Show and tell: farmer field days and learning about inputs with heterogeneous yield effects

Abstract

Organic inputs can be effective in reversing soil degradation and improving crop yields, but are often underused in a developing country context. This study seeks to determine whether farmer field days (FFDs) are effective in disseminating information about novel organic inputs, and the extent to which they increase demand for these products. Using experimental auctions to measure willingness to pay (WTP) among smallholder farmers in Western Kenya, we find that those farmers exposed to information from FFDs related to biochar and vermicompost, novel organic inputs, have lower WTP for the products. We present evidence that this is likely driven by two factors in particular: changes in perceptions of input profitability and heterogeneity in yields across demonstration plots within field day sites.

1. Introduction

Crop yields remain highly variable and far below potential in many developing countries, which is linked to significant levels of food insecurity and rural poverty (Barrett and Bevis, 2015). Organic inputs, such as farmyard manure, compost and crop residues, can be effective at improving soil health and increasing agricultural yields, but are often underused. For this reason, researchers have continued to seek the most effective means for diffusing information on agricultural technologies and practices related to organic inputs that enhance farm productivity. One method is farmer field days (FFDs), which are often led by farmers trained in new agricultural technologies or practices, such that FFDs can potentially aid in diffusing this information to their communities in developing countries. During farmer-led FFDs, these trained farmers hold on-farm events seeking to transfer information to community members who participate (FAO, 2011). While FFDs do not include the advantages of long-term experiential and discussion-based learning found in other programs, such as farmer field schools, they can aid in the cost-effectiveness of these programs through broadening the numbers of individuals exposed to new information in a particular area (Amudavi et al., 2009). Moreover, FFDs empower trained participants by showcasing their accomplishments to the community and demonstrating to local government officials the benefits of new technologies and methodologies, which may increase future support (FAO, 2011).

In this research, we contribute towards our understanding of best practices for FFDs by analysing the effectiveness of a FFD program on farmers’ demands for new organic agricultural inputs in western Kenya. This example is of particular interest given the growing recognition of the importance of soil health to maintaining agricultural productivity in Sub-Saharan Africa. We measure willingness to pay (WTP) by conducting experimental auctions for the organic inputs discussed in the FFD program with a random sample of individuals from the villages involved. Some of these individuals had exposure to the products discussed at the FFDs through attendance at the FFD event, by casually seeing the demonstration plots outside of the formal FFD, or through informal conversations with the FFD host or other community members. To causally identify the relationship between exposure to these novel organic inputs and WTP, we use the GPS-determined distance between a farmer’s homestead and the FFD location. Farmers who live closer to the FFD location are significantly more likely to have information about these novel agricultural inputs. We explore and defend this instrument in significant detail below.

Our analysis shows that individuals who had exposure to information about biochar prior to the experimental auctions bid 52.6 KSh (about 28 per cent of average WTP, p < 0.05) less for organic inputs than those without exposure to the inputs prior to the experimental auction. Prior exposure to vermicompost led to a 61.6 KSh decrease (33 per cent of average WTP, p < 0.05). Because these were completely novel types of organic inputs, farmers could only have any information about these inputs from the inception of the FFD program in the village, either through the FFD host farmer, demonstration plots, field day activities, or indirect information originating from these other sources. The causes for this decrease in WTP among those who had learned about these inputs likely include at least one of two factors: a decrease in expected profitability of the inputs and heterogeneity in information signals received by attendees at field days. Indeed, we present suggestive evidence that proximity to field day sites with lower average yields and those with higher variance in yields among demonstration plots with organic inputs are associated with lower WTP among those living nearby.

While FFDs are often integral aspects of these extension strategies that have the potential to multiply impacts throughout a community, there are relatively few economic studies that have focused specifically on FFDs and their impact on information diffusion. One, by Ricker-Gilbert et al. (2008) in Bangladesh, finds that FFDs are particularly cost-effective as they can reach a large number of farmers at low cost. More recently, Emerick and Dar (2021) show that NGO-led field days in India increase the adoption of improved seed by 40 per cent. Maertens, Michelson and Nourani (2021), however, find lower planned adoption among FFD participants compared to those in season-long, farmer-led programs in Malawi. Perhaps closest to this study, Morgan, Mason and Maredia (2020) in Tanzania find no effect from a lead-farmer extension program on farmer WTP for improved seed varieties. In light of these mixed results, the relative lack of rigorous studies analysing the effectiveness of farmer-led FFDs is notable given the large numbers of studies that focus on technology diffusion in developing countries.

