 2018). Integrated system technologies are typically knowledge-intensive, as they require the understanding of at least basic underlying biological functions and processes and the adaptation of practices to local agri- environmental conditions (Vanlauwe et al. 2015; Jayne et al. 2019). While information and knowledge constraints are frequently cited barriers to the adoption of agricultural innovations in general (Foster and Rosenzweig 1995; Magruder 2018), they are likely to play a key role in explaining incomplete or non-adoption of complex system technologies (Takahashi, Muraoka, and Otsuka 2019). Based on a randomized controlled trial, this study analyzes the effects of ‘farmer-to-farmer extension’ on knowledge and adoption of an integrated system technology. The effectiveness of information spillovers is analyzed as a key principle of farmer-to-farmer extension models by estimating differential effects for those who actively participate in the extension activities and those who at most benefit indirectly. The study further evaluates whether an additional intervention in form of a video can counterbalance incomplete information diffusion likely to occur in farmer-to-farmer extension set-ups and thus, foster the wider adoption of the system technology. Agricultural extension services aim at transferring knowledge to farmers in order to bridge knowledge and capacity gaps. In recent decades, agricultural extension approaches in SSA have become more decentralized, where extension agents train a small number of selected farmers 2 (often referred to as ‘model farmers’) in the application of new techniques. These model farmers are then expected to pass their knowledge on to other farmers in the village, who are usually organized in groups to facilitate participatory learning processes. This goes along with a shift in perspective from a ‘top-down’ to a more inclusive ‘bottom-up’ strategy and is often referred to as ‘farmer-to-farmer extension’ (Takahashi, Muraoka, and Otsuka 2019). Eventually, expo- sure to on-farm demonstrations, trained model farmers and group members is supposed to spur broader technology adoption in the community (Gautam 2000; Swanson 2008). A growing body of literature has analyzed the effectiveness of decentralized extension mod- els in facilitating innovation and knowledge diffusion. There is now substantial evidence that directly training selected farmers spurs knowledge and adoption among them (Godtland et al. 2004; Davis et al. 2012; Kondylis, Mueller, and Zhu 2017), and some evidence that subsequent diffusion to other farmers takes place (Fisher et al. 2018; Nakano et al. 2018; Takahashi, Mano, and Otsuka 2019). On the other hand, several studies conclude that knowledge gains among trained individuals hardly trickle down to neighboring farmers (Rola, Jamias, and Quizon 2002; Feder, Murgai, and Quizon 2004; Tripp, Wijeratne, and Piyadasa 2005), and that increased technology adoption among trained farmers does little to change the behavior of non-trained peers (Van den Berg and Jiggins 2007; Kondylis, Mueller, and Zhu 2017). A relatively new strand of research focuses more explicitly on the determinants of diffusion processes in farmer-to-farmer extension set-ups. These studies find that successful diffusion is shaped by model farmers’ motivation and familiarity with the technology (Fisher et al. 2018), incentives attached to information dissemination (BenYishay and Mobarak 2019; Shikuku et al. 2019), the social distance between communicators and target farmers (BenYishay and Mo- barak 2019; Shikuku 2019) as well as other context-specific forms of social capital (Pamuk, Bulte, and Adekunle 2014). In addition, some studies suggest that farmers need to learn from multiple sources before they adopt (Beaman et al. 2018; Fisher et al. 2018). Most of these stud- ies, however, focus on the adoption of (several) individual practices, while recent extension 3 efforts in SSA increasingly concentrate on integrated system technologies (Takahashi, Mu- raoka, and Otsuka 2019). This paper contributes to the literature by focusing on the integrated adoption of a complex system technology, rather than on the uptake of individual practices, an issue that is largely understudied to date (Sheahan and Barrett 2017; Jayne et al. 2019). The study was implemented in the context of a large-scale farmer-to-farmer extension program promoting ‘Integrated Soil Fertility Management’ (ISFM) in three rural regions of Ethiopia. ISFM is a knowledge-inten- sive system technology widely promoted in SSA as a strategy to sustainably intensify agricul- tural productivity, combat land degradation and enhance rural livelihoods (Jayne et al. 2019). A fundamental feature of ISFM is the integrated use of improved seeds together with inorganic and organic soil amendments. In addition, ISFM aims at generally improving agronomic tech- niques adapted to local conditions (Place et al. 2003). By estimating differential effects for direct and indirect beneficiaries of the extension inter- vention, this paper adds to the scarce literature on information spillovers in the context of inte- grated system technologies. For the case of complex technologies, previous studies document substantial information losses along the transmission chain from extension agents to farmers (Niu and Ragasa 2018; Maertens, Michelson, and Nourani 2020). Yet, literature also suggests that reminders of commonly neglected knowledge dimensions can help to offset incomplete information diffusion (Hanna, Mullainathan, and Schwartzstein 2014; Niu and Ragasa 2018). This could potentially be achieved through video messages, since previous research has shown that video as information delivery channel can increase knowledge and technology adoption in farming communities (Bernard et al. 2016; Van Campenhout, Spielman, and Lecoutere 2020). This study expands the existing evidence on the effectiveness of video interventions, par- ticularly focusing on their role in offsetting incomplete information diffusion in farmer-to- farmer extension settings. In this context, the video intervention is intended to remind farmers of commonly neglected knowledge dimensions, in particular emphasizing the importance of the 4 holistic concept of the system technology (joint application of practices), and additionally ex- plains the underlying principles of the components. While extension activities often aim at providing awareness for improved practices and instructions on how to implement them, they frequently disregard the importance of providing sufficient information on why certain practices are beneficial (Rogers 1995; Anderson and Feder 2007). This study explicitly focuses on the role of different types of knowledge, including knowledge on the underlying principles, as po- tential drivers of adoption. The paper is structured as follows. Section 2 describes the experimental design, section 3 the empirical data and estimation strategy. Section 4 presents the main results; section 5 dis- cusses the implications of the findings, and section 6 concludes. 2. Experimental design The study builds on a randomized controlled trial (RCT) with two treatment arms and a control group. The first treatment consists of an extension intervention; the second treatment combines the extension intervention with a video intervention. Microwatersheds (mws) were used as units of randomization, which are common implementation units for natural resource related inter- ventions in Ethiopia.1 The extension intervention was part of the ‘Integrated Soil Fertility Man- agement Project ’ (ISFM+ project) carried out by the German Agency for International Cooper- ation (GIZ) in three rural regions of Ethiopia; Amhara, Oromia, and Tigray. 2 During the initial phase of the ISFM+ project from 2015 to 2018, the use of five so-called ‘quickwin technolo- gies’ including compost, blended fertilizers3, improved seeds, line seeding (as opposed to broadcasting), and lime application was promoted for the major cereal crops maize, wheat, teff, 1 Microwatersheds are the smallest hydrological entity, i.e., subunits of major watersheds, defined by the topography of the land, and typically consist of around 250 to 300 households in one or several communities that share a common rainwater outlet. 2 The ISFM+ project is a component of GIZ’s contribution to the Ethiopian ‘Sustainable Land Management Programme’ (SLMP) and only operates in districts, where physical land rehabilitation measures (stabilization of hillsides, erosion control measures) have been successfully introduced by the SLMP. Beginning of 2018, the SLMP has been replaced by the successor program named ‘Sustainable Use of Rehabilitated Land for Economic Development’ (SURED). 3 Blended fertilizers are inorganic fertilizers composed in accordance to a specific location’s soil type. 5 barley and sorghum. Lime is promoted only in regions where soils suffer from acidity, which applies to Amhara and Oromia, but not to Tigray. The project encourages the joint application of these individual components as a key feature of ISFM. Treatment description The core elements of the extension intervention are the following: In each treatment mws, three so-called ‘farmer research and extension groups’ (FREG) were formed, each consisting of 16 or 17 members, leading to a total of around 50 FREG members per mws. FREG farmers were selected in a non-random manner by extension agents and village heads, based on farmers’ interest and social involvement. The FREGs conduct regular meetings, typically once or twice per month, to discuss agricultural topics. Each group is led by three of its members, called ‘model farmers’, whose central activity is the establishment and maintenance of demonstration plots. For this purpose, model farmers receive trainings on ISFM from public extension agents and are provided with all necessary inputs. Demonstration plots are on-farm trials on which the package of ISFM practices is applied, next to plots that are managed according to traditional farming practices. Hence, the benefits of ISFM in comparison to traditional practices, such as yield improvements, become clearly visible to farmers (cp. Supplementary Online Appendix S1, available with this article at The World Bank Economic Review website). In each mws, ‘farmer field days’ are conducted twice per harvest cycle; at critical stages around mid-season and just before harvest. During these field days, model farmers share and discuss their experi- ence with FREG members; extension agents are present to complement information. Field day activities are mainly targeted at FREG members, although in some communities, other farmers do also participate. Overall, the extension treatment aims at creating awareness and know-how about ISFM through a knowledge sharing process from extension agents to model farmers, and from model farmers to other FREG members. FREG members are then encouraged to share the information with the broader population of farmers in the communities. Hence, this model 6 heavily relies on peer-to-peer learning. While information diffusion to model farmers and other FREG members is facilitated through regular meetings, farmer field days, and visits to demon- stration plots, information sharing with the broader community is not formalized or incentivized in any particular way. The video intervention has been designed to provide an additional stimulus for adoption by exposing farmers to information about the ISFM concept, in order to overcome potential knowledge gaps on key dimensions of the approach. It provides farmers with information on why each component is important, that is, explanations about the underlying principles and mechanisms of ISFM, and emphasizes the positive synergy effects of applying the practices jointly on the same plot. The movie is composed of two parts: A narrative and documentary part that presents the example of a farmer couple who has successfully implemented the ISFM quickwin technologies and visibly increased yields, serving as (potential) role models for treated farmers. Given cultural, linguistic and agroecological differences between Tigray, Am- hara and Oromia, three different farmer couples were featured in the versions for the respective region. Beyond smaller local adaptations, all three versions strictly follow the same script in order to convey the same messages. The second component of the film consists of animations that visualize processes taking place in the soil – such as hydrological and soil nutrient cycles – and how these relate to ISFM. Ultimately, farmers should gain a better understanding of why the integrated use of all techniques is important to improve soil fertility and productivity. Sampling and randomization strategy Since the participatory extension approach draws on the establishment of community-based farmer groups and demonstration sites, the study applied a cluster randomization approach us- ing microwatersheds (mws) as sampling units. The full sampling frame consists of 161 mws located in 18 districts (in Ethiopia called Woredas), equally distributed among the three regions 7 Tigray, Amhara and Oromia.4 From this list, a sample of treatment mws was drawn randomly – stratified by region and Woreda – so that in each Woreda four mws were selected, resulting in a total of 72 treatment mws. Half of the 72 treatment mws were assigned to the additional video intervention. Consequently, 36 mws received the extension treatment only (in the follow- ing referred to as T1), and another 36 mws received the extension treatment plus the additional video intervention (T2). The remaining 89 mws serve as control group (C). For the survey, complete lists of households living in the respective mws were compiled using administrative village lists of registered households. From these lists, 18 households were randomly selected in each mws; 15 to be interviewed and three as potential replacements. 5 Thus, in treatment mws, the proportions of non-FREG and FREG farmers in the sample should on average represent their distribution in the population. The full original sample consists of 2,416 households and is graphically depicted in Figure 1. Figure 1. Diagrammatic illustration of the full sample. Source: Authors’ illustration. Note: MWS stands for microwatershed, HH for household. 4 The list of target mws was compiled based on the criteria (i) benefiting from the Sustainable Land Management Programme and (ii) no/minimal previous exposure to soil fertility interventions. 5 Households were only replaced if they refused to participate in the survey or could not be found. If household members were currently unavailable, the enumerator would return at a later time for the interview. 8 Treatment implementation Figure 2 depicts a timeline of the treatment implementation and the survey data collection. While the ISFM+ project was launched by the German Agency for International Cooperation (GIZ) in mid-2015, the above-described extension activities only started in the following main cropping season, i.e., in April 2016. 6 Since then, extension activities in the treatment mws are on-going and implemented each year during the main growing season. The video screenings were conducted in T2 mws in early 2017, around six weeks prior to the start of the main growing season. The video was shown in public spaces such as farmer training centers, health posts or schools, and followed by group discussions facilitated by ex- tension agents. In each of the T2 mws, the 15 households from the sample were invited by village heads a few days prior to the screenings orally and with written invitation cards. In the case of double-headed households we invited both spouses, otherwise only household heads. Farmers participating in the video screenings were registered by name to document attend- ance.7, 8 Figure 2. Timeline of treatment and survey implementation. Source: Authors’ illustration. Note: FFD stands for farmer field day. 6 In the first months, the project’s main activities were establishing partnerships and conducting planning workshops. 7 Individuals from other mws were not allowed to attend the screenings. However, this situation did not occur in the field, because the video screenings were not broadly announced. 8 After the endline data collection, the video became freely available for extension staff to be used in T1 as well as control communities. 