How Do Farmers Learn From Extension Services?: Evidence From Malawi

Agricultural extension services can play an important role in increasing farmer yields and incomes yet evidence of the effectiveness of extension services in Sub-Saharan Africa has been mixed. We study farmers learning about agricultural technologies using a (quasi) randomized controlled trial in which farmers differ in their exposure to commonly used extension methods that range in their intensity of interaction. We find that farmers who participated in season-long farmer-led demonstration plot cultivation learn about the critical adoption details of production processes and adopt more components of a new, multi-component technology. Farmers invited to attend farmer field day events learn considerably less about production process details. Building on qualitative interviews, we then develop a two-stage learning process in which farmers first form yield expectations and then choose how much to invest in learning the details of the production processes subject to yield beliefs and the learning costs. We test this model using detailed data on beliefs, knowledge, adoption and constraints and find evidence that farmers yield beliefs hinge around observed yield, and these observed yields affect learning efforts.

If learning about the production process comes at a signi®cant cognitive cost, farmers might engage in what has been termed "rational inattention" (see Ghosh 2016 for a theoretical approach on rational inattention; and Gabaix 2017 for an overview). This might imply a two-stage learning process, as in Nourani (2017), where farmers ®rst establish a belief of the pro®tability, and only when this belief meets a particular cut-oV or set of conditions, proceed with learning the production processes. Moreover, farmers using attention strategically may focus on dimensions of the technology where the perceived bene®ts might be most likely to exceed the perceived costs. For instance, credit-constrained farmers might focus on the more labor-intensive dimensions of a new production process.
The contribution of this paper is to exploit an opportunity that introduced variation into etc.; only a yield-draw from a plot which shares these conditions is likely to impact their own thinking and operations (on the implications of heterogeneity for social learning, see, among others, Munshi  farmer exposure to diVerent extension methods instructing farmers about a single (complex) agricultural technology. We study the eVects of farmer ®eld-days and farmer demonstration plots on farmer learning and adoption and we present a model of farmer learning based on the insights from the evaluation. We exploit rich data on soil conditions, demonstration plot performance, and agronomic outcomes to understand how farmers direct their attention under time constraints and how spatial variability in growing conditions impacts farmer learning and adoption. To be clear at the outset: this paper is not a horserace between extension models. Instead, we are interested in contrasting ®eld-days and demonstration plots to gain broader and more generalizable insights into farmer learning.
The literature on the eVects of extension struggles with two primary empirical challenges.
First, farmers who seek out and receive extension services might be more skilled and motivated than farmers who do not seek such services. 4 Moreover, areas that attract extension services are also often areas with better agronomic potential. Because such factors are often unobserved by researchers, they can cause omitted variable bias, threatening the causal interpretation of estimated parameters. A second challenge is that although an extension program may be successful in terms of knowledge diVusion, adoption among farmers may be in¯uenced by other factors (market failures, logistical challenges, etc.); and learning may not always translate into adoption. As standard household surveys often do not detail the learning process, studies have often faced challenges discerning whether such failures reside in the education process itself, or in other circumstances down the line.
We designed our study to meet these challenges. We worked in partnership with the Clinton Development Initiative (henceforward CDI) in Malawi. CDI has set up a program of farmer-led demonstration plots and ®eld-days aiming to disseminate information about ISFM -with a focus on the maize and soybean cropping systems. We selected 250 villages and randomized access to CDI's ®eld-days. This process eliminates biases originating from between-village unobserved variation when establishing the eVects of access to ®eld-days.
Using detailed village-level data we note that the villages in which CDI placed demonstration plots (which, unlike the ®eld-days, were not randomized) were comparable to the villages which merely received access to the ®eld-days. This observation allows us to use a regression approach when establishing the impacts of demonstration plots. We complement this quasi-randomized design with a household panel survey documenting the adoption of ISFM technologies, as well as farmer knowledge of ISFM technology processes and yield expectations.
We begin our analysis by establishing the impacts of demonstration plots and ®elddays. We ®nd that farmers who participate in demonstration plots plan to adopt around, on average, 22 percent more of the recommended ISFM technologies one year after the program's start compared to similar farmers in control villages. Farmers who participate in a ®eld-day, on the other hand, do not plan to adopt more of the ISFM technologies relative to similar farmers in the control villages. This lack of adoption might be due to a lack of learning: One year after the program was introduced, farmers who participate in demonstration plots score, on average, 8 percent higher on a test measuring knowledge of ISFM, compared to similar farmers in control villages. Farmers who attend a ®eld-day do not score any higher compared to similar farmers in control villages.
To gain a better understanding of the learning process, we conducted focus group interviews. Farmers who attended ®eld-days report being impressed by the yields on the ®eld-day plots they visited, and reported being convinced that pesticides, and more generally, "modern inputs" are important. However, when we inquired for instance about speci®cs related to pesticide use, few of the ®eld-day participants knew brand names, where to purchase these inputs, or how to prepare and apply the products. In other words, though farmers did acquire some knowledge about improving yields, they retained little speci®c information about production practices. Farmers who participated in a demonstration plot, on the other hand, tended to recall details about the inputs and production practices. For instance, these farmers were able to recount the inoculation process for soybean seeds, from preparing the inoculant to covering the seeds and planting them on ridges.
This narrative is consistent with the presence of a rational, but costly, two-step learning process in farmers ®rst assess pro®tability, and only then invest in learning subject to We present a learning model based on these insights and test its implications in our data.
First, the model predicts that yield expectations should respond to observed yields, and respond more so if the conditions of the extension demonstrations more closely match the farmer's own conditions. We test this prediction in the demonstration plot villages where we collected and analyzed soil samples from the demonstration plot and the farmers; as well as demonstration plot yield data. We ®nd that the farmers' yield expectations correlate more strongly with the observed yields of soybean and maize if the farmers' soil is more similar to the demonstration plot's soil.
Second, this model predicts that the amount of cognitive eVort farmers commit to learning the production process responds to these yield expectations. We again test this prediction in the demonstration plot villages using rainfall data. We note a positive correlation between demonstration plot yields and the farmers' knowledge of ISFM production practices. As this correlation could be attributed to reverse causality: demonstration plot farmers with increased knowledge more accurately followed CDI's guidelines, and hence, these demonstration plots had a higher yield; we repeat the regression analysis using the rainfall data instead of the yields. This analysis builds on the observation that germination rate, i.e., the percentage of seeds which germinate about 3 weeks after planting, correlate strongly with yields, but are largely determined by rainfall patterns at the start of the season. We con®rm that a later start of the rains, and additional¯ood days (days with more than 50 mm/day), yields. For farmers who were invited to participate in ®eld-days and who arguably face a signi®cantly higher cognitive cost of learning, we note a similar correlation between being credit constrained and learning about soybean cultivation (it is notable that we ®nd no such result for hybrid maize, for which recommended technologies are more labor rather than credit-intensive). This is consistent with the accounts of the ®eld-day participant who noted that, because they were credit-constrained, they preferred to focus on learning about (labor-intensive) technologies for maize, such as mulching, and optimal planting practices.
These results give a new meaning to the "rational but poor" farmers thesis originally proposed by Schultz (1964) and, subsequently tested by many, among others, Hopper (1965).
Indeed, while farmers in our study might appear "irrational" at ®rst in that they show evidence of inattention to technologies which might be bene®cial, once one takes into account constraints imposed by heterogeneity and cognitive resources, the learning process appears rational. In eVect, farmers in our study appear to decide how much eVort to put into a learning experience, and to focus on speci®c aspects which they ®nd important. These aspects are determined by individual farmers' expectations and constraints including market constraints.
The structure of the rest of this paper is as follows. In the next section, we introduce CDI's CDI de®ned best practices on the soybean-maize demonstration plots as follows: 7 • Maize: Use of a high-yielding variety, SC719, optimal plant spacing and seeding practices, regular herbicide (harness and roundup) and fertilizer application (23:21:0+4S and urea), a mulch of crop residues and other transferred biomass, and, on some subplots, the use of fertilizer trees (Tephrosia) and rotation with soybean.
All best practice subplots featured ridges, and included the incorporation of organic material through the addition of compost-manure (a mix of decomposed plant residues and livestock manure).
In addition to best practice subplots, each demonstration plot also included control subplots and farmer practice subplots. The latter are understood as "the local method to cultivate a crop" 8 , while the former aimed to provide a benchmark for the best practice subplots (using the same variety and planting techniques, but did not include any other external inputs, such as innoculants, fertilizer, pesticides and herbices). Appendix A provides details of the layout. 7 The best practices for groundnut and common bean are de®ned as follows: (1) Groundnut: Use of a high-yielding variety, CG7, optimal plant spacing and seeding practices, regular fertilizer (D-compound, single superphosphate and gypsum), pesticide (karate) and herbicide (harness and roundup) applications., and (2) Common bean: Use of a high-yielding variety, Kholophete, optimal plant spacing and seeding practices, regular fertilizer (D-compound), pesticide (cypermethrine) and herbicide (harness and roundup) applications. 8 In practice, however, these guidelines were not always adhered to, and several farmer practice subplots resembled either control subplots or best practice subplots.

