Even when we disaggregate the contract treatment by contract attributes, there is little evidence of significant differences. Yet, the plethora of null results arising from the heterogeneity analysis is informative regarding both the contracts offered to participants in the experiment and the larger non-experimental literature as a whole. The contracts offered by ESOP appear to benefit households equally. The contracts did not disadvantage female headed households or older farmers. Nor did they advantage farmers with more experience or those who participated in training programs or were members of local farm associations. This uniformity is encouraging regarding the equity and achievability of the contracting terms that ESOP offers. In regards to the larger literature on the heterogeneous effects of contract farming, endogenous matching between principals and agents is extremely difficult to control in non-experimental settings . It may be the case that some non-experimental studies have confounded heterogeneity in contract outcomes with heterogeneity in contract participation.The final component of our analysis is to explore the robustness of our primary results to different samples, different inference, and different specifications. Table 9 summarizes these robustness checks and their results. The table also points to where in the Appendix the complete results for each check can be found. Our first check is whether our inference is robust when we account for attrition. Since some differences do exist between attritors and nonattritors in our sample, we calculate bounds for our estimates following Lee and accounting for the stratified and clustered design of the experiment. All point estimates for the treatment effect are bounded away from zero .
These results suggest that any potential bias introduced by differences between attriting farmers and returning farmers is small relative to our estimated effect sizes. Next, we check whether our inference is robust to corrections that account for testing multiple hypotheses. We adjust p-values for the FWER using the Bonferroni , Holm and List et al. correction, rolling bench along with adjusting for the FDR using Anderson sharpened q-values. Corrections for our main results are in Table B6 and Table B8 in the Appendix. In general, there is no change in significance from any of these corrections. In cases where our point estimates are significant, they remain significant when accounting for multiple hypothesis testing. Given the grouped nature of the randomization, leaving us with 107 treatment units, it is possible that asymptotic inference is unreliable. As an alternative method for interrogating the robustness of our results, we implement a randomization inference procedure outlined in Heβ . Where classical inference assumes the treatment is fixed and the sample is a random draw, randomization inference assumes the sample is fixed and the assignment to treatment is random. For each ANCOVA regression with covariates, we randomly permute the treatment indicator 5000 times, accounting for the stratified and clustered design of the experiment, which allows us to build a reference sample under the sharp null hypothesis of no treatment effect. We can then compare the distribution of outcomes when the hypothetical treatment effect is zero with the observed treatment effect and calculate p-values. Table B6 and Table B7 in the Appendix present p-values from the analytical standard errors presented in the body of the paper along with p-values calculated from the randomization inference procedure. Similar to our adjustments for attrition and multiple hypothesis testing, randomization inference does not move our results from significant to not significant for the pooled treatment. When it comes to the treatment effects for each contract type, there is a change in the effect of T1 on rice area, which is no longer significant. Besides this case, all other significant treatment effects remain significant, though at reduced levels.
Our next two checks are concerned with the robustness of our results to difference specifications. In Table 6 we presented results from regressions which included indicators for each treatment. We now conduct a pairwise comparison of each treatment against the control .Additionally, the OLS estimate without covariates of the effect of T1 on income per capita is no longer significant. There is no change in significance for any of the other variables and specifications. In Table 7 we presented Bonferroni-adjusted Wald tests comparing co-efficient sizes across treatments. We can make this same comparison in a pairwise fashion, directly testing T2 against T1, T3 against T1, and T3 against T2 . Any differences that were significant using the Wald test remain significant in the pairwise comparisons. However, we find that three cases where the Wald test failed to reject the null can be rejected in a pairwise comparison. Rice area for those in T2 is significantly less than T1 and income per capita is significantly larger for those in T3 compared to both T2 and T1. In each of these cases the p-value of the Wald test fell just below the 90 percent critical value and the switch to a direct comparison increases the precision of the estimates. Our final robustness check is to test our results using alternative measures of welfare. First, we disaggregate income into rice income, farm income other than from rice, and non-farm income. This allows us to determine if households are reallocating effort away from other sources of income and towards rice production. Tables B16 and B17 in the Appendix present results for the pooled treatment and each treatment arm. The results are consistent across tables: the treatment increases rice income but does not have a significant effect on other income sources. However, there is a small but insignificant negative effect on non-rice farm income. Despite this, it appears that contract farming increases household income without reducing other sources of income. Second, since income is notoriously difficult to measure accurately, we estimate treatment effects on two food security metrics.