9 3. Data and empirical strategy Data In order to assess the interventions’ impact, two rounds of survey data were collected (cp. Fig- ure 2). A baseline survey took place in early 2016, immediately before the start of the 2016 main growing season. The endline survey was carried out in early 2018, i.e., before the start of the 2018 main growing season, among the same rural households. Data in both rounds were collected through tablet-based face-to-face interviews with the household head or spouse, using a structured questionnaire. The attrition rate was remarkably low, since 2,382 (98.6%) of the 2,416 baseline sample households could be re-interviewed during endline, and we cannot detect any non-random patterns in this. The surveys collected data on household socio-demographic characteristics, income and assets, farming practices, agricultural production and exposure to agricultural extension. Data was elicited retrospectively for the last twelve months; thus, the agricultural data covers the preceding cropping season, i.e., the baseline survey refers to the 2015 and the endline survey to the 2017 season. Accordingly, the impact assessment measures adoption in the 2017 cropping season while extension activities were ongoing, and after farmers had been exposed to one full cycle of extension activities during the 2016 growing season. The endline survey assessed most information in the exact same way as in the baseline to maintain comparability. Yet, for the adoption of ISFM practices we had to adapt the mode of measurement slightly, because the information collected at baseline was not detailed enough for the intended analyses. While at baseline, agricultural information was elicited at farm level, we obtained plot-level information in the endline, which allows to assess the integrated adop- tion of ISFM at the plot level. In addition, the endline survey contained measures to increase the reliability of the self-reported data. For compost, enumerators inspected and measured the compost pit or heap, and assessed detailed questions on the production process and quality of 10 the end product to avoid simple yes/no questions that tend to be more susceptible to response bias (cp. Online Appendix S2). For inorganic fertilizers, enumerators used pictures to identify the type of fertilizer used by the farmer and in addition checked the labels on fertilizer bags. 9 For lime, they also inspected the respective bags. In order to verify the use of improved seeds, the survey contained detailed questions on the original source of the seed and for how long it had been reused. If seeds had been reused for more than four seasons, they are not considered improved seeds in the analysis, because improved traits get lost over generations. 10 The endline survey further comprised questions on the participation in ISFM+ extension activities as well as a detailed knowledge test on ISFM technologies. Enumerators first assessed farmers’ awareness by asking them which ISFM components they actively remembered, and in a second step, asked them which practices they recognized when read to them off a list by name.11 Subsequently, questions on their underlying principles and purpose (‘principles knowledge’) as well as their mode of implementation (‘how -to knowledge’) were posed. The survey combined different types of knowledge questions, including open-ended and multiple- choice questions, to minimize fatigue effects (for details of the knowledge exam, see Online Appendix S3). For the survey, a team of 34 enumerators was recruited and intensively trained during a ten- day training period. Questionnaire contents were carefully translated into the three local lan- guages and pre-tested in several rounds. In addition to the farm household survey, the survey team administered two community level questionnaires to key informants at the Woreda and mws levels, in order to collect data on infrastructure, availability of extension services, rainfall and temperature, as well as other contextual characteristics. 9 Even if bags are empty, farmers usually keep them for other purposes, e.g., to sit on them or to store other things. 10 Please note that visual inspections in the field were not possible for the practices improved seeds and line seeding, because at the time of the survey, packaging material of seeds was usually not available anymore and fields had already been harvested. 11 Inspired by (Kondylis, Mueller, and Zhu 2015), this list included a placebo practice (“seeding in circles”) to get a sense for possible response bias, which does not appear to threaten the results since yes-answers regarding this practice are close to zero. 11 Table 1 depicts descriptive statistics for selected variables at baseline using data from the bal- anced panel of 2,382 households, including tests for covariate balancing between the three treat- ment groups to verify the success of the randomization process. 12 Overall, households in the three groups seem largely balanced on a series of socio-demographic and economic indicators. Yet, they exhibit a few differences regarding agricultural production-related characteristics, which need to be considered in the outcome estimation framework. Moreover, ISFM practices are not necessarily new to farmers, since some were used prior to the interventions, though mostly to a modest extent. Table 1 about here Key outcome variables Since the key concern of this study is the increase in ISFM adoption, we measure how many ISFM practices households adopted in the preceding harvest cycle (2017), and therefore assess the number of ISFM technologies adopted, ranging from zero to five. We use this variable de- spite the fact that lime is not relevant in one of the regions (and thus, no farmer in Tigray reaches a value of five), but provide robustness checks verifying that implications do not change, if we exclude lime and employ a zero to four measure instead. Since the complementary use of the practices is pivotal to ISFM, the second main outcome is the integrated adoption of the full ISFM package, assessed with a binary variable that measures whether a farmer has used all four quickwin practices in combination on at least one plot. This variable excludes the use of lime, since adoption would otherwise always be zero in Tigray. 13, 14 Furthermore, the adoption of the individual ISFM quickwin components is of interest, which are compost, blended fertilizer, 12 Table 1 shows those variables that are used as additional covariates in the adoption and knowledge regressions, Table S4.1 in Online Appendix S4 presents further balance checks on selected household, farming and community characteristics 13 To check sensitivity of the results with respect to this definition, we also use several alternative measures. Firstly, a variable that equals one when at least four out of five practices (including lime) are adopted jointly. Secondly, a measure for the joint adoption of compost, blended fertilizer and line seeding, that is, excluding improved seeds, which might possibly be concen- trated around a certain crop type. And lastly, a region-specific measure that requires all five practices to be adopted in Amhara and Oromia, but only four in Tigray. 14 These analyses are restricted to plots planted with main crops, i.e. wheat, maize, teff, barley and sorghum, which are the main focus of the interventions. 12 improved seeds, line seeding and lime. For each technology, a dummy variable taking the value of one defines if the household has applied the respective practice in the 2017 main cropping season on any plot and for any crop type. We construct an overall ISFM knowledge score based on questions on each of the individual ISFM components (excluding questions on lime). It ranges from zero to one, with one standing for full knowledge, i.e., having answered all questions correctly. 15 Since the aim of the video treatment was to increase farmers’ knowledge on why ISFM is important (and not on how to implement it), we also construct two individual indicators for principles and how-to knowledge, depending on whether a question was on the purpose (the ‘why’) of a technology, or its mode of implementation (the ‘how-to’), which also range from zero to one. For the how-to score, twice the weight is given to knowledge on how to correctly produce compost, since this is a more complicated process than the implementation of the other ISFM practices. The principles score gives equal weights to all ISFM components, and includes one indicator for the general understanding of the necessity to integrate organic and inorganic soil inputs. 16 For the analyses, the household heads’ knowledge score is used. Estimation strategy for intent-to-treat effect using RCT design In order to assess the effects of the experimental interventions on ISFM adoption and knowledge, regressions of the following form are estimated: = + 1 + 2 + є (1) where denotes the respective outcome variable for household or individual i in mws j, meas- ured at endline. 1 is a dummy variable indicating whether the mws was assigned to the ex- tension intervention, and 2 indicates whether the mws was assigned to the combined 15 Since the number of questions is not the same for all practices, we first calculate a knowledge score for each component individually and then combine it to an overall score, so that each ISFM dimension is included with the same weight. Farmers who were not aware of a practice in the first place, were immediately given a value of zero in the respective follow-up questions. 16 This is based on the respondent selecting the following statement as being correct: “The soil needs both organic and inorganic inputs to be healthy and fertile”. 13 extension + video treatment; є represents an individual level error term that is clustered at the mws level to allow for arbitrary correlation of households and individuals within clusters. Although treatment indicators should be orthogonal to further explanatory variables due to randomization, all models are re-estimated including additional covariates in order to increase precision of the estimates and to control for small-sample imbalances: 0 0 = + 1 + 2 + + + є + (2) 0 In these models, represents a vector of control variables related to farmer and household characteristics captured at baseline, while , adds community-level indicators for rainfall, temperature and remoteness. When available, the inclusion of the baseline level of an outcome 0 in the equation should reduce the overall variance, since some degree of path dependency on previously gained experience with a technology is assumed. This treatment effect model is appropriate in this case, since for some outcomes, baseline and endline measures are not com- pletely identical, or baseline data is not available at all. 17 In addition, this specification has been shown to be more powerful than the difference-in-difference estimator in the presence of rela- tively low autocorrelation (McKenzie 2012a; De Brauw et al. 2018), which can at least be stated for some of the outcome variables. The main estimators of interest are β and λ, capturing the intent-to-treat effect (ITT). One-sided equality tests evaluate the additional impact of the video intervention, by testing whether the coefficient of T2 is significantly larger than that of T1. 17 Baseline data is available for adoption of compost, blended fertilizer, improved seeds and lime. Regarding blended fertilizer, we additionally control for ex-ante use of any inorganic fertilizer, since during time of baseline, blended fertilizer was largely unavailable; instead, farmers used the widely available DAP fertilizer (cp. Table 1). In the two years between baseline and endline, NPS/NPK fertilizer blends largely replaced other inorganic fertilizer types. Line seeding can only be proxied, since it was assessed on a more general level during baseline, asking farmers how they usually plant crops, but not at the plot level. Knowledge variables were not measured in the baseline survey. 14 Differential effects for members and non-members of ‘farmer research and extension groups’ Beyond average effects, it is also interesting to assess potentially differential effects of the treat- ments on the primary beneficiary group, i.e., FREG members, and those who might only benefit indirectly, i.e., non-FREG farmers. Since FREG membership is not randomized, the estimation strategy needs to consider that FREG members and non-members are likely to be systematically different. Propensity score matching should reduce potential bias resulting from self-selection into FREGs, (see e.g. Cameron and Trivedi 2005; Frölich and Sperlich 2019). In a first step, a probit regression is estimated to predict FREG membership: = + + (3) where Z denotes a vector of farmer and household covariates assumed to influence farmer ′ participation , and µ is the error term. Based on this estimation, a propensity score is calcu- lated, which is then used to match each real FREG member in the treatment group with their most similar counterfactual from the control group (referred to as ‘matched controls’) via one- to-one nearest neighbor matching without replacement. The analysis proceeds analogously for the non-members. Subsequently, treatment effects on the core outcome variables within the two mutually exclusive subsamples are re-estimated: the FREG sample, consisting of actual FREG members and their matched controls (in treatment and control mws, respectively) and the non- FREG sample, consisting of actual non-FREG members in treatment mws and their matched controls. Applying this matching algorithm to the sample leads to a high level of common support, since only one treated observation is off support (see Online Appendix S4).18 Balance checks between the treatment groups for the constructed FREG and non-FREG samples show that 18 See Table S4.2 for first-stage propensity score matching regression results and Figures S4.1 and S4.2 for histograms of the estimated propensity scores. 15 samples are well balanced, with only few exceptions (see Tables S4.3 and S4.4 in Online Ap- pendix S4). Propensity score matching allows to control for selection on observables as well as a portion of unobserved heterogeneity that is correlated with these observables. Nonetheless, FREG and non-FREG members may still differ in terms of unobserved characteristics that were not captured by the matching procedure and are therefore also not reflected in the selected matches from the control group. Matching results can also be sensitive to variable selection and the choice of the matching algorithm. Therefore, a number of robustness checks are carried out.19 Findings reveal that results remain robust to alternative model specifications and match- ing algorithms (see Tables S4.5 to S4.10 in Online Appendix S4). 4. Results Treatment participation Among the two treatment groups, 82 farmers (8% of T1 and T2) were active model farmers in the 2017 cropping season; that is, they were leading members of a FREG and responsible for the implementation and maintenance of an ISFM demonstration plot, for which the project pro- vided them with inputs. In addition to model farmers, 120 farmers (around 12% of T1 and T2) are active FREG members, meaning they belong to a FREG and have participated in field day activities along the course of the preceding season. In addition, 77 (8%) of the treatment farmers who are no FREG members state to have participated in a field day in 2017, plus 39 (3%) of control group farmers. Regarding the visit of demonstration plots, 55 treatment farmers (6% of T1 and T2) not belonging to a FREG report to have visited a demonstration site on their own behalf, i.e., independently of a field day, in addition to 39 farmers in the control group (3% of C). Consequently, although to a low extent, there are indications of treatment spillovers both 19 Firstly, a different set of matching variables is used, namely the same set of covariates used in the ITT regressions. Secondly, we perform non-parametric kernel matching for both FREGs and non-FREGs with the control group, and use the calculated kernel weights in the outcome regressions. And lastly, instead of FREG membership, field day attendance is used to construct the two subsamples. 16 within and across groups, which also means that ITT estimates might suffer from a slight down- ward bias due to ‘contamination’ of the control group. Compliance in the video intervention was remarkably high, 499 (94%) of T2 households attended the screenings. Considering that in double-headed households both spouses were in- vited to the sessions, compliance at the individual level was 83%, equal to 804 participants. None of the households from T1 or the control group were present during the video screenings. ISFM adoption decision Aggregated adoption measures The ITT effects of the two treatment arms on the first core outcome, the number of adopted ISFM quickwin technologies (0-5) are obtained with three different regression specifications (cp. Columns (1) to (6) of Table 2). Since the dependent variable essentially is a count variable, a Poisson model is estimated. Yet, considering that it can also be perceived as either an under- lying continuous or ordered process, a linear as well as an ordered probit model is run to under- line the robustness of the findings. Table 2 about here The results of all models indicate positive and highly statistically significant effects of both treatments on the number of adopted practices, which are robust to the inclusion of baseline control variables. Furthermore, all models lead to larger point estimates for T2 compared to T1. Findings from the linear model show that farmers in T1 adopt an additional .448 practices, while households in T2 communities on average adopt .569 more practices than those in the control group (Column (2)).20 Yet, across specifications, p-values of the test that T1 >= T2 are 20 Result of the Poisson and the ordered probit regressions are well in line with those of the OLS specifications and will therefore not be explicitly discussed. 