Similarly, there are few studies that analyse the effectiveness of information dissemination regarding usage of organic inputs. Analysing the case of biochar in Kenya, Crane-Droesch (2018) shows how its adoption can spread through social networks. In Vietnam, Vu et al. (2020) find that informational videos of similar farmers describing their success using organic inputs can increase their adoption. Grimm and Luck (2023) in Indonesia find that a 3-day intensive training program for organic farming among smallholder farmers increases their experimentation with organic inputs. Our results contribute to this literature by providing evidence that FFDs can cause a statistically significant and economically meaningful impact on farmer behaviour with respect to organic inputs and offsets a common preconception that more information necessarily increases product valuation.

This study also makes important contributions to the general literature addressing cost-effective methods of information diffusion in developing countries. Traditional agricultural extension is relatively costly, and evidence has been mixed as to its efficacy (Anderson and Feder, 2004). As a result, there has been a search for alternative methods of information diffusion, such as using network analysis to target optimal entry points of information into social networks in order to maximize information diffusion from key farmers (Beaman et al., 2021). A significant literature has emerged analysing the impacts of learning from fellow farmers in developing countries (Foster and Rosenzweig, 1995; Conley and Udry, 2010; Ben Yishay and Mobarak, 2019). Because research shows that social learning can be highly effective for information diffusion, farmer-led FFDs represent a potentially strong though understudied tool for information dissemination.

This study is organized in the following way: we first explore our data and the project background, which includes a discussion of our calculation of the value of the organic inputs used in the study. We then discuss the empirical methods and identification strategy, focusing on our spatial instrument, following which we detail our results that show the significant impact of the FFDs on farmer WTP. We then discuss potential theoretical explanations for our findings as well as various robustness and falsification checks are next discussed, followed by our concluding remarks.

Background and data

The food security and incomes of farmers living across the sub-humid mid-altitude highlands in western Kenya depend to a large extent on the production of maize and the common bush bean. Yields of these two crops, however, have stagnated or decreased over the past decade (FAO, 2018) due to limitations caused by various abiotic and/or biotic factors. In partnership with the International Institute of Tropical Agriculture (IITA), multi-locational field trials analysed the effects of vermicompost¹ and biochar² agricultural inputs on the production of the common bush bean in Kakamega, Bungoma and Busia counties in western Kenya. These inputs were new to these villages and prior to these field days, were completely unknown.

In this project, 21 farmers across a wide area of western Kenya engaged in participatory demonstration trials with researchers, which involved learning how to prepare and use biochar and compost on their farm plots.³ In 2016, these trained farmers planted common bush beans on eight demonstration plots on each of their own farms. These plots included a control plot (no inputs), a plot where only biochar was used, another with only vermicompost, one with an agricultural inoculant, one with NPK fertilizer, and the remaining plots with combinations of these inputs. Results from these farmer-managed plots are given in Table 1, and demonstrate that most plots using biochar and vermicompost performed better than the control plot on the same farm, although yield impacts were heterogeneous both within and between farms.⁴ As expected, plots with NPK fertilizer generally performed better than plots with only biochar or vermicompost added (given the nutrient density of NPK fertilizer), although plots with biochar and NPK fertilizer were often especially effective.

Later in 2016 (shortly before harvest time), each trained host farmer held a single-day FFD for local farmers, showcasing management practices and the differences in crop yields between the control and treatment plots. These host farmers distributed IITA research results to FFD attendees, informing them about the impacts of both inorganic and organic input practices over varying agro-ecological conditions. Attendees also learned information regarding composting and about cookstoves that generate biochar, which provided insights into organic input generation.

In our study, these 21 trained farmers resided in 18 different villages, which are broadly representative of rural communities in western Kenya—generally poor and agriculture-based. In this research, we collected village-level lists of all household heads in these villages from local chiefs and village elders. A subset of the household heads was randomly selected for participation in the surveys and experimental auctions, resulting in a total sample of 884 individuals in 548 households. The survey instrument included questions on demographics, assets/income, agricultural production, market activity, organization membership and activity, self-reported attendance at a FFD, as well as questions that measured the level of knowledge about the products there were discussed at the field days. As can be seen in our summary statistics in Appendix Table A1, the majority of the individuals surveyed identify their primary occupation as farming, cultivating an average of about one acre of land. Of respondents across all villages, 24 per cent stated that they had attended a FFD organized through this project, although only 19 per cent and 15 per cent of farmers had heard of biochar and vermicompost, respectively, prior to the visit by survey enumerators (22 per cent had heard of at least one). This suggests some respondents who stated they attended the field days either did not attend, or attended and did not pay attention. Importantly for our empirical identification, host farmers were not chosen by village or location within village. This means that the FFD locations were as good as randomly located within the borders of a particular village (we will return to this point again below).⁵ Surveyed households were on average approximately 0.53 km from the nearest FFD location.⁶

Given that agronomic research generally shows positive long-run effects of organic inputs on soil health and crop yields, we seek to determine the value that farmers place on organic inputs such as biochar and vermicompost that have been introduced to farmers in western Kenya. More specifically, we are interested in whether those with any exposure to biochar or vermicompost have a significantly different WTP distribution for these organic inputs compared to those without, prior to the experimental auction. To elicit our WTP estimates, we use an experimental auction methodology after Becker, DeGroot and Marschak (1964) (BDM) to determine farmer WTP for several agricultural inputs. An advantage of the BDM auction is that it is incentive-compatible (i.e. represents true WTP), as it penalizes individuals for making bids outside of their true preferences, thus accurately aligning preferences with WTP measurements (Shogren, 2005).⁷ Moreover, the BDM auction can be implemented with each participant separately, ensuring that there is no bias from the presence of any other individuals.

