Enhancing the diffusion of Information About Agricultural Technology

Enhancing the diffusion of information about agricultural technology Abstract We introduced a new seed variety to a subset of farmers across 100 rural Indian villages. Half of the villages were randomly allocated to receive an information treatment known as the farmer field day. The new seed variety was explained, selected adopters shared their experiences, and villagers were taken to the field to observe the crop during these events. We find that farmer field days improve the learning process and increase adoption in the following year by around 40 percent. Further analysis shows that the field days were both cost effective and more impactful for poorer farmers. 1 1 Introduction The adoption of improved technologies is an important engine of growth for smallholder farmers in developing countries. Yet, levels of adoption often remain disappointingly low and lack of credible information is an oft-cited barrier (Jack, 2011). Several studies show that peers are credible sources of information about new technology (Foster and Rosenzweig, 1995; Bandiera and Rasul, 2006; Conley and Udry, 2010; Krishnan and Patnam, 2014). Yet, transmitting information generates little private benefit for farmers. This begs the question of how to enhance the process of farmer-to-farmer learning? One approach is to strategically select the earliest adopters to maximize information flow (Beaman et al., 2015). Another proven solution is to directly incentivize early adopters based on future adoption in their community (BenYishay and Mobarak, 2015). A different approach — and the one considered in this paper — is to facilitate the learning process by explicitly creating a venue for information sharing. We ask whether adoption of an improved technology can be increased by simply organizing a meeting where the technology is explained and adopters are invited to share their experiences with nonadopters. This approach is better known as the “farmer field day” because the meeting also involves visiting a field where the technology is being used. The event is intended to disseminate information and demonstrate technology in an environment that mirrors the production conditions of local farmers. Improving the speed of learning is important for policy. Low productivity in the agricultural sector is tightly associated with low levels of aggregate productivity (Restuccia, Yang, and Zhu, 2008). At the same time, modernizing agricultural technology is an important driver of structural transformation and hence long-term growth (Yang and Zhu, 2013; Bustos, Caprettini, and Ponticelli, 2016). The policy challenge that arises is how to make proven agricultural technologies diffuse faster? If farmers consider neighboring farmers to be reliable sources of information — as the literature suggests — but individual farmers gain little from demonstrating technology for others, then information flow will be inefficiently low. Policy interventions to enhance learning are warranted in this context. This paper tests one such intervention where a venue for peer-to-peer learning and information sharing is explicitly created. Importantly, there is no rigorous evidence on the effectiveness of this approach to agricultural extension. Our findings were generated by a two-step randomized experiment in the Indian state of Odisha. In the first step we introduced 25 kilograms of a new high-yielding and floodtolerant rice variety called Swarna-Sub1 in 100 villages. Importantly, the technology has been shown to dominate existing technology, indicating that it should diffuse rapidly in the 2 absence of barriers to adoption.1 The farmers receiving seeds were chosen using one of three methods. In one third of the villages we used the status-quo approach of delivering the seeds to locally-elected village officials — ward members in the Gram Panchayat — who then chose how to further distribute the seeds amongst villagers.2 In another third of the villages we used a participatory meeting where villagers were invited to jointly determine how the seeds should be allocated. Finally, we used village meetings with local women’s groups (Self Help Groups or SHG’s) in the remaining villages. This randomization allows us to test whether the learning intervention is more (or less) effective under different methods of identifying demonstrators In the second step of the experiment our partner NGO carried out the farmer field days in 50 villages. These field days occurred approximately four months after the seeds were introduced and while the crops of the demonstrators were approaching time for harvest. The field days were simple two hour events where the variety was described, demonstrators spoke about their experience, questions were answered, and then attendees were taken to observe the crop of one or more of the demonstrators. We then offered the new variety for sale after harvest and immediately prior to planting for the next growing season. The sales teams went door-to-door and asked a random sample of 15 households per village whether they were interested in buying seeds. Importantly, we fixed the price to be near the prevailing market price of the variety.3 Therefore, offering seeds directly to farmers in this way allows us to measure demand in the absence of both supply constraints and high subsidies. We have three main results. First, we show strong effects of farmer field days on uptake. Adoption increased by around 12 percentage points from 30 to 42 percent in field day villages. This effect is concentrated more heavily on adoption of a single seed package: purchases of one five kilogram packet rose by 59 percent, while purchases of two or more packages rose by only 25 percent.4 In addition, we show that the effect of field days on demand is the largest for the poorer farmers in the village. More specifically, the effects of farmer field days on buying a single package of seeds is significantly larger for farmers that are in lower caste 1 The variety was released in 2009. It offers flood tolerance without reducing yield during non-flood years (Dar et al., 2013). The technology also leads to significant welfare gains by inducing farmers to modernize production (Emerick et al., 2016) 2 The method of delivering a small amount of seeds for testing and knowledge creation is a popular approach in South Asia. India’s National Food Security Mission (NFSM) program uses these seed minikits and relies on members of the Gram Panchayat to help identify beneficiaries. 3 The new variety was not yet available at the government block offices where most farmers purchase subsidized seeds. One company in the area was selling seeds of this variety at a price higher than the subsidized price. We refer to this as the prevailing market price. 