Examining results of farming contracts on the other three variables of interest, we also find consistently positive and significant effects. Focusing on the ANCOVA estimates with covariates, being offered a farming contract increases yields by 473 kg per hectare, a 29 percent increase in yields compared to the control. Given that we offered three types of contracts, this result does not immediately reveal what contract attributes most contributed to the yield gains. What is clear is that farmers did not simply fulfill their contracts by increasing the amount of land planted to rice. Rather, their productivity per unit of land increased in response to signing a farming contract. Not unexpectedly, farmers with production contracts increase their market participation by selling 35 percentage points more of their rice harvest, a 140 percent increase above farmers without contracts. This result may appear tautological, as farmers with contracts are expected to sell the contracted quantity to ESOP. However, even with contracts, farmers sell well less than 100 percent of their rice crop, implying that farmers produce enough rice to meet the terms of their contract and are able to decide how to dispose of the excess quantity, either by consuming the rice or saving it as seed for next year.One concern in the existing literature on contract farming is that by signing a contract, farmers reallocate land and labor to the contracted crop. Thus, while farmers may increase their production on one crop, the overall income effect may be zero or negative . We find that farmers in the treatment earned $140 more per person, an increase of 52 percent or about four tenths of a standard deviation above the mean for control households. This is a substantial income gain in a country where GDP per capita is around $800. In considering how farmers increased yields and income, Table B3 in the Appendix presents results from ANCOVA estimates of treatment on seed, fertilizer, pesticide, herbicide, and labor.

Treatment significantly increases the use of each input, indicating that the contracts resulted in an intensification of rice cultivation,mobile grow rack in addition to the extensification show in the regressions of rice area on treatment.19 Overall, our results, the first from an RCT, provide consistent evidence that contract farming has a positive and significant impact on several measures of farm productivity and household welfare. At least for rice growing households in Benin, contract farming appears to be a mechanism that encourages vertical coordination and can contribute to rural transformation.Given these positive results, it is particularly important to understand which contract attributes matter most in increasing yield and income. To do this, we randomly assigned treated households into one of three contract types. Fig. 3 summarizes the effect of each of the three types of contracts by drawing distributions of post-experiment values for each outcome. To the distributions we add vertical lines to mark the unconditional mean for each outcome variable by contract type. Visual inspection shows some heterogeneity in outcomes based on contract attributes. We also present regression analysis of these treatment effects, which not only allows us to test for differences between each treatment and the control but also test for differences between one treatment and another. Results from these regressions are presented in Table 6, with Bonferroni-adjusted Wald tests for differences between co-efficients on the treatment dummies in Table 7.However, testing for differences between the magnitudes of the co-efficients reveals that the effect of T1 is not significantly different from the effect of T2 or T3 . By comparison, the effect of T2 on area planted to rice is significantly lower than the effect of T3. While one could expect that the provision of input loans lowers the per unit cost of production, allowing farmers to expand area planted to rice without increasing their total farm production costs, it is less obvious why farmers with a contract that only guaranteed a price planted a similar sized area.

It may be that farmers who were to receive the extension services decided to focus effort on applying their training to a more circumscribed area. For those in T3, the addition of the input loan to the training may have reduced costs enough for farmers in this group to increase their area planted to an amount similar to those in T1. However, we lack the detailed farm production data needed to test this hypothesis. Turning to each contract’s effect on yield, we find that all three have a positive and significant impact. The magnitude of the impact varies slightly, from about 450 kg per hectare for farmers in T2 to about 500 kg per hectare for farmers in T1 and T3. A Wald test for differences between each of these co-efficients fails to reject the null of equality . One possible explanation for this results is that, given the variance in yields, we lack power to detect significant differences across treatment arms. A second possible explanation is that farmers gained little in terms of productivity by receiving extension training or input loans. Table B4 in the Appendix provides ANCOVA estimates of each contract on input use. Each contract significantly increases seed, fertilizer, and labor use, though contracts tend to have a null effect on pesticide and herbicide use. Bonferroni-adjusted Wald tests for differences between co-efficients on each contract indicator are never significant, indicating that across treatments farmers used about the same level of inputs. While far from conclusive, we take this as suggestive evidence that simply resolving price risk was sufficient to allow farmers to increase their use of inputs and thereby substantially increase yield. All three contracts have a positive and significant impact on market participation. However, unlike yield, in which each contract’s effect size was statistically similar, the impact of each contract on market participation significantly differs from each other. Conforming with our priors, effect sizes are greater for contracts that offer more services to the farmer. Those in the T1 treatment sell just under 50 percent of their rice harvest into the market , while those in T2 sell 57 percent and those in T3 sell 66 percent. The effects of using contracts to integrate farmers into the market are clear.

