Any treatment effects on farmer-level uptake might occur simultaneously with supply responses by dealers.To measure this, we surveyed seed dealers around the same time as the farmer survey. We timed the survey to be in September so that seed purchases would be recently completed and easier to recall for dealers. Dealers were asked which varieties they carried for the 2017 season, how much of each was sold, and whether they were selling seeds from private companies or from the state’s seed corporation. Our sample consists of 613 dealers from the list of dealers obtained prior to the experiment.A large fraction could not be located or were no longer selling rice seeds. Specifically, 22.8 percent of them could not be reached. Of the 473 dealers located, 274 were selling rice seeds in the 2017 season. In results that follow, we show effects both for all dealers that were reached and those that remained in the seed business. Table A2 shows that the likelihood of being located and the probably of selling rice seeds during the 2017 season are uncorrelated with treatment. Focusing on the treatment blocks, about 42 percent of the dealers surveyed received the intervention. In addition to these dealer sales, we obtained data on the physical location of seed production. Seeds are grown by registered farmers that contract with the state to produce seeds that meet minimum certification standards. OSSC then collects, processes,hydroponic equipment and bags these seeds before selling them to farmers during the next season. The average block in our study had 32 seed growers per season from 2014 to 2019.
We use records from a publicly available database that gives the location of each seed grower, the contracted area, the variety they produced, and the amount that was collected and processed. Seed growers tend to be large farmers. They have incentives to produce the most profitable varieties for their land — just like farmers.As such, their production of a new variety depends on them being convinced of its potential. We therefore aggregate seed production at the block-season level and estimate the effect of the dealer treatment on the amount of Swarna-Sub1 produced in the block. Returning to farmer-level information, we use remote sensing data to approximate flooding risk. These data help us predict which farmers are expected to benefit the most from Swarna-Sub1. Being able to observe a key determinant of returns makes it possible to test for heterogeneous treatment effects according to a proxy for predicted benefits. More simply, is there a trade off between intervening with private-sector agents and a technology reaching the right people? Or, does involving input suppliers in the diffusion of information cause technology to diffuse to high-return individuals? We have GPS coordinates of the houses for 83 percent of the farmers that we surveyed in 2017.These coordinates are matched to daily images of flooded areas from June to October for the period 2011 to 2017. We consider a household as exposed to flooding on a given day if their house is within one kilometer of any flooded area.We then aggregate the total number of days of flood exposure across the 7 years as a measure of flooding risk — and hence as a proxy for the return to Swarna-Sub1. The online appendix shows three characteristics of this variable. First, it varies substantially across the sample . About 30 percent of households were not exposed to flooding. In contrast, 10 percent of households had flooding for 40 days or more. Second, this variation is partly driven by geographic characteristics. Particularly, Figure A4 shows that flooding is more frequent in lower-elevation areas that are closer to rivers.
These correlations provide verification that our measure at least partly reflects underlying determinants of flooding risk — not just recent flood shocks. Third, farmers exposed to more flooding tend to be smaller, poorer, and belong to low-caste social groups . Informing private input dealers and providing them with seeds to test leads to greater adoption by farmers when compared to conventional extension approaches used by the public sector. Table 2 shows this result. Starting with Column 1, farmers in treatment blocks are 3.5 percentage points more likely to have adopted Swarna-Sub1 a year after the treatment, compared to farmers in control blocks. Given an adoption rate of 6.3% in the control group, this implies the treatment leads to a 56% increase in uptake. The treatment also caused acreage cultivated to increase: farmers in treatment blocks planted an average of 0.06 more acres with Swarna-Sub1 compared to farmers in control blocks, a 69% increase . This adoption effect operates on both the extensive and intensive margins: private agrodealers also caused cultivated area of adopters to increase. Focusing specifically on the 329 adopters in treatment blocks, they cultivated 10% more of their land with Swarna-Sub1 compared to the 210 adopters in control blocks . Table A4 shows that decomposing the intensive and extensive margins more formally with a tobit model leads to the same conclusions. Our specifications in Table 2 use only the random variation created in the experiment. Table A5 verifies that controlling for the large set of covariates included in the balance test does not change the result. The point estimates stay similar when including these additional explanatory variables. Table A6 shows that the level of contact with extension agents or with cluster demonstrations is very low, even with our reinforced extension service in control blocks and that farmers in treatment blocks were no less likely to be in contact with extensions workers,or to have observed a demonstration of Swarna-Sub1, compared to control farmers.In other words, we do not find evidence of displacement at the expense of other traditional channels. Following up on the idea of displacement, we look at whether the treatment displaced other new varieties, potentially lowering welfare if it caused a shift away from high-quality seeds. We find no such evidence. Table A7 shows that the treatment had a negative effect on adoption of only two seed varieties — both of which were released over three decades ago. It does not appear that the increase in adoption caused by agrodealers corresponds to a shift away from newly released technologies. Finally, we find no evidence that the SMS messages increased adoption . They also did not change the effectiveness of the dealer treatment. The adoption gains from the dealer treatment cannot be obtained with a “lighter touch” SMS messaging intervention, at least in our context. The evidence on average adoption rates shows that helping private agrodealers learn is more effective than conventional approaches used in the public sector.