17 insignificant or only marginally significant, and thus do not provide consistent evidence of T2 having a larger effect on the number of adopted ISFM practices. 21, 22 With respect to the integrated adoption of the full quickwin package, the estimated ITT effects are also positive and statistically significant (Columns (7) and (8) of Table 2). House- holds in T1 are on average 8.4 percentage points more likely than control group households to adopt the full set of practices on the same plot, while the likelihood for farmers in T2 is 10.9 percentage points above the control group mean. However, again the difference between the effect sizes of T1 and T2 is not statistically significant.23 In order to test whether the estimated treatment effects might be driven by the 82 model farmers in the sample that have been trained by extension agents and provided with inputs, ITT models on the two adoption outcomes are re-estimated excluding these 82 model farmers. All treatment effects remain highly statistically significant with only a slight decrease in magnitude suggesting that the interventions affect farmers in treatment communities beyond model farm- ers (cp. Table S4.13, Online Appendix S4). Adoption of individual components To further explore which components are the main drivers of increased ISFM adoption, ITT effects on the decision to adopt each of the five quickwin practices individually are analyzed. Average marginal effects presented in Table 3 indicate that both the extension-only and the combined intervention exert positive and statistically significant effects on the decisions to adopt compost, improved seeds, line seeding and lime. In contrast, effects on blended fertilizer 21 Rough power calculations show that the minimum detectable effect size in the comparison T1 versus T2 is about 24% of a std. deviation. Accordingly, the estimated difference between the two coefficients is likely too small to detect significant addi- tional effects of the video intervention. 22 Implications of these results do not change when using the 0-4 measure (excluding lime), see Table S4.11 in Online Appendix S4. 23 Table S4.12 in Online Appendix S4 shows results for the three alternative specifications. Although effect sizes naturally vary with the choice of this measure, results remain qualitatively unaltered. 18 are not significant (T1) or do not remain significant with the inclusion of additional controls (T2). To account for the probability of false discoveries in multiple outcome testing, we follow Sankoh et al. (1997) and Aker et al. (2016) and use a version of the Bonferroni correction, which corrects for inter-outcome correlations for families of outcomes (cp. Online Appendix S5) (McKenzie 2012b). With this form of adjustment, p-values of the estimated coefficients of both T1 and T2 increase (respectively remain), above the .10 threshold for blended fertilizer and improved seeds, while results for compost, line seeding and lime remain significant for both treatment arms. For these robust results, the estimated effect sizes of the extension-plus-video intervention are larger than those of the extension-only intervention, which is in line with the findings on the aggregated ISFM adoption measures. However, again there are no statistically significant differences between the effects of T1 and T2 on technology adoption (cp. adjusted p-values of one-sided equality tests in Columns (2), (8) and (10)). In summary, results indicate significant ITT effects of the extension intervention on the adoption of ISFM, both on aggregated measures as well as on some of its individual compo- nents. Yet, despite consistently larger point estimates, findings provide little significant evi- dence for an additional ‘video effect’. Table 3 about here Differential effects for FREG members Next, we are interested in whether the extension treatment has an effect on ISFM adoption beyond FREG membership – or whether the estimated ITT is solely concentrated among FREG farmers –, and whether the additional video treatment might influence FREG members and non- FREG members differently. This is analyzed using the matching approach outlined above. 19 In both subsamples, treatment effects of the two interventions regarding the number of adopted ISFM practices remain positive and statistically significant (Table 4). Yet, both the linear and the Poisson specification indicate that the effects of the treatments are substantially larger in the FREG than in the non-FREG sample. While in the non-FREG sample, being as- signed to T1 on average increases the number of applied technologies by .278, this coefficient is 1.232 in the FREG sample. Similarly, T2 is estimated to increase average adoption by .483 practices in the non-FREG sample, but by 1.117 technologies in the FREG sample (OLS results in Columns (2) and (4)). Further, for non-FREGs, coefficients of the combined treatment are significantly larger than those of the extension-only treatment, pointing towards a reinforcing effect of the video for this group of farmers (cp. Columns (2) and (6)). Regarding integrated adoption of the ISFM package, estimation results show that both in- terventions significantly increase the likelihood of adoption in the FREG sample in comparison to their matched controls by 28 and 23 percentage points, respectively (Column (12) of Table 4). In contrast, in the non-FREG sample, the extension-only treatment does not lead to a signif- icant increase in integrated adoption. Only if extension is complemented by the video (T2), results show a significant treatment effect for non-FREG farmers, with an increase in the like- lihood to adopt the full package of 8 percentage points compared to their matched controls (Column (10)). These findings lead to the following two conclusions: Firstly, the effect of the extension treatment is substantially larger for FREG members – even after taking into account that they may be the better farmers anyways. This is expected because they are the farmers directly ben- efitting from the extension activities. Yet, the extension intervention does show a positive in- fluence also on non-FREG farmers when it comes to the number of adopted ISFM practices at the household level, pointing towards the presence of diffusion effects. However, most inter- estingly, the findings indicate that extension alone does not significantly affect non-FREG farmers when it comes to integrated adoption, i.e., using the practices together on the same plot. 20 By contrast, it seems that the video intervention has a significant complementary effect for non- FREG farmers, in particular when it comes to the integrated adoption of the practices. Table 4 about here ISFM knowledge Treatment effects on knowledge Positive and significant ITT estimates show that T1 on average seems to increase overall ISFM knowledge by around 3.6 percentage points, while T2 increases farmers’ overall knowledge score by almost 7 percentage points in comparison to the control group mean (Column (2) of Table 5). The p-value of .008 indicates that extension-plus-video has a significantly stronger effect on knowledge than extension alone and thus, points towards an additional effect of the video regarding ISFM knowledge formation. We also assess the ITT effects on the two distinct domains, principles and how-to knowledge. Regarding principles knowledge, the positive co- efficient of extension alone does not remain statistically significant with the introduction of further covariates, whereas extension combined with video on average increases this knowledge score by 5.4 percentage points on a highly significant level (Column (4) of Table 5). Thus, the video, that explicitly sought to explain the underlying principles of ISFM, was successful in conveying some of this knowledge. How-to knowledge seems to be positively affected by both T1 and T2, with no statistical difference regarding their effect sizes (cp. Column (6)). Table 5 about here In order to understand the contribution that these gains in ISFM knowledge make to the adop- tion decision, a formal causal mediation analysis is conducted (cp. Online Appendix S6). Find- ings show that gains in knowledge are indeed a relevant driver of adoption. Increases in how- to knowledge account for 16% to 23% of the treatment effects on adoption. Gains in principles knowledge explain 6% to 7% of the T2 effect on adoption, but no portion of the T1 effect (Table 21 S6.1). This makes sense, given that the video had a particular focus on conveying principles knowledge, and indeed seems to positively affect this knowledge domain. Differential effects for FREG members Next, we follow the earlier approach and disaggregate the sample into a FREG and a non-FREG sample. The significant difference between the effect sizes of T1 and T2 on overall knowledge persists in the non-FREG sample, but is much less pronounced in the FREG sample (cp. Table 6). While in the FREG sample, both extension-only and the combined intervention significantly increase the overall knowledge score, in the non-FREG sample, only the combined intervention is highly significant. Again, the overall knowledge score is split into principles and how-to knowledge. The results for principles knowledge look very similar to those from the full sample. Only the combined intervention significantly increases principles knowledge in the non-FREG sam- ple (cp. Columns (5) and (6) of Table 6). In the FREG sample, both T1 and T2 have positive and significant effects on principles knowledge, with some indication of a statistically stronger effect of the extension-plus-video treatment (cp. Columns (7) and (8)). Regarding knowledge on how to implement ISFM, treatment effects are positive and sig- nificant in both subsamples. In the case of FREG members, differences in the effect sizes of T1 and T2 are not significant (cp. Columns (11) and (12)). For non-FREG farmers, the effect of the combined treatment seems to increase how-to knowledge significantly stronger than exten- sion-only (cp. Column (9) and (10)). Further analyses reveal that this effect mainly stems from improved knowledge on how to produce compost among this group of farmers. This is fairly surprising, since the video did not convey any information on how to implement any of the practices. Yet, it may be that increased awareness and understanding of why ISFM is beneficial induced further knowledge-seeking processes on the mode of compost production among non- FREG farmers. 22 Table 6 about here 5. Discussion The present results show that the farmer-to-farmer extension approach under study has signifi- cantly increased ISFM knowledge and adoption of individual components and the full package in the treated communities. Moreover, the study finds that ISFM knowledge and adoption – at least of individual practices – also increases among farmers not actively participating in exten- sion activities. This points towards the existence of information spillovers from FREG farmers to their peers, that occur either through active information-sharing or through observation and imitation. These results provide support for the rationale of farmer-to-farmer extension models and contradict previous research finding weak evidence for diffusion effects (Rola, Jamias, and Quizon 2002; Feder, Murgai, and Quizon 2004; Kondylis, Mueller, and Zhu 2017). The study finds robust significant treatment effects on the adoption of the more knowledge- intensive ISFM components compost, line seeding and lime, but not on blended fertilizer and improved seeds. The latter is likely due to the fact that blended fertilizers and improved seeds have been heavily promoted by the public extension system in Ethiopia. As a result, adoption rates of these two practices are rather high in the research area, also in control mws. Further increasing these already high adoption rates may then hinge on relaxing capital constraints (Ambler, de Brauw, and Godlonton 2020), rather than knowledge constraints. Results indicate that overall, the extension approach is particularly effective in the case of unfamiliar (like lime in the research area) or knowledge-intensive technologies (like the production of good-quality compost or the benefits of line seeding). Additional analyses show that the extension treatment has a significantly positive impact on the quality of the compost farmers produce (see Table A1.1 in the Appendix), suggesting that farmers do not only gather more knowledge via the 23 interventions, but are also able to put this knowledge into practice. 24 The causal mediation anal- ysis showed that how-to and – albeit to a smaller extent – principles knowledge are significant underlying drivers of adoption, suggesting that enhancing both types of knowledge – how and why practices work – can foster more widespread adoption of complex system technologies. While there is evidence of information spillovers from FREG farmers to other farmers in the community, we also find that increases in ISFM knowledge and adoption are substantially lower for non-FREG farmers compared to FREG farmers. When it comes to the integrated adoption of all practices on the same plot in particular, the extension-only treatment seems to do little for non-FREG farmers. Furthermore, for non-FREG farmers, knowledge increases through extension-only are modest and mostly limited to gains in how-to knowledge (as op- posed to principles knowledge). This suggests that some information loss occurs in the knowledge transmission process from actively participating farmers to other farmers in the community, so that only some pieces of knowledge are passed on or picked up, likely leading to incomplete adoption. One possible explanation for these findings is based on selective attention theory (Schwarz- stein 2014). ISFM is a complex system technology that requires learning about several individ- ual practices as well as about the importance of applying them jointly. A resource-constrained farmer might not have the time or mental capacity to learn about all aspects, and hence, pay attention in a selective manner (Hanna, Mullainathan, and Schwartzstein 2014; Niu & Ragasa 2018). Farmers might base their decision which components to focus on – consciously or un- consciously – on their beliefs about the yield effects of a certain practice, on their complexity, but also on how feasible they consider the adoption of a technology in their particular context (Maertens, Michelson, and Nourani 2020). 24 This is assessed with a compost quality index ranging from zero to nine, based on farmers’ self-reported information. The index is composed of six questions on the compost production process and three questions on the compost end product (see Online Appendix S2 for details) 24 The video that was shown to farmers in T2 could potentially help to reduce information losses. Findings suggest that for the full sample, there is no strong additional effect of the video on adoption. Yet, for non-FREG farmers, the video has a significant complementary effect on knowledge as well as on adoption, which is particularly pronounced for the integrated adoption of ISFM practices, compared to extension-only. One possible explanation is that the video in- tervention indeed contributed to counterbalance incomplete information transmission by draw- ing farmers’ attention to dimensions of the ISFM technology package they did not notice before, or that are not transmitted via farmer-to-farmer extension at all. The data shows, for instance, that for the group of non-FREG farmers, the additional video intervention triggered gains in knowledge on the process of compost production, even though no explicit how-to messages were conveyed in the video. A possible explanation is that the video increased awareness of the benefits associated with this complex practice and thereby spurred how-to knowledge seeking processes (cp. Maertens, Michelson, and Nourani 2020). This interpretation is in line with se- lective attention theory; additional information is especially important for more complex prac- tices that farmers may disregard if not sufficiently convinced of their benefits. Another potential explanation is that the video intervention increases knowledge and adop- tion among non-FREGs by providing a platform for exchange and communication, rather than through the video content itself. If this is the case, one would expect higher adoption rates among non-FREG farmers not only with respect to the five quickwin practices promoted in the video, but also more generally with respect to other ISFM-related practices. Analyzing adoption of four individual practices, i.e., urea, green manure, intercropping, and planting of forage crops, which are promoted by the ISFM+ project, but not featured in the video, indicates some evidence for such a ‘gathering effect’.25 This suggests that non-FREG farmers in T2 mws may 25 We find that (1) treatment effects of T1 and T2 are positive and significant indicating that adoption of these ISFM-related practices is higher in treatment than in control mws; (2) the same applies to FREG farmers, but (3) among non-FREG farmers, only T2 has robust significant effects on adoption for two of the four practices (see Tables S4.14 to S4.16 in Online Appendix S4). 