4 Almost all farmers that purchased seeds bought only one or two packages. This amount of seeds is enough to cultivate around 10-30 percent of their land. 3 groups and farmers that are below the poverty line (BPL), as defined by the government’s well-known entitlement program. Second, the effect of field days is no larger in villages where demonstrators were selected with meetings. If anything, the effects of field days are largest under the status-quo approach where demonstrators are selected in consultation with locally elected officials. However, the differences in the effect of field days across the three methods of introducing the seeds are not statistically significant. Nonetheless, we can rule out that field days are more effective when demonstrators are selected via participatory meetings. In other words, the field day is effective regardless of the identities of the farmers cultivating the new variety. This lack of heterogeneity is not due to inability of meetings to select a different pool of demonstrators. The meetings succeeded in reducing patronage. More concretely, the demonstrators are 31 percent and 62 percent less likely to be close family or friends of the ward member in village meeting and SHG meeting villages, respectively. In addition, 11 percent of recipients in ward member villages are elected representatives (including the ward member him or herself). These percentages fall to 5 percent and 3.2 percent in village meeting and SHG meeting villages, respectively. Despite this reduction in patronage, our third result is that selecting demonstrators via meetings had no effect on adoption. Differences in adoption between ward member, SHG meeting, and village meeting villages are small and statistically insignificant. Our estimates are generally precise enough to rule out large effects such as the 40 percent increase in adoption caused by field days. The consequences of patronage in distribution of free seeds for demonstration appears to be minimal: farmers appear to learn equally well from demonstrators identified by local politicians compared to demonstrators identified with participation of the village more broadly. The strong effect of farmer field days could arise because they generate demand by farmers that do not interact with demonstrators and thus otherwise would not have had the opportunity to learn about the variety.5 In other words, the field days make up for the fact that networks are incomplete, i.e. not all farmers interact with each other. The data are generally incompatible with this explanation. Using either common surnames or geographic distance as measures of connectedness, we show that the effect of field days is no larger for farmers that are less connected to the demonstrators. Focusing on the SHG villages only, we find that if anything, field days only have an effect on demand for SHG member households and households that are friends or family of the president of the SHG. These results suggest that field days either convey additional information or re-enforce 5 As shown by Golub and Jackson (2012), diffusion of information between individuals is limited by certain network structures where individuals only interact with people that have similar characteristics. 4 existing knowledge rather than introduce a technology to farmers that lack connections with demonstrators. The field days are a cost effective method of improving learning and increasing adoption. Our approximate estimate of the one-year revenue gains from field days is 416 dollars per village. This compares to the costs — including time cost of attendance — of around 230 dollars per village. This favorable benefit-cost ratio is encouraging, but is also driven partly by the profitability of the technology being promoted. Field days to provide information about less profitable technologies may generate similar gains in adoption, but less favorable benefit-cost ratios. Our paper adds to the literature on learning in developing-country agriculture. Numerous studies have established the importance of peers as sources of information.6 However, there is no reason to expect that using farmers as disseminators of information will maximize learning. Farmers may not internalize the benefits of spreading information or be hesitant to promote a technology without using it over several years. Based on this idea, BenYishay and Mobarak (2015) use an experiment in Malawi to show that compensating early adopters for transmitting information can be effective at increasing future adoption. The positive effects of compensation suggest that further intervention to encourage peer-to-peer learning can be impactful. Our experiment shows that the farmer field day serves as one such intervention that improves learning and stimulates demand. By doing so, we also show that information is indeed an important constraint that limits the diffusion of improved agricultural technology. This paper also adds to the literature on the efficacy of agricultural extension. There is a widespread view that agricultural extension in developing countries underperforms (Anderson and Feder, 2007). Building on this, recent work has focused on testing new methods of providing information to farmers. As one example, there has been a recent emphasis on mobile phones as an extension tool (Aker, 2011; Fafchamps and Minten, 2012; Cole and Fernando, 2014). We instead focus on the more traditional method of selecting “entry points” and relying on these farmers to demonstrate technology for others. The notable difference between our experiment and this traditional model is that we test whether farmer field days can improve this process by explicitly encouraging learning. To our knowledge, BenYishay and Mobarak (2015) is the only other study that experiments with methods of improving peer-to-peer learning once demonstrators have been identified.7 6 Conley and Udry (2010) is the leading example where peers are shown to be an important source of information about pineapple cultivation in Ghana. 7 The “farmer field school” is a related extension technique, but is distinct from the method we study. Farmer field schools are usually carried out over multiple days, provide general information about agriculture rather than details about a specific technology, and don’t engage demonstrators to leverage peer-to-peer learning. Studies evaluating farmer field schools have relied on observational data and non-experimental methods such as matching (Godtland et al., 2004; Davis et al., 2012). 5 The rest of this paper is organized as follows. In section 2 we give further background on the particular seed variety. We focus on why the variety is suitable for a study on learning and diffusion. We also walk through the conceptual motivation for the experiment. Section 3 discusses the experimental design and some basic characteristics of the population we study. Section 4 discusses our main results. Finally, section 5 offers concluding remarks. 2 Background and conceptual framework In this section we start by giving more details on the specific technology used in the experiment. We focus on its key benefits and how they make it a worthwhile technology for studying diffusion. We then discuss the conceptual reasoning behind the farmer field days intervention. 