During implementation, we first conducted two rounds of practice auctions with each participant for small food products (e.g. cookies) to ensure that the participants understood the auction methodology. Project staff then presented 1 and 5 kg packs of DAP (diammonium phosphate—a common inorganic fertilizer),⁸ biochar, vermicompost, cow manure,⁹ and combinations of these inputs to each participant in a random order, and the participant bid on each one. After bidding for each of the agricultural inputs, the enumerator’s tablet computer selected a random input and price. If the participant had bid at least that random price for that input, they paid that price and received the input, otherwise they kept the full cash endowment.¹⁰

Because of potential liquidity constraints, in this auction, farmers received a cash endowment from the enumerators totalling 70 KSh (0.69 USD at that time) for each of two practice auctions and 700 KSh (6.90 USD) at the beginning of the primary auction for the agricultural inputs.¹¹ Participants who did not use all of their cash endowment for the practice auction could carry it over to be used in the primary auction. Because many of the participants were unfamiliar with these inputs, prior to making bids, the enumerators read a short description of the agricultural inputs to all participants (included in  Appendix B), which focused on the composition of the inputs and potential benefits for crops that may occur from their use. All participants therefore had at least basic knowledge of the inputs prior to the experimental auctions.

In Appendix Table A2, we present WTP results elicited from the experimental auction methodology divided between those who state that they had exposure to biochar or vermicompost (defined as having at least heard of these inputs), which is the primary explanatory variable in our analysis. In general, we do not find significant differences in bids between the two groups in these raw results with the exception of bids for farmyard manure, where those with exposure to either of the novel organic inputs have higher average bids.

Empirical method and identification

Using the results from our experimental auctions that elicit WTP for organic inputs, we seek to determine the causal relationship between exposure to information about novel organic inputs (i.e. biochar and vermicompost) and the valuation of these inputs by the randomly selected participants. However, establishing a causal relationship is complicated by the fact that exposure to information about these organic inputs is potentially endogenous with WTP: there are numerous observed and unobserved variables that may influence exposure to these inputs, such as an individual’s experience, inherent motivation, or relationships with the FFD host or other attendees. Summary statistics in Appendix Table A3 show significant differences in observable characteristics between those with and without exposure to the inputs, such as education, farm size, past fertilizer usage and contacts with NGOs. This suggests that selection into exposure to these organic inputs may significantly bias OLS estimates of the relationship between exposure and WTP for the inputs. If we assume an unobserved variable, such as motivation, is positively correlated with both the choice of exposure to the inputs and WTP for the organic inputs, then OLS estimates of the effect of exposure on organic input WTP could be significantly upwardly biased. There may also be other unobserved factors that influence WTP for these organic inputs, such as constraints on obtaining complementary inputs, that could bias OLS coefficients.

To identify this relationship between exposure to information about these inputs and organic input bids, and to eliminate the potential sources of bias, we include distance from an individual’s household to the nearest FFD location as an instrumental variable (IV). This is a highly relevant instrument, as those who live closer to the location of a FFD are more likely to have heard something about these inputs prior to the experimental auctions. This is shown in Table A3, as those who have heard about biochar or vermicompost on average live 0.37 km from the nearest field day site, while those who have not heard of either live an average of 0.57 km from the site (difference significant at the 1 per cent level). There are many ways that an individual could have heard about the inputs: by attending a field day, casually observing the demonstration plots, speaking with the host farmer, or through indirect information. The precise way(s) that an individual was exposed to the information cannot be determined through our data, but they are more likely to have heard of the inputs by living closer to the field day site.¹²

