The contract included all the attributes of T2 and added the provision of inputs, on loan, from ESOP. The contract stipulated the amount of seed and fertilizer to be provided as well as the price for the inputs. At the end of the season, the total cost of inputs provided would be deducted from the price paid to the farmers. ESOP did not charge any interest on the loans but only recouped the cost of the inputs.It is natural to wonder about the degree to which farmers in our sample are subject to price risk and credit constraints, and their level of efficiency in rice production. Regarding price risk, we obtained monthly data from the Ministry of Agriculture for the period 2010–2019. Fig. 2 graphs the monthly price and overlays gray bars to designate the harvest season. Prices show a high degree of seasonality, with prices falling by as much as 30 percent in the periods immediately after harvest. To this graph is added a dashed line that represents the guaranteed price provided by the contracts. The price of 150 CFA offered by ESOP is slightly above the eight-year monthly average of 144 CFA .12 While the contract price is typically below the seasonal high, in each of the last eight years it has been above the seasonal low, which occurs at harvest.As part of the baseline survey, we asked farmers about the credit they received from both formal and informal sources. Seventy-eight percent of farmers reported that they lacked access to credit. For those who did have access, credit utilization was low. While the mean dollar amount received was US$99.36 the median value was zero. Levels of rice production were also low. Africa Rice estimates that rice producers in Benin should be able to average 2000 kg per hectare in rainfed production and 3500 kg per hectare under irrigation.

According to the most recent data available from FAOSTAT, which aggregates rainfed and irrigated rice, vertical farming racks average rice production in Benin was 3140 kg per hectare in 2014. In our baseline data, the average yield was 822, with a median of 400. While these values are simply descriptive, they do offer prima facie evidence of farmers in the sample operating well below the technical efficiency frontier.In the baseline survey we asked farmers about their previous experience with contract farming. Contract farming was relatively well known among participants, with 60 percent aware of the existence of contract farming and 74 percent of these having engaged in at least one contract for crop production . The majority of these contracts, 67 percent, were written agreements. Around 79 percent of farmers had engaged in a contract that set the price while 74 percent of farmers had engaged in a contract that set the quantity. The most common type of contract was one that included terms governing quality. About 70 percent of contracts included input loans and 36 percent included some aspect of production-management. Note that most farming contracts were for cotton, which is a cash crop in the surveyed area. We investigate contract farming’s impact on four key outcome variables. Rice area is measured as the total land in hectares area cultivated with rice. Yield is measured as the total amount of rice harvested in kg divided by area cultivated. Market participation is measured as the share of harvested rice that was sold, either to a rice processor, such as a parboiler, or into the market for paddy rice.Rice that is not marketed is either kept for consumption or saved for seed. Household income is measured as the sum of income from rice, from other crop production and livestock sales, and any income from non-farm activity. Non-farm activity consists of salaried employment, small time trading, wage labor, and remittances.

Crop production not sold is valued at market prices, though we do not value livestock consumed by the household or on-farm household labor as we lack detailed data on these inputs. Total income is then divided by household size to arrive at income per capita. Our analysis relies on self-reported measures for all variables. In the last couple of years, a number of studies using observational data have shown that several stylized facts in the development literature, such as the inverse productivity-size relationship, are in reality a result of bias introduced by non-classical measurement error . The lack of objective measures in our experiment is not ideal, though in our context measurement error likely introduces only noise, not bias, into our estimates. This is because any measurement error in our self-reported outcome variables will be orthogonal to our randomized treatment, absent any Hawthorne effects.If farmer recall of any of our outcome variables is better remembered when under contract, Hawthorne effects may exist. Such effects are not a given, though it is possible that farmers with contracts report yields with less measurement error than those not under contract. To check whether those assigned to treatment remember better than those in the control, we calculate the co-efficient of variation across groups at endline. If Hawthorne effects do exist through focusing treated farmers on outcomes, we would expect less variation in treated outcomes. As can be seen from Table B1 in the Appendix, co-efficients of variation are similar for those in the control compared to those in the treatment. We take this as evidence that any measurement error in our variables of interest is classical in nature, being uncorrelated with random assignment to treatment.Farmers randomly assigned to the pooled treatment had a significantly larger area planted to rice prior to the experiment. Control farmers tended to plant 0.62 ha to rice while treatment farmers tended to plant 0.86 ha to rice.