2.1 Details about technology Swarna-Sub1 — the rice variety introduced as part of the experiment — offers flood tolerance as its key benefit. Swarna-Sub1 is otherwise similar to Swarna, which is a popular type of rice cultivated throughout eastern India and Bangladesh. The technology was developed by moving a group of flood tolerance genes from a traditional rice landrace to the popular variety Swarna.8 Plant breeders were able to rely on modern breeding techniques to create the improved variety without introducing other undesirable characteristics (Xu et al., 2006). This is important because it guarantees that the technology offers an added benefit without any obvious disadvantages during normal years. Previous randomized experiments have conferred two channels through which this new variety improves farmer welfare. First, the variety does improve output under flooding without compromising output when there is no flooding (Dar et al., 2013). Second, SwarnaSub1 induces farmers to invest more in inputs, particularly at or near the time of planting. These effects likely arise due to the risk-reducing property of the technology (Emerick et al., 2016). As a result, the new variety improves outcomes even in years when flooding does not occur. Swarna-Sub1 is appropriate for our study because it dominates existing technology and should diffuse rapidly in a perfect environment with no information barriers or supply constraints. This dominance makes it a worthwhile technology for studying different mechanisms to encourage diffusion. In addition, we can be confident that adoption is the appropriate 8 The biological mechanism is that Swarna-Sub1 suppresses the plant’s natural response of elongation during flooding. This allows the plant to retain the necessary carbohydrates for regeneration after the flooding is over (Voesenek and Bailey-Serres, 2009). 6 outcome variable because there are no groups in the population for which adoption is strictly unprofitable. 2.2 Conceptual motivation The farmer field day intervention is meant to create a venue where farmers can learn about a new variety, hear the experiences of demonstrators, and observe the variety in the field. A more passive approach would be to rely on demonstrators to directly provide this information to other farmers. However, any model that requires farmer-to-farmer learning suffers from the reality that farmers gain little from sharing information with others — i.e. demonstration and information-sharing create spillover benefits for others. This necessitates some method of encouraging demonstrators to spread information — direct incentives in the case of BenYishay and Mobarak (2015). In our case the field day allows demonstrators to stand out as leaders in the village by sharing their experience with a new technology. Consistent with this, demonstrators seldom declined to share their experience despite not being directly incentivized for participation. There are two mechanisms through which field days could induce adoption. First, field days could provide information to farmers that otherwise would not have learned because they would not talk to demonstrators. Put differently, some people will lack information because they are not connected to demonstrators. Field days give these people an opportunity to learn and make up for the fact that real-world networks are incomplete. Second, field days could reinforce knowledge that was already transmitted. Or, field days could make existing knowledge more precise by aggregating the experiences of several adopters, rather than just a single adopter sharing a connection with a given non-adopting farmer. One example would be a farmer that learns about the new variety from a demonstrator and thus possesses some signal of its effectiveness on her land. In this case the field day would aggregate the experiences of other adopters that this farmer would not have otherwise interacted with. If the farmer needs to observe the experience of more than one demonstrator to adopt herself — as in Beaman et al. (2015) — then field days increase the likelihood of crossing the threshold of communicating with more than one demonstrator. The first explanation delivers a simple prediction on who should benefit from field days: farmers that have no connection to demonstrators should benefit the most because they otherwise would not have heard about the experiences of these demonstrators. In contrast, there is no reason to expect this heterogeneity if farmers that were already connected to demonstrators also learn from the field day. We test this directly in the analysis that follows. The alternative methods of selecting demonstrators allow us to further test whether field 7 days are more effective when demonstrators are chosen by villagers rather than hand-picked by locally elected officials. There are two competing effects. One the one hand, farmers may learn better from similar individuals, particularly when returns are heterogeneous (Munshi, 2004; Tjernström, 2015; BenYishay and Mobarak, 2015). On the other hand, locally elected officials may hold better information about productivity and use that information in selecting demonstrators.9 If more productive farmers are chosen as demonstrators then field days could send a stronger signal about potential output of a new technology under optimal management. 3 Details of the experimental design In this section we give specific details on the experimental design. We give more information on how the treatments were carried out and the timing of data collection. We also describe some basic summary statistics as well as randomization balance. The experiment took place in 100 villages in Balasore — the northernmost district in the state of Odisha. The villages are located in three blocks — an administrative unit two levels above villages — where our partner NGO frequently works. We randomly selected these villages from the subset of villages that were affected by flooding for at least 2 days in 2011 or 2013. We used satellite images of flooding during these years to classify which villages were affected. The sample is focused on flood-prone areas to ensure that adoption of Swarna-Sub1 is a profitable outcome and therefore diffusion is important to study. We next describe the timing of events, which we also display graphically in Figure 1. We administered a baseline survey to 10 farmers in each village prior to carrying out any activities in the village.10 Enumerators first visited a local village leader and identified the names of 10 villagers that cultivate land in flood-affected areas. These farmers were then administered with a short survey. The main purpose of the baseline was to measure whether farmers in the sample villages had any past experience with Swarna-Sub1. Past experiences were indeed limited: only two farmers that were surveyed at baseline had cultivated SwarnaSub1 during the previous season in 2013. In contrast, 74 percent of farmers were cultivating Swarna, the variety that is otherwise similar except for flood tolerance. This makes the technology ideal for the experiment because it is profitable relative to the most popular variety but unknown to farmers, making learning an important consideration. Shortly after completion of the baseline in May 2014, enumerators from our partner NGO 9 As one example, Basurto, Dupas, and Robinson (2015) show that chiefs in Malawi account for productivity differences when allocating subsidized inputs to villagers. 10 Enumerators were unable to carry out the baseline survey in one of the 100 villages. 