Along with the instrument being highly relevant (i.e. strongly correlated with the hearing about the agricultural inputs), we argue that the identification strategy is valid through plausible exogeneity both in homestead and FFD locations within a particular village. As long as the homestead locations were not chosen to be close to the FFD site, we can assume that the homestead location is exogenous in our model. The vast majority of individuals in the sample (88 per cent) have either inherited their land or married into its possession, therefore their homestead location is usually not a choice. For the distance between the FFD site and homestead to be a valid instrument, the FFD location itself must also be exogenous. It is important to note that we use village-level fixed effects in the analysis, so we need rely only on within-village exogeneity; validity rests on the assumption that the field day site is effectively random within a village. In our case, we are fortunate in that IITA did not select the host farmer based on any consideration of their location within the village. In fact, IITA did not even have data on the village of the host farmer prior to this study, and as mentioned above, villages in this region are contiguous and their boundaries arbitrary; often only separated by a paved road. We wish to emphasize this point: due to the spatial randomness of field day host farmers within the de jure borders of a village in this densely populated area of western Kenya, FD locations are spatially random within the de jure borders of a village. Because the sampling frame consisted of households within the de jure borders, a household is randomly located with respect to the nearest field day site. We show an example in Appendix Figure A1, in which the field day site is located in the north-west portion of the village, with indicators showing that those in homesteads closer to the FFD site are more likely to attend the FFD. We present additional evidence and robustness checks below to further support our claim of the exogeneity of homestead location and FFD and thus the validity of the instrument.

Further evidence supporting our argument that the field day sites are effectively random within the administrative boundaries of a village comes from regressions of the distance between an individual’s homestead and the nearest FFD site on variables that are unlikely or impossible to have changed as a result of the field day program, such as age, years of education, land size, gender and level of assets. We also include a specification with soil chemistry variables, as there may be concern that farms closer to field day sites may have better (or worse) soil quality compared to other farms. The results in Appendix Table A4 show that only one variable is weakly correlated with the distance between a homestead and field day site across all specifications—the respondent’s primary occupation being farming. This weak statistical significance, though, and the absence of any other statistically significant variables contribute to our confidence that the distance variable is exogenous, satisfies the exclusion restriction and is appropriate for use as an instrument. The reader should, however, please note that additional placebo and falsification tests for this instrument are located in the Robustness Checks section of this paper.

Another component designed to strengthen our empirical strategy is the use of auction data on an unrelated good. As Dizon-Ross and Jayachandran (2022) demonstrate, the use of bids on unrelated benchmark goods in a BDM auction framework can improve the accuracy and precision of WTP, specifically by controlling unrelated variation such as day/time effects or social desirability bias. The authors show that an optimal choice for this benchmark good is one unrelated to the good of interest (i.e. in this study, the novel organic inputs) such that the preferences for the two goods should be orthogonal. Secondly, the benchmark good should be inexpensive or have weak income effects on demand, so as not to overcontrol for income effects. An optimal good in our context is cookies, which were used in the practice auction rounds. As described above, we held two practice rounds to familiarize the participants with the auction methodology. We include the bid for cookies from the second practice auction round in these estimations, as the bids in the first practice round are likely less accurate given that the respondents were just becoming familiar with the auction methodology.

We thus conduct two-stage least square (2SLS) estimations, with β₁ in Equation 1 measuring the average impact of exposure to information on the novel organic inputs on WTP:

where Bid_ik is the bid by individual i for organic input k, which contains four observations for each individual i: bids for 1 and 5 kg quantities of vermicompost and biochar, respectively, and variables I_k are input-quantity level controls (e.g. 1 kg biochar, 5 kg biochar, etc.). Variable Info is an instrumented, binary variable (described above) indicating whether the individual stated that they had heard of biochar or vermicompost prior to the visit by project enumerators. Variable M is whether the individual purchased an item in the practice auction round; variable C_i is the individual’s bid for the benchmark item, cookies, in the second practice auction round; and ϑ represents village, enumerator and survey month fixed effects. In addition, vector X_n represents an optimal configuration of controls that are selected from a list of exogenous characteristics, including quadratic transformations, using a machine learning algorithm (double LASSO) (Urminsky, Hansen and Chernozhukov, 2016; Chernozhukov et al., 2017).¹³ Additional specifications of estimated Equation 1 also include the respondent’s perception of their soil’s health (self-reported on a scale of 1 to 5) and soil chemistry (nitrate, phosphate and organic carbon) measured through soil tests on the respondent’s farm.

Results

Estimations of Equation 1 demonstrate significant differences in WTP distributions between those with and without prior exposure to information about the organic inputs. Recall that we identify these estimations by using distance between the farmer’s homestead and the FFD site (an average distance of 0.53 km), which we treat as effectively random within a given village (see further tests for this assumption below). Results from our first-stage results are in Table 2 and are of expected directions and magnitudes: a 1 km increase in distance from the field day site is correlated with a 33 percentage point (pp) decrease in the likelihood of hearing any information about biochar (Column 2), a 28 pp decrease in the likelihood of hearing any information about vermicompost (Column 4), and a 33 pp decrease in the likelihood of hearing information about either (Column 6).¹⁴ Moreover, the results are highly statistically significant, with first-stage test statistics on the instrument ranging from 31.5 to 39.0.