Sample, randomization and data collected
In 2014, CDI was planning an expansion of their program into two districts in central Malawi: Farmers formed clubs in 91 out of 125 treatment villages (in our study sample, this will be 47 out of 55). Observing this club formation process it appears that villages who formed clubs tend to be further away from markets and have a history of community action (we will con®rm this observation, and discuss its implication, in the next section).
Seventeen of the 91 villages received farmer-led demonstration plots during the 2014-15 growing season. These 17 villages were selected by CDI. By CDI's own account, these were villages with some familiarity with agricultural extension services, located in an accessible location and where people were in "unity". All but ®ve villages received a soybean/maize demonstration plot, the remaining ®ve received either a groundnut/maize plot or a common bean/maize plot. Appendix table 2 compares the demonstration plot villages with the other treatment villages, and reports few signi®cant observable diVerences between these two sets. 11 9 As CDI works through farmer clubs and the functioning of these clubs requires a minimum village size, we excluded the villages with less than 50 households. 10 Appendix Table 1 compares village-level descriptive statistics for the control villages (Columns (4) through (6)) and the treatment villages (Columns (7) through (9)). Column (10) presents the P-values of a t-test with unequal variances. We note three areas of diVerences between the two groups: ethnicity, size of civil organisations, and daily wage rates. Note that the sample includes only the villages who were revisited one year after baseline (see the data collected section below). 11 Appendix Table 2

Data collected
We collected data at baseline, before the treatment villages participated in the program activities (in Fall 2014), and one year later (in Fall 2015). The baseline was conducted in all 250 villages in the sample, while the data collection the following year included 100 villages. 12 Before collecting baseline data, we generated a census of all households in the 250 villages as well as a census of all CDI club members in the treatment villages. We used these two census lists to draw a sample of 10 households for each village: In the control villages and the treatment villages without a club, we randomly selected 10 households from the village census. In the treatment villages with CDI clubs, we strati®ed the sample and sampled ®ve households not participating in a CDI club and ®ve participant households. 13 One of the ®ve households sampled was the household of the lead farmer of the club, whom serves as the point of contact between CDI and the club. The other four CDI households were randomly selected from the list of households who belong to the CDI club. We discuss self-selection into clubs in the next Section. Club members are wealthier, more educated and better connected than non-club members.
At baseline, we conducted a village survey, a household survey and collected and analyzed soil samples from farmers' plots and demonstration plots. One year later, we followed up with household surveys; creating a household panel dataset. 14 Between these two rounds the 10 additionally selected demonstration plot villages (see the data collected section below). 12 These 100 villages were selected as follows: First, we selected selected 90 villages randomly from the 250 sample villages, strati®ed by EPA and treatment status. These 90 villages included 7 villages with demonstration plots. Then, we included an additional 10 villages which had been selected as demonstration plot villages by CDI (as to include all 17 villages which were selected as a demonstration plot locations). 13 In case of multiple CDI clubs in a village, we selected the club to be included in the study randomly. In terms of the treatment, all CDI clubs are invited to the farmer ®eld day while only one club was engaged per demonstration plot (this would be the same club which we interviewed). 14 The attrition rate is 5% -speci®cally, there were 51 households who were present at the baseline who were not present in the follow up survey. The households who left the sample are uniformly distributed of data collection, we collected agronomic data at the demonstration plot sites on a weekly basis. We also conducted a series of focus group interviews and interviewed extension agents.
We discuss these data sources below.

Village survey
We administered a village questionnaire at baseline in each of the 250 villages with a knowledgeable individual, often the village head or the secretary to the village head. This village questionnaire covered information on the village's distance to paved roads, national highways, (seasonal) markets and other services (such as banks). In addition, we collected demographic information (number of residents, ethnic distribution), and information on access to government and NGO extension, civic organizations and the price of casual agricultural labor in diVerent seasons. We noted the location of the village center using GPS.

Household survey
We conducted a household survey among 2500 households in 250 villages at baseline, and among a subset of 1000 households in 100 villages one year later. The survey was collected in the months of October and November, about ®ve months after harvest and right before planting for the next season. We interviewed the head of the household.
At baseline, we collected data on household composition, groups, networks and information sources, landholding, marketing, subsidies and credit 15 , and assets. At both baseline and one year later, we collected information on the adoption of ISFM technologies and yield expectations. One year after baseline, we also collected data on knowledge of ISFM technologies. Given the focus of this study, we detail the latter three modules below.
Adoption of ISFM technologies At baseline, we collected information on current use of ISFM technologies using an input-output plot-level questionnaire (pertaining to the previous, 2013-14 season). We focus on the technologies introduced by CDI and include geographically and in terms of treatment status. The households who left the sample have household heads who are slightly younger (0.01 years -signi®cant at the 5% level) and slightly more educated (0.05 years -signi®cant at the 10% level) but do not diVer in terms of household composition and asset wealth. To keep the sample size intact, these 51 households were replaced in the follow up survey using the random sampling methods outlined above. 15 17 Respondents could also opt for a "don't know" response, which we coded as incorrect. and Maertens (2017) to elicit yield expectations at both base and endline. We focus on soybean, groundnut and maize. At baseline, we asked the respondent: "Imagine that you would cultivate maize this coming year (and that maize is the only crop on the ®eld, i.e., no inter-cropping), how much maize do you think you would harvest on one acre of land" We recorded the answer in 50 kg bags of shelled or unshelled maize. We then repeated these questions for soybean (in 50 kg bags of shelled soybean) and groundnut (in 50 kg bags of unshelled, dried groundnut). 1819

