David Murphy is a PhD Candidate at Cornell’s Dyson School of Applied Economics and Management.
In a May 2016 blog post, I described a plan for fieldwork in western Kenya that would investigate whether providing soil testing results together with individualized agricultural input recommendations would impact farmer behavior towards input usage. The motivation for this work stems from the recognition that small-scale farmers in Sub Saharan Africa (SSA) are often risk-averse and this is recognized as one component causing underinvest in agricultural inputs given the low and variable yield responses to fertilizers and other inputs (Kebede et al., 1990). Past studies have shown that farmer perceptions of the health of their soils is primarily based on previous crop yields (Marenya et al., 2008; Berazneva et al., 2016), which may not be accurate or sufficient for optimal agricultural decision making. This is because nutrient deficiencies are difficult to differentiate solely based on crop yield information, which can impact the effectiveness of inputs. For example, inorganic fertilizers containing necessary nutrients such as Nitrogen and Phosphorus may be ineffective at improving yields if organic matter in the soil is low (Marenya and Barrett 2009a,b) or the soil is acidic (Burke et al., 2017).
To test whether providing personalized soil test results and recommendations to farmers is a beneficial and cost effective agricultural development strategy in SSA, this project sampled soil from 550 small-scale farms in 18 villages in Western Kenya and brought the results to these households between July and November of 2016. Household heads were randomly selected from official village rosters, and selected household heads and their spouses (if alive/available) were individually surveyed, providing a total sample size of 884. We used individual, two-round experimental auctions based on Becker-DeGroot-Marshack to determine whether farmer valuations for various quantities of six different inorganic and organic inputs changed after receiving the soils information and recommendations. Following practice auctions, a cash endowment, and a baseline auction that measured the willingness to pay for each input by each farmer, each participant was randomly assigned by the enumerator’s tablet computer into one of the following treatments:
Treatment 1 (Input Recommendation [IR]): Enumerators presented the participants with their soil test results, explaining them thoroughly and providing tailored fertilizer recommendations developed for this project by the International Institute of Tropical Agriculture.
Treatment 2 (Village Comparison [VC]): Enumerators showed the participant a chart that compared their soil test results with other anonymized test results in their village and a village average. The enumerator pointed out their placement on the village distribution but provided no specific fertilizer recommendations.
Treatment 3 (Combined treatment [IR&VC]): Participants in this treatment received both treatments 1 and 2 together.
Control: Participants received no information transfer between auction rounds.
Immediately after the information treatment, an identical second auction round was held during the same sitting. Thus, the only change between the two auction rounds was the information treatment. After the second auction round, a random price, random input, and random auction round were chosen as binding. If the respondent bid the random price or higher for that input in the chosen round, they paid the random price and received the input. Otherwise, the respondent kept the full cash endowment. After the conclusion of the experiment, all farmers received the full set of information about soil tests and input recommendations, regardless of the treatment.
We are interested in two major research questions connected with the information transfers. First, did the information transfer affect farmer valuation of these inputs by updating their beliefs regarding which inputs are most effective for their own soils? Secondly, conditional on receiving an input recommendation, did farmers increase (decrease) their bids for inputs that were recommended (not recommended) for their soils?
We’ll address the second question in an upcoming post. The first question is addressed by looking at the distribution of the change in bids before and after each treatment. We would expect that treatments would be most effective at influencing behavior if the distribution becomes more disperse (flatter) – the treatment leads to more widespread updating of farmer beliefs. Below, Figure 1 is a box plot (boxes as 75th percentile, lines 95th percentile, and dots as observations outside of the 95th percentile) of the distribution in the percent change in bid for each treatment for one example input (biochar). This figure suggests that the treatment distributions appear to be more disperse than the distribution of the control.
Women are of particular interest in this study because they are more active in the agricultural sector than men in Kenya, yet they tend to adopt technology at lower rates than men (Ndiritu et al., 2014). In Figure 2, we see that women are more likely to change their bids in the control group (after receiving no information) compared to men (Table 1 shows sample sizes for these groups). If we interpret the level of effectiveness of the treatment as the amount of dispersion caused by the treatment minus the dispersion of the control, then it seems that the treatments were less effective for women than for men.
Preliminary random coefficient regressions with estimated standard deviations (SD) that include multiple levels of controls and fixed effects appear to support this finding. The SD of the control group for men (7.7) differs at the 1% level from the SD of the control group for women (23.5) Meanwhile for women, there is no statistical difference in the SD between the control group (23.5) and the treatments (25.89, 26.16, 18.69 respectively for IR, VC, and IR&VC), indicating that the treatments were not more effective in changing behavior than the control group for women. On the other hand, for men we see statistical significance at 1% between the control group SD and the SD for each treatment, suggesting that the treatments were effective in changing behavior in their case.
Table 1: Sample Size by Group*
|T3. IR & VC||128||77||205|
*Uneven distribution among treatments due to random assignment by tablet computer at time of auction.
As we move ahead in our analysis, we are particularly interested in learning why the treatments appear to have had such a differential impact between men and women. Some initial ideas include a larger Experimenter’s bias or Hawthorne Effect for women, or less confidence in their own baseline bids. Please keep an eye on the blog as we continue to learn more from our data!
Funding for this project provided by: The Atkinson Center for Sustainable Development, a U.S. Borlaug Fellowship, and several grants and fellowships from Cornell University including: A Richard Bradfield Research Award, a Frosty Hill Fellowship, an Andrew Mellon Fellowship, and an Einaudi Travel Grant. Crucial support for the project also comes from the Cornell project “Improving bean yields by reversing soil degradation and reducing soil borne pathogens on small-holder farms in western Kenya” funded by a USDA NIFA grant, and from the International Institute of Tropical Agriculture (IITA).