Our sample was drawn from the main maize growing areas of Ethiopia. We randomly generated four 10 x 10 km sampling grids within each of the four main maize growing zones (Jimma, Bako, West Gojjam, and East Shewa), with each grid subdivided into 100 1km2 grid cells. Within each grid, eight grid cells were chosen randomly. If a randomly selected grid-cell could not be included (either because it was physically inaccessible or if there was no maize production taking place at or near that point), then a replacement location was drawn from the same 10km x 10km grid.
Within each of the randomly selected grid cells, 6 farmers were identified, using the following protocol. First, we identified the farm household closest to the selected point. For example, if the point fell within a field, we identified the farmer who owns this field. If this farmer grew maize in the current year, then this farmer entered the sample, as farmer number 1 for that location. If the farmer did not grow maize in the current year, then we identified the nearest neighbor to that farmer and repeated until farmer number 1 is identified for that location. Second, from farmer number 1 (for that location), 5 neighboring farmers were identified on the basis of spatial proximity and direction. We started with the nearest farmer in the direction due North (0 degrees), and proceeded to the nearest farmer in a clockwise direction at 72 degree intervals. Once it was confirmed that they grew maize in the current season, they were added to the sample. If any of these farmers did not grow maize, their nearest neighbor (not otherwise already included in the sample) was evaluated for suitability, until a total of 6 farmers for the selected point location were identified. If, at the time of the survey, a sample farmer was not available (or unable to be enumerated due to death, leaving the village or no longer planting maize), a replacement was made following the same spatial proximity rules as used in the initial selection. This replacement farmer was given the same household identification number as the drop-out farmer. In our baseline line data, around 27 originally selected households (3.6 percent of the sample) were replaced.
The sample size was chosen on the basis of power calculations carried out for our main outcome variables. These are the absolute difference between recommended and actual fertilizer use in kilogram per hectare (kg/ha) for the maize area planted, total maize production in kg per hectare, and average per-capita household income. Randomization took place at the grid-cell level with each grid-cell randomly assigned to one of three groups. We stratified by four blocks, defined by administrative zones. Since the total number of grid-cells cannot be evenly assigned to the three groups, we first randomly assigned an extra grid-cell to one of the three groups. The first treatment arm was randomly selected to get one extra grid-cell. This means that treatment arm one has 44 grid-cells, whereas treatment arm two and the control group have 43 grid-cells each.
In addition, while each zone should have an equal number of grid-cells across the three groups, the uneven number of grid-cells prevents this. Hence, we randomly allocate the extra grid-cell in each zone to a particular group. Two zones take an extra grid-cell in treatment group one given that there is an extra grid cell assigned to this group, with the Bako and East Shewa zones randomly selected to have an extra grid-cell. Jimma and West Gojjam are randomly selected to have an extra grid-cell in treatment group two and the control group, respectively.
Households in the first treatment group received site-specific agronomic information (SSI), consisting of a recommended amount and blend of fertilizer to use on a particular maize plot, the optimal timing of the application of that fertilizer, and the expected yield outcome for that recommendation. The information was provided both verbally and on paper. Comparing outcomes ex-post with the control group will allow us to test whether fertilizer recommendations are more likely to be adopted when accompanied by agronomic information.
Similarly, households in the second treatment group received the same SSI as in treatment arm one and insurance cover to mitigate the risk associated with potential crop failure. For each farmer in this group, we purchased crop insurance from Oromia Insurance Company (OIC) and informed them that they were insured while we provided the site-specific recommendation. The cover includes any crop failure associated with drought, flood, excess rain, fire, storms, and hail for the 2018 agricultural season. This information was explained face to face to farmers during the planting period. As discussed above, the aim of this treatment is to allow us to test whether the downside risk associate with fertilizer investment plays a role in the take-up of the site-specific fertilizer use recommendation. During our sample period, around 25.6 percent of farmers were experienced biophysical shocks (drought, flooding, pests and diseases). Moreover, in the 2017 agricultural season, farmers lost around 14.7 percent of their crop income due to weather related shocks. As such, weather related risks are salient for these farmers. Despite this, the rate of crop insurance in rural farm households in Ethiopia is very low. From the baseline Agronomic Panel Survey (APS) data, less than one percent of households bought crop insurance in the main maize producing areas of Ethiopia.