Focus group interviews and interviews of extension agents
We conducted focus group discussions in ten villages before the CDI program, and one and two years after the program. We interviewed CDI clubs whom had been invited to a ®eld-day and clubs who managed demonstration plots. We followed best practices (see Morgan 1996 and Krueger and Casey 2008) and focused on learning about agricultural technologies, the club's activities and challenges faced, and relationships with extension agents.
We conducted semi-structured interviews with two government extension agents and two CDI extension agents in our study area, before the program and one and two years after the program. We focused on the constraints and opportunities facing extension agents, and their relationship with the farmers. 18 In our baseline data, while mono-cropping was the norm (with 75% of the plots mono-cropped), intercropping is common. Due to the complexity in generating per-acre beliefs on inter-cropped ®elds, we asked the respondent to imagine a mono-cropped ®eld. The unit was determined in qualitative interviews preceding the data collection as most common unit people think about for the crop. In addition, we recognise the diYculty in imagining the exact size of one acre of land, and in the formulation of this question we often referred to a 70 by 70 feet area or provided a comparison ®eld in the village. However, we do expect measurement error due to the lack of ability to imagine exactly the size of one acre, and also asked the respondent for the expected yields on a particular ®eld, instead of a per-acre basis (see also Bevis and Barrett 2016). 19 At endline, we expanded this module. To obtain a probability distribution, we ®rst asked the respondent to describe the best growing conditions, average conditions and worst conditions he/she could imagine for maize. In this round, we distinguished between hybrid maize and local maize. Respondents, in response, often noted variance in weather, pest pressure etc. Then, we asked him/her to state how much maize he/she would harvest under the best condition, the average condition and the worst condition, respectively. Finally, we asked the respondent to distribute ten equal size stones (each representing a 10 percent probability) in three equal-sized circles drawn on the ground, the ®rst circle representing the best condition, the second the average conditions and the third the worst conditions. We repeated these questions for soybean and groundnut.

Field observations on demonstration plots
We visited the demonstration plots two weeks after planting to record germination and record activities and inputs used to up to that date. Data on agronomic practices were recorded via a phone call with the lead farmer on a weekly basis between planting and harvesting.
During this weekly phone call we recorded any activity that had taken place, such as applying fertilizer or other inputs, and the number of club members and other visitors present for the activity (including whether the CDI extension agent was present). Rainfall gauges were mounted on each demonstration plot and the lead farmer was trained to record rainfall on a daily basis.
At harvest, we visited the demonstration plots and collected crop yield data. We recorded the stand count at harvest, the total biomass, grain yield and stover or leafy biomass. Grain moisture content was determined using a Mini GAC plus moisture meter. It is important to note that the club members were present during these on-®eld activities, and hence, are expected to have good idea of the planting and harvesting counts.

Soil sampling and analysis
The key indicators of land fertility in the study area are soil pH and organic matter content (see Snapp 1998). We collected soil samples from a total of 225 farmers' ®elds in addition to When dry, we sieved them through a 2mm sieve and recorded the soil texture using the hand feel method. 20 The farmers were selected as follows. First, we selected all ten sample farmers who live in villages where a CDI demonstration plot was set up. Second, we randomly selected 20 treatment villages and 9 control villages, and approached all 10 farmers in each village for soil sampling. Third, we selected 10 villages purposefully, for their relatively higher share of female-headed households and collected a soil sample from all ten sample households in these villages. This results in a total of 560 farmers, of which 225 live in villages which were covered by our follow up survey. As many farmers cultivate more than oen ®eld, we asked farmers to identify the ®eld they would be most likely to try new technologies on. Farmers are more likely to select ®elds they own, ®elds of mixed soil texture, and ®elds with a higher incidence of soil erosion and nutrient depletion (Regression results are available on request).
We use the SoilDoc program to analyze the sample pH, nitrate nitrogen (NO -), inorganic phosphorus (P), sulfur (S), exchangeable potassium (K) and electrical conductivity (EC), and active carbon (C). See Gatere at al. (2013) and Weil and Gatere (2015) for an introduction to SoilDoc. Note that we did not measure the total organic carbon matter, a measure of carbon contained within the soil organic matter, and generally accepted to be a good summary measure of overall soil fertility. Instead, we measured active carbon, which compared to total organic carbon, is more sensitive to management eVects, and more closely related to soil productivity and biologically mediated soil properties, such as respiration, microbial biomass and aggregation (Weil et al. 2003).

Descriptive statistics
Appendix Table 1 introduces the 100 villages which were re-visited one year after the baseline (in Columns (1) through (3) The households in our study area are land-poor and dependent on rainfall agriculture.  On average, households own 3.5 acres of land. While this ®gure excludes outliers above the 95 percentile, it might still appear high. It is important to note however that the median ®eld size is small, 1.5 acres, and likely to be an over-estimate (these are self-reported acreages, which are often over-estimated in the case of smaller plots, see Bevis and Barrett 2016 for a discussion).
Plots of land are small and population density is high in the area, according to the respondents' own account often the result of generations of plots being sub-divided for inheritance.
A lack of land can further reduce land quality, as leaving ®elds fallow, or using crop rotation, might no longer be options for many households. Soil fertility in the study area is low, and declining. Soils are classi®ed as Ferralsols, Lixisols and Plinthosols (FAO Harmonized World Soil Database). 21 A common feature of these soil types is that they depend on the addition of organic and inorganic matter to improve soil structure and overall fertility. Lacking these, soil fertility will be low. Our respondents report soil fertility problems (see Table 2  The soil sample analysis results, summarized in Table 1 -Panel C -con®rm these perceptions. Soils have an average Ph of 6.12 (with 70 percent of soils tested within the optimal range, between 5.5 and 6.5), so only slightly acidic. Active carbon is, on average, 423 mg/kg soil, and is low to very low in 30 percent of soils tested (with a concentration of less than 350 mg/kg soil), and at a medium level in another 40 percent (with a concentration between 350 and 500 mg/kg soil). This would indicate that organic carbon is mostly suYcient to maintain soil structure, but still low. These ®ndings are consistent with widespread nutrient de®ciencies. All soils tested are Nitrogen (N) de®cient (less than 42 mg NO -/kg soil), 77 percent are Sulphur de®cient (with concentration is less than 10 mg/kg soil), 52 percent are Phosphorus (P) de®cient (with concentration is less than 0.3 mg/kg soil) and 33 percent are Potassium (K) de®cient (with concentration less than 20 mg/kg soil). All-in-all, over 50 percent of soils were de®cient in three or more nutrients. The intra-class correlation between observations of the same village is below 0.5 for measures of Nitrogen and Potassium, indicating signi®cant within village variation in these measures. subsidy program targeting small-holder farmers). Animal manure had been used by 30 percent (but compost was not common), 14 percent had incorporated crop residue in the soil, and 9 percent had planted fertilizer trees. The use of other inputs is not very common.
Less than 5 percent had used pesticides, herbicide or fungicide; and only two soy growing farmers had inoculated the seed.
Finally, Table 1-Panel E reports on the farmers' baseline yield expectations for maize and soybean. Farmers expect to harvest, on average, 3,480 kg/ha (or 29 kg bags per acre) of shelled maize. This is signi®cantly larger than the average yield on the farmers' plots in 2013-14, which was 1,750 kg/ha for mono-cropped plots. However, it is important to keep in mind that: (i) the yield expectation distribution is not normal, with a long left tail -in eVect the median of the distribution is 3,088 kg/ha, and (ii) the beliefs are re¯ect also perceived acreage, which, in our data, are over-estimated for smaller plots and under-estimated for larger plots (we know this as for a sub-set of the plots, we also have the GPS measured acreage). Farmers expect to harvest, on average, 1,608 kg/ha or (13 kg bags per acre) of shelled soybean. This is larger than the average actual yield in 2013-14 (312.5 kg/ha, on mono-cropped plots). Again, the same disclaimers apply and it should be noted that the median yield expectation is 1,235 kg/ha.
As a comparison, Table 2 presents the yields obtained on the demonstration plots in 2014-15. We focus on the maize and soybean subplots. Maize grain yield was variable and ranged from 452 kg/ha under control treatment to 8,990 kg/ha under best practice agronomy. The latter is within the range of potential yields for maize, ranging from 6,000 to 14,000 kg/ha (depending on the variety, see MAIFS, 2012). Overall, the use of best practice agronomy practices increased maize yield by 62 percent and 25 percent over the control and farmer practice treatments respectively. 22 However, no diVerences were observed between best practice agronomy treatments with and without fertilizer trees.
DiVerences in grain yield of soybean between the treatments and sites are also signi®cant.
ields range from 0 (crop failure) to 2,218 kg/ha. The use of best agronomy practices increased the yield of soybean by 50.4 percent over the control. 23 Overall, the yields of soybean are somewhat lower than the potential yield of 2000-4000 kg/ha, but in the range of the attainable yields on smallholder farms (1,500-2,500 kg/ha) with good agronomic practices (MAIFS, 2012). 22 Based on within demonstration plot analysis, excluding outliers. 23 Based on within demonstration plot analysis, excluding outliers. To summarize, respondents were well aware of the low soil fertility in the region, and note signi®cant nutrient depletion in their own soils. Accordingly, yield expectations are low, and well below both attainable yields (albeit considerably above the actual yields of the previous season, 2013 -14). Despite this awareness, adoption of ISFM technologies has been relatively low, especially of certain information intensive technologies such as the use of pesticides, herbicides, fungicides, fertilizer trees, compost, and inoculation. Our central hypothesis is that this low uptake is partially due to a lack of knowledge, and that exposure to demonstration plots and ®eld-days can remedy this lack of knowledge.
Note that a demonstration plot eVectively combines three sources of learning (self, social and from external agents), and during the baseline focus group interviews farmers in CDI clubs, by their own account, noted to be keen on establishing a demonstration plot in their village to learn about production processes. Demonstration plots have been central to much of Malawi's extension history (see Knorr et al. 2007), and farmers, prior to the program, stressed that the best way to learn is to work on a demonstration plot together.
When we spoke with the farmers (during the focus groups) one year, and even two years after they had begun working on a demonstration plot together, many were able to recall the exact names of the ISFM inputs used, report the amounts used and explain how the inputs should be applied. They stated feeling "comfortable" with the techniques. Farmers who had only attended the ®eld-days stated that they were impressed by the productivity of the crops presented at the ®eld-day, and reported that they learned about the importance of herbicide and pesticide (and in some cases inoculants) for soybean, as well as plant spacing and the importance of using crop residuals for mulching and plant spacing. In contrast, few ®eld-day participants were able to recall the details: which input was used; the amounts required; and the method of application. This is consistent with what farmers told us before the CDI program started in the region: while learning from extensions services at ®eld-days and the radio is common, they also noted that they rarely immediately adopted the new technologies after visiting a ®eld-day or hearing something on the radio, as through these channels they are less aware of the intricacies of how the technologies work and less certain as to how they can be applied on their ®elds.
We tested the knowledge of farmers after the treatment group had participated in the program. In Table 3 we present summary statistics for this knowledge test (recall, the test was administered one year after baseline). The overall knowledge score is 7.87 -out of 20 (with a standard deviation of 2.30). We note that while most respondents are aware of the general bene®ts of soy, fewer know the details of the production process in terms of which pesticides and fungicides one should apply following best practices. The share of correct answers drops even further -to under 10 percent -when we ask the respondent to tell us about the details of soybean input preparation and application. For maize, a crop with which farmers have extensive experience, farmers seem to be aware of certain ISFM technologies, such as the use of crop residues and fertilizer trees but have limited knowledge of the details of the production process as well.
In the next section, we will test for the impacts of the demonstration plots and ®eld-days on knowledge and adoption plans.