As discussed earlier, naive OLS estimations will likely be significantly biased due to unobserved variables correlated with both exposure to the novel organic inputs and their WTP. We include results of these estimations in Appendix Table A6, which show negative relationships between hearing about biochar, vermicompost, or either novel organic input and WTP, although the coefficient magnitudes are relatively small. We would expect, however, that many unobserved variables are positively correlated with both exposure to information about the inputs and WTP (e.g. motivation), which could lead to upward bias on the OLS coefficient estimates.

We therefore show our 2SLS results in Table 3. Across all specifications, we find that exposure to information about these novel organic inputs, as measured by having heard about any of them prior to the survey, leads to lower WTP relative to those without prior exposure. We find that exposure to information about biochar decreases WTP for organic inputs by 52.6 KSh (Column 2, p < 0.05) relative to those who had not heard of biochar prior to the auctions—a 28 per cent decrease relative to the sample mean. Similar estimations in Columns 3 and 4 show similar results for vermicompost: exposure to information about vermicompost leads to a 61.6 KSh fall (Column 4, p < 0.05) relative to those with no prior exposure—a 33 per cent decrease relative to the sample mean. Lastly, we analyse the effect on WTP for the novel organic inputs from hearing about either biochar or vermicompost. Column 6 shows that exposure to either input decreases WTP by 52.0 KSh relative to those without exposure (p < 0.05).

Other results from Table 3 include the effect of a farmer’s soil perception on WTP for organic goods. Recall that for this variable, the respondent indicated their perceived soil quality on a scale of 1–5 (5 being highest quality). We might expect that those who have lower soil quality might be especially interested in these inputs given their potential to improve their soils. Columns 2, 4 and 6 show that perceived soil quality has a positive relationship with WTP for the product, although the estimated magnitude is fairly small and not statistically significant, suggesting participants were not linking their perceived soil quality levels to WTP for the inputs. We likewise find no statistically significant relationship between a farmer’s soil chemistry and WTP for the organic inputs (Columns 1, 3 and 5). Also, as expected, we find high correlation between bids for the benchmark good (cookies) and bids for the novel organic inputs, indicating the utility of this variable to control for social desirability bias and other statistical noise from the experimental auctions.

Potential mechanisms

In this section, we discuss three potential mechanisms for our findings as well as limitations with each: updating perception of expected profitability, updating perception of production costs, and the high variance of information signals received.¹⁵ To better understand the third mechanism, we describe a model that supports our findings. We also describe empirical evidence that supports two of these potential mechanisms.

Expected profitability

The first potential mechanism explaining the negative relationship between information exposure and WTP for these inputs is perhaps the most straightforward: if farmers who learn about these organic inputs realize that their mean effects on yields are lower than expected, they will have lower average WTP compared to those who have little information about the inputs. While we cannot causally determine whether this is indeed a mechanism, we can obtain some supporting evidence by analysing the relationship between average yields from the nearest FFD demonstration plots and WTP for the organic inputs. We present results from this exercise in the first two columns of Appendix Table A7. The dependent variable is the WTP for organic inputs, as in Equation 1 above. We measure mean yields by taking the mean of the yields of the five demonstration plots at each field day site that included organic inputs (as shown in Table 1), and subtracting the yield from the control plot at that site.

The results show that there is a statistically significant relationship between mean yields at the nearest field day site and WTP: an increase in mean yields by 1 kg per hectare is associated with an increase in the WTP for the organic inputs by 0.07 KSh on average (Columns 1 and 2; p < 0.01). However, we must interpret these results with caution. First, this is not a causal analysis. Secondly, variation in the data come from only 19 distinct field day sites (of a total of 21).¹⁶ With these caveats in mind, these results suggest a significant relationship exists that supports this mechanism and indicate that those nearby field days with lower average yields also have lower WTP for the organic inputs.

Production costs

A second potential mechanism is related to changes in farmers’ perceptions of production costs associated with these organic inputs. It is possible that those who learn more about these inputs also learn that the opportunity costs associated with producing these inputs are relatively high. Indeed, own production of both biochar and vermicompost involves substantial labour costs, as well provision of raw material. These high production costs may lead farmers to be less interested in these inputs.

However, we believe that the actual effect of learning about true production costs on WTP is ambiguous. While high associated labour and raw material costs may lead some farmers to become less interested in the inputs, it also would increase a farmer’s implicit price (i.e. shadow price) of the input. This would potentially lead to higher personal valuation of the input, increasing bids. Unfortunately, we do not have data to measure the effect of this particular mechanism, but look forward to future research that investigates this research question.