8 returned to each village to distribute seeds to demonstrators. Each village was provided with 25 kilograms — an amount that is sufficient to cultivate one or two acres. But more importantly, the seeds were already packaged into five kilogram packages to encourage that at least five farmers be selected as demonstrators. Villages were randomly assigned to one of three methods for identifying demonstrators. The seeds were delivered to the locally elected village ward member in 33 villages. The ward member is elected to represent the village in the local Gram Panchayat, or the next administrative unit above villages. A representative from our partner NGO delivered the seeds directly to the ward member and informed them that the NGO was giving the seeds to the village. The enumerator then asked the ward member to identify the most suitable cultivators of the new seeds. In all cases the seeds were left to the ward member and she independently decided about their further distribution, including whether to keep some for herself. This approach simulates the common approach of both government and NGO’s of using local political figures to distribute seed minikits, as in Bardhan and Mookherjee (2006, 2011). We used two types of village meetings to identify demonstrators in the remaining villages. In 34 meeting villages NGO staff first visited the village and informed as many villagers as possible that they were carrying out a short meeting to describe a new flood-tolerant rice variety. Enumerators were specifically instructed to put the seed minikits at the front of the meeting and describe the benefits of the new variety relative to Swarna. Importantly, enumerators instructed villagers to jointly decide on who would be best to cultivate the variety. In all cases villagers were able to come to an agreement and all 25 kilograms of seed were distributed to farmers that were willing to plant. We used a process that was nearly identical to this in the remaining 33 villages. The only difference was that only Self Help Group (SHG) members were invited to the meeting. This approach makes the meeting inclusive of entirely women. Enumerators returned to all villages in September 2014 to survey all of the demonstrators. This short survey had two main purposes. First, the survey allows us to compare characteristics of demonstrators across the three different treatment arms. Second, we collected information on how much area was planted with Swarna-Sub1, the current status of the crop, and the GPS boundaries for the plots of farmers that actually transplanted the seedlings.11 Overall, we have plot locations for 452 (67 percent) of the farmers that received seeds. 11 Some farmers were affected by flooding during the nursery stage before seedlings had been transplanted in the main field. No rice variety is tolerant to submergence at this stage. As a result, 29 percent of farmers lost their seedlings and were unable to plant the crop in the main field. 9 Farmer field days were then carried out in 50 randomly selected villages during the month of November 2014. The field days were purposefully timed to take place slightly before harvest when the adopters had built some experience with the technology but while the crop was still in the ground for demonstration. The field days were short. Staff from our partner NGO again described Swarna-Sub1 and its properties, spoke about seed quality and how to distinguish quality seeds, and then gave the demonstrators an opportunity to discuss their experience with Swarna-Sub1. The meetings were fairly well attended: an average of 41 farmers — or 59 percent of rice-farming households — attended the field days. Table A1 shows that household characteristics in general do not strongly predict attendance. The field days appear to have been attended by a broad group of villagers and not just the wealthiest or most elite farmers. We then carried out a survey with approximately 15 farmers in each village in order to measure knowledge about Swarna-Sub1. We refer to this group as the non-adopting farmers, i.e. those that were not in the group of demonstrators. The surveys took place during February to March 2015. We used the list of households from the 2002 Below the Poverty Line (BPL) census to randomly select households.12 We removed the demonstrators before randomly selecting the households. Each respondent was asked several questions to measure their knowledge of Swarna-Sub1. These included whether they knew about it at all, knowledge of its main benefit, which areas are suitable for cultivation, and duration (time from planting to harvest). Our NGO sent a new team of staff members to each village in May 2015. Each farmer that was surveyed in February-March was visited and given the opportunity to purchase Swarna-Sub1 seeds. There was only one other NGO selling Swarna-Sub1 to farmers for a price of around 20 rupees per kilogram. Our price was set to 20 rupees in order to mimic this market price. Thus, farmers benefitted mostly from free delivery when given this purchasing opportunity. Most of the farmers in our sample did not know how to obtain Swarna-Sub1.13 The door-to-door sales were used to eliminate these barriers and allow us to reveal demand at the market price and in the absence of supply barriers. We observed a strong demand for the technology in the door-to-door sales: 36 percent of farmers bought at least one package of seeds. The inability to record adoption from other sources — largely other farmers in the village — is the disadvantage of measuring adoption with the door-to-door sales. Swarna-Sub1 is 12 We selected all households in the villages where there were fewer than 15 non-adopting households. One of the main questions that came up during the field days was how to obtain the seed. Private seed dealers do not operate in this area and the seeds were generally not available at the local block office where most farmers buy their seeds. There was only one other NGO with access to seeds and most farmers were unaware of this NGO. 13 10 an inbred rice variety that can be multiplied, reused, and traded with other farmers. Many estimates indicate that this informal seed system of either reusing one’s own seed or obtaining from neighbors accounts for a large portion of seed use in South Asia. We would miss this source of adoption with our door-to-door sales. We remedied this issue by carrying out a door-to-door adoption census starting in July 2015. Survey teams went door-to-door in each village and asked each household a small set of questions, including whether they were currently cultivating Swarna-Sub1. A total of 6,511 households were surveyed during these door-to-door visits. The data do show the importance of supply barriers. Only 14 percent of all households adopted Swarna-Sub1.14 This compares to a 36 percent adoption rate in the sample of farmers that received door-todoor visits. This additional dataset allows for measurement of adoption from all potential sources, not just our door-to-door sales. Table 1 shows summary statistics. We focus on the sample of approximately 15 farmers per village that we use for much of the analysis. Farm sizes of just over two acres are roughly representative of farm sizes in coastal flood-prone parts of the state. Swarna is also a popular variety: over half the farmers in the sample cultivated Swarna during the season before the survey. The table also considers covariate balance by regressing each characteristic on the field day indicator and block fixed effects. This is the same form of regression used in the main analysis that follows. The two experimental groups are roughly comparable on observable characteristics. Table A2 shows that household characteristics vary little across the three different methods of choosing demonstrators. Characteristics are roughly balanced even considering that the estimation sample is a random subset of farmers that were not chosen as demonstrators. Part of the reason we introduced only 25 kilograms was to limit any selection into the group of non-demonstrators. More concretely, around 6 to 8 farmers — or 10 percent of the ricefarming population — were selected as demonstrators in most villages. Thus, the sample of non-demonstrating farmers represents most of the village. We also show adoption effects for the entire village — including demonstrators, which effectively shuts down the possibility that this type of selection influences any of the estimates. 