25 not only have learned about the video content, but also about other ISFM-related practices from discussions with FREG members and extension agents present during the screening event.26 These results point towards the existence of information spillovers from FREG farmers to their peers, and thus provide support for the rationale of farmer-to-farmer extension models to promote even complex agricultural technologies. Yet, to judge the cost-effectiveness of the approach, studies on the actual benefits derived from technology diffusion are needed in order to weigh the project’s costs against its benefits. Arguably, in the present study area, it may be too early to conduct such assessment, given that the endline data only covers short-term effects measured during the second season of project interventions. An analysis of treatment effects conditional on adoption of the integrated ISFM package shows that adopters in treatment mws have significantly higher yields than their closest matches in control mws. 27 Results on net crop income are not as strong, but point in the same direction. These results are robust when exclud- ing model farmers from the analysis, who benefited from free input provision (cp. Table A1.2 in the Appendix), and are also supported by previous survey-based studies that have docu- mented positive productivity and welfare effects of ISFM adoption (Adolwa, Schwarze, and Buerkert 2019; Hörner and Wollni 2021). However, the effectiveness of the extension approach hinges on a combination of both welfare impacts and adoption rates. Analyzing the current data does not reveal significant intent-to-treat effects of the extension interventions on net crop in- comes or yields (cp. Table A1.3 in the Appendix), implying that on the average, the project has not (yet) made farmers in treatment mws better off. This is in line with previous research, that has reported treatment effects of extension on adoption, but not on yields or profits (Cole and Fernando 2020). One important observation in this context is that currently, adopters of the integrated ISFM package in the research area apply it to only 30% of their agricultural land 26 We thank an anonymous reviewer for suggesting this alternative interpretation and the respective tests. 27 Given that adoption is a choice variable and therefore subject to selection bias, the same matching approach as above is used to match farmers in treatment mws who adopted the integrated ISFM package with their closest matches from control mws. 26 planted with main cereal crops, on the average. Accordingly, further increasing the intensity of adoption, preferably on the basis of successful experimentation, remains an important goal. 6. Conclusion The study shows positive effects of a farmer-to-farmer extension intervention on the adoption of a complex agricultural technology. The effects are stronger for direct beneficiaries of the intervention, but do also trickle down to other community members, suggesting the existence of spillover effects. While this is promising, more research is needed that evaluates the longer- term impacts on adoption intensity, yields and welfare outcomes to assess the cost-effectiveness of the extension approach. In this context, it is important to gain a better understanding of how to further strengthen information exchange between direct beneficiaries of extension and other farmers in the com- munity. A growing body of literature is dedicated to the investigation of mechanisms that coun- teract incomplete diffusion. This study showed that more complex information, e.g., on the importance of applying practices jointly and on the process of compost production, was only transmitted to the wider community if they were additionally exposed to a video intervention. On the one hand, the video provides an opportunity to highlight the more complex aspects of the technology package that may otherwise be ignored or get lost along the information chain. On the other hand, the video could potentially also reach a larger share of the population in the community and provide them with a platform for information exchange with direct beneficiar- ies. While previous studies have mostly looked at mechanisms that incentivize model farmers to increase knowledge dissemination, findings of this study suggest that interventions that acti- vate knowledge-seeking processes among non-model farmers can be effective and represent a promising avenue for further research. 27 Tables Table 1. Baseline descriptive statistics and balance between treatment groups. T1 T2 C Overall (extension T1 - T2 T1 - C T2 - C (extension) + video) (control) Panel A: Household characteristics Age HH head (in years) 47.03 46.27 47.32 47.22 -1.05 -0.95 0.10 [14.61] [14.61] [14.54] [14.64] (1.02) (0.85) (0.94) Gender HH head (1=male) 0.85 0.86 0.84 0.85 0.02 0.01 -0.00 (0.02) (0.02) (0.02) Education HH head 2.15 2.19 2.42 2.03 -0.23 0.15 0.39 (grades compl.) [3.36] [3.36] [3.61] [3.24] (0.37) (0.32) (0.28) Non-farm family business 0.19 0.21 0.18 0.18 0.03 0.02 -0.00 (1=yes) (0.04) (0.03) (0.03) Off-farm wage employment 0.19 0.18 0.23 0.18 -0.05 -0.00 0.05 (1=yes) (0.03) (0.02) (0.03) No. of HH members 3.06 3.08 3.11 3.03 -0.03 0.05 0.08 over age 14 [1.31] [1.34] [1.31] [1.29] (0.11) (0.10) (0.08) No. of organizations involved 4.47 4.53 4.38 4.49 0.15 0.05 -0.1 (0-12) [1.87] [1.91] [1.78] [1.90] (0.21) (0.19) (0.18) Basic assets score (0-4) 1.84 1.79 1.91 1.83 -0.13 -0.04 0.09 [0.89] [0.84] [0.90] [0.90] (0.10) (0.09) (0.09) No. of TLU owned 3.39 3.26 3.48 3.42 -0.22 -0.16 0.06 [2.83] [2.61] [2.93] [2.87] (0.29) (0.26) (0.25) Radio owned (1=yes) 0.29 0.27 0.29 0.30 -0.02 -0.03 -0.01 (0.03) (0.03) (0.03) Cellphone owned (1=yes) 0.52 0.53 0.53 0.51 -0.00 0.02 0.02 (0.04) (0.03) (0.04) Contracted any credit (1=yes) 0.34 0.38 0.35 0.32 0.03 0.06** 0.03 (0.04) (0.03) (0.06) Eligible for formal credit 0.73 0.71 0.73 0.75 -0.02 -0.04 -0.02 (1=yes) (0.05) (0.04) (0.05) Food insecurity score (0-12) 2.34 2.29 2.08 2.46 0.21 -0.18 -0.39 [2.99] [3.05] [2.71] [3.07] (0.31) (0.27) (0.26) Walking dist. to nearest FTC 33.30 33.34 32.94 33.42 0.40 -0.09 -0.48 (min) [25.55] [26.21] [25.31] [25.40] (3.84) (3.03) (3.29) Walking dist. to nearest 27.36 25.56 22.34 30.14 3.22 -4.58 -7.80** (all-season) road (min) [29.47] [23.66] [25.98] [32.50] (3.34) (3.02) (3.56) Walking dist. to nearest market 74.17 67.03 67.91 79.65 -0.88 -12.63* -11.74* (min) [48.15] [44.82] [46.86] [49.31] (7.84) (6.96) (6.47) Panel B: Agricultural production characteristics Total land size (in ha) 1.34 1.37 1.40 1.30 -0.03 0.07 0.10 [1.11] [1.16] [1.18] [1.06] (0.18) (0.15) (0.14) Grows main crop (1=yes) 0.94 0.94 0.95 0.93 -0.01 0.01 0.02* (0.01) (0.02) (0.01) No. of adopted quickwins (0-5) 1.40 1.51 1.53 1.30 -0.02 0.21* 0.23** [0.99] [1.00] [0.97] [0.99] (0.14) (0.13) (0.11) Compost applied (1=yes) 0.36 0.34 0.39 0.37 -0.05 -0.03 0.02 (0.06) (0.05) (0.05) Blended fertilizer applied 0.014 0.009 0.021 0.014 -0.011 -0.004 0.007 28 (1=yes) (0.011) (0.007) (0.011) Improved seeds used (1=yes) 0.57 0.64 0.59 0.53 0.04 0.11** 0.07 (0.06) (0.05) (0.05) Plants crops usually in lines 0.45 0.52 0.52 0.39 -0.00 0.13* 0.13* (1=yes) (0.09) (0.07) (0.07) Lime applied (1=yes) 0.008 0.009 0.009 0.007 0.000 0.002 0.003 (0.007) (0.005) (0.005) DAP applied (1=yes) 0.70 0.76 0.74 0.66 0.02 0.09* 0.07 (0.05) (0.05) (0.05) Used irrigation (1=yes) 0.19 0.17 0.19 0.20 -0.03 -0.03 -0.00 (0.05) (0.04) (0.04) Last season was bad (1=yes) 0.48 0.46 0.45 0.51 0.02 -0.04 -0.06 (0.08) (0.07) (0.07) No. of times talked to 5.53 5.76 6.42 5.07 -0.66 0.69 1.35 extension agent in last year [10.97] [11.06] [14.35] [9.20] (1.23) (0.90) (1.06) Attended agric. training 0.27 0.30 0.34 0.23 -0.03 0.08** 0.11*** in last year (1=yes) (0.05) (0.04) (0.04) Panel C: Community-level characteristics Mean annual temperature 2017 20.56 20.45 20.40 20.68 0.05 -0.23 -0.28 (°C) [4.10] [4.21] [4.24] [4.00] (1.00) (0.82) (0.83) Mean annual rainfall 2017 1108.84 1140.53 1140.11 1083.12 0.42 57.40 56.98 (mm) [396.23] [381.88] [380.26] [406.66] (90.84) (77.59) (77.13) Distance to Woreda capital 14.62 13.66 15.52 14.65 -1.86 -0.99 0.87 (km) [15.42] [16.31] [13.69] [15.69] (3.67) (3.27) (2.85) N 2,382 539 532 1,311 1,071 1,850 1,843 Source: Authors’ analyses based on baseline survey data. Note: HH stands for household. Basic asset score comprises the following: HH has modern roof, improved stove, modern lighting, toilet facility. TLU stands for tropical livestock unit. Calculation of food insecurity score based on self-experienced events of food insecurity, based on Household Food Insecurity Access Scale (HFIAS). FTC stands for farmer training center. Main crops are maize, wheat, teff, barley, sorghum. Temperature and rainfall assessed at endline. For means, standard deviations in brackets; for mean comparisons, robust standard errors in parentheses, clustered at the mws level. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. 29 Table 2. ITT effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package. Number of ISFM technologies adopted Integrated adoption OLS Poisson Oprobit of full ISFM package (1) (2) (3) (4) (5) (6) (7) (8) T1 (extension) 0.683*** 0.448*** 0.688*** 0.468*** 0.542*** 0.529*** 0.103** 0.084*** (0.184) (0.080) (0.178) (0.088) (0.136) (0.085) (0.043) (0.025) p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.016 0.001 T2 (extension + video) 0.840*** 0.569*** 0.822*** 0.551*** 0.671*** 0.671*** 0.137*** 0.109*** (0.174) (0.079) (0.164) (0.088) (0.134) (0.086) (0.043) (0.024) p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.002 0.000 Test T1 >= T2 (p-value) 0.216 0.087 0.216 0.177 0.206 0.080 0.242 0.185 Endline control mean 2.222 0.152 Additional controls No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.071 0.531 0.016 0.122 0.025 0.223 0.023 0.271 Observations 2,382 2,382 2,382 2,382 2,382 2,382 2,160 2,160 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson and probit models (Columns (3), (4), (7) and (8)) show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 5. Integrated adoption of full ISFM package is a dummy variable. Additional baseline control variables at household level are age, gender and education (in completed years) of HH head; whether HH participated in off-farm work or a non- farm business activity; number of HH members above age 14; walking distances to nearest farmer training center, paved road and market (in minutes); number of local organizations involved; use of irrigation, total land size in ha, tropical livestock units (TLU), a basic assets score, a food insecurity score, whether HH is eligible for formal credit and has contracted a credit in the last farming season; whether HH had a below-average preceding farming season; number of times HH had contact with an extension agent and whether HH has participated in agricultural training; whether HH grew main crops (teff, wheat, barley, maize, sorghum) and used any kind of inorganic fertilizer. Community level covariates are rainfall, temperature, and distance to Woreda capital (in km). Two region dummies for Oromia and Amhara included. One-sided equality tests of T1 and T2 are F-tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. 30 Table 3. ITT effects on adoption of individual ISFM components. Adopted Adopted Adopted Adopted Adopted compost blended fertilizer improved seeds line seeding lime (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) T1 (extension) 0.150*** 0.143*** 0.097 0.039 0.129** 0.065* 0.157*** 0.091*** 0.222*** 0.214*** (0.057) (0.028) (0.062) (0.042) (0.058) (0.036) (0.057) (0.029) (0.042) (0.029) Unadjusted p-value 0.008 0.000 0.118 0.348 0.027 0.070 0.006 0.002 0.000 0.000 Adjusted p-value 0.025 0.000 0.334 0.749 0.092 0.225 0.022 0.007 0.000 0.000 T2 (extension + video) 0.219*** 0.192*** 0.111** 0.046 0.129** 0.067* 0.204*** 0.112*** 0.254*** 0.239*** (0.054) (0.025) (0.055) (0.037) (0.058) (0.040) (0.057) (0.030) (0.042) (0.028) Unadjusted p-value 0.000 0.000 0.043 0.218 0.027 0.093 0.000 0.000 0.000 0.000 Adjusted p-value 0.000 0.000 0.132 0.539 0.092 0.291 0.000 0.000 0.000 0.000 Robust to Adjustment? Yes No No Yes Yes Test T1 >= T2 (unadjusted p-value) 0.141 0.058 0.421 0.442 0.501 0.467 0.245 0.250 0.226 0.194 Test T1 >= T2 (adjusted p-value) 0.382 0.173 0.829 0.848 0.914 0.891 0.638 0.647 0.548 0.488 Endline control mean 0.405 0.596 0.574 0.624 0.040 Additional controls No Yes No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.027 0.361 0.009 0.220 0.014 0.257 0.033 0.388 0.137 0.336 Observations 2,382 2,382 2,382 2,382 2,382 2,382 2,382 2,382 1,464 1,464 Source: Authors’ analyses based on baseline and endline survey data. Note: Average marginal effects (AME) of binary probit models. For lime, Tigray is excluded since it is not recommended in this region and adoption is zero. Additional control variables identical to those listed in notes of Table 2. Bonferroni-adjusted p-values taking into account correlations between outcomes. One-sided equality tests of T1 and T2 are Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 31 Table 4. ITT effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, FREG- and non-FREG samples separately. Number of ISFM technologies adopted Integrated adoption of full package OLS Poisson Non-FREG sample FREG sample Non-FREG sample FREG sample Non-FREG sample FREG sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.322* 0.278*** 1.257*** 1.232*** 0.329* 0.293*** 1.260*** 1.228*** 0.030 0.034 0.273*** 0.280*** (0.193) (0.080) (0.200) (0.123) (0.194) (0.088) (0.193) (0.128) (0.048) (0.024) (0.058) (0.048) p-value 0.097 0.001 0.000 0.000 0.090 0.001 0.000 0.000 0.528 0.161 0.000 0.000 T2 (extension + video) 0.550*** 0.483*** 1.140*** 1.117*** 0.540*** 0.478*** 1.161*** 1.131*** 0.079* 0.084*** 0.231*** 0.231*** (0.181) (0.082) (0.184) (0.137) (0.173) (0.086) (0.183) (0.147) (0.046) (0.025) (0.066) (0.043) p-value 0.003 0.000 0.000 0.000 0.002 0.000 0.000 0.000 0.082 0.001 0.001 0.000 Test T1 >= T2 (p-value) 0.145 0.009 0.711 0.803 0.144 0.014 0.712 0.775 0.176 0.040 0.719 0.795 Endline control mean 2.444 2.775 2.444 2.775 0.168 0.225 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.031 0.496 0.224 0.570 0.006 0.094 0.030 0.077 0.007 0.285 0.059 0.299 Observations 1,606 1,606 400 400 1,606 1,606 400 400 1,606 1,606 400 400 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson and probit models (columns (5) to (12)) show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 5. Integrated adoption of full ISFM package is a dummy variable. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws, and their matched controls. Additional control variables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are F-tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 32 Table 5. ITT effects on different knowledge outcomes. Overall knowledge Principles knowledge How-to knowledge (1) (2) (3) (4) (5) (6) T1 (extension) 0.050*** 0.036*** 0.030* 0.020 0.068*** 0.052*** (0.015) (0.013) (0.016) (0.014) (0.015) (0.013) p-value 0.001 0.006 0.062 0.152 0.000 0.000 T2 (extension + video) 0.082*** 0.068*** 0.063*** 0.054*** 0.091*** 0.073*** (0.013) (0.011) (0.014) (0.011) (0.016) (0.013) p-value 0.000 0.000 0.000 0.000 0.000 0.000 Test T1 >= T2 (p-value) 0.014 0.008 0.018 0.007 0.102 0.103 Endline control mean 0.448 0.522 0.382 Additional controls No Yes No Yes No Yes R-squared 0.034 0.219 0.012 0.155 0.048 0.221 Observations 2,334 2,334 2,334 2,334 2,334 2,334 Source: Authors’ analyses based on baseline and endline survey data. Note: OLS regressions of household heads' knowledge scores, ranging from 0 to 1, and calculated based on the number of correct answers relative to the total number of questions in a respective domain. Additional control variables are age, gender, education (in completed years), whether respondent had access to off-farm work or a non- farm family business, whether hh adopted the ISFM quickwin package at baseline, whether hh has a cell phone and radio, number of times hh had contact with an extension agent, whether it has participated in agricultural training, number of local organizations involved, and walking distance to nearest farmer training center. Two region dum- mies for Oromia and Amhara included. One-sided equality tests of T1 and T2 are F-tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 33 Table 6. ITT effects on different knowledge outcomes, FREG- and non-FREG samples separately. Overall knowledge Principles knowledge How-to knowledge Non-FREG sample FREG sample Non-FREG sample FREG sample Non-FREG sample FREG sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.013 0.014 0.129*** 0.124*** -0.004 0.001 0.089*** 0.079*** 0.027** 0.027** 0.178*** 0.177*** (0.014) (0.013) (0.025) (0.021) (0.017) (0.016) (0.029) (0.026) (0.013) (0.012) (0.027) (0.024) p-value 0.351 0.284 0.000 0.000 0.803 0.974 0.003 0.002 0.049 0.025 0.000 0.000 T2 (extension + video) 0.049*** 0.049*** 0.149*** 0.152*** 0.035** 0.042*** 0.129*** 0.126*** 0.052*** 0.052*** 0.160*** 0.166*** (0.013) (0.012) (0.021) (0.020) (0.016) (0.014) (0.027) (0.024) (0.014) (0.013) (0.020) (0.019) p-value 0.000 0.000 0.000 0.000 0.030 0.004 0.000 0.000 0.000 0.000 0.000 0.000 Test T1 >= T2 (p-value) 0.006 0.005 0.201 0.099 0.014 0.006 0.095 0.049 0.048 0.046 0.750 0.689 Endline control mean 0.464 0.510 0.541 0.578 0.397 0.440 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes R-squared 0.013 0.136 0.148 0.252 0.005 0.102 0.071 0.178 0.017 0.134 0.203 0.307 Observations 1,540 1,540 390 390 1,540 1,540 390 390 1,540 1,540 390 390 Source: Authors’ analyses based on baseline and endline survey data. Note: OLS regressions of household heads' knowledge scores, ranging from 0 to 1, and calculated based on the number of correct answers relative to the total number of questions in a respective domain. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws, and their matched controls. Additional control variables identical to those listed in notes of Table 5. One-sided equality tests of T1 and T2 are F-tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 34 Appendix A1: Additional analyses Table A1.1. ITT effects on compost quality. Compost quality (1) (2) T1 (extension) 0.547** 0.423*** (0.216) (0.126) p-value 0.013 0.001 T2 (extension + video) 0.614*** 0.443*** (0.211) (0.124) p-value 0.004 0.001 Test T1 >= T2 (p-value) 0.393 0.444 Endline control mean 4.365 Additional controls No Yes (Pseudo) R-squared 0.030 0.277 Observations 1,127 1,127 Source: Authors’ analyses based on baseline and endline survey data. Note: Subsample of compost producers only. Compost quality is an index ranging from 0 to 9. One-sided equality tests of T1 and T2 are F- tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 35 Table A1.2. ITT effects on crop productivity and net crop income, including adopters in treatment mws and their closest matches in control mws. Crop productivity Net crop income in kg/ha in ETB/ha Matched sample of adopters in treatment mws Including model farmers W/o model farmers Including model farmers W/o model farmers (1) (2) (3) (4) (5) (6) (7) (8) T 455.61** 369.41*** 467.03** 410.30*** 2,402.72 2,041.35* 2,611.61 2,179.20* (extension and (183.91) (122.10) (198.19) (131.38) (1,710.80) (1,100.85) (1,779.81) (1,152.31) extension + video) p-value 0.015 0.003 0.020 0.002 0.163 0.066 0.145 0.061 Endline control mean 2,057.38 2,064.98 16,770.18 16,608.65 Additional controls No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.035 0.389 0.034 0.385 0.012 0.389 0.014 0.392 Observations 552 552 454 454 552 552 454 454 Source: Authors’ analyses based on baseline and endline survey data. Note: Outcomes refer to productivity and income obtained from the main crops maize, wheat, teff, barley and sorghum. Sample includes treatment farmers adopting the full integrated ISFM package on at least one plot, and their matched controls. T subsumes T1 and T2. Additional control variables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are F-tests. Robust standard errors in parentheses, clustered at the mws level. Signifi- cance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 36 Table A1.3. ITT effects on crop productivity and net crop income. Crop productivity Net crop income in kg/ha in ETB/ha (1) (2) (3) (4) T 172.320 89.55 161.82 -2.37 (extension and (129.40) (78.93) (1,110.26) (703.02) extension + video) p-value 0.185 0.258 0.884 0.997 Endline control mean 1,832.24 14,875.23 Additional controls No Yes No Yes (Pseudo) R-squared 0.005 0.273 0.000 0.284 Observations 2,160 2,160 2,160 2,160 Source: Authors’ analyses based on baseline and endline survey data. Note: Outcomes refer to productivity and income obtained from the main crops maize, wheat, teff, barley and sorghum. T subsumes T1 and T2. Additional control variables identical to those listed in notes of Table 2. Robust standard errors in parentheses, clus- tered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 37 References Adolwa, Ivan Solomon, Stefan Schwarze, and Andreas Buerkert. 2019. “Impacts of Integrated Soil Fertility Management on Yield and Household Income: The Case of Tamale (Ghana) and Kakamega (Kenya).” Ecological Economics 161 (July): 186–92. https://doi.org/10.1016/j.ecolecon.2019.03.023. Aker, Jenny C, Rachid Boumnijel, Amanda McClelland, and Niall Tierney. 2016. “Payment Mechanisms and Antipoverty Programs: Evidence from a Mobile Money Cash Transfer Ex- periment in Niger.” Economic Development and Cultural Change 65 (1): 1–37. Ambler, Kate, Alan de Brauw, and Susan Godlonton. 2020. “Cash Transfers and Management Advice for Agriculture: Evidence from Senegal.” World Bank Economic Review 34(3): 597– 617. doi: 10.1093/wber/lhz005 Anderson, Jock R., and Gershon Feder. 2007. “Agricultural Extension.” In Handbook of Agri- cultural Economics, edited by Robert E. Evenson and Prabhu Pingali, 3:2343 –78. Amster- dam and New York: North Holland. https://doi.org/10.1016/S1574-0072(06)03044-1. Barrett, Christopher B., and Leah E. M. Bevis. 2015. “The Self-Reinforcing Feedback between Low Soil Fertility and Chronic Poverty.” Nature Geoscience 8 (12): 907–12. https://doi.org/10.1038/ngeo2591. Beaman, Lori, Ariel BenYishay, Jere my Magruder, and Ahmed Mushfiq Mobarak. 2018. “Can Network Theory-Based Targeting Increase Technology Adoption?” NBER Working Paper 24912. Cambridge, MA: National Bureau of Economic Research. BenYishay, Ariel, and A. Mushfiq Mobarak. 2019. “Social Learning and Incentives for Exper- imentation and Communication.” Review of Economic Studies 86: 976–1009. https://doi.org/10.1093/restud/rdy039. 38 Berhane, G., Ragasa, C., Abate, G. T., & Assefa, T. W. (2018). “The State of Agricultural Extension Services in Ethiopia and their Contribution to Agricultural Productivity. ” Addis Abeba and Washington D. C.: International Food Policy Research Institute (IFPRI). Bernard, Tanguy, Simrin Makhija, Kate Orkin, Alemayehu Seyoum Taffesse, and David J. Spielman. 2016. “Video -Based Agricultural Extension - Analysis of a Pilot Project in Ethio- pia.” Project Note. Addis Abeba and Washington D. C.: International Food Policy Research Institute (IFPRI). Cameron, A. Colin, and Pravin K. Trivedi. 2005. Microeconometrics: Methods and Applica- tions. Cambridge: Cambridge University Press. Cole, Shawn A., and A. Nilesh Fernando . 2020. “'Mobile'izing Agricultural Advice: Technol- ogy Adoption, Diffusion and Sustainability.” Harvard Business School/University of Notre Dame. Davis, K., E. Nkonya, E. Kato, D.A. Mekonnen, M. Odendo, R. Miiro, and J. Nkuba. 2012. “Impact of Farmer Field Schools on Agricultural Productivity and Poverty in East Africa.” World Development 40 (2): 402–13. https://doi.org/10.1016/j.worlddev.2011.05.019. De Brauw, Alan, Patrick Eozenou, Daniel O Gilligan, Christine Hotz, Neha Kumar, and J V Meenakshi. 2018. “Biofortification, Crop Adoption and Health Information: Impact Path- ways in Mozambique and Uganda.” American Journal of Agricultural Economics 100 (3): 906–30. https://doi.org/10.1093/ajae/aay005. Feder, Gershon, Rinku Murgai, and Jaime B. Quizon. 2004. “The Acquisition and Diffusion of Knowledge: The Case of Pest Management Training in Farmer Field Schools, Indonesia.” Journal of Agricultural Economics 55 (2): 221–243. Fisher, Monica, Stein T. Holden, Christian Thierfelder, and Samson P. Katengeza. 2018. “Awareness and Adoption of Conservation Agriculture in Malawi: What Difference Can 39 Farmer-to-Farmer Extension Make?” International Journal of Agricultural Sustainability 16 (3): 310–25. https://doi.org/10.1080/14735903.2018.1472411. Foster, Andrew D., and Mark R. Rosenzweig. 1995. “Learning by Doing and Learning from Others: Human Capital and Technical Change in Agriculture.” Journal of Political Economy 103 (6): 1176–1209. Frölich, Markus, and Stefan Sperlich. 2019. Impact Evaluation: Treatment Effects and Causal Analysis. Cambridge: Cambridge University Press. Gautam, Madhur. 2000. “Agricultural Extension - The Kenya Experience. An Impact Evalua- tion.” Washington D.C.: The World Bank. http://documents.worldbank.org/cu- rated/en/972111468758711518/pdf/multi0page.pdf. Godtland, Erin M., Elisabeth Sadoulet, Alain de Janvry, Rinku Murgai, and Oscar Ortiz. 2004. “The Impact of Farmer Field Schools on Knowledge and Productivity: A Study of Potato Farmers in the Peruvian Andes.” Economic Development and Cultural Change 53 (1): 63– 92. Hanna, Rema, Sendhil Mullainathan, and Joshua Schwartzstein. 2014. “Learning through No- ticing: Theory and Evidence from a Field Experiment.” Quarterly Journal of Economics 129 (3): 1311–53. https://doi.org/10.1093/qje/qju015. Hörner, Denise, and Meike Wollni. 2021. “Integrated soil fertility management and household welfare in Ethiopia.” Food Policy 100. https://doi.org/10.1016/j.foodpol.2020.102022 Hörner, Denise, Adrien Bouguen, Markus Frölich, and Meike Wollni. 2019. “The Effects of Decentralized and Video-based Extension on the Adoption of Integrated Soil Fertility Man- agement – Experimental Evidence from Ethiopia. ” AEA RCT Registry. July 04. https://doi.org/10.1257/rct.4406-2.0. 40 Jayne, Thomas S., Sieglinde Snapp, Frank Place, and Nicholas Sitko. 2019. “Sustainable Agri- cultural Intensification in an Era of Rural Transformation in Africa.” Global Food Security 20 (March): 105–13. https://doi.org/10.1016/j.gfs.2019.01.008. Kondylis, Florence, Valerie Mueller, and Siyao J. Zhu. 2017. “Seeing Is Believing? Evidence from an Extension Network Experiment.” Journal of Development Economics 125: 1–20. https://doi.org/10.1016/j.jdeveco.2016.10.004. Kondylis, Florence, Valerie Mueller, and Siyao J. Zhu. 2015. “Measuring Agricultural Knowledge and Adoption.” Agricultural Economics 46 (3): 449–62. https://doi.org/10.1111/agec.12173. Maertens, Annemie, Hope Michelson, and Vesall Nourani. 2020. “How do Farmers Learn from Extension Services? Evidence from Malawi. ” American Journal of Agricultural Economics 103(2): 569–595; doi:10.1111/ajae.12135 Magruder, Jeremy R. 2018. “An Assessment of Experimental Evidence on Agricultural Tech- nology Adoption in Developing Countries.” Annual Review of Resource Economics 10 (1): 299–316. https://doi.org/10.1146/annurev-resource-100517-023202. McKenzie, David. 2012a. “Beyond Baseline and Follow-up: The Case for More T in Experi- ments.” Journal of Development Economics 99 (2): 210–21. https://doi.org/10.1016/j.jdeveco.2012.01.002. ———. 2012b. “Tools of the Trade: A Quick Adjustment for Multiple Hypothesis Testing.” World Bank Development Impact Blog. 2012. https://blogs.worldbank.org/impactevalua- tions/tools-of-the-trade-a-quick-adjustment-for-multiple-hypothesis-testing (accessed April 11, 2021). Nakano, Yuko, Takuji W. Tsusaka, Takeshi Aida, and Valerien O. Pede. 2018. “Is Farmer -to- Farmer Extension Effective? The Impact of Training on Technology Adoption and Rice 41 Farming Productivity in Tanzania.” World Development 105 (May): 336–51. https://doi.org/10.1016/j.worlddev.2017.12.013. Niu, Chiyu, and Catherine Ragasa. 2018. “Selective Attention and Information Loss in the Lab- to-Farm Knowledge Chain: The Case of Malawian Agricultural Extension Programs.” Agri- cultural Systems 165: 147–63. https://doi.org/10.1016/j.agsy.2018.06.003. Noltze, Martin, Stefan Schwarze, and Matin Qaim. 2012. “Understanding the Adoption of Sys- tem Technologies in Smallholder Agriculture: The System of Rice Intensification (SRI) in Timor Leste.” Agricultural Systems 108: 64–73. https://doi.org/10.1016/j.agsy.2012.01.003. Pamuk, Haki, Erwin Bulte, and Adewale A. Adekunle. 2014. “Do Decentralized Innovation Systems Promote Agricu ltural Technology Adoption? Experimental Evidence from Africa.” Food Policy 44: 227–36. https://doi.org/10.1016/j.foodpol.2013.09.015. Place, Frank, Christopher B. Barrett, H. Ade Freeman, Joshua J. Ramisch, and Bernard Van- lauwe. 2003. “Prospects for Integrated Soil Fertility Management Using Organic and Inor- ganic Inputs: Evidence from Smallholder African Agricultural Systems.” Food Policy 28 (4): 365–78. https://doi.org/10.1016/j.foodpol.2003.08.009. Rogers, Everett M. 1995. Diffusion of Innovations. 4th ed. New York: Free Press. Rola, Agnes, Serlie Jamias, and Jaime Quizon. 2002. “Do Farmer Field School Graduates Re- tain and Share What They Learn? An Investigation in Iloilo, Philippines.” Journal of Inter- national Agricultural and Extension Education 9 (1): 65–76. https://doi.org/10.5191/jiaee.2002.09108. Sankoh, A. J., M. F. Huque, and S. D. Dubey. 1 997. “Some Comments on Frequently Used Multiple Endpoint Adjustment Methods in Clinical Trials.” Statistics in Medicine 16 (22): 2529–42. 42 Schwartzstein, Joshua. 2014. “Selective Attention and Learning.” Journal of the European Eco- nomic Association 12 (6): 1423–52. https://doi.org/10.1111/jeea.12104. Sheahan, Megan, and Christopher B. Barrett. 2017. “Ten Str iking Facts about Agricultural In- put Use in Sub-Saharan Africa.” Food Policy 67: 12–25. https://doi.org/10.1016/j.food- pol.2016.09.010. Shikuku, Kelvin Mashisia. 2019. “Information Exchange Links, Knowledge Exposure, and Adoption of Agricultural Technologies in Northern Uganda.” World Development 115: 94– 106. https://doi.org/10.1016/j.worlddev.2018.11.012. Shikuku, Kelvin Mashisia, Janneke Pieters, Erwin Bulte, and Peter Läderach. 2019. “Incentives and the Diffusion of Agricultural Knowledge: Experimental Evidence from Northern Uganda.” American Journal of Agricultural Economics 101 (4): 1164–80. https://doi.org/10.1093/ajae/aaz010. Swanson, Burton E. 2008. Global Review of Good Agricultural Extension and Advisory Service Practices. Rome: Food and Agriculture Organization of the United Nations (FAO). Takahashi, Kazushi, Yukichi Mano, and Keijiro Otsuka. 2019. “Learning from Experts and Peer Farmers about Rice Production: Experimental Evidence from Cote d’Ivoire.” World Development 122 (October): 157–69. https://doi.org/10.1016/j.worlddev.2019.05.004. Takahashi, Kazushi, Rie Muraoka, and Keijiro Otsuka. 2019. “Technology Adoption, Impact, and Extension in Developing Countries’ Agriculture: A Review of the Recent Literature.” Agricultural Economics, November, agec.12539. https://doi.org/10.1111/agec.12539. Tripp, Robert, Mahinda Wijeratne, and V. Hiroshini Piyadasa. 2005. “What Should We Expect from Farmer Field Schools? A Sri Lanka Case Study.” World Development 33 (10): 1705– 20. https://doi.org/10.1016/j.worlddev.2005.04.012. 43 Van Campenhout, Bjorn, David J. Spielman, and Els Lecoutere. 2020. “Information and Com- munication Technologies to Provide Agricultural Advice to Smallholder Farmers: Experi- mental Evidence from Uganda.” American Journal of Agricultural Economics 103(1): 317– 337. doi:10.1002/ajae.12089 Van den Berg, Henk, and Janice Jiggins. 2007. “Investing in Farmers - The Impacts of Farmer Field Schools in Relation to Integrated Pest Management.” World Development 35 (4): 663– 86. https://doi.org/10.1016/j.worlddev.2006.05.004. Vanlauwe, B., K. Descheemaeker, K. E. Giller, J. Huising, R. Merckx, G. Nziguheba, J. Wendt, and S. Zingore. 2015. “Integrated Soil Fertility Management in Sub-Saharan Africa: Unrav- elling Local Adaptation.” SOIL 1 (1): 491–508. https://doi.org/10.5194/soil-1-491-2015. Ward, Patrick S., Andrew R. Bell, Klaus Droppelmann, and Tim G. Benton. 2018. “Early Adop- tion of Conservation Agriculture Practices: Understanding Partial Compliance in Programs with Multiple Adoption Decisions.” Land Use Policy 70 (January): 27–37. https://doi.org/10.1016/j.landusepol.2017.10.001. 44 Supplementary Online Appendix to “Knowledge and adoption of complex agricultural technologies – Evidence from an extension experiment – ” by Denise Hörner, Adrien Bouguen, Markus Frölich, and Meike Wollni 1 Online Appendix S1: Demonstration plot pictures Figure S1.1. ISFM demonstration plot for maize, next to traditional practices. Source: GIZ-ISFM+ project Ethiopia. Figure S1.1. ISFM demonstration plot for wheat, next to traditional practices. Source: GIZ-ISFM+ project Ethiopia. 2 Figure S1.2. ISFM demonstration plot for teff, next to traditional practices. Source: GIZ-ISFM+ project Ethiopia. 