Average yields in the baseline vary between 820 and 980 kg per hectare but with large standard deviations and no differences across treatment and control. For market participation we see some differences across multiple treatments. Farmers randomly assigned to the control and T1 sold about 30 percent of their pre-experiment rice production into the market. By comparison, farmers randomly assigned to the other two contracts sold about 45 percent of their pre-experiment rice production in the market. Despite this greater share of market participation prior to the experiment, per capita income was no different across the four groups, with average income being about $220 per person. To control for where we lack baseline balance in our dependent variables, we use an Analysis of Covariance estimator. Among our control variables, the average farm household had eight members with the head of the household aged 40 years. Around 60 percent of households were male headed with the household head having grown rice for around eight years. Only ten percent of household heads had even a primary education while 90 percent of households listed farming as their primary business or activity. Nearly 100 percent of household heads were members of a farming association. Farmers varied in whether or not they had participated in training on rice production. While only 20 percent of farmers randomized into T1 had participated in training, around 55 percent of farmers in the control and other two treatments had training in rice production. In addition to checking balance by examining the correlation between treatment assignment and each individual outcome variable or household characteristic, we also regress treatment assignment on the complete set of outcome variables and covariates. Table 3 presents the results from these six regressions as well as the F-statistic from a test of joint significance. In general, both of our balance checks suggest that our randomization was effective, though differences do exist across a small number of variables. Where the F-statistic is significant, this is typically due to significant differences in control variables, such as past participation in rice training, and not due to differences in the outcome variables. In general, differences do not appear to be indicative of systematic variation across multiple treatments and we employ an empirical strategy that allows us to control for where differences do exist.Our experimental design involved a baseline survey prior to randomization, random assigned prior to planting, and an endline survey seven months later, after harvest. Because of this time delay we did experience attrition among the farmers in our experiment. Of the 953 farmers interviewed at baseline, 98 farmers dropped out, an attrition rate of ten percent. To test for the presence of attrition bias, we compare outcome variables and covariates at baseline across the returning and attriting farmers.

We also check for systematic differences between attritors and returners within each treatment arm.16 As in our balance check, we regress each variable on an indicator for if the farmer was an attritor. Columns 1 and 2 of Table 4 present means and standard deviations for attriting and returning farmers. The following six columns present co-efficients and standard errors, clustered at the farmergroup-level, from OLS regressions. For example, the third column displays co-efficients and standard errors on an indicator equal to one if the farmer attrited for the sub-population of farmers randomized into T1 or the pure control. We find that attriting farmers had significantly lower income per capita prior to the experiment than returning farmers. Attritors also tended to be older and less educated, suggesting that they may be less adept at farming than returning individuals. However, significantly more attritors reported that farming was their primary activity. To ensure our results are not due to attrition bias, we calculate bounds on our primary outcomes.We focus on estimating the direct impacts of randomly assigned farming contracts on four measures of rural transformation: rice area , yields , market participation ,vertical rack system and income per capita . To estimate these impacts, we compare outcomes for treated farmers with the outcomes in the absence of the treatment. We are not only interested in the effect of being offered a farming contract but the effects of each contract characteristic. As such, we present a large complement of results comparing treatment to control, comparing each contract to control, and comparing differences in outcomes between the various treatment groups. We expect any contract that reduces price risk, increases technical efficiency, or eases capital constraints to positively and significantly affect all four outcome variables. When it comes to expected differences between the impacts of each contract, the effect size will be heterogeneous, depending on where the largest gains are to be had for each individual farmer. That said, a priori we expect contracts that address more of the limitations facing farmers to have larger impacts. Because of this, our prior is that T3, which combines a fixed-price, production-management, and input-supply contract, will result in larger and more significant impacts compared to either of the other two treatments.

Similarly, T2, which includes the price guarantee and the extension training , should have larger impacts than T1, which only includes the price guarantee.Table 5 presents the treatment effects of a farmer being randomly assigned a production contract on four measures of rural transformation. We present results from OLS and ANCOVA regressions, without and with covariates. Panel A presents treatment effects on rice area, measured in hectares; Panel B presents treatment effects on yields, measured as kg of paddy rice harvest per hectare; Panel C presents treatment effects on market participation, measured as the percentage of harvested rice sold into the market; and Panel D presents treatment effects on income per capita, measured as the total value of farm and non-farm income in U.S. dollars divided by household size. Farmers randomly selected to receive a farm contract were provided with the written and signed contract prior to planting, which gave them time to reallocate their own land or bring in more land if they desired. Both the OLS and ANCOVA estimates reveal that farmers with a contract did plant a significant amount of additional land with rice compared to control farmers without a contract. Despite land being a lumpy input, farmers with contracts planted 23 percent more land with rice than control farmers, a half standard deviation shift above the control mean. Anecdotal evidence from the study reveals that farmers planted rice in lowland areas, so as to minimize the cost of irrigation. These lowland areas are typically used to cultivate home vegetable gardens or are left fallow. Farmers appear not to have substituted away from maize and cotton, the region’s primary crops, in order to plant rice but rather brought new marginal land into cultivation.