4 Results This section is divided into separate components for each of the main results. We first outline the empirical framework, which is simple given the experimental design. The next subsection documents the effect of field days on knowledge. We then turn to our main estimates on 14 This figure includes the demonstrators from the previous year. 11 technology adoption revealed with door-to-door sales. In addition to these main effects, two sources of heterogeneity are considered: the different methods of choosing demonstrators and the wealth / social status of the household. The analysis concludes by showing suggestive evidence on whether the field days informed, and thus increased demand, of farmers that were less likely to interact with demonstrators. 4.1 Regression framework Our experimental design affords us a simple estimating equation. The reduced-form effect of field days is estimated with yivb = β0 + β1 F ieldDayvb + β2 Xivb + αb + εivb , (1) where yivb is some outcome (usually seed adoption) for household i located in village v and block b. Our main estimate is β1 , the average treatment effect of the field days. In some specifications we verify that our results are unaffected when including household covariates Xivb . The villages in the experiment were spread across three blocks (an administrative unit which was a stratification variable) and therefore we include blocked fixed effects, i.e. the αb terms. Finally, εivb is a random error term which we allow to be correlated within villages but assume to be uncorrelated across villages. The variation induced by the experiment also identifies the effects of field days for each of the three methods of choosing demonstrators: yivb = β0 + β1 F ieldDayvb + β2 F ieldDayvb ∗ M eetvb + β3 F ieldDayvb ∗ SHGvb (2) + β4 M eetvb + β5 SHGvb + β6 Xivb + αb + εivb . The main coefficients of interest are β2 and β3 which measure whether field days are more (or less) effective in villages where demonstrators were identified with meetings. 4.2 Effects of field days on knowledge about new technology A useful question before turning to demand effects is was there any measurable learning from the farmer field days? Prior to the door-to-door sales, approximately 15 household in each village were surveyed to assess their knowledge of Swarna-Sub1. Farmers were asked several questions, starting with whether they had ever heard of Swarna-Sub1 and how many farmers they had spoken to about the variety. We then asked several multiple choice questions such as the two differences between Swarna-Sub1 and Swarna15 , the length of flooding that Swarna15 In addition to flood tolerance, Swarna-Sub1 has a white husk, making it distinguishable from Swarna. 12 Sub1 can tolerate, and the duration of Swarna-Sub1 (days from planting to harvesting). Table 2 shows some modest effects of the field days on these observable measures of learning. Starting with column 1, farmers are six percentage points more likely to have ever heard of Swarna-Sub1 in field day villages. However, this is compared to a fairly high base: almost 80 percent of farmers knew of Swarna-Sub1 in control villages. Column 2 shows that farmers in field day villages report talking to an additional 0.12 famers about Swarna-Sub1 — an approximate 20 percent effect. Knowledge on attributes of the technology in columns 3 through 6 are somewhat mixed. The strongest effect is in column 4 where field days led to an approximate 55 percent increase in knowledge of how long Swarna-Sub1 can survive when flooded. Column 6 shows that farmers in field day villages were slightly more likely to know the duration of Swarna-Sub1, although this effect is also modest because of high knowledge in the control group. The inability to measure all aspects of learning is an important caveat of this exercise. There are certainly some attributes of the technology that we do not measure. It is further difficult to know what type of information would be transmitted since most of the field days involved an unstructured discussion. These results on information therefore represent a more intermediate check before considering the more important outcome of adoption. 4.3 Effects of field days on seed adoption The experiment was designed with the policy objective of increasing the diffusion of a proven agricultural technology. Adoption is therefore the most important outcome of interest. We measure adoption directly with our door-to-door sales experiment that was carried out approximately one year after seeds were originally introduced. Table 3 shows the main result that field days had positive and large effects on adoption. Focusing on column 1, the field days led to a 12.2 percentage point increase in adoption. The rate of adoption in control villages was 29.7 percent. The point estimate therefore indicates that this relatively simple method of bringing farmers together to discuss a new technology leads to a 41 percent gain in adoption. Interestingly, column 2 shows a larger effect on adoption of a single package of seeds. Adoption of just one five kilogram package increases by 8.6 percentage points — or 59 percent. On the other hand, adoption of two packages increases by only 3.6 percentage points and this effect is statistically insignificant. Our data do not allow us to pinpoint an exact reason for this difference. Nonetheless, one possibility is that the field days provided additional information to farmers that were near the threshold of simply trying the new seed. But field days were less impactful for farmers that had already decided to plant the variety on a larger share of their land. 13 Including several household control variables does not change the main result (columns 4 through 6). The point estimates are almost identical to those that only use the variation created by the experiment. This is not surprising given that the randomization was successful at achieving balance between the experimental groups. Figure 2 helps further understand this effect by showing the distribution of the villagelevel adoption rates for the two types of villages. Two things stand out from the figure. First, the field days decreased the frequency of little or no adoption at the village level. 35 percent of control villages had adoption rates lower than 10 percent. In contrast, only 12 percent of field day villages had such low adoption rates Second, the distribution for field day villages puts much more mass on adoption rates greater than 50 percent. 