3 Online Appendix S2: Compost quality index Compost production process, comprises six components: (1) Input mix: considered appropriate if farmers use at least three different materials to produce their compost, of which at least one is rich in nitrogen and one in carbon (2) No. of times compost was turned: correct if compost was turned three times (3) Days waited until compost was turned for the first time: correct is 30 day (acceptable range 25 to 35 days) (4) Days until compost was finished: correct is 90 days (acceptable range 75 to 120 days) (5) Stick or tube used for aeration of compost pit/heap? correct if yes (6) Compost pit/heap covered? correct if yes Compost end product, comprises three components: (1) Compost color: correct if dark brown or black (2) Compost odor: correct if described as good or neutral smell (on a scale from 1=very bad smell to 4=good smell (like good quality soil)) (3) Compost texture: correct if almost or fully decomposed (on a scale from 1=not at all decomposed to 5=fully decomposed) 4 Online Appendix S3: Knowledge exam Known by memory 1.) What are the most important components of integrated soil fertility management? (open question) Known by name 2.) Which of the following technologies have you heard of before this interview? (list of several ISFM technologies read out) How-to knowledge 3.) Imagine you buy improved seeds for wheat. For how many cropping seasons could you re- use them until you should purchase new ones? Up to four cropping seasons (correct) Five to eight cropping seasons I can use them endlessly, no need to purchase again Don’t know 4.) What are the three most important ingredients if you want to produce good-quality com- post? (open question; correct if mentions at least one nitrogen- and one carbon-rich material) 5.) What is the optimal sequence of layers to produce improved compost? (choose the correct out of three pictures) 6.) In order to produce good-quality compost, how many days should you wait at least until you turn the material? (open question; correct: 30; acceptable range 25 to 35) 7.) In order to produce good-quality compost, how many times should you turn the materials in the pit or heap until the composting is finished? (open question; correct: 3) 8.) If you seed maize in lines, how wide should the distance between lines usually be? (open question, assessed with measurement tape; correct: 75 to 80 cm; acceptable range: 65 to 90 cm) 9.) If you seed faba beans in lines, how wide should the distance between lines usually be? (open question, assessed with measurement tape; correct: 30 to 40 cm; acceptable range: 25 to 45 cm) Principles knowledge 10.) For which purpose/benefit should you use improved seeds? (open question; correct if mentions at least two correct points, i.e. one beyond "higher crop yield") 11.) What are the major advantages of blended fertilizer (NPS+/NPK+) over DAP fertilizer? Which statements are correct? K11_1.) Blended fertilizer contains a greater number of nutrients than DAP. (correct) K11_2.) Nutrient supply is better balanced in blended fertilizer than in DAP. (correct) K11_3.) Blended fertilizer directly improves soil structure. K11_4.) Blended fertilizer is more suitable for your location's soil type than DAP. (correct) K11_5.) Blended fertilizer controls weeds and pathogens. 12.) Why is it important to use compost/organic fertilizer? (open question; correct if mentions at least three correct points, i.e. two beyond "higher crop yield") 5 13.) What are the major advantages of line seeding over broadcasting? Which statements are correct? K13_1.) Line seeding reduces the crops’ competition for space, nutrients and water. (correct) K13_2.) Seeding a crop in lines is faster than broadcasting. K13_3.) Line seeding reduces soil acidity. K13_4.) With line seeding usually less seeds are needed. (correct) K13_5.) With line seeding less fertilizer is needed because it can be targeted directly to the roots. (correct) K13_6.) Line seeding makes weeding and harvesting easier. (correct) K13_7.) Line seeding has no advantages. 14.) Why is it important to use inorganic fertilizer and compost at the same time? Which statements are correct? K14_1.) It is always better to apply inorganic fertilizer only. K14_2.) Because the soil needs both organic and inorganic nutrient sources to be healthy and fertile. (correct) K14_3.) Less seeds are needed when using inorganic and organic fertilizer at the same time. 15.) What are the important characteristics of a fertile soil? (open question; correct if mentions at least three correct points) 16.) What are the benefits of applying inorganic fertilizer in lines or by band/microdosing? Which statements are correct? K16_1.) It has no benefits. K16_2.) It is faster than broadcasting. K16_3.) It leads to less leaching of nutrients because they are directly targeted to the roots. (correct) 6 Online Appendix S4: Additional analyses and robustness checks Table S4.1. Further baseline and endline descriptive statistics and balance between treatment groups. T1 T2 C Overall (extension T1-T2 T1-C T2-C (extension) + video) (control) Panel A: Household Characteristics (baseline) HH head married (1=yes) 0.831 0.835 0.831 0.830 0.004 0.005 0.001 (0.375) (0.372) (0.375) (0.376) (0.028) (0.024) (0.023) Literacy HH head (1=yes) 0.561 0.532 0.560 0.574 -0.028 -0.041 -0.013 (0.496) (0.499) (0.497) (0.495) (0.036) (0.029) (0.032) No. of HH members 5.340 5.269 5.342 5.368 -0.073 -0.099 -0.026 (2.071) (2.105) (2.074) (2.057) (0.187) (0.156) (0.154) TV owned (1=yes) 0.016 0.011 0.019 0.017 -0.008 -0.006 0.002 (0.125) (0.105) (0.136) (0.128) (0.010) (0.006) (0.010) Received remittances (1=yes) 0.102 0.089 0.100 0.108 -0.011 -0.018 -0.008 (0.302) (0.285) (0.300) (0.310) (0.020) (0.015) (0.019) Received support from social programs (1=yes) 0.204 0.167 0.212 0.217 -0.045 -0.050 -0.004 (0.403) (0.373) (0.409) (0.412) (0.060) (0.046) (0.057) Food insecure (1=yes) 0.279 0.263 0.250 0.297 0.013 -0.034 -0.047 (0.449) (0.441) (0.433) (0.457) (0.045) (0.039) (0.040) Panel B: Agricultural Production Characteristics (baseline) Reduced tillage practiced (1=yes) 0.067 0.091 0.051 0.064 0.040 0.027 -0.013 (0.250) (0.288) (0.220) (0.245) (0.032) (0.031) (0.020) Manure applied (1=yes) 0.487 0.505 0.479 0.482 0.025 0.023 -0.003 (0.500) (0.500) (0.500) (0.500) (0.042) (0.037) (0.038) Urea applied (1=yes) 0.639 0.701 0.664 0.603 0.038 0.098* 0.060 (0.480) (0.458) (0.473) (0.489) (0.060) (0.055) (0.052) Intercropping applied (1=yes) 0.169 0.187 0.175 0.159 0.013 0.029 0.016 (0.375) (0.391) (0.380) (0.365) (0.047) (0.040) (0.039) Grown green manure crops (1=yes) 0.022 0.030 0.026 0.018 0.003 0.012 0.009 (0.148) (0.170) (0.160) (0.131) (0.014) (0.010) (0.011) Experienced shock in last season (1=yes) 0.818 0.803 0.793 0.834 0.010 -0.030 -0.040 (0.386) (0.398) (0.405) (0.372) (0.045) (0.038) (0.035) Av. perception of change in soil fertility 1.881 1.876 1.903 1.875 -0.028 0.001 0.029 (1-decreased, 2-same, 3-increased) (0.686) (0.694) (0.691) (0.681) (0.096) (0.080) (0.074) Panel C: Community Level Characteristics (endline) MWS receives support in agriculture from other 0.337 0.364 0.306 0.339 0.057 0.025 -0.032 development organizations (1=yes) (0.473) (0.481) (0.461) (0.473) (0.112) (0.095) (0.092) No. of agricultural trainings offered in mws 3.244 3.180 3.976 2.973 -0.796 0.207 1.002 (apart from ISFM+) (2.625) (2.748) (3.532) (2.019) (0.746) (0.504) (0.625) Agri-input dealer in Kebele (1=yes) 0.628 0.612 0.699 0.606 -0.087 0.007 0.094 (0.483) (0.488) (0.459) (0.489) (0.112) (0.097) (0.093) Seed enterprise in Kebele (1=yes) 0.118 0.139 0.109 0.114 0.030 0.025 -0.005 (0.323) (0.346) (0.312) (0.318) (0.078) (0.067) (0.062) N 2382 539 532 1311 1071 1850 1843 Source: Authors’ analyses based on baseline and endline survey data. Note: HH stands for household. Calculation of food insecurity status based on self-experienced events of food insecurity, based on House- hold Food Insecurity Access Scale (HFIAS). MWS stands for microwatershed. Kebele is the lowest administrative unit in Ethiopia. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 7 Table S4.2. Probit regression results for calculation of propensity score to predict FREG membership, used for matching with the most similar observations in control group. FREG membership Gender HH head (1=male) -0.464** (0.221) Age HH head (in years) -0.003 (0.004) No. of months per year hh head away -0.041 (0.058) HH head married (1=yes) 0.260 (0.216) Education HH head (completed grades) 0.035** (0.014) HH head participates in off-farm wage employment (1=yes) 0.096 (0.138) No. of HH members 0.017 (0.027) No. of organisations involved (0-12) 0.005 (0.026) Father of HH head important in community (1=yes) 0.132 (0.103) Walking distance from homestead to nearest FTC in minutes -0.000 (0.002) No. of times talked to extension agent in the last 12 months 0.012*** (0.004) Attended agric. training in the last 12 months (1=yes) 0.774*** (0.095) Basic assets score (0-4) 0.015 (0.055) Radio owned (1=yes) -0.177* (0.105) Food insecurity score (0-12) -0.071*** (0.018) Received support from social programs (1=yes) 0.168 (0.117) Total land size (in ha) 0.125*** (0.040) No. of tropical livestock unit (TLU) owned -0.025 (0.019) No. of adopted quickwins (0-5) 0.173*** (0.046) Grows main crop (1=yes) 0.257 (0.315) Constant -2.103*** (0.424) Pseudo R-squared 0.180 Observations 1,513 Source: Authors’ analyses based on baseline and endline survey data. Note: Probit regression results for calculation of propensity score for FREG membership. FREG stands for ‘Farmer research and extension group’. For variable definitions see notes of Table 1. Robust standard errors in parentheses. Baseline variables used as covariates. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. 8 Table S4.3. Balance between treatment groups in FREG sample, composed of actual FREG members in treatment communities and their matched controls. T1 T2 C (extension T1 vs. T2 T1 vs. C T2 vs. C (extension) + video) (control) Gender HH head (1=male) 0.90 0.91 0.91 0.00 -0.01 0.00 (0.04) (0.03) (0.02) (0.05) (0.05) (0.04) Age HH head (in years) 42.02 45.79 43.18 -3.77* -1.16 2.61 (1.43) (1.34) (0.98) (1.95) (1.72) (1.64) No. of months HH head away 0.05 0.12 0.08 -0.07 -0.03 0.04 (0.03) (0.09) (0.05) (0.09) (0.06) (0.10) HH head married (1=yes) 0.87 0.93 0.90 -0.06 -0.03 0.03 (0.06) (0.02) (0.02) (0.06) (0.06) (0.03) Education HH head (completed grades) 3.49 3.57 3.48 -0.08 0.01 0.09 (0.59) (0.47) (0.31) (0.75) (0.66) (0.56) HH head has off-farm wage employment (1=yes) 0.15 0.16 0.14 -0.01 0.00 0.02 (0.03) (0.04) (0.02) (0.05) (0.04) (0.05) No. of HH members 5.73 5.83 5.91 -0.10 -0.17 -0.07 (0.27) (0.21) (0.13) (0.34) (0.30) (0.25) No. of organisations involved (0-12) 5.51 5.00 5.39 0.51 0.12 -0.40 (0.35) (0.24) (0.19) (0.42) (0.39) (0.30) Father of HH head important in community (1=yes) 0.78 0.73 0.78 0.05 0.00 -0.05 (0.05) (0.05) (0.03) (0.07) (0.06) (0.06) Walking dist. from homestead to nearest FTC (min) 31.11 30.75 33.44 0.35 -2.34 -2.69 (3.54) (4.06) (2.12) (5.35) (4.10) (4.53) No. of times talked to extension agent in the last 12 months 9.85 11.75 9.23 -1.90 0.62 2.52 (2.24) (1.69) (1.37) (2.79) (2.61) (2.16) Attended agric. training in the last 12 months (1=yes) 0.63 0.61 0.60 0.01 0.03 0.01 (0.07) (0.05) (0.04) (0.08) (0.08) (0.06) Basic assets score (0-4) 1.97 2.21 2.10 -0.24* -0.13 0.11 (0.10) (0.09) (0.08) (0.13) (0.12) (0.11) Radio owned (1=yes) 0.30 0.43 0.34 -0.14** -0.05 0.09 (0.05) (0.05) (0.04) (0.07) (0.06) (0.06) 9 Food insecurity score (0-12) 1.49 0.96 1.23 0.53 0.26 -0.27 (0.35) (0.28) (0.17) (0.44) (0.38) (0.33) Received support from social programs (1=yes) 0.19 0.23 0.22 -0.03 -0.03 0.01 (0.05) (0.06) (0.03) (0.08) (0.06) (0.07) Total land size (in ha) 1.48 1.83 1.65 -0.35 -0.18 0.17 (0.15) (0.30) (0.11) (0.33) (0.19) (0.31) No. of TLU owned 3.78 4.68 4.22 -0.90 -0.43 0.47 (0.31) (0.45) (0.28) (0.54) (0.41) (0.53) No. of adopted quickwins (0-5) 2.66 2.71 2.63 -0.05 0.02 0.07 (0.21) (0.17) (0.10) (0.27) (0.23) (0.20) Grows main crop (1=yes) 0.99 1.00 1.00 -0.01 -0.01 0.00 (0.01) (0.00) (0.00) (0.01) (0.01) (0.00) N 94 106 200 200 294 306 Source: Authors’ analyses based on baseline and endline survey data. Note: Total no. of observations N = 400. FREG stands for ‘Farmer research and extension group’. For variable definitions see notes of Table 1. Robust standard errors in parentheses, clustered at the mws level. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. 10 Table S4.4. Balance between treatment groups in non-FREG sample, composed of actual non-FREG farmers in treatment communities and their matched controls. T1 T2 C (extension T1 vs. T2 T1 vs. C T2 vs. C (extension) + video) (control) Gender HH head (1=male) 0.88 0.88 0.88 0.00 0.00 -0.01 (0.02) (0.02) (0.01) (0.03) (0.02) (0.02) Age HH head (in years) 45.83 46.11 46.30 -0.27 -0.46 -0.19 (0.74) (0.85) (0.59) (1.12) (0.94) (1.03) No. of months HH head away 0.12 0.10 0.09 0.01 0.02 0.01 (0.05) (0.03) (0.02) (0.06) (0.05) (0.04) HH head married (1=yes) 0.86 0.86 0.86 0.00 0.00 -0.01 (0.02) (0.02) (0.01) (0.03) (0.02) (0.02) Education HH head (completed grades) 1.98 2.28 2.07 -0.30 -0.09 0.21 (0.27) (0.25) (0.16) (0.37) (0.31) (0.30) HH head has off-farm wage employment (1=yes) 0.12 0.18 0.15 -0.06 -0.03 0.03 (0.02) (0.03) (0.02) (0.03) (0.02) (0.03) No. of HH members 5.32 5.39 5.40 -0.08 -0.08 -0.01 (0.15) (0.15) (0.09) (0.21) (0.17) (0.17) No. of organisations involved (0-12) 4.40 4.31 4.36 0.09 0.04 -0.05 (0.15) (0.16) (0.12) (0.22) (0.19) (0.20) Father of HH head important in community (1=yes) 0.67 0.66 0.65 0.01 0.02 0.01 (0.03) (0.04) (0.03) (0.05) (0.04) (0.04) Walking dist. from homestead to nearest FTC (min) 33.30 33.33 34.79 -0.03 -1.49 -1.46 (2.85) (2.87) (1.79) (4.02) (3.34) (3.36) No. of times talked to extension agent in the last 12 months 5.12 4.88 5.22 0.24 -0.10 -0.34 (0.68) (0.77) (0.50) (1.02) (0.84) (0.91) Attended agric. training in the last 12 months (1=yes) 0.25 0.28 0.26 -0.04 -0.02 0.02 (0.04) (0.04) (0.02) (0.05) (0.04) (0.04) Basic assets score (0-4) 1.77 1.87 1.83 -0.10 -0.06 0.04 (0.08) (0.08) (0.05) (0.12) (0.09) (0.10) Radio owned (1=yes) 0.28 0.26 0.28 0.02 0.00 -0.02 (0.03) (0.03) (0.02) (0.04) (0.03) (0.04) 11 Food insecurity score (0-12) 2.41 2.37 2.48 0.04 -0.08 -0.12 (0.25) (0.25) (0.16) (0.35) (0.29) (0.29) Received support from social programs (1=yes) 0.16 0.20 0.18 -0.05 -0.02 0.02 (0.04) (0.05) (0.03) (0.07) (0.05) (0.06) Total land size (in ha) 1.35 1.28 1.32 0.06 0.03 -0.03 (0.14) (0.10) (0.06) (0.17) (0.15) (0.11) No. of TLU owned 3.20 3.32 3.30 -0.13 -0.10 0.03 (0.24) (0.18) (0.15) (0.30) (0.28) (0.24) No. of adopted quickwins (0-5) 2.25 2.20 2.24 0.05 0.01 -0.05 (0.13) (0.11) (0.09) (0.17) (0.16) (0.14) Grows main crop (1=yes) 0.95 0.96 0.97 -0.01 -0.02 -0.01 (0.01) (0.01) (0.01) (0.02) (0.01) (0.01) N 416 387 803 803 1,219 1,190 Source: Authors’ analyses based on baseline and endline survey data. Note: Total no. of observations N = 1,606. FREG stands for ‘Farmer research and extension group’. For variable definitions see notes of Table 1. Robust standard errors in parentheses, clustered at the mws level. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. 12 Table S4.5. ITT effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, FREG- and non-FREG samples separately; using a different set of matching variables. Number of ISFM technologies adopted Integrated adoption of full package OLS Poisson Non-FREG sample FREG sample Non-FREG sample FREG sample Non-FREG sample FREG sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.296 0.286*** 1.057*** 1.091*** 0.302 0.301*** 1.056*** 1.091*** 0.046 0.047* 0.228*** 0.235*** (0.189) (0.076) (0.194) (0.126) (0.190) (0.083) (0.186) (0.127) (0.047) (0.024) (0.065) (0.050) p-value 0.120 0.000 0.000 0.000 0.112 0.000 0.000 0.000 0.326 0.055 0.000 0.000 T2 (extension + video) 0.527*** 0.494*** 0.940*** 0.987*** 0.517*** 0.492*** 0.954*** 1.005*** 0.093** 0.096*** 0.184** 0.200*** (0.177) (0.081) (0.178) (0.126) (0.169) (0.083) (0.175) (0.130) (0.045) (0.025) (0.073) (0.043) p-value 0.003 0.000 0.000 0.000 0.002 0.000 0.000 0.000 0.040 0.000 0.011 0.000 Test T1 >= T2 (p-value) 0.141 0.010 0.711 0.777 0.141 0.013 0.712 0.757 0.184 0.038 0.719 0.742 Endline control mean 2.465 2.975 2.465 2.975 0.153 0.275 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.029 0.489 0.185 0.503 0.005 0.089 0.020 0.056 0.010 0.300 0.037 0.262 Observations 1,604 1,604 400 400 1,604 1,604 400 400 1,604 1,604 400 400 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson and probit models (columns (5) to (12)) show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 5. Integrated adoption of full ISFM package is a dummy variable. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws, and their matched controls. Matching variables are those listed in notes of Table 2 (ITT regressions). Additional control variables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are F- tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 13 Table S4.6. ITT effects on different knowledge outcomes, FREG- and non-FREG samples separately; using a different set of matching variables. Overall knowledge Principles knowledge How-to knowledge Non-FREG sample FREG sample Non-FREG sample FREG sample Non-FREG sample FREG sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.006 0.007 0.115*** 0.110*** -0.006 -0.003 0.071** 0.064** 0.024* 0.025** 0.169*** 0.168*** (0.014) (0.013) (0.023) (0.020) (0.017) (0.016) (0.028) (0.027) (0.014) (0.012) (0.026) (0.023) p-value 0.684 0.572 0.000 0.000 0.711 0.826 0.013 0.017 0.090 0.047 0.000 0.000 T2 (extension + video) 0.043*** 0.045*** 0.136*** 0.136*** 0.036** 0.041*** 0.117*** 0.112*** 0.052*** 0.052*** 0.146*** 0.149*** (0.013) (0.012) (0.018) (0.017) (0.016) (0.014) (0.025) (0.023) (0.015) (0.013) (0.018) (0.017) p-value 0.001 0.000 0.000 0.000 0.024 0.004 0.000 0.000 0.001 0.000 0.000 0.000 Test T1 >= T2 (p-value) 0.005 0.003 0.195 0.107 0.010 0.004 0.063 0.045 0.040 0.033 0.803 0.801 Endline control mean 0.470 0.524 0.540 0.595 0.398 0.450 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes R-squared 0.011 0.135 0.139 0.236 0.006 0.102 0.062 0.135 0.016 0.129 0.192 0.306 Observations 1,534 1,534 396 396 1,534 1,534 396 396 1,534 1,534 396 396 Source: Authors’ analyses based on baseline and endline survey data. Note: OLS regressions of household heads' knowledge scores, ranging from 0 to 1, and calculated based on the number of correct answers relative to the total number of questions in a respective domain. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws, and their matched controls. Matching variables are those listed in notes of Table 2 (ITT regressions). Additional control variables identical to those listed in notes of Table 5. One-sided equality tests of T1 and T2 are F-tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 14 Table S4.7. ITT effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, FREG- and non-FREG samples separately; using kernel matching weights in the outcome regressions. Number of ISFM technologies adopted Integrated adoption of full package OLS Poisson Non-FREG sample FREG sample Non-FREG sample FREG sample Non-FREG sample FREG sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.304 0.269*** 1.165*** 1.141*** 0.310 0.277*** 1.165*** 1.147*** 0.036 0.036 0.260*** 0.256*** (0.189) (0.077) (0.192) (0.118) (0.190) (0.084) (0.184) (0.122) (0.047) (0.024) (0.060) (0.049) p-value 0.111 0.001 0.000 0.000 0.102 0.001 0.000 0.000 0.447 0.138 0.000 0.000 T2 (extension + video) 0.532*** 0.471*** 1.047*** 1.037*** 0.522*** 0.459*** 1.064*** 1.064*** 0.085* 0.085*** 0.211*** 0.205*** (0.177) (0.080) (0.174) (0.121) (0.169) (0.083) (0.171) (0.131) (0.045) (0.025) (0.067) (0.041) p-value 0.003 0.000 0.000 0.000 0.002 0.000 0.000 0.000 0.061 0.001 0.002 0.000 Test T1 >= T2 (p-value) 0.145 0.010 0.711 0.778 0.144 0.016 0.711 0.742 0.176 0.042 0.745 0.830 Endline control mean 2.463 2.868 2.463 2.868 0.163 0.251 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.029 0.495 0.196 0.520 0.008 0.289 0.049 0.282 Observations 1,958 1,958 1,352 1,352 1,958 1,958 1,352 1,352 1,958 1,958 1,352 1,352 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson and probit models (columns (5) to (12)) show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 5. Integrated adoption of full ISFM package is a dummy variable. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws, and control observations weighed with Kernel matching weights according to their similarity with FREGs and non-FREGs, respectively. Additional control variables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are F-tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 15 Table S4.8. ITT effects on different knowledge outcomes, FREG- and non-FREG samples separately; using kernel matching weights in the outcome regressions. Overall knowledge Principles knowledge How-to knowledge Non-FREG sample FREG sample Non-FREG sample FREG sample Non-FREG sample FREG sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.009 0.009 0.122*** 0.116*** -0.007 -0.004 0.078*** 0.072*** 0.022* 0.023* 0.174*** 0.170*** (0.014) (0.013) (0.023) (0.018) (0.017) (0.015) (0.026) (0.023) (0.013) (0.012) (0.026) (0.021) p-value 0.532 0.471 0.000 0.000 0.678 0.787 0.003 0.002 0.094 0.055 0.000 0.000 T2 (extension + video) 0.044*** 0.045*** 0.147*** 0.149*** 0.032** 0.037*** 0.130*** 0.127*** 0.048*** 0.047*** 0.152*** 0.154*** (0.013) (0.012) (0.018) (0.016) (0.015) (0.013) (0.024) (0.021) (0.014) (0.013) (0.018) (0.015) p-value 0.001 0.000 0.000 0.000 0.031 0.005 0.000 0.000 0.001 0.000 0.000 0.000 Test T1 >= T2 (p-value) 0.006 0.005 0.145 0.060 0.015 0.006 0.046 0.026 0.051 0.047 0.792 0.760 Endline control mean 0.468 0.513 0.541 0.581 0.401 0.446 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes R-squared 0.011 0.134 0.144 0.260 0.005 0.103 0.069 0.166 0.014 0.128 0.195 0.305 Observations 1,915 1,915 1,324 1,324 1,915 1,915 1,324 1,324 1,915 1,915 1,324 1,324 Source: Authors’ analyses based on baseline and endline survey data. Note: OLS regressions of household heads' knowledge scores, ranging from 0 to 1, and calculated based on the number of correct answers relative to the total number of questions in a respective domain. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws, and control observations weighed with Kernel matching weights according to their similarity with FREGs and non-FREGs, respectively. Additional control variables identical to those listed in notes of Table 5. One-sided equality tests of T1 and T2 are F-tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 16 Table S4.9. ITT effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, FD- and non-FD samples separately. Number of ISFM technologies adopted Integrated adoption of full package OLS Poisson Non-FD sample FD sample Non-FD sample FD sample Non-FD sample FD sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.340* 0.301*** 1.133*** 1.206*** 0.348* 0.308*** 1.127*** 1.208*** 0.045 0.041* 0.227*** 0.237*** (0.192) (0.077) (0.187) (0.137) (0.193) (0.085) (0.179) (0.141) (0.046) (0.025) (0.059) (0.049) p-value 0.079 0.000 0.000 0.000 0.072 0.000 0.000 0.000 0.331 0.093 0.000 0.000 T2 (extension + video) 0.573*** 0.521*** 0.954*** 0.943*** 0.563*** 0.510*** 0.973*** 0.950*** 0.074 0.074*** 0.249*** 0.247*** (0.174) (0.085) (0.201) (0.128) (0.166) (0.091) (0.198) (0.134) (0.045) (0.026) (0.070) (0.044) p-value 0.001 0.000 0.000 0.000 0.001 0.000 0.000 0.000 0.100 0.004 0.000 0.000 Test T1 >= T2 (p-value) 0.135 0.006 0.789 0.954 0.135 0.010 0.790 0.961 0.289 0.130 0.388 0.427 Endline control mean 2.465 2.789 2.465 2.789 0.167 0.227 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.034 0.503 0.172 0.545 0.006 0.094 0.023 0.075 0.006 0.282 0.053 0.310 Observations 1,616 1,616 388 388 1,616 1,616 388 388 1,616 1,616 388 388 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson and probit models (columns (5) to (12)) show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 5. Integrated adoption of full ISFM package is a dummy variable. FD stands for field day. FD- and non-FD samples consist of actual field day participants respectively non-participants in treatment mws, and their matched controls. Additional control variables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are F-tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 17 Table S4.10. ITT effects on different knowledge outcomes, FD- and non-FD samples separately Overall knowledge Principles knowledge How-to knowledge Non-FD sample FD sample Non-FD sample FD sample Non-FD sample FD sample (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) T1 (extension) 0.013 0.014 0.112*** 0.117*** -0.003 -0.001 0.076** 0.083*** 0.025* 0.027** 0.156*** 0.160*** (0.015) (0.013) (0.019) (0.017) (0.018) (0.017) (0.032) (0.024) (0.014) (0.012) (0.024) (0.023) p-value 0.368 0.298 0.000 0.000 0.878 0.969 0.021 0.001 0.066 0.026 0.000 0.000 T2 (extension + video) 0.043*** 0.044*** 0.135*** 0.133*** 0.032* 0.036** 0.126*** 0.122*** 0.044*** 0.045*** 0.139*** 0.136*** (0.013) (0.012) (0.020) (0.019) (0.016) (0.014) (0.027) (0.024) (0.014) (0.013) (0.020) (0.016) p-value 0.001 0.000 0.000 0.000 0.053 0.015 0.000 0.000 0.002 0.001 0.000 0.000 Test T1 >= T2 (p-value) 0.016 0.013 0.143 0.229 0.070 0.019 0.079 0.098 0.115 0.103 0.730 0.836 Endline control mean 0.468 0.522 0.541 0.584 0.404 0.454 Additional controls No Yes No Yes No Yes No Yes No Yes No Yes R-squared 0.010 0.135 0.134 0.264 0.004 0.101 0.068 0.219 0.012 0.136 0.165 0.282 Observations 1,550 1,550 382 382 1,550 1,550 382 382 1,550 1,550 382 382 Source: Authors’ analyses based on baseline and endline survey data. Note: OLS regressions of household heads' knowledge scores, ranging from 0 to 1, and calculated based on the number of correct answers relative to the total number of questions in a respective domain. FD stands for field day. FD- and non-FD samples consist of actual field day participants respectively non-participants in treatment mws, and their matched controls. Additional control variables identical to those listed in notes of Table 5. One-sided equality tests of T1 and T2 are F-tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 18 Table S4.11. ITT effects on number of adopted ISFM technologies, using an alternative measure. Number of ISFM technologies adopted (0-4) OLS Poisson Oprobit (1) (2) (3) (4) (5) (6) T1 (extension) 0.541*** 0.326*** 0.543*** 0.341*** 0.446*** 0.397*** (0.169) -0.078 (0.165) -0.087 (0.139) -0.093 p-value 0.002 0.000 0.001 0.000 0.001 0.000 T2 (extension + video) 0.662*** 0.418*** 0.650*** 0.406*** 0.558*** 0.527*** (0.157) -0.079 (0.149) -0.089 (0.135) -0.094 p-value 0.000 0.000 0.000 0.000 0.000 0.000 Test T1 >= T2 (p-value) 0.241 0.141 0.241 0.234 0.231 0.109 Endline control mean 2.199 Additional controls No Yes No Yes No Yes (Pseudo) R-squared 0.051 0.505 0.010 0.107 0.018 0.225 Observations 2,382 2,382 2,382 2,382 2,382 2,382 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson models show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 4 (excluding lime). Additional control variables identical to those listed in notes of Table 2. One- sided equality tests of T1 and T2 are F-tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. Table S4.12. ITT effects on integrated adoption of the full ISFM package, using alternative measures. Integrated adoption of full ISFM package At least 4 out of 5 3 out of 3 Region-specific (1) (2) (3) (4) (5) (6) T1 (extension) 0.152*** 0.123*** 0.113** 0.095*** 0.062*** 0.062*** (0.044) (0.027) (0.048) (0.028) (0.017) (0.014) p-value 0.001 0.000 0.017 0.001 0.000 0.000 T2 (extension + video) 0.188*** 0.148*** 0.156*** 0.124*** 0.075*** 0.064*** (0.042) (0.024) (0.047) (0.025) (0.021) (0.014) p-value 0.000 0.000 0.001 0.000 0.000 0.000 Test T1 >= T2 (p-value) 0.233 0.189 0.215 0.176 0.241 0.457 Endline control mean 0.157 0.185 0.033 Additional controls No Yes No Yes No Yes (Pseudo) R-squared 0.039 0.274 0.025 0.269 0.042 0.238 Observations 2,160 2,160 2,160 2,160 2,160 2,095 Source: Authors’ analyses based on baseline and endline survey data. Note: Average marginal effects (AME) of probit models. In columns (1) to (2), integrated adoption of full ISFM package is a dummy variable defined as adopting at least four out of five practices together on the same plot (including lime). In columns (3) to (4), full package is defined as adopting all three practices (compost, blended fertilizer, line seeding). In columns (5) to (6), full package is defined as adopting all five practices in Amhara and Oromia (including lime), but only four in Tigray, since lime is not relevant there. Additional control vari- ables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 19 Table S4.13. ITT effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, excluding model farmers. Number of ISFM technologies adopted Integrated adoption of full OLS Poisson Oprobit ISFM package (1) (2) (3) (4) (5) (6) (7) (8) T1 (extension) 0.582*** 0.390*** 0.582*** 0.404*** 0.460*** 0.463*** 0.077* 0.066*** (0.183) (0.076) (0.176) (0.085) (0.136) (0.082) (0.042) (0.024) p-value 0.002 0.000 0.001 0.000 0.001 0.000 0.068 0.005 T2 (extension + video) 0.737*** 0.526*** 0.717*** 0.508*** 0.597*** 0.629*** 0.106** 0.091*** (0.178) (0.081) (0.166) (0.086) (0.138) (0.090) (0.044) (0.025) p-value 0.000 0.000 0.000 0.000 0.000 0.000 0.015 0.000 Test T1 >= T2 (p-value) 0.221 0.064 0.220 0.117 0.196 0.051 0.276 0.189 Endline control mean 2.222 No Yes Additional controls No Yes No Yes No Yes 0.152 (Pseudo) R-squared 0.054 0.525 0.012 0.121 0.019 0.221 0.015 0.271 Observations 2,300 2,300 2,300 2,300 2,300 2,300 2,078 2,078 Source: Authors’ analyses based on baseline and endline survey data. Note: Poisson and probit models (columns (3), (4), (7) and (8)) show average marginal effects (AME). Number of ISFM technologies adopted ranges from 0 to 5. Integrated adoption of full ISFM package is a dummy variable. 82 model farmers from treatment groups excluded. Additional control variables identical to those listed in notes of Table 2. One-sided equality tests of T1 and T2 are F-tests or Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 20 Table S4.14. ITT effects on other ISFM-related practices. Adopted Adopted Adopted Adopted urea green manuring intercropping forage crops (1) (2) (3) (4) (5) (6) (7) (8) T1 (extension) 0.149*** 0.091*** 0.055** 0.048*** 0.099** 0.072*** 0.060** 0.053** (0.043) (0.024) (0.022) (0.014) (0.041) (0.028) (0.028) (0.023) Unadjusted p-value 0.001 0.000 0.015 0.001 0.016 0.009 0.030 0.021 Adjusted p-value 0.003 0.000 0.048 0.003 0.052 0.029 0.097 0.068 T2 (extension + video) 0.142*** 0.093*** 0.053*** 0.050*** 0.094** 0.063** 0.064* 0.040 (0.043) (0.027) (0.018) (0.013) (0.042) (0.030) (0.033) (0.026) Unadjusted p-value 0.001 0.001 0.003 0.000 0.026 0.035 0.053 0.120 Adjusted p-value 0.003 0.003 0.010 0.000 0.083 0.111 0.167 0.348 Test T1 >= T2 (unadjusted p-value) 0.562 0.470 0.536 0.437 0.539 0.604 0.453 0.694 Test T1 >= T2 (adjusted p-value) 0.921 0.858 0.919 0.847 0.922 0.953 0.867 0.981 Endline control mean 0.740 0.016 0.204 0.150 Additional controls No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.034 0.321 0.056 0.330 0.011 0.145 0.007 0.118 Observations 2,382 2,382 2,382 2,382 2,382 2,382 2,382 2,382 Source: Authors’ analyses based on baseline and endline survey data. Note: Average marginal effects (AME) of probit models. Additional control variables identical to those listed in notes of Table 2. Bonferroni-adjusted p-values taking into account correlations between outcomes. One-sided equality tests of T1 and T2 are Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 21 Table S4.15. ITT effects on other ISFM-related practices, FREG sample. FREG sample Adopted Adopted Adopted Adopted urea green manuring intercropping forage crops (1) (2) (3) (4) (5) (6) (7) (8) T1 (extension) 0.134*** 0.136*** 0.153** 0.139*** 0.207*** 0.208*** 0.310*** 0.310*** (0.048) (0.029) (0.059) (0.037) (0.072) (0.053) (0.048) (0.046) Unadjusted p-value 0.005 0.000 0.010 0.000 0.004 0.000 0.000 0.000 Adjusted p-value 0.014 0.000 0.033 0.000 0.013 0.000 0.000 0.000 T2 (extension + video) 0.177*** 0.182*** 0.139*** 0.160*** 0.177*** 0.212*** 0.221*** 0.220*** (0.060) (0.046) (0.053) (0.040) (0.057) (0.045) (0.066) (0.058) Unadjusted p-value 0.003 0.000 0.008 0.000 0.002 0.000 0.001 0.000 Adjusted p-value 0.009 0.000 0.027 0.000 0.006 0.000 0.003 0.000 Test T1 >= T2 (unadjusted p-value) 0.235 0.132 0.600 0.300 0.654 0.467 0.904 0.939 Test T1 >= T2 (adjusted p-value) 0.534 0.332 0.953 0.697 0.968 0.870 1.000 1.000 Endline control mean 0.850 0.040 0.245 0.190 Additional controls No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.123 0.478 0.078 0.356 0.034 0.203 0.074 0.171 Observations 400 400 400 400 400 400 400 400 Source: Authors’ analyses based on baseline and endline survey data. Note: Average marginal effects (AME) of probit models. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws and their matched controls. Additional control variables identical to those listed in notes of Table 2. Bonferroni-adjusted p-values taking into account correlations between outcomes. One-sided equality tests of T1 and T2 are Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 22 Table S4.16. ITT effects on other ISFM-related practices, non-FREG sample. Non-FREG sample Adopted Adopted Adopted Adopted urea green manuring intercropping forage crops (1) (2) (3) (4) (5) (6) (7) (8) T1 (extension) 0.067 0.039 0.031 0.024** 0.062 0.056* -0.022 -0.007 (0.041) (0.024) (0.021) (0.012) (0.043) (0.029) (0.033) (0.026) Unadjusted p-value 0.101 0.100 0.139 0.047 0.154 0.053 0.498 0.805 Adjusted p-value 0.293 0.290 0.407 0.155 0.446 0.175 0.912 0.997 T2 (extension + video) 0.075* 0.071** 0.028* 0.024** 0.063 0.047 0.005 0.010 (0.043) (0.028) (0.016) (0.010) (0.051) (0.035) (0.034) (0.027) Unadjusted p-value 0.076 0.011 0.074 0.021 0.216 0.185 0.880 0.713 Adjusted p-value 0.227 0.035 0.236 0.072 0.577 0.515 0.999 0.988 Test T1 >= T2 (unadjusted p-value) 0.430 0.130 0.562 0.506 0.495 0.590 0.233 0.285 Test T1 >= T2 (adjusted p-value) 0.839 0.364 0.944 0.915 0.911 0.957 0.608 0.694 Endline control mean 0.819 0.017 0.224 0.173 Additional controls No Yes No Yes No Yes No Yes (Pseudo) R-squared 0.012 0.265 0.025 0.293 0.005 0.130 0.001 0.105 Observations 1,606 1,606 1,606 1,606 1,606 1,606 1,606 1,606 Source: Authors’ analyses based on baseline and endline survey data. Note: Average marginal effects (AME) of probit models. FREG stands for ‘Farmer research and extension group’. FREG- and non-FREG samples consist of actual FREG respectively non-FREG farmers in treatment mws and their matched controls. Additional control variables identical to those listed in notes of Table 2. Bonferroni-adjusted p-values taking into account correlations between outcomes. One-sided equality tests of T1 and T2 are Wald tests. Robust standard errors in parentheses, clustered at the mws level. Significance levels indicated as following: *** p<0.01, ** p<0.05, * p<0.1. 23 Figure S4.1. Histogram of estimated propensity score used for matching FREG members in treatment mws with control observations, using one-to-one nearest neighbor matching. Source: Authors’ illustration based on analyses with baseline and endline survey data. Note: FREG stands for ‘Farmer research and extension group’. Figure S4.2. Histogram of estimated propensity score used for matching non-FREG members in treat- ment mws with control observations, using one-to-one nearest neighbor matching. Source: Authors’ illustration based on analyses with baseline and endline survey data. Note: FREG stands for ‘Farmer research and extension group’. 