Without a contract to produce rice, households sell about a quarter of their rice production and keep the remaining three quarters. Under the most complex contract, farmers nearly reverse this ratio, selling almost 70 percent of their rice into the market and retaining only 30 percent. The evidence for each type of contract’s impact on income per capita is less obvious than when we simply compare all contracts to no contract, as in Table 5. While the effects of all three contracts are positive relative to the control, only the effect of the input-supply contract is consistently significant. Yet, when we conduct Wald tests for differences between co-efficients, we consistently fail to reject the null of equality. We speculate that this is due to a lack of power sufficient to detect differences in effect size in a notoriously noisy variable such as income and not evidence of a true null. We base this on the similarity in the size of co-efficients and standard errors across the income regressions in Tables 5 and 6. Overall, we find a curious degree of variation in impacts based on the terms of the contract. Contrary to our priors, it is not always the case that the effect size of T1 is smaller than T2, which is smaller than T3. Instead, we find that the fixed-price contract increases rice area to the same extent as the input-supply contract , while the production management contract has a smaller effect. All three contracts have similar effects on yields, meaning that the provisioning of extension training and/or input loans does not result in increased yield relative to the contract the only provides a price guarantee. For income per capita we again find that the added elements of T2 and T3 do not seem to provide much additional value over the simple fixed-price contract. Throughout the analysis, we frequently find that the magnitude of the co-efficient on the T2 treatment is the smallest of the three treatment arms, while the magnitude of the co-efficient on the T1 treatment is only slightly less than that on the T3 treatment. In fact, the only outcome variable that conforms to our prior is market participation, where farmers with the production-management contract sell significantly more rice than farmers with the fixed-price contract, and farmers with the input supply contract sell significantly more rice than the other two.Columns display ANCOVA results for the four outcome variables as the dependent variable. Each row designates which covariate is interacted with the treatment indicator. Cells report the co-efficient and standard error on the interaction term of household covariate and treatment indicator on the dependent variable .

We find almost no evidence of heterogeneous treatment effects by baseline characteristics. We fail to reject the null that any of the covariates mitigate or accentuate the effect of contract farming on the area of land put into rice production. For yields and market participation, a marginally significant degree of heterogeneity exists based on a farmer’s previous training in rice production. For income per capita, the only interactions that are significant are household size with the contract and experience producing rice with the contract. In both cases, larger households and more experienced rice producers had lower income with the contract than similar households without the contract. To provide a more detailed exploration regarding these three potential sources of heterogeneity, we graph the marginal effects of each interaction term on our outcome variables. Fig. 4 plots the marginal effects and 95 percent confidence intervals for the interactions between household size and indicators for each type of production contract. Panels document the effect on one of the four outcome variables. As was evident from Table 8, there is a lack of heterogeneity in household size on rice area,ebb and flow table yield, and market participation. For income per capita, we find that smaller households offered the input-supply contract have higher income than control households of similarly small size. As household size increases, income per capita for all groups decreases until there are no significant differences across treatments. It appears that when households are relatively small , they are better able to take advantage of being offered the input-supply contract and convert it into more income for each member. This difference diminishes for households with more than eight members and is not significant for the other treatment arms. Fig. 5 presents a similar set of margin plots for the effect of experience in rice production, measured in years. As with household size, there is little evidence of heterogeneous treatment effects on rice area, yield, or market participation. Where significant evidence does exist is for income per capita. Without a contract, more experienced farmers have higher income than less experienced farmers. This heterogeneity based on experience disappears for farmers randomly assigned to a production contract. Regardless of the type of contract, less experienced farmers have approximately the same amount of income as more experienced farmers. Contract farming helps inexperienced farmers earn incomes comparable to that earned by much more experienced farmers. It takes farmers without a contract a decade or more of experience to earn similar levels of income. Finally, Fig. 6 graphs the marginal effects of each treatment interacted with an indicator for whether or not the farmer had participated in training in the last 12 months. Because the household characteristic is now a binary variable, we graph each contract along the horizontal axis and the lines represent if the farmer participated in training. Here again we find little evidence of heterogeneity. As was evident in Table 8, households with training and a contract had higher yields and greater market participation than households with training in the control. But there are no significant differences across treatment arms and no significant differences within treatment arms across training/no-training. To some extent, our heterogeneity analysis appears fruitless. Across a number of different pre-specified covariates and a number of different outcome variables, we fail to find much evidence of heterogeneity in treatment effects.