Impact of the program on adoption and knowledge
To estimate the impact of the CDI program, one would ideally run a regression such as speci®cation (1) linking outcomes Y ij of farmer i from village j, on whether or not the farmer participated in demonstration plots D ij or ®eld-days F ij : However, while being invited to a farmer ®eldday is randomized at the village level, a critical aspect of participation is a choice: Farmers have to sign up for CDI clubs in order to become eligible for the CDI activities. As one is unlikely to be able to control for all relevant confounding factors -many are unobservable to the researcher such as, climatic factors and personal attributes -we might expect a correlation between ij and F ij and D ij , resulting in omitted variable bias.
Appendix Tables 2 and 4 shed light on this participation decision. Recall that 47 out of 55 treatment villages formed clubs. Appendix Table 2 presents, in Columns (8)  In Appendix Table 3, Column (11), we show the result of t-tests testing the baseline diVerences between club and non-club members in the villages which formed clubs. Clubmembers are diVerent from non-club members in many dimensions, and that this diVerence is both statistically and economically signi®cant. Compared to non-club members, clubmembers are better educated, have larger families and more land, and are also more likely to take agricultural credit. 2425 Our estimation strategy takes this self-selection into account by constructing two comparable samples: the sample of households which received the CDI program and a comparable sample of households, in the control villages, which do not received the program. Using this approach, we essentially assume that the treatment villages are comparable to the control villages; and we can use the latter to construct a similar sample (we presented evidence of the similarity between treatment and control villages in Appendix Table 1).
To do this, we ®rst predict the probability of club membership using the treatment villages sample only: The control variables used in this ®rst step include all village level characteristics included in Appendix Table 1 (X j ), all household level characteristics included in Appendix Table 3 (and the square terms of the non-binary variables) and baseline adoption indicators (X ij ).
We do a reasonable job predicting club membership -with about 70% correctly classi®ed.
We use the estimated coeYcients to then do an out-of-sample prediction for the control villages. We summarize the results of this process in Appendix Figure 1 where we present the resulting kernel distribution of this predicted probability for both the treatment villages and the control villages. As expected, the mass of the distribution of the control group is situated to the left of the median of the treatment group: This is due to our sampling design -we strati®ed the sample in the treatment villages and likely over-sampled club members.
We then create two groups in each set of villages: (would be) club members and (would be) non-club members using the cut-oV value of the predicted probability of 0.5. This includes 241 individuals in the treatment group and 173 individuals in the control group.
Appendix Figure 2 presents the distribution of the predicted probability of these individuals.
This time, the distributions are comparable. A two-sample Kolmogorov-Smirnov test for equality of distribution functions cannot reject equality between them with a P-value of 0.22.
We now compare the club members in the treatment villages with the would-be club members in the control villages using: Where y ij is the outcome variable of farmer i from village j; T ij = 1 if the individual is in a club which managed a demonstration plot, and = 0 otherwise; T ij = 1 if the individual is in a club which was invited to a ®eld-day, and = 0 otherwise. We clustered the errors at the village level, and used a bootstrap procedure to account for the two stages in the estimation process (see Abadie and Imbens, 2009). 26 The dependent variables y ij include the planned adoption (2015-16 season) of: soybean, inoculation of soybean, groundnut, hybrid maize, herbicide, pesticide, fungicide, inorganic fertilizer, fertilizer tree, intercropping, animal manure, crop residue and compost and whether or not each one of the questions in the knowledge test was answered correctly. We also compute an adoption score (out of 13) and knowledge score (out of 20).
Selection into T ij (as opposed to T ij ) might still be a concern. However, as noted earlier, the demonstration plot villages are comparable to the other treatment villages (see Appendix Table 2). Appendix Table 4 shows that this lack of selection also holds at the individual level: demonstration plot farmers are comparable to the CDI club members who did not manage demonstration plots.
Tables 4 and 5 present the results. Table 4 shows that demonstration plot participation increases adoption of ISFM practices by 0.79 points, which is about 22 percent (Column (1)), while being invited to a ®eld-day does not produce such (statistically signi®cant) result.
Columns (2) through (14) present the results of a series of linear probability models. We note that being a member of a demonstration plot club increases the chances of inoculating soybean (at the 5 percent level), using hybrid maize (at the 1 percent level) and planting fertilizer trees (at the 5 percent level). Participating in a demonstration plot improves knowledge. Table 5 reports an increase by 0.63 points, which is about 8 percent (P-value of 0.11), with statistically signi®cant increases in knowledge of inoculation and pesticides in particular (note that we do spot a negative eVect on question (14) -but the negative, tricky, formulation of that question cautions against over-interpretation). In contrast, being invited to a ®eld-day does not (statistically signi®cantly) alter one's knowledge.
The results so far seem to suggest that the ®eld-days were not very eVective. Results of focus group discussions however indicate that ®eld-day participants learned about the production processes of some of the more labor-intensive ISFM technologies, such as mulching and optimal plant spacing. Farmers reported that mulching, for instance, was a useful technology to combat striga (a common weed in maize ®elds which can cause heavy crop losses), and also the very common drought spells. This suggests that these farmers, who are likely constrained as to what they can focus on during one day, focus on these technologies that they are most likely to successfully implement. Many of the recommended inputspesticides, herbicides and inoculants -are not available in local markets. As few households own cars or motorbikes and public transport is limited, distance could represent an additional barrier to participation in input markets. Many farmers also noted that, even if such inputs were readily available, they would not have the funds available to purchase these inputs.
Hence, ®eld-day participants, by their own accounts, often focused on the labor-intensive technologies, rather than the credit-intensive technologies. This would suggest that both groups are learning, but learning should be considered a choice, a choice which is constrained by factors such as credit and time.
In the next section, we present a learning model which follows this narrative, capturing the motivation and constraints to learn about ISFM technologies.
As a robustness check, Appendix Tables 5 and 6 present the results including the ®rststage set of control variables in the second stage equation (3) and Appendix Table 7 presents the results of a farmer ®xed eVects estimation. The results of these speci®cations are broadly consistent with the ones presented in Table 4 and 5.
The validity of our estimation depends on our ability to create a comparable control group of would-be club members in the control villages. Systematic errors in our ®rst stage could result in over-classi®cation (placing people in the comparable group whom would not have joined, possible biasing our estimates upwards), or under-classi®cation (placing systematically better people in the comparable group, possibly biasing our estimates downwards).
However, the fact that the predicted distributions of treatment and control group largely overlap (Appendix Figure 2), and that adding additional control variables does not oppose our results is encouraging. 27 We conclude this section with a note on participation in the program. Tables 4 and 5 present the intent-to-treat eVects. In the case of demonstration plots, this does not matter much: all CDI club farmers who were invited to participate in demonstration plot activities actually did. In the case of ®eld-days, the estimated eVect should be interpreted as the intent-to-treat eVect, as not all clubs invited to the farmer ®eld-days participated.
In Appendix Table 4 we presents the P-values of a t-test, comparing farmers who attended ®eld-days, which those CDI club members who did not attend ®eld-days (Column (7)). We note considerable diVerences: Farmers who attend ®eld-days are higher educated, own more land and are overall wealthier and better connected, and are more likely to use credit. This is consistent with the observation that in most clubs, it is the lead farmer who attended the ®eld-days. Our estimates should be interpreted as the average eVects, across all club members, immediately after the intervention. As CDI requested the club members who attended the ®eld-day to share the information with the other members, this, in a way, is also a measure of the success of this process. As we discuss in the concluding section, in future research, which we aim to conduct ®ve years after the ®rst intervention, we aim to provide a more nuanced and complete picture of this social learning component. Yields We introduce three production technologies: a capital-intensive technology in-27 Appendix Tables 5 and 6 are demanding speci®cations; and many of the bootstrap rounds did not converge (see also Horowitz 2001). 28 We abstract away from the notion that knowledge decays over time, which would empirically result in a false acceptance of the null hypotheses for in particular demonstration plot farmers who might have learned earlier.