38 percent of field day villages had adoption rates of 50 percent or higher, while only 19 percent of the control villages had at least half of the farmers adopt. Figure 3 verifies that the effect of field days is strongly correlated with attendance. The adoption rates of farmers in treatment villages that did not attend the field days are almost identical to adoption rates in control villages. In contrast, adoption rates are about 50 percent higher for attendees in treatment villages. Attendance is certainly non-random and perhaps correlated with unobservables that we do not measure in Table A1. Nonetheless, the result presents useful — but certainly suggestive — evidence that whatever learning happened at the field days likely did not spill over to non-attendees. The data suggest that the field days were most effective for poorer farmers. Two indicators of poverty status are readily available in our data. Around a third of the sample belongs to the scheduled castes or tribes, the most disadvantaged castes in the country. Members of scheduled castes and tribes obtain less education and earn lower incomes relative to higher caste individuals (Munshi and Rosenzweig, 2003). Column 1 in Table 4 shows that the marginal impact of field days on adoption for higher caste farmers is 8.3 percentage points. This impact increases to 20.1 percentage points for farmers belonging to the scheduled castes and tribes — however the large differential effect is not statistically significant (p=0.14). Column 2 shows that this differential effect is largely driven by inducing lower caste farmers to purchase a single package of seeds. The effect of field days on adoption of a single package is only 4.6 percentage points for higher caste farmers. In contrast, the effect is over three times larger for scheduled castes and tribes and the differential effect is statistically significant at the 10 percent level. Column 3 shows that there is virtually no differential effect for scheduled castes and tribes on the probability of purchasing two packages. In addition to caste status, we also use possession of a “below the poverty line” card, which is meant to deliver various social assistance benefits to poor households. About 62 14 percent of our sample holds one of these cards which are available to households that are determined to be poor by a proxy means test. We observe a similar pattern where field days are more likely to induce BPL households to purchase a single package of seeds. In particular, column 5 shows that the effect of field days on adoption of five kilograms is larger by 10.1 percentage points for BPL households. In combination, the results suggest that field days increase equity by delivering the largest gains in adoption for the poorest farmers. These gains exist despite a lack of evidence that poor farmers learn better from field days. Tables A3 and A4 show that the effect of field days on observed learning is not significantly larger for either ST/SC or BPL households. Therefore, the differential effects on adoption must arise for a reason different than differential learning — at least for the attributes measured by our survey. 4.4 Does the effect of field days vary by method of selecting demonstrators? We first establish that demonstrators identified by either village or SHG meetings differ from demonstrators selected by ward members. In short, the meetings reduced patronage and decreased the likelihood that demonstrators would be close family and friends of ward members. We estimate regressions of the form yivb = β0 + β1 M eetvb + β2 SHGvb + εivb , (3) where y is a characteristic of the demonstrator or their household and M eetvb and SHGvb are indicators for meeting and SHG meeting villages, respectively.16 We show these coefficient estimates for various characteristics that were collected when our survey teams returned to villages after seed distribution to gather additional information from demonstrators. Table 5 shows the results. The main result is that patronage exists when seeds are distributed via ward members. The first row in column 1 shows that 30.5 percent of demonstrators are either the ward member or her family / close friend in villages where demonstrators were selected by ward members. This falls by 9.6 and 18.8 percentage points in meeting and SHG meeting villages, respectively. Turning to row 2, part of this effect is the ward member keeping seeds for him or herself. 11 percent of demonstrator households had an elected panchayat representative in ward member villages. The share of households having an elected panchayat representative was significantly lower in village meeting and SHG 16 These regressions are meant to measure selection. The block fixed effects are dropped from the regression to avoid absorbing any selection effects. 15 meeting households.17 In addition to proximity to the ward member, other characteristics of demonstrators vary across the treatments. Not surprisingly, demonstrators in SHG meeting villages come from entirely SHG households. This is a consequence of using SHG meetings which consist entirely of women. Row 4 shows that early adopters in SHG villages are significantly more likely to be family or close friends with the SHG president — relative to village meeting villages. Focusing on the other rows, adopters in SHG villages are more likely to be part of agricultural cooperatives, are younger, and are less likely to come from disadvantaged castes. Differences in livestock ownership appear to be the only wealth differences across the treatments. How were the demonstrators chosen at the meetings? Table A5 shows results where we aggregate various outcomes to the village level in order to study whether the seed was demonstrated differently in the different types of villages. There are three noteworthy results. First, almost all 25 kilograms were distributed and planted in each village.18 Second, village meetings — but not SHG meetings — were over twice as likely to result in more than five demonstrators relative to selection via ward members. In other words, the five kilogram bags were more likely to be opened and redistributed in smaller quantities in village meeting villages. Third, some demonstrators were unsuccessful with the variety because the seedlings were planted in an area affected by flooding and the technology is not tolerant to flooding when seedlings have just emerged.19 This was slightly more likely with demonstrators selected from SHG meetings relative to village meetings. Despite the differences in characteristics and numbers of demonstrators, the field days were no more effective when early adopters were identified with meetings. Table 6 shows the complete specification in (2) where the field day indicator is interacted with the indicators for village and SHG meetings. The second two rows show that if anything, the field days were less effective when early adopters were selected by meetings. The coefficient on the interaction term between field days and SHG meetings is negative, somewhat large, but statistically insignificant. Similarly, the coefficient on the interaction between field days and village meetings is also negative and imprecisely estimated. Nonetheless, our findings are 17 We removed the ward member households from the data and re-estimated the regression in the first row. The coefficient for village meeting villages decreases and becomes statistically insignificant, but the coefficient for SHG meeting villages remains negative and statistically significant. In addition, both variables jointly explain the likelihood that the adopter is connected to the ward member (p-value of F statistic = 0.05). This helps ensure that this effect is not entirely driven by the ward member keeping seeds for their own household. 