Variance of information signals

A third mechanism, which we explore in detail below, is that the number of demonstration plots at each field day site leads to noisy information about the effectiveness of the organic inputs, which can decrease WTP. To analyse this mechanism, we apply a model developed by Johnson and Myatt (2006) and Liaukonyte, Streletskaya and Kaiser (2015). In our context, we assume a new input k with profitability |$\pi_k(E_k)$|⁠, where for simplicity we assume π_k is primarily a function of the unit increase in yields for a unit increase of the input applied per hectare (or agronomic efficiency), E (Vanlauwe et al., 2011). Profitability for input k, π_k, is not constant, but varies by soil condition, rainfall, etc., and we assume a normal distribution for π_k defined by |$N(\mu_k, \sigma_k^2)$|⁠. The variance σ² is small when the profitability of k is similar for all farmers, and large when there is significant heterogeneity in profitability. Given the large degree of variation in soil nutrient levels in western Kenya, for many inputs, σ² is often large. Before receiving any information through a FFD or enumerator visit, for simplicity we assume that an individual i has a prior on µ and σ².¹⁷ The farmer receives an information signal, ω_k, regarding the profitability of input k, where |$\omega_k\sim N\left(\pi_k,\zeta_k^2\right)$|⁠.

In this study, there are two groups (g) of farmers: group A, which is composed of farmers who have some exposure to the products prior to the experimental auction through, for example, a FFD held in their village. The second group (N) is composed of farmers who are completely unaware of the novel organic inputs; |$g=\{A,N\}$|⁠.¹⁸ Farmers in each group g receive different information signals about these novel inputs, such that |$\omega_{Ak}\neq\omega_{Nk}$|⁠. Importantly, we assume that the variance of the information signals to those with exposure, ω_Ak, is |$\zeta^2_{Ak}(var(E_{k}))$|⁠, that is, a function of the variance of the agronomic efficiency of input k. This is because in viewing demonstration plots with the novel organic inputs, these farmers receive complex and potentially noisy information signals regarding the input’s profitability at a field day site as it varies with its agronomic efficiency E (itself a function of soil nutrient levels and other site-specific conditions, which vary based on field day site). Indeed, as Table 1 illustrates, in the demonstration plots in this study there were heterogeneous effects of the organic inputs on crop yields, not only across host farms, but between the eight demonstration plots on each farm. The information signal for those unaware of the products (group N), however, is simply the brief information provided to them at the time of the survey by the enumerators, such that we assume |$\zeta^2_{Nk}$| to be smaller and only a function of individual enumerator characteristics (which are controlled for in our estimations).

With Bayesian updating, a farmer’s updated posterior of input profitability becomes:

If we assume that farmers are maximizing their utility under particular risk preferences, we can arrive at their WTP by inserting the results of Equation 1 into a certainty equivalent |$E[\pi_k|\omega_{gk}]-\lambda var[\pi_k|\omega_{gk}]/2$|⁠, where λ is a parameter indicating level of risk aversion (assumed, for simplicity, to be uniform across participants) (Featherstone and Moss, 1990):

This, we can see, is a weighted average of the original distribution of π_k and that of the information signal. If we assume that realized information signals are distributed |$\omega_{gk}\sim N(\mu_k,\sigma_k^2+\zeta_{gk}^2)$| and are linear in ω, then, following Liaukonyte, Streletskaya and Kaiser (2015), we have the following distribution for WTP:

which provides us with the following partial derivative of |$\zeta_{gk}^2$|⁠:

The partial derivative is negative, indicating that the mean of the WTP distribution falls with increases in ζ², the variance of the information signal. Recall that we assume that for those with exposure to input k (i.e. by talking to the host farmer, casually viewing the demonstration plots, or attending a field day), ζ² is a function of the variance of the agronomic efficiency of input k, and thus |$\zeta_{Ak}^2\gt\zeta_{Nk}^2$|⁠. A wider array of information signals received may decrease the attractiveness of the input (i.e. increasing uncertainty), and can potentially decrease mean WTP. The eight demonstration plots at each site, each with different combinations of inputs and varying yields, leads to a high variability of information signals and may contribute to uncertainty about the effectiveness of the inputs.

In Columns 3 and 4 of Appendix Table A7, we provide suggestive evidence for this mechanism. We take the standard deviation of the yields on demonstration plots with organic inputs from each field day site (as presented on Table 1). We then analyse the relationship between the standard deviation from the nearest field day site and a farmer’s WTP for the organic inputs. The results suggest a strongly inverse relationship: an increase in the standard deviation by 1 kg per hectare is associated with a decrease in the WTP for the inputs by 0.26 KSh (Column 4; p < 0.01). We apply the same caveats, however, to this analysis as above when discussing the effects of mean yields on WTP above—this is not a causal analysis and the variation is limited to differences in 19 field day sites.

Extensions and robustness checks

There are several additional aspects that we consider with regard to our specifications and estimation results.

Other measures of input knowledge

In our primary estimations (Table 3), we use variables indicating whether the participant has heard of the novel organic input as a proxy for exposure to information about the product. However, hearing of a product could include a range of knowledge levels, from simply hearing the name of the product to having detailed information. Additionally, enumerators probed the respondents to determine how much information they had about the novel inputs. Thus, as another variable, we use one that indicates whether the respondent could describe the input. This group represents a subset of the sample who only heard about the product. For example, in our random sample, 19 per cent of the sample had heard of biochar, but only 13 per cent could describe biochar. Similarly, 15 per cent had heard of vermicompost, but only 9 per cent could describe it. Overall, 22 per cent had heard of at least one of the new products, while 14 per cent could describe at least one.