A model of learning
We also do not model the fact that demonstration plot participants might share the proceeds of the demonstration plot, which might further incentive them to learn.
dexed K, a labor-intensive technology indexed L and a traditional technology which represents a risk-free technology. Each risky technology K and L has average per-acre payoVs (yields) of µ j (j Î {K, L}) -yields associated with the risk-free technology are normalized to one. Furthermore, we assume that the yields from the capital-intensive technology are higher than the labor-intensive technology: µ K > µ L > 1.
Two-stage approach The farmer ®rst establishes a belief of the yield (we assume that the farmer does not need to learn about prices), and then invests resources to generate knowledge, i.e., learn about the production process, and makes decisions as to which technologies to adopt.
Learning about yields We assume that the true value of µ j is unknown to the farmer.
Let the prior belief about µ j , µ j , be normally distributed, centered around the true value, with variance s µ . So each farmer's prior represents a draw of this distribution.
When observing yields on the demonstration plot, either in the village, or at a ®eld-day, the farmer receive an unbiased information signal j . This signal is the sum of the true yield µ j plus a normally distributed noise term: The variance in the noise term, s h , can be farmer-dependent, and signi®cant, if the farmer believes that the soil and climatic conditions of the demonstration plot are dissimilar to his own conditions. However, to maintain simplicity, we will abstract from this farmerdependency in our notation. Assuming the farmer uses Bayesian updating, then noisier signals will be down-weighted in posterior beliefs. Posterior beliefs, µ p j are characterized by: Thus, the posterior beliefs will decrease (relative to the prior) if the signal received during the ®eld day is less than the prior and increase otherwise. However, the degree of change in posterior beliefs depends on the size of j and the farmer's perception of the relative noisiness of the ®eld-day signal.
Equation (6) implies that the posterior belief represents the weighted average between the prior belief and the signal received. In regression terms, this would imply that for farmer i the posterior beliefs of technology j can be expressed as: Production process Production requires the use of inputs (e.g., amounts and timing of fertilizer, herbicide, labor). If inputs are inaccurately applied farmers incur a knowledge penalty. Speci®cally, let q * j indicate the optimal amount of input required for technology j. If the farmer applies input q j instead, he incurs a (per-unit) loss equal to (q jq * j ) < 0 for all q j = q * j . Learning about the production process The optimal input use , q * j , associated with technology j is unknown. Let the belief, q j , again be normally distributed, centered around the true value (q * j ): Note the dependency of the belief on learning eVort, denoted . In particular, the beliefs are more precise if the farmer applies a discrete learning eVort, j = 1 (compared to the situation where the farmer applies no learning eVort, i.e., j = 0): Note that when the farmer's belief over the target input is imprecise, the knowledge penalty will be large in expectation (E[(q jq * j ) ] = s q ( j )) -and as the farmer gains production knowledge, his knowledge penalty decreases in expectation.
Payo s The farmer holds initial wealth w and is tasked with choosing the optimal amount of wealth to invest in each production technology, x j , each unit of which costs p j to purchase. At harvest, the farmer receives the following payoV: where j represents the cognitive cost associated with gaining knowledge of production technique j and only contributes to payoVs when learning eVort is applied (1 represents the indicator function). The amount invested in the traditional technology is denoted x , and its cost p (recall that the yield of the traditional technology was normalized to 1).
Expected payo maximization Given the presence of credit market imperfections, the farmer's problem will be one of maximization of expected payoVs given a budget constraint.
The farmer chooses values of x j , q j , and j given values of µ p j and p j for both j Î {K, L}. The choice of q j is straightforward if the farmer selects a positive value for x j : he will select q j = q j to minimize square loss in expectation. Thus, the farmer will need to decide the amount of x j to use in production and whether to place eVort in learning about j's production process ( j ). Let P represent expected payoVs. 29 The farmer's problem can now 29 We abuse notation slightly in expression (11) where P refers to the expected payoV, and not the of the expression. be summarized as: Solution Given the discrete nature of learning eVort in our setup, we can ®nd the solution to problem (11) by backward induction. We ®rst determine the optimal value of x j given each choice of j ; and then plug this value back into the objective function of problem (11) to determine which levels of eVort result in the highest payoV.
First order conditions on x j yield the following demand function: Note that, x * j is determined by values of µ p j and j . Expression (12) intuitively shows that demand for production technology x j is increasing in the net (perceived) returns of the production method and knowledge of the production process (recall, increased knowledge indicates a smaller value for s j ( j )). Demand is decreasing with the penalty associated Notice that P (x * K ( K ), x * L ( L ), L , K ) can be calculated for each of the four discrete choices a farmer can make. Thus, the farmers optimal learning eVort vector, * , can be characterized by: From (11) it is obvious that the farmer will not excert any eVort in learning the new technologies if the yields (net of costs) are smaller than 1 -the yield of the traditional technology. The choice of eVort, in all other circumstances will depend on the net bene®t from learning about technology j relative to alternative technologies as described in equation (13); i.e., the bene®t that comes from decreasing uncertainty about the optimal input level net of the cost of learning.
Holding the price of inputs and cost of learning eVort ®xed, this would imply ®nding a positive relationship between posterior yield beliefs and knowledge with any given technology.
In regression terms, we would estimate the following equation: Choosing what to learn A simple association between posterior beliefs and knowledge may be insuYcient to explaining the nuanced ways in which farmers choose to learn about new technologies. To analyze the solution further, we make simplifying assumptions to the parameters in the model. Recall that we assumed that the capital-intensive technology generate higher average returns, i.e. µ K > µ L . We now, in addition, assume that they are also more expensive to purchase, or: p K > p L . Furthermore, we denote j , the average pro®t gain from an added unit of technology j, (i.e., j = µ jp j ) and we assume that K > L . Finally, we assume that the cost of learning about technology L is the same as that of learning about technology K and the knowledge-bene®t is similarly equivalent: When the budget constraint binds, then > 0, and (11) can be solved by entering optimal values of x * j into the budget constraint and equating the left and right-hand sides. In other words, we obtain a solution for when w -jÎ{K,L} p j x * j = 0 by replacing expressions of x * j with the expression in equation (12).
As can be seen, the Lagrangian multiplier, or borrowing-constraint penalty, , is decreasing in wealth and exhibits a complex relationship between input price and knowledge.
Speci®cally, there is a cross-technology knowledge-uncertainy trade-oV that manifests itself in the multiplication of the knowledge-penalty of one technology with the price of the second technology. Depending on the underlying parameter space, this trade-oV will lead to selective learning about one technology over the other if the wealth constraint is binding.
We can now compute the expected payoVs for the optimal solution using equation (11).
In particular, when borrowing constraints do not bind, then = 0 and x j can be computed for all combinations of learning eVort using equation (12). Plugging this information back into (11)) will, by comparing across the four alternatives, yield the optimal combination of eVort and technology uptake; and resulting expected payoV.
When borrowing constraints do bind, then is given by equation (15) and we can similarly compute expected payoV for all combinations of learning eVort. Then, for any parameter combination, we compare expected payoVs across the four alternatives and identify P * , as de®ned by equation (13), as the maximal value across the alternatives. Figure 1 shows the relationship between the optimal expected payoV, which we denoted by , and the farmer's initial wealth w for each of the four possible learning combinations.
We graph each learning combination separately. The largest expected payoV is determined by the particular value of initial wealth each farmer holds. Notice that expected payoVs are monotonically increasing in wealth but that there are thresholds at which farmers may choose to learn about neither technology K or L (lowest wealth category), either one of K or L (mid-tier wealth), or both K and L (unconstrained by wealth). Thus, we should only expect wealthy farmers to learn about the most capital-intensive components of new technologies. However, this is strongly contingent on the assumption that k >> L. If beliefs about k are not suYciently high, then even a wealthy farmer will choose not to learn about technology K because it is preferable to specialize in the labor-intensive mode of production. 30 The regression implications of Figure 1 could be captured by: where iK captures the pro®t gain observed by farmer i and w i captures farmer i's wealth.
j is our coeYcient of interest: If farmers have no reason to believe that an capital-intensive method of production will generate higher pro®ts, then they will not choose to learn about this method of production. However, if farmers believe the new method will be pro®table, we only expect learning to take place when farmers possess suYcient wealth. Thus, we hypothesize that K will be positive.