18 At this stage we define planting as accepting the seeds and planting them in a nursery bed. 19 The common planting technique in our sample is transplanting, i.e. planting seedings in a small area and then uprooting these seedlings after three weeks and transplant them to the larger field where the crop will be grown. 16 not compatible with the idea that farmers gain more from field days when demonstrators are more representative of the village and less likely to be connected to local political figures. Farmers appear to gain just as much — if not more — from participating in field days when demonstrators are identified by locally elected officials. The level effects of using meetings to identify demonstrators are also small and statistically insignificant. Using SHG meetings in the absence of field days led to 7.3 percentage point increase in adoption, but this effect is imprecisely estimated. The effect of using village meetings is closer to zero and also imprecisely estimated. The meetings were ineffective in field day villages. The impact of selecting demonstrators with SHG meetings is a 5.2 percentage point decrease (-.052 = .073 - .125) in adoption. The effect of village meetings in field day villages is also negative and of a similar order of magnitude. Table 7 clarifies this result by showing the average effect of both types of meetings. The average effect of using meetings across both field day and control villages is effectively zero. The point estimate in column 1 of -.005 is estimated with more precision since more observations are pooled together. In particular, the 95 percent confidence interval allows us to rule out effects larger than 8.8 percentage points or 24.7 percent relative to adoption in villages where ward members selected demonstrators. Columns 2 through 6 show that this conclusion changes little when splitting adoption into purchases of one or two bags or controlling for several household characteristics. In combination, we don’t find any evidence that engaging villagers to participate in selection of demonstrators drives future technology diffusion. These estimated null effects are only interesting alongside the result that meetings reduce patronage by generating a pool of early adopters that are on average less connected to local political officials. In practice the result means that although relying on elected officials to identify early program beneficiaries does induce favoritism, this favoritism has no consequences on the rate at which the new technology diffuses to other villagers. This result contrasts somewhat with the literature on targeting of anti-poverty programs. Engaging villagers in meetings has been shown in that context to improve the efficiency at which anti-poverty benefits are targeted to the poor (Alatas et al., 2012). While targeting is a different objective than encouraging learning and diffusion, our results are less favorable for the use of participatory meetings when it comes to identifying demonstrators of new technology. Finally, we show that results change little when measuring adoption from all sources — not just our door-to-door sales. In addition to the sales, farmers could have obtained seeds directly from demonstrators. Adoption from our door-to-door sales would present an inaccurate picture if either of the two meeting types were more successful at identifying early adopters that are better seed distributors or if field days induced more farmer-to-farmer 17 sharing of seeds. The results are similar when estimated with the entire village. Column 1 in Table 8 shows the specification for all villagers — including demonstrators. The coefficients for both village and SHG meetings remain small and statistically insignificant. However, we continue to estimate large positive effects of field days on seed demand at the village level. Field days caused in increase in adoption of 6.2 percentage points, or around 60 percent.20 Column 2 shows that we still fail to detect significant interaction effects between meetings and field days in this larger sample. Measuring adoption for the entire village eliminates any concern that selection into the sample of demonstrators is responsible for the results. The relatively small number of demonstrators — relative to village size — seems to limit this particular type of sample selection. 4.5 Are field days a cost effective method of agricultural extension? A back-of-the-envelope calculation suggests that field days are likely cost effective. The average village in our sample has 69 rice-farming households. Thus, a field day would be expected to generate around 8.3 additional adopters of Swarna-Sub1. In Emerick et al. (2016) we find that this technology leads to revenue gains of 10 percent or around 2,969 rupees. This effect arises largely due to the crowd-in effect of inducing farmers to use more inputs. Thus, field days generate one-year revenue gains of around 24,643 rupees, or around 410 dollars. Our partner NGO required approximately 200 dollars per village to execute the farmer field days. A rough estimate of the time cost to farmers of attending the field days is 29.7 dollars.21 Therefore, farmer field days appear to pay for themselves after just a single growing season. One the one hand this calculation is very encouraging because the one-year benefits are an extreme lower bound on the flow of benefits that farmers will receive from continued adoption. On the other hand, the calculation has to be interpreted with caution for two reasons. First, it is unclear whether there would be dis-economies of scale with wider implementation. There would likely be additional costs of incorporating a larger number of villages outside of the coverage area where our partner NGO operates. Second, our estimated adoption effects are conditional on the absence of other supply barriers. We can only think of field days as being 20 The much lower adoption rate at the village level is indicative of supply constraints. 35 percent of our sample that received door-to-door sales adopted, but this falls to around 10 in control villages. 21 This calculation is based on daily wages of 174 rupees (2.9 dollars), i.e. the wage in the central government’s labor guarantee program. We multiply this by 0.25 since the field days took approximately two hours. Finally, an average of 41 farmers attended the field days. 18 cost effective when seeds are readily available to farmers at market prices. 