Using these alternative independent variables based on whether they could describe the variable, we conduct additional 2SLS estimations.¹⁹ The results of these new estimations, shown in Appendix Table A10, reveal that using the variables that indicate an ability to describe increases the magnitude of the effects on WTP for the novel organic inputs, but also increases the standard errors leading some lower statistical significance in some specifications. This appears to be driven by a slightly less strong first stage for the instrument. Nonetheless, the large and statistically significant impacts using these alternative measures of information exposure provide further support the findings that exposure to information from the field day program led to lower WTP.

Wild bootstrapping standard errors

There may be some concern that standard errors may be artificially small given the relatively few village clusters (18). To mitigate these concerns, we conduct Wild Bootstrap estimations after Cameron, Gelbach and Miller (2008), which corrects for the small number of clusters. p-values from the original estimations in Table 3 are shown together with p-values from the bootstrapped estimations in Appendix Table A11. Although larger, the wild bootstrapped p-values are not highly dissimilar than those in our original estimations. For some estimations, there is a drop of one significance level (i.e. from 5 per cent level to 10 per cent level), but all maintain statistical significance indicating that the relatively small number of clusters are not leading to Type 1 errors.

Placebo test for instrument

Along with knowledge of organic inputs taught during the field days, in the survey we also asked questions to gauge farmers’ knowledge of agricultural nutrients and conduct a placebo test. We would be concerned if there were a similarly statistically significant relationship between distance to the field day site and knowledge of any of these nutrients, which would suggest that distance to the field day site was non-random. The enumerator asked the respondent to list as many agricultural nutrients as they knew (without prompting). Nutrients listed by a non-negligible number of respondents include nitrogen (of which 27 per cent of respondents stated they knew), phosphorus (13 per cent), potassium (10 per cent), calcium (11 per cent) and carbon (2 per cent). We then conduct estimations similar to those for the first stage of the instrument above: regressing each of the binary variables indicating stated knowledge of the nutrient on distance to the FFD site along with LASSO selected controls and fixed effects. We are reassured by the results presented in Appendix Table A12. None of the estimated correlations are significant at conventional levels, thus suggesting that distance to the field day site is not correlated with farmers who are generally more informed about properties of agricultural inputs.

Falsification tests for instrument

Earlier, we explained that the validity of our IV (distance from homestead to field day site) is plausibly exogenous in our empirical model due to the effectively random location of the field day within each village. However, we conduct additional tests on our IV strategy to provide further assurance of its validity.

As an initial falsification test, we follow a randomization inference framework and seek to determine whether it is not the distance from a farmer’s homestead to the FFD site that matters, but distance from a farmer’s homestead to another random farmer’s homestead in the village. We test this possibility by running simulations with randomly sampled homesteads in the village serving as simulated FFD sites. For each simulation, we test whether the randomly chosen homestead in each of the villages provide us with a significant 2SLS first stage (distance between simulated FFD site and FFD attendance). If we find a large percentage of these simulated distances provide us with significant results, rather than distance from a farmer’s homestead to FFD site driving the strength of the instrument, it would suggest that it was distance between farmers’ homesteads in general that is providing the strength for the instrument. If this is the case, we would need to develop alternative explanations to defend the exclusion assumption of our instrumental variable, which could threaten our identification strategy.

In each of 1,000 simulations, a randomly chosen homestead in each village was designated as the simulated FFD location. Using the distance between each homestead and the simulated FFD location, each simulation conducted a first stage estimation (specifically that in Column 6 of Table 2 with the selected LASSO controls), and reports calculated F-statistics and p-values for the distance to the simulated FFD variable. The results located in Appendix Table A13 calculate a randomization inference p-value and present additional statistics to show that only a small share of these estimations had statistically significant results. Only 5 per cent of simulations resulted in an F-statistic above 5.05 or a p-value less than 0.038. This is compared to F-statistics of 31.5 and p-value of 0.0 in our first-stage estimation using distance to the actual FFD site. These results demonstrate that distances to random homesteads that serve as simulated field day locations are generally not strongly correlated with field day attendance, suggesting that our instrument genuinely represents a strong correlation between distance to the field day site.