Analysis of the learning process
In this ®nal section, we present the results of the regression speci®cations presented in the previous section, further documenting the learning process.

Correlates of yield expectations
In Table 6 we estimate regression speci®cation (7) and regress endline yield expectations for soybean, hybrid maize and local maize (in kg/ha) on yield expectations at baseline (in kg/ha) 30 This is demonstrated in Appendix Figure 4, which relaxes this assumption and shows that farmers will varies the diVerence between and while holding ®xed at a suYciently high level (allowing adoption of the capital-intensive technology). Notice that the farmer will never choose to learn about when is suYciently small -and certainly will never learn about when and the performance of the local demonstration plots (in particular, the mean diVerences between BPA and control subplots; also in kg/ha, adjusted for the moisture content) and a series of control variables. We split the sample according to absolute diVerence in soil quality between the local demonstration plot and the farmer's plot (measured by active carbon in mg/kg): Columns (1), (3) and (5)

Correlates of knowledge
In Table 7 we estimate regression speci®cation (14). We use demonstration plot yields, rather than self-reported beliefs, to capture the farmer's beliefs about pro®tability. We plot the relationship between demonstration plot yields and learning, as measured through the crop-speci®c knowledge score for the demonstration plot participants (again, including 31 We focus on analysis of the CDI farmers in demonstration plot villages here for three reasons: (i) There is little variation in the demonstration plot performance observed by the ®eld-day participants as they attended only one of two ®eld-day sites (and we collected data only at one site which was within our study area), (ii) We did not collect soil data among all farmers in the ®eld-day villages, and (iii) We noted that the quality of the yield expectation data was poorer in the non-demonstration plot villages. We attribute this lack of accuracy to the fact that farmers were asked to imagine the yield on a plot of one acre. As noted earlier, for some farmers, one acre is diYcult to imagine. Farmers who attended the demonstration plot, having measured out the demonstration plot, and also having had one of their own plots measured, appeared to have an easier time with this type of question. Still, in interpreting the results it is important to recall that yield expectation question asked did not explicitly refer to median, average, or another moment -and hence the interpretation was up to the respondent. a series of control variables). We ®nd a statistically signi®cant, and positive, relationship between the yields measured, and the average learning of the demonstration plot participants (soybean is almost statistically signi®cant at the 10 percent level) -in Columns (1) and (3) for both soybean and hybrid maize. The magnitude indicates that an increase in 50 kg/ha increases the knowledge by approximately 5 percent (relative to the average score).
To interrogate the causal interpretation of the results, Columns (2) and (4)  In Appendix Figure 3 we plot the rainfall distributions for demonstration plots. The rainfall distribution is quite variable. We de®ned three statistics of the distribution: start of the rainy season, the total amount of rainfall, and the number of¯ood days (de®ned as > 50 mm/day). The latter, in particular at the start of the season, can be quite damaging for germination (Wenkert et al. 1981, Martin 1991, Githiri et al. 2006). We note the anticipated correlation between rainfall patterns and knowledge of soy (but no such correlation for hybrid maize, which suggests that either rainfall and germination rates were not as closely related, or farmers' learning was more uniform across the various plots, perhaps because maize is a historically important crop).
In Table 8 we present the relationship between learning, also measured using the knowledge score, and yield expectations at endline for farmers were invited to participate in ®elddays. As in previous analyses, we split the knowledge score into knowledge related to soybean and knowledge related to maize. But this time, based on discussion of Figure 1, we include only the credit-intensive technologies in the soybean score, and the labor-intensive technologies in the maize score. We focus on one control variable: farmer wealth. We use "having obtained input credit in the previous season" as a proxy for the relevant wealth variable.
If the farmer answers no to this question, the farmer is likely more credit constrained. We note a positive correlation between the endline beliefs and the knowledge score for soybean.
The coeYcient on our credit-measure suggests that farmers who are not credit constrained are more likely to learn something about the (credit-intensive) soybean technologies. The knowledge of soybean best practice conditional on a set of individual characteristics. 32 According to our hypothesis, only wealthy farmers on the right-hand side panel should exhibit higher levels of learning, which we are able to con®rm.
Note that Figure 2 considers soybean only. This relies on the assumption that soybean production is more capital-intensive and, therefore, more expensive. However, as a cash crop, it also potentially generates higher pro®ts. We do not expect knowledge scores for hybrid maize to change with wealth since its production does not require as costly of inputs as soybean production under CDI's demonstrated best-practices (and eVectively, this is what we ®nd in Table 8).
It is notable that credit and wealth matter both for ®eld-days and demonstration plots.
Field-day participants likely face a signi®cantly higher cost of learning (within that day); and for them, the trade-oV between which technology to focus on likely much more salient.
Focus group participants noted the rushed nature of the ®eld-day and 'information overload' repeatedly. But even for the demonstration plot participants, wealth appears to matter, and learning might be concentrated among the wealthy farmers motivated by observed demonstration plot success.