A concern may be that, as private agents, dealers optimize behavior based on their own expected sales and profits; in contrast with government extension agents who can factor in equity and may be better at targeting farmers who have high expected returns to adoption. It is however not obvious whether profit maximizing dealers will deliver inferior targeting. profit maximization strategies and farmers benefiting from adoption could coincide and may lead to similar outcomes, especially if we consider the repeated interactions between dealers and farmers over time. In our context, being exposed to frequent flooding gives an easy-to-observe measure of potential returns — given the flood tolerance property of the variety.We show that treatment dealers were successful at targeting Swarna-Sub1 to farmers who could benefit the most from the new technology, i.e. farmers who live in flood prone areas. Figure 2 separates the sample by the satellite-based measure of past flooding and shows that treatment effects only exist in approximately half the sample where there were at least 3 flood days from 2011 to 2017. Conversely, the dealer treatment had little or no effect on adoption in the bottom half of the sample. In Table 3, we show how the treatment effect is heterogeneous based on the number of flood days. Two results stand out in the table. First, control farmers who live in flood prone areas are less likely to adopt Swarna-Sub1. This negative relationship is true whether flood risk is measured in days of flooding or as a binary variable separating the sample into high- and low-risk farmers based on the median number of flood days . Indeed,vertical grow table being a high-risk control farmer is associated with a 6% lower likelihood of adoption compared to low-risk control farmers. But it is important to emphasize that this estimate is merely a correlation.Second, and more importantly, the dealer treatment was only effective in flood-prone areas, i.e. the interaction between treatment and flooding exposure is positive. The interaction term in Column 1 is less precise, likely because the heterogeneity in Figure 2 did not appear to be linear. But Column 2, which corresponds most closely with the figure, shows that the dealer treatment targets high-risk farmers increasing their adoption by 6.4%, while the effect of the treatment is only 0.8% for low-risk farmers . The difference between the two treatment effects is statistically significant at the 10% level. As another piece of evidence, Table A9 shows that the average adopter in treatment blocks is more exposed to flooding. Specifically, they are more than twice as likely to be above the median in terms of flood exposure.
There is no evidence that informing dealers prioritizes adoption by the wealthiest farmers, which might have been expected if agrodealers cater more to larger and wealthier farmers. In particular, Table A10 shows that there is no treatment-effect heterogeneity according to farm size. Adoption is more likely by larger farmers, but this is equally true in treatment and control blocks. We also find no heterogeneity according to being below the poverty line or in a marginalized caste group. Recall that we only treated a fraction of the dealers in each block. More precisely, 42% of sample dealers in treatment blocks received seeds and information . These dealers were not randomized. Hence, our dealer-level analysis compares all private dealers in treatment blocks to those in control blocks. We therefore capture any direct effect of receiving the seeds and information and any spillovers — which of course could be either negative or positive. There is some evidence that the treatment caused dealers to increase the availability of Swarna-Sub1. Columns 1-4 in Table 4 show results from one year after the treatment . Focusing on all dealers — including those that were no longer operating — the treatment has a small positive effect on the likelihood of carrying Swarna-Sub1 at any time during the season and the total amount the dealer reported selling throughout the year . But both of these estimates are very imprecise, partly due to some dealers no longer being in business. Amongst the subset of active dealers, those in treatment blocks were 6.2 percentage points more likely to carry Swarna-Sub1, a 17 % increase . Column 4 shows that dealers in treatment blocks sold 3.7 additional quintals, which represents a 59% increase in volume sold. But again, while larger, neither of these results are close to statistically significant. Anticipating on an intervention done in year 3 , we find large and precise effects on stocking behavior . 19.3% of dealers in control blocks had Swarna-Sub1 in stock when visited by the secret shopper.This increases by 11.4 percentage points in treatment blocks. This large effect is being observed two years after the treatment. It also comes from a direct observation of what the dealer had available on a certain day, rather than a noisy estimate from what they recalled after the season. This result could be driven by a number of things. First, it could come directly from the dealers that were treated and had their information sets updated. Second, dealers talk to farmers. Any increase in knowledge of farmers could spread to other dealers, not only those that were treated. Third, dealers were provided with several minikits for testing. They could have shared those in a way that increased local knowledge. We cannot distinguish between these effects in the analysis. We next test whether the treatment changed the extent of local seed production. Our data here amount to six observations per block: three from the period before our treatment could have triggered a production response and three from the post treatment period .