24 Online Appendix S5: Formula p-value correction = 1 − (1 − ())() Where () = (1−(.)) , with as the number of tested outcomes in a family, (. ) as mean correlation among all outcomes other than outcome k, and () as the unadjusted p-value for the kth outcome. Source: McKenzie (2012b), based on Sankoh et al. (1997) and used in Aker et al. (2016). 25 Online Appendix S6: Causal mediation analysis Our main findings provide evidence that both extension-only as well as extension-plus-video increase farmers’ knowledge about ISFM. The video has triggered additional gains in knowledge, especially among those farmers that do not actively participate in the activities of the extension intervention. In particular, the video has shown success in understanding why the ISFM practices are important. Therefore, we seek to understand the contribution that these gains in ISFM knowledge make to the treatment effect on adoption. To do so, we conduct a causal mediation analysis, in which we use first the overall, and subsequently both the how-to and the principles knowledge indicator as mediating variables. Estimation strategy We apply a causal mediation analysis, following De Brauw et al. (2018), Frölich and Huber (2017) and Imai et al. (2011). The aim is to estimate the average effect of our treatments that is occurring through changes in knowledge as a mediating variable ( ) that are triggered by the treatment. The causal mediation effect can be written as () ≡ (, (1)) − (, (0)) (S6.1) in which t = 0, 1 denotes the treatment status. By holding the treatment status otherwise constant at t and therefore eliminating all other causal mechanisms, () isolates the change in the outcome that stems from changing the mediator from the control to the treatment con- dition. The direct effect of the treatment (), that is, the portion of the treatment effect not ex- plained by the mediator, can be identified by changing the treatment status from 0 to 1, while fixing the effect of the mediator at t: () ≡ (1, ()) − (0, ()) (S6.2) 26 When averaging over all observations, the average causal mediation effect (ACME) is given by (), and the average direct effect (ADE) is estimated by (), while the sum of the two () + () represents the total average treatment effect. Two ignorability assumptions have to be made in order to estimate the ACME and the ADE (Frölich and Huber 2017). Firstly, treatment assignment is assumed to be independent of po- tential outcomes and mediators, conditional on baseline confounders. This exogeneity assump- tion holds due to randomization. The second imposes a selection-on-observables assumption, which states that when we control for actual treatment status and observed pre-treatment char- acteristics, the mediating variable is statistically independent of potential outcomes. This as- sumption is called sequential ignorability and implies that no unobserved confounders exist that affect both our outcome and the mediator. Subsequently, we can estimate the ACME and ADE by sequentially estimating the following equations: = 2 + 2 1 + 2 2 + 2 0 + 2 0 + є2 + 2 (S6.3) = 3 + 3 1 + 3 2 + + 3 0 + 3 0 + є3 + 3 (S6.4) ̂, where 2 represents the effect of T1 on the ̂2 The ACME of knowledge for T1 is given by ̂ ̂ 2 mediator variable, and the effect of the mediator on the outcome measure. Similarly, gives the ACME of knowledge for T2.1 Due to sequential ignorability, non-correlation between the error terms є2 and є3 is assumed, denoted by = 0. Yet, since we can reasonably think of potential unobservable confounders that affect both knowledge and adoption (e.g. farmers’ level of motivation or commitment) and would consequently bias our ACME estimates, we perform sensitivity tests in which we relax the assumption of = 0 and re-estimate equations (S6.3) and (S6.4) for different hypothetical values of . 1 Note that this formal description of causal mediation analyses assumes to fit linear regressions, in which both the outcome and the mediating variable are continuous measures. When the outcome is binary (as it is the case for the integrated adoption of the full ISFM package), the product of coefficients does not correspond to the ACME (Hicks and Tingley 2011; Imai, Keele, and Tingley 2010). Methods to correctly estimate mediation effects for binary outcomes and continuous mediators have been developed and are applied accordingly (Hicks and Tingley 2011). 27 Results and discussion Panel A of Table S6.1 suggests positive and highly statistically significant effects of all knowledge variables on both the number of adopted technologies as well as integrated adoption, while the effects of T1 and T2 remain highly statistically significant. Looking at the coefficient sizes, how-to knowledge appears to have a stronger effect on adoption than principles knowledge. Panel B presents the estimated ACME and ADE of T1 for all three mediators and both adoption outcomes separately, Panel C the corresponding effects for T2. Regarding overall knowledge, which comprises knowledge on both why and how to implement ISFM, Columns (2) and (8) show that for both outcome variables, on average around 11% of the treatment effect of T1 and 16 to 17% of the effect of T2 on adoption can be explained by gains in knowledge. Looking at the two different knowledge types, on average, gains in principles knowledge do not seem to significantly contribute to the effect of T1 on both adoption variables (Panel B, Columns (3), (4), (9) and (10)). In contrast, for the combined treatment (T2), the ACME is significant, albeit relatively small, for both outcome indicators. Estimates show that on average, around 6.9% respectively 6.4% of the effect of T2 on the number of adopted practices and on integrated adoption are driven by an increase in principles knowledge (Panel C, Columns (4) and (10)). The ACME for how-to knowledge is robustly significant for both treatments and both out- come variables (Columns (5), (6), (11) and (12)). The effect sizes indicate that on average, increases in how-to knowledge triggered through T1 account for 16.1% respectively 23.0% of its total effect on adoption, while 17.5% respectively 23.3% of the effect of T2 seem to be transmitted through how-to knowledge gains (Columns (6) and (12)). Hence, in line with the results presented earlier (Table 5), both extension-only and extension combined with video induce increases in understanding how to implement ISFM, and these increases partly account for higher ISFM adoption. In contrast, only the combined treatment 28 leads to robustly significant gains in understanding why ISFM works, which accounts for a small, but significant portion of the T2 effect on adoption. Since the sequential ignorability assumption we made to identify causal mediation effects is unjustifiably strong, we perform a sensitivity test to assess how severely our ACME estimates may be biased due to potential correlation ≠ 0 of the error terms of equations (S6.3) and (S6.4). Figures S6.1 to S6.12 depict the ACME for both mediators and both treatment variables as functions of varying values for . Results show that only relatively large negative correla- tions between the error terms would imply a strong impact of the knowledge mediators on both adoption outcomes. Yet, a positive correlation of error terms appears far more plausible, since unobservables determining additional unexplained knowledge should also positively affect un- explained adoption. In fact, when we estimate the correlation between error terms of equations (S6.3) and (S6.4) for both knowledge and both adoption variables, we find positive, but fairly small correlations never exceeding = .003 for the number of adopted practices, and = .089 for integrated adoption. Hence, our estimated ACMEs should be considered upper bounds. 29 Table S6.1. ITT and knowledge effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, ADE of treatments and ACME of overall, principles and how-to knowledge as mediating variables. Number of ISFM technologies adopted Integrated adoption of full ISFM package (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel A: Coefficient Estimates T1 (extension) 0.546*** 0.406*** 0.643*** 0.437*** 0.480*** 0.380*** 0.084** 0.074*** 0.098** 0.081*** 0.070* 0.066*** (0.163) (0.075) (0.175) (0.080) (0.162) (0.075) (0.042) (0.025) (0.042) (0.026) (0.041) (0.024) T2 0.609*** 0.490*** 0.744*** 0.542*** 0.568*** 0.479*** 0.108** 0.097*** 0.125*** 0.106*** 0.098** 0.089*** (extension + video) (0.161) (0.074) (0.171) (0.077) (0.157) (0.074) (0.043) (0.024) (0.043) (0.024) (0.042) (0.023) Overall knowledge 2.909*** 1.443*** 0.398*** 0.254*** score (0.180) (0.147) (0.050) (0.057) Principles knowledge 1.663*** 0.760*** 0.227*** 0.114*** score (0.142) (0.109) (0.039) (0.039) How-to knowledge 3.102*** 1.422*** 0.483*** 0.344*** score (0.193) (0.157) (0.056) (0.057) (Pseudo) R-squared 0.214 0.559 0.144 0.545 0.223 0.557 0.052 0.282 0.038 0.274 0.065 0.293 Table S6.1. ITT and knowledge effects on number of adopted ISFM technologies and integrated adoption of the full ISFM package, ADE of treatments and ACME of overall, principles and how-to knowledge as mediating variables (continued). Number of ISFM technologies adopted Integrated adoption of full ISFM package Overall Principles How-to Overall Principles How-to (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Panel B: ACME and ADE Estimates of T1 ACME (T1) 0.147*** 0.051*** 0.050* 0.015 0.212*** 0.073*** 0.019** 0.009** 0.004 0.002 0.034*** 0.020*** (0.044) (0.019) (0.027) (0.011) (0.049) (0.020) (0.008) (0.004) (0.005) (0.002) (0.010) (0.006) ADE (T1) 0.553*** 0.410*** 0.651*** 0.441*** 0.488*** 0.383*** 0.094** 0.077*** 0.107** 0.085*** 0.079* 0.069*** (0.153) (0.080) (0.165) (0.085) (0.153) (0.078) (0.045) (0.027) (0.045) (0.028) (0.045) (0.026) Total effect (T1) 0.700*** 0.460*** 0.700*** 0.456*** 0.700*** 0.457*** 0.109** 0.085*** 0.111** 0.086*** 0.108** 0.086** (0.163) (0.080) (0.168) (0.083) (0.161) (0.079) (0.045) (0.028) (0.045) (0.028) (0.045) (0.027) Share of T1 effect explained by 20.9% 11.0% 7.1% 3.2% 30.2% 16.1% 17.3% 11.1% 4.0% 2.1% 31.6% 23.0% knowledge Panel C: ACME and ADE Estimates of T2 ACME (T2) 0.240*** 0.097*** 0.105*** 0.041*** 0.281*** 0.102*** 0.038*** 0.020*** 0.017*** 0.008** 0.049*** 0.028*** (0.042) (0.019) (0.025) (0.010) (0.052) (0.022) (0.008) (0.006) (0.005) (0.003) (0.011) (0.007) ADE (T2) 0.616*** 0.493*** 0.751*** 0.546*** 0.575*** 0.483*** 0.125*** 0.106*** 0.142*** 0.115*** 0.115** 0.097*** (0.152) (0.079) (0.161) (0.082) (0.148) (0.078) (0.048) (0.027) (0.049) (0.027) (0.048) (0.026) Total effect (T2) 0.856*** 0.590*** 0.856*** 0.586*** 0.855*** 0.585*** 0.155*** 0.122*** 0.155*** 0.121*** 0.154** 0.120*** (0.160) (0.078) (0.165) (0.080) (0.159) (0.080) (0.048) (0.027) (0.049) (0.027) (0.048) (0.026) Share of T2 effect explained by 28.0% 16.4% 12.3% 6.9% 32.8% 17.5% 24.9% 16.9% 11.2% 6.4% 32.1% 23.3% knowledge Additional controls No Yes No Yes No Yes No Yes No Yes No Yes Observations 2,334 2,334 2,334 2,334 2,334 2,334 2,116 2,116 2,116 2,116 2,116 2,116 Source: Authors’ analyses based on baseline and endline survey data. Note: Causal mediation estimates for overall, principles, and how-to knowledge variables on number of practices adopted (OLS specification) and integrated adoption of full ISFM package (probit specification, AME shown). ACME stands for average causal mediation effect, ADE for average direct effect. Additional controls identical to those listed in notes of Tables 2 and 5. Robust standard errors in parentheses, clustered at the mws level. Significance levels: *** p<0.01, ** p<0.05, * p<0.1. 31 All in all, these results suggest that more complete knowledge indeed fosters adoption, since increases in ISFM adoption are at least partly channeled through gains in ISFM knowledge triggered by the interventions. Both knowledge types, how-to and principles, contribute to ex- plaining treatment effects on adoption. Hence, a better understanding of ISFM as a package consisting of several important and complementary technologies might indeed have positively influenced the decision to adopt the full package in an integrated manner, especially for those farmers that are excluded from the extension activities but did take part in the video interven- tion. Yet, we also find that increases in knowledge only partially explain the effects of our ex- perimental treatments, which is in line with previous studies finding effects of extension on adoption that are only modestly explained by gains in knowledge (Kondylis, Mueller, and Zhu 2017; De Brauw et al. 2018). The limited explanatory effect of knowledge as impact channel might to some extent be attributed to imperfect measurement that knowledge assessments are prone to, in the sense that with our questions we might have missed to capture some adoption- relevant dimensions of knowledge, which the treatments may have altered (Laajaj and Macours 2017). Yet, as Kondylis et al. (2017) argue in the context of adoption of sustainable land man- agement practices, knowledge constraints might simply not be the most decisive barrier to adop- tion, but rather a lack of awareness of their productivity benefits. In line with this, our treatments may have played a more crucial role in influencing farmers’ awareness of the environmental and in particular the yield-enhancing benefits of ISFM, which has been shown to be an im- portant driver of adoption (Knowler and Bradshaw 2007; Maertens, Michelson, and Nourani 2020). Testimonies of the farmers about improvements of yields and their livelihoods presented in the video might have further increased the credibility of information obtained via the exten- sion intervention. 32 Figure S6.1. Sensitivity test ACME overall knowledge (T1), no. of adopted practices. Figure S6.2. Sensitivity test ACME overall knowledge (T2), no. of adopted practices. Source: Authors’ illustration based on analyses with baseline and endline survey data. Source: Authors’ illustration based on analyses with baseline and endline survey data. Figure S6.3. Sensitivity test ACME overall knowledge (T1), integr. adoption. Figure S6.4. Sensitivity test ACME overall knowledge (T2), integr. adoption. Source: Authors’ illustration based on analyses with baseline and endline survey data. Source: Authors’ illustration based on analyses with baseline and endline survey data. 33 Figure S6.5. Sensitivity test ACME prin. knowledge (T1), no. of adopted practices. Figure S6.6. Sensitivity test ACME prin. knowledge (T2), no. of adopted practices. Source: Authors’ illustration based on analyses with baseline and endline survey data. Source: Authors’ illustration based on analyses with baseline and endline survey data. Figure S6.7. Sensitivity test ACME prin. knowledge (T1), integr. adoption. Figure S6.8. Sensitivity test ACME prin. knowledge (T2), integr. adoption. Source: Authors’ illustration based on analyses with baseline and endline survey data. Source: Authors’ illustration based on analyses with baseline and endline survey data. 34 Figure S6.9. Sensitivity test ACME how-to knowledge (T1), no. of adopted practices. Figure S6.10. Sensitivity test ACME how-to knowledge (T2), no. of adopted practices. Source: Authors’ illustration based on analyses with baseline and endline survey data. Source: Authors’ illustration based on analyses with baseline and endline survey data. Figure S6.11. Sensitivity test ACME how-to knowledge (T1), integr. adoption. Figure S6.12. Sensitivity test ACME how-to knowledge (T2), integr. adoption. Source: Authors’ illustration based on analyses with baseline and endline survey data. Source: Authors’ illustration based on analyses with baseline and endline survey data. 35 Additional References Online Appendix S6 Frölich, Markus, and Martin Huber. 2017. “Direct and Indirect Treatment Effects -Causal Chains and Mediation Analysis with Instrumental Variables.” Journal of the Royal Statistical Society B 79 (5): 1645–66. https://doi.org/10.1111/rssb.12232. Hicks, Raymond, and Dustin Tingley. 2011. “Causal Mediation Analysis.” The Stata Journal 11 (4): 605–19. Imai, Kosuke, Luke Keele, and Dustin Tingley. 2010. “A General Approach to Causal Media- tion Analysis.” Psychological Methods 15 (4): 309–34. https://doi.org/10.1037/a0020761. Imai, Kosuke, Luke Keele, Dustin Tingley, and Teppei Yamamoto. 2011. “Unpacking the Black Box of Causality: Learning about Causal Mechanisms from Experimental and Obser- vational Studies.” American Political Science Review 105 (4): 765–89. https://doi.org/10.1017/S0003055411000414. Knowler, Duncan, and Ben Bradshaw. 2007. “Farmers’ Adoption of Conservation Agriculture: A Review and Synthesis of Recent Research.” Food Policy 32 (1): 25–48. https://doi.org/10.1016/j.foodpol.2006.01.003. Laajaj, Rachid, and Karen Macours. 2017. “Measuring Skills in Developing Countries.” Policy Research Working Paper 8000. Washington D.C.: The World Bank. https://openknowledge.worldbank.org/bitstream/hadle/10986/26250/WPS8000.pdf?se- quence=1&isAllowed=y 36