Conclusion
We studied farmers' learning about agricultural technologies based on diVerential exposure to commonly used extension methods that range in their intensity of interaction. We ®nd that farmers who participated in farmer-led demonstration plots learn about the production ®nd that what farmers learn is conditional on the degree to which they are credit-constrained. 32 We apply a control-function approach to variable construction by using the predicted error term on the vertical access after regressing farmer knowledge against gender, age, years of education, household size We re-iterate that this study is not an evaluation of the relative eVectiveness of ®elddays versus demonstration plots and our results do not suggest that ®eld-days should be discarded as a strategy. We make suggestions below building on what we have learned about the learning process to suggest improvements in the ®eld-days.
First, farmer ®eld-days may provide too much information in too short a time period, giving farmers insuYcient chance to absorb the details. This implies that, at ®eld-days, farmers should be given tools which will allow them to learn the information presented more eVectively. Examples might include pamphlets with pictures of the inputs used and measuring spoons to measure the correct amounts of inputs. 36 Second, the fact that farmers' learning appears to be constrained by markets suggests that agricultural extension might need a re-coupling with market activities, and in particular, credit interventions in order to be eVective. In Malawi, extension agents used to perform an additional role as regional credit oYcers. While con¯ict of interest should be avoided, Finally, it may be that ®eld-days could be used in sequence with demonstration plots or other more intensive methods of teaching farmers. The ®eld-days could serve to introduce a new technology and to focus on its broad features, demands, and processes and this initial introduction could be followed by methods employing more detailed exposure, perhaps based on farmer demand.
We conclude with a note on further research. While the limited time frame of this study

Economics of Land Degradation and Improvement -A Global Assessment for Sustain-
able Development. Springer, pp. 215-260.

Mapping Adoption of ISFM Practices Study: The Case of Keny,a Rwanda and Zambia.
IFPRI AND IITA Report.
[93] Nourani, . 2016. "Social Network EVects of Technology Adoption: Investing with Family, Learning from Friends and Reacting to Acquaintances". Working paper.  We dropped farmers with more than 13 acre (95% percentile) for these statistics. 2: We asked the respondent about the three main sources of information about agriculture, if the government extension agent was mentioned, we coded the first answer = yes (no otherwise); and if another farmer in the village was mentioned, we coded the second answer = yes (no otherwise). 3: We asked the respondent whether he/she took any loans in 2013-14. Note 4: We elicited characteristics of each field and averaged the responses across fields for each farmer (Note that the sample only includes respondents who own at least one field). 5: Panel C only includes the households who had a soil sample analysis done, 252 farmers.

Panel E -Yield expectations
Harvest of maize expected (in kg/ha) 1,000 3,599 2,288 Harvest of soy expected (in kg/ha) 993 1,608 1,336 Notes: Inoculation statistics are conditional cultivating soybean; hybrid maize use is conditional on cultivating maize. Recall that yield expectations were elicited in 50 kg bags per acre, or 50 kg bags per plot. Due to issues with farmers' estimation of acreage, we used the former here, which we converted to kg/ha, excluding outliers above the 95 percentile.  Table 2 focus on two crops: Soybean and Maize. Appendix A presents the layout of the demonstration plots. One will note that one would have a total of 24 Soybean Control sub-plots and 24 Soybean Best Practice sub-plots in this design. Our sample size is 22 -as one of the demonstration plots did not correctly follow the guidelines. For Maize, which is planted on all three types of demonstration plots, one would have 38 Control sub-plots, 7 Farmer Practice, 7 Best Practice sub-plots, and 48 Best Practice with fertiliser trees sub-plots. Our sample design only has 37 Control sub-plots as, again, one of the demonstration plots did not correctly follow the guidelines.   13), soy (binary variable), inoculation soy (binary variable), groundnut (binary variable), hybrid maize (binary variable), herbicide (binary variable), pesticide (binary variable), fungicide (binary variable), inorganic fertilizer (binary variable), fertilizer tree (binary variable), intercropping (binary variable), animal manure (binary variable), crop residue (binary variable), and compost (binary variable). These refer to planned adoption in the 2015-16 season. The independent variables are whether or not the individual is in a club which managed a demonstration plot, and whether or not the individual is in a club which was invited to a farmer field day. The estimation uses a two-step procedure. The first steps uses the reported club membership at endline in the treatment villages to predict who would be most likely to join a CDI club. The second step uses all individuals in both treatment and control group whose predicted probability is larger than 0.5. This includes 241 individuals in the treatment group and 173 individuals in the control group. The control variables used in this first step include all village level characteristics included in Appendix Table 1, all household level characteristics included in Appendix Table 4 (and the square terms of the non-binary variables) and baseline adoption indicators. Bootstrapped clustered errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Appendix Table 6 present the results including the control variables in the second step.  20), and the 20 knowledge questions listed in Table 3 (binary variables). The independent variables are whether or not the individual is in a club which managed a demonstration plot, and whether or not the individual is in a club which was invited to a farmer field day. The estimation uses a two-step procedure. The first steps uses the reported club membership at endline in the treatment villages to predict who would be most likely to join a CDI club. The second step uses all individuals in both treatment and control group whose predicted probability is larger than 0.5. This includes 241 individuals in the treatment group and 173 individuals in the control group. The control variables used in this first step include all village level characteristics included in Appendix Table 1, all household level characteristics included in Appendix Table 4 (and the square terms of the non-binary variables) and baseline adoption indicators. Bootstrapped clustered errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Appendix Table 7 present the results including the control variables in the second step.  (1) and (2) refer to soybean, Columns (3) and (4) to hybrid maize and Columns (5) and (6) to local maize. We split the sample according to absolute difference in soil quality between the local demonstration plot and the farmer's plot (measured by active carbon in mg/kg): Columns (1), (3) and (5) include the observations below the median of this distribution ("similar"), while Columns (2), (4) and (6) include the observations above the median of this distribution ("different") (the median absolute difference is around 120-130 mg/kg). The yield expectations at baseline refer to the baseline expectations of soy in the case of Columns (1) and (2) and maize in the case of the other Columns. Similarly, the differences between BPA and control subplots refer to the mean differences between these two treatments on the local demonstration plot. Other control variables included but not reported: Gender household head, age household head, education household head (years), number of household members, number of adult household members, maximum education level in the household, acreage of land owned, value of all assets (excluding land) and whether the household cultivated hybrid maize in 2013-15 (for Columns (3) through (6) only). Sample includes the club farmers in the demonstration plot villages. Whether or not farmer is in a club is determined by the self-reported club status at endline. Village-clustered errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1  (3) and (4) to hybrid maize. The knowledge score for soybean is a number out of 6, while the knowledge score for hybrid maize is a number out of 8. Columns (1) and (3) consider use yield on the BPA subplot as the main independent variable of interest, which refers to the maximum yield on the BPA subplots on the local demonstration plot. Columns (2) and (4) (2) to hybrid maize. The knowledge score for soybean is a number out of 5 (we excluded the first generic question on soybean), while the knowledge score for hybrid maize is a number out of 3 (we focus on the labour-intensive techniques). Other control variables included but not reported: Gender household head, age household head, education household head (years), number of household members, number of adult household members, maximum education level in the household, acreage of land owned, value of all assets (excluding land), relevant yield expectations at baseline, and whether the household cultivated hybrid maize in 2013-15 (for Columns (3) through (6) only). Sample includes the club farmers in the treatment villages (excluding the demonstration plot villages). Whether or not farmer is in a club is determined by the self-reported club status at endline. Village-clustered errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1 Parameters for the model are the following. π L = 5; π K = 5.8; σ(0) = 3; σ(1) = 1; e = 3.