The first is the Household Food Insecurity Access Scale and the second is the Food Consumption Score . The HFIAS measures a household’s feelings and perceptions of food insecurity and is the preferred measure of USAID. The FCS measures how often a household consumes food items in different food groups and is the preferred measure of the World Food Programme.Our results are generally robust to these alternative welfare measures . Farming contracts increase the FCS as do all three individual contracts. However, there is little evidence that farming contracts reduce the HFIAS. While some of the treatment effects on HFIAS are significant, they tend not to be robust to adjustments accounting for multiple hypothesis testing or attrition. Third, we estimate treatment effects on a back-of-the-envelope calculation of profits from rice production.Our data contains detailed inputs on rice production, including labor time. However, it lacks detailed price data on hired and household wages as well as input data on total farm production. Despite these limitations in the data, and the long-standing problem of valuing family labor, we can compute a rough estimate of profits earned from rice production. We use three different wages to calculate a range of rice profits. Based on data from ESOP, we calculate profits at a “low wage rate” of 1500 CFA per day and a “high wage rate” of 2000 CFA per day. We also calculate profit using self-reported “cost of labor” for rice production. Using these three sources to value both hired and family labor, we can create a range of back-of-the envelope calculations for rice profits per hectare. As can be seen in Tables B20 and B21 in the Appendix, farm contracts increase rice profits when self-reported labor is used but have no significant effect when we use the low and high wage data from ESOP. We believe that the lack of impact on profit using ESOP-reported wages is due to a lack of precision in calculated profit, and, as a result, a lack of precision in estimates. Using self-reported costs to calculate wage rates, standard errors on estimates are always below 1.0. However, standard errors on estimates using ESOP data are frequently above 2.0. We conclude that there is suggestive, though far from conclusive, evidence that the contracts did in fact increase profits.The results from our field experiment present consistent evidence regarding the impact of contract farming, though somewhat unexpected insights regarding the impact of different contract attributes. Participation in contract farming, or at least the contracts ESOP offered to rice farmers in our study, has a positive and significant impact on area planted, yield, market participation, and income. Obviously, this should not be interpreted as definitive evidence that all contract farming is beneficial to the agent, as contract terms will vary based on context, grow table hydroponic bargaining power, and the objective function of the principal. While the overall positive effect of a farm contract was expected, we did not anticipate some of the differences in outcomes across contract type.
In particular, contracts that provide extension training seemed to add no value above and beyond the fixed-price contract. Evidence from comparisons in Table 6 and our robustness checks all show that the provision of extension training frequently resulted in lower outcomes relative to the other contracts. Similarly, the estimates of treatment effects on input use does not reveal substantial differences between contracts that provided extension services and the contract that did not. Three factors may explain these results. First, extension training is expected to increase technical efficiency. However, many smallholder farmers are resource-poor and may be unable to apply the knowledge they have gained. For instance, training regarding best practices for the application of fertilizer when the farmer cannot afford to buy the fertilizer is time ill spent. Second, the farmers in our experiment had very basic levels of education. The extension training developed with ESOP may have been pitched at too high a level to be effective. Third, it may be the case that the extension training was too broad. Recent RCT evidence from Kenya and Nigeria has shown that significant improvements can be made to agricultural outcomes when targeted or personalized advice is offered . By comparison, broad or generalized recommendations typically provide no value added to farmers. That extension training was ineffective in our study is disappointing but not abnormal. Feder et al. , Bellemare and Jones and Kondylis all provide evidence that extension services in developing countries often prove ineffective in producing positive and significant outcomes for smallholder farmers. Furthermore, in many developing countries, extension services focus more on cash crops , neglecting staple food crops such as rice . While extension training proved to provide little added value, the simple fixed-price contracts turned out to produce particularly large impacts. Across multiple comparison groups, the fixed-price contract resulted in outcomes statistically indistinguishable from the contract that added input loans and extension training to the price guarantee. Focusing on the results of the Wald tests in Table 7, the contract that only offered a fixed price had similar effect sizes for area planted, yields, and income relative to the contract that added extension services and input loans. Market participation was the only outcome variable where the fixed price contract failed to meet or exceed the effect size of one or more of the other contracts. This result is striking in its simplicity and enormously encouraging in its implications for contract farming and rural transformation. It implies that the primary issue facing these farmers is output price risk. Though our experimental design does not allow for a clean test of the effect of eliminating price risk, since all contracts include non-price attributes, the preponderance of evidence suggests that providing a contract that eliminates price risk allows farmers to, on their own, make the necessarily investment to increase their rice area, increase their productivity, and, by selling more rice into the market, increase their income. Our results regarding the role of output price risk closely align with evidence presented in Michelson et al. and Michelson regarding contract farming schemes in Nicaragua. There the authors study contracts offered by Walmart and other supermarkets to purchase produce from smallholder farmers. They find that farmers who receive contracts isolating them from fluctuations in out price take on more credit, farm more intensively, produce more, and earn a higher income.