4.6 Do field days only generate demand by farmers that don’t interact with demonstrators? One interpretation of our findings is that field days simply transmit information to farmers that do not regularly communicate with demonstrators. Put differently, learning from peers is imperfect and slow because some people simply do not communicate with each other. If this explanation drives our findings, then field days should be most effective for farmers that otherwise would not have interacted with demonstrators. We test this idea using several different measures of how likely non-adopting farmers are to communicate with demonstrators. Additional heterogeneity analysis is generally not compatible with this mechanism. Starting with column 1 in Table 9, we interact the field days indicator with the number of demonstrators sharing the same surname.22 We would expect the interaction effect to be negative if field days are most effective for the relatively unconnected farmers. The point estimate on the interaction term is effectively zero. Columns 2 and 3 combine GIS data on the plot locations of demonstrators and the houses of farmers in the estimation sample. We calculate for each farmer the number of Swarna-Sub1 plots of demonstrators within a given radius from their household. We allow this radius to vary from 250 to 500 meters.23 We again find that the field days do not appear to be more effective for farmers that lived the furthest from the places where the new technology was being demonstrated. These additional tests do not favor the idea that field days simply substitute for peer-to-peer learning. Additional results in the online appendix show that this conclusion is no different when we relax the assumption that the field day effect varies linearly with the number of socially proximate demonstrators.24 We next show additional analysis suggesting that if anything, the field days were complementary with having an opportunity to learn from peers. More specifically, we focus our estimation on the SHG villages and test whether the field days were less effective for households having an SHG member. Essentially all of the demonstrators were SHG member households in SHG meeting villages. Therefore, we would expect field days (with the whole 22 Surnames are a common marker of social connection in rural India, largely because of their tight association with caste. The average farmer in the sample shares the same surname with one demonstrator. Figure A1 shows the distribution of this measure. 60 percent of farmers have a surname that is common with zero demonstrators. At the maximum, one farmer in the sample has the same surname as 11 demonstrators. 23 Figure A2 shows the distribution of the two measures across the entire sample. 24 Figure A3 shows separate treatment effects for farmers that are “close” to 0, 1, or 2 or more demonstrators. There is no clear pattern where the effect of field days is smallest for farmers that are close to more demonstrators. 19 village) to be the least effective for SHG households if field days only transmit information to unconnected farmers. The interaction effect between field days and the SHG household indicator in column 4 of Table 9 is large and positive, but not statistically significant (p=0.22). Column 5 estimates a larger — and marginally significant — interaction effect when we use an indicator for households that report being a friend or family member of the SHG president. In short, the field days were only effective for households that are close friends or family of the SHG president. Table A7 shows that these two interaction effects are robust to controlling for household characteristics. The coefficients actually increase in size and significance with the addition of several household controls. More importantly, we also estimate the same two models for the ward member and village meeting villages. The field days had no differential effect for SHG households in these villages. These “triple difference” results suggest that the differential effects observed in SHG villages are unlikely to result from unobserved correlates of SHG membership that cause these households to benefit more from field days. In sum, we interpret these findings to suggest that the field days do more than just transmit information to people that don’t know about new technology. The data seem to suggest that if anything the field days instead enhanced the learning process for farmers that were more likely to be linked to demonstrators and know about the technology.25 Or, farmers gain more from field days when demonstrators are similar to them. 5 Concluding Remarks There are several explanations why agricultural technology diffuses slowly. Slow learning is one such explanation. It simply takes time to observe a new seed variety or agricultural practice under the multiple states of nature where that variety or practice would be beneficial. The ability to learn from nearby farmers has been shown to improve this process and thus likely makes technology diffuse faster. The important next question is are there ways to improve how farmers learn from each other about new technology? We have shown that the farmer field day — where the experiences of early adopters are explicitly shared with other farmers — is one way to improve learning and increase technology adoption in a cost effective manner. The magnitude of this effect is non-trivial: field days increased adoption rates of an improved technology by 40 percent. This result suggests that learning is a key barrier that slows the diffusion of agricultural technology. Consequently, it is not enough to rely on select farmers to demonstrate and spread information. Rather, there 25 Table A6 shows evidence that independently of field days, SHG households possess more knowledge about Swarna-Sub1 in villages where SHG meetings were used to identify demonstrators. 20 is room to increase adoption by intervening to encourage farmers to better learn from each other’s experience, particularly if peers will continue to be the dominant source of learning about new agricultural technology. We showed suggestive evidence that the mechanism behind this effect is not that the field days simply engaged farmers that otherwise were unlikely to learn from demonstrators. We found that the field days were equally effective or if anything, they enhanced the learning ability of those that were the most likely to interact with demonstrators. More concretely, the field days appear to be a complement to learning from peers rather than a substitute. In addition, we considered whether ex-ante the selection of the first generation of adopters can be improved by seeking the input of various farmers through village meetings. We found that these meetings do change the composition of the group of demonstrators. More specifically, using meetings shifts the pool of demonstrators away from friends and family of the locally elected political figures. However, this has no meaningful effect on technology adoption one season later. This result in no way means that careful selection of demonstrators is unimportant. Indeed, Beaman et al. (2015) and BenYishay and Mobarak (2015) suggest that improved selection of demonstrators can increase adoption. Instead, our results suggest that using meetings to engage villagers in this selection process does little to drive adoption. Thus, future work is needed to identify the most policy relevant and scalable methods to improve the selection of demonstrators. The policy lesson from the experiment is straightforward. It is important to consider methods to enhance or optimize the ways in which farmers learn about new technologies. Our results suggest than an easy method does not involve much of a deviation from traditional agricultural extension. Rather than exploiting social learning alone, improved extension models could combine social learning from selected demonstrator farmers with simple interventions to encourage learning and sharing of results from demonstration. Put differently, there appears to be inefficiently low levels of information-sharing in the absence of outside intervention. 21