Figure 1: Predicted relationship between expected payoff and wealth (through applying various levels of effort)
Notes: Includes demonstration plot farmers only.     (3)) and the complementary set of non-demonstration plot villages (Columns (4) through (6)). Column (7) presents the P-value of a t-test with unequal variances between these two groups. The rest of the table includes the villages which formed CDI clubs (Columns (8) through (10)) and the complementary set of villages which did not form CDI clubs (Columns (11) through (13)). The sample includes all villages which were revisited after one year and belong to the treatment group, N=55.  (3)) and the complementary set of non-demonstration plot villages (Columns (4) through (6)). Column (7) presents the P-value of a t-test with unequal variances between these two groups. The rest of the table includes the villages which formed CDI clubs (Columns (8) through (10)) and the complementary set of villages which did not form CDI clubs (Columns (11) through (13)). The sample includes all villages which were revisited Appendix  Table 3), for villages without clubs (Columns 4 through 6), the non-club households in the villages with clubs (Columns 7 through 9) and the club households in the villages with clubs (Columns (10) through (12)). Column (13) presents the P-value of a t-test with unequal variances between the first and second sub-group; Column (14) presents the P-value of a t-test with unequal variances between the second and third sub-group. Note that this table only includes the households in the villages in the random sample, and not in the purposefully selected sample of 10 demonstration plot villages. Total sample size = 450 households in 45 villages.

Treatment villages
Villages without clubs Appendix  Table 3), for villages without clubs (Columns 4 through 6), the non-club households in the villages with clubs (Columns 7 through 9) and the club households in the villages with clubs (Columns (10) through (12)). Column (13) presents the P-value of a t-test with unequal variances between the first and second sub-group; Column (14) presents the P-value of a t-test with unequal variances between the second and third sub-group. Note that this table only includes the households in the villages in the random sample, and not in the purposefully selected sample of 10 demonstration plot villages. Total sample size = 450 households in 45 villages.

Villages with clubs
Villages with clubs Non-club members Club members Appendix  (7), we distinguish between households who participated in fielddays and households who do not participate in fielddays. Column (7) presents the P-value of a t-test with unequal variances between these two groups. In Columns (2) through (14) we distinguish between households who participate in demo plots and households who do not participate in demo plots. Column (14) presents the P-value of a t-test with unequal variances between these two groups. The full sample corresponds to all self-declared club members in the treatment villages (including Tongolele) in the 100 vilages.
Do not participate in fieldday Participate in fieldday Appendix  (7), we distinguish between households who participated in fielddays and households who do not participate in fielddays. Column (7) presents the P-value of a t-test with unequal variances between these two groups. In Columns (2) through (14) we distinguish between households who participate in demo plots and households who do not participate in demo plots. Column (14) presents the P-value of a t-test with unequal variances between these two groups. The full sample corresponds to all self-declared club members in the treatment villages (including Tongolele) in the 100 vilages.

Do not participate in demo plot Participate in demo plot
Electronic copy available at: https://ssrn.com/abstract=3321171 Appendix  13), soy (binary variable), inoculation soy (binary variable), groundnut (binary variable), hybrid maize (binary variable), herbicide (binary variable), pesticide (binary variable), fungicide (binary variable), inorganic fertilizer (binary variable), fertilizer tree (binary variable), intercropping (binary variable), animal manure (binary variable), crop residue (binary variable), and compost (binary variable). These refer to planned adoption in the 2015-16 season. The independent variables are whether or not the individual is in a club which managed a demonstration plot, and whether or not the individual is in a club which was invited to a farmer field day. The estimation uses a two-step procedure. The first steps uses the reported club membership at endline in the treatment villages to predict who would be most likely to join a CDI club. The second step uses all individuals in both treatment and control group whose predicted probability is larger than 0.5. This includes 241 individuals in the treatment group and 173 individuals in the control group. The control variables used in both steps include all village level characteristics included in Appendix Table 1, all household level characteristics included in Appendix Table 3 (and the square terms of the non-binary variables) and baseline adoption indicators. Bootstrapped clustered errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Electronic copy available at: https://ssrn.com/abstract=3321171 Appendix  20), and the 20 knowledge questions listed in Table  3 (binary variables). The independent variables are whether or not the individual is in a club which managed a demonstration plot, and whether or not the individual is in a club which was invited to a farmer field day. The estimation uses a two-step procedure. The first steps uses the reported club membership at endline in the treatment villages to predict who would be most likely to join a CDI club. The second step uses all individuals in both treatment and control group whose predicted probability is larger than 0.5. This includes 241 individuals in the treatment group and 173 individuals in the control group. The control variables used in both steps include all village level characteristics included in Appendix Table 1, all household level characteristics included in Appendix Table 3 (and the square terms of the non-binary variables) and baseline adoption indicators. Bootstrapped clustered errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1.
Electronic copy available at: https://ssrn.com/abstract=3321171 Adoption score (score out of 13). The independent variables are whether or not the individual is in a club which managed a demonstration plot, and whether or not the individual is in a club which was invited to a farmer field day. We include, but do not report, the farmers who live in treatment villages but do not belong to the treatment clubs in this regression. We use adoption of 2013-14 as elicited at baseline (in 2014), and planned adoption elicited in 2005 referring to the 2015-16 season. Standard errors clustered at the village level in parentheses. Whether or not farmer is in a club is determined by the self-reported club status in 2015.*** p<0.01, ** p<0.05, * p<0.1

Appendix Table 7: The impact of the CDI program on (planned) adoption of ISFM technologies -Farmer Fixed Effects Approach
Electronic copy available at: https://ssrn.com/abstract=3321171 Appendix Figure 1: Kernel density plot of the predicted probability of CDI club membership in the treatment and control villages (see also Appendix Figure 2: The distribution of the probability of CDI club membership, by treatment and control villages, for the sample for which the probability estimated is larger than 0.5 (see also