In another test, we analyse the potential impact of outlying homesteads in our estimations. While we described above that field day locations were effectively randomly located within each village (as IITA did not choose location based on village boundaries) and not necessarily near the centre of a village, there may still be concern that the instrument is being driven by those who are located far from the village centre. The argument may be that those located far from the village centre may be both less likely to hear about the novel inputs and also have fundamental differences in their relative WTP for the inputs. If this were the case, it would violate the identification assumptions. We test this by re-estimating our first stage with distance between homestead and village centre as our hypothetical instrumental variable. Our results, reported in Appendix Table A14, are reassuring. There is no statistically significant correlation between a homestead’s distance from the village centre and hearing about either biochar or vermicompost. There is mild to moderate correlation between hearing of vermicompost and distance to the village centre, however, the first-stage test statistic is low (3.57–4.32). Moreover, unlike the strong negative correlation we found between distance from the field day site and hearing about a novel organic input, the correlation here is small and positive, suggesting that, if anything, those farther from the village centre were more likely to hear about vermicompost. We conclude, then, that outlying homesteads in the sampled villages are likely not a threat to the validity of the instrument used in our primary estimations. Overall, the robustness checks presented above generate confidence in the instrument and estimation results.

Discussion and conclusion

Farmer field days (FFDs) are a potentially effective method to diffuse productivity-enhancing information or technology within developing country communities. FFDs are one-day programs anchored on a demonstrator of a technology that is hosted by a trained farmer and attended by their peers in the farmer’s village. Such an outreach model is useful to magnify the impacts of information and technologies incorporated in intensive, participatory, on-farm programs such as a farmer field schools or contact farmer programs facilitated through extension services, which are costly to implement. However, to this point, there have been few evaluations measuring the impacts from these FFD programs. This is in part due to identification problems: attendance at a FFD is a choice by the farmer, which is influenced by numerous unobserved factors such as motivation and connections with the FFD host.

We identify the relationship between information disseminated through a field day program and WTP by collecting data on whether an individual has been exposed to information introduced through the field day program. We then use the GPS-measured distance between a farmer’s homestead and the nearest FFD location (the host farmer’s farm), the location of which is effectively random within each village. Using this distance as an instrument, we find that exposure to information decreases the farmer’s WTP for the featured organic inputs relative to WTP of farmers with no prior exposure.

These results suggest that farmer-led FFDs affect farmer demand for agricultural products—in this study’s context, organic inputs. While most of the yield effects on treatment plots were positive relative to control plots, we show causal evidence that knowledge of the inputs led to lower WTP. We find suggestive evidence that supports two of three key mechanisms. First, evidence suggests that lower mean yields on demonstration plots using organic inputs at a nearby field day site lower a farmer’s WTP for the product. Second, we illustrate how noisy information attributable to heterogeneous yield effects of the organic inputs on a field day site and the uncertainty they generated could have contributed to the decrease in average farmer WTP for the organic inputs. Data are unavailable to test a third potential mechanism, that improved knowledge of the costs of organic input production affected WTP.

Because of the highly degraded soils in much of sub-Saharan Africa, crop yields are often below potential, resulting in food insecurity and poverty among many small-scale farmers. Therefore, rigorous evaluations of methods for disseminating information on productivity-enhancing technology are needed. One method of information diffusion, FFDs, can potentially be a cost-effective method of widespread information diffusion. This study demonstrates both the potential and limitations of FFDs, which continue to be a widely used mechanism of information diffusion in developing countries.

Acknowledgements

Dries Roobroeck, Research Associate, International Institute of Tropical Agriculture. David R. Lee: Professor Emeritus, Cornell University. Special thanks to Janice Thies, Jura Liaukonyte, Johannes Lehmann, Amanda Kerr, Edmundo Barrios and AAEA annual conference attendees for valuable comments and suggestions. We also thank Ray Weil and Cheryl Palm for the use of the SoilDoc soil testing system, and James Agwa Otieno, Samuel Omollo, Ezekiel Avedi, Harim Awiti, Martin Wakoli and Khisa Nanyama for assistance in collecting the data on which this study is based.

Funding

Funding for this study was provided by Cornell University’s Atkinson Center for a Sustainable Future, a U.S. Borlaug Fellowship and several grants from Cornell University. Other major support for this research came from a USDA National Institutes for Food and Agriculture grant, from the International Institute of Tropical Agriculture and from the World Agroforestry Centre. None declared.

Conflict of interest

None declared.

Footnotes

References

Appendix B Description of novel organic inputs provided to all respondents

The following descriptions of the novel organic inputs included in the village farmer field days were read to all participants during the experimental auctions.

‘Biochar’ is a type of charcoal that is produced from left-over plant material of field crops on farm like maize cobs and stovers, rice husks and haulms, sugarcane bagasse, coconut shells and others. If applied to soil at the correct rate, biochar helps to improve crop production by increasing the uptake of fertilizers, manure and water.

‘Vermicompost’ is the end-product of the breakdown of organic matter by an earthworm, also called worm castings. It is compost produced using earthworms. If applied to the soil in the correct rate vermicompost will improve crop production because it contains substantial amounts of nutrients, has a large water holding capacity and enriches the soil with micro-organisms.