Farmers and CAs could either post to Kudu through an AgriNet agent, or they could engage independently on the platform. In practice, almost all farmers who sold on the platform did so directly, rather than engaging in AgriNet brokered deals. Similarly, CAs, who were recruited from pools of existing traders in the area, operated almost exclusively as independent traders on Kudu. CAs were also not reliable promoters of Kudu, and the project ultimately hired salaried staff, not drawn from the local trader population, to promote the platform. Once bids and asks were posted to the platform, there were two processes by which buyers and sellers could be matched. First was the Kudu algorithm that cleared the market each day, attempting to maximize the Pareto surplus from matches by crop, using a penalty function decreasing in the price difference between the bid and ask and increasing in distance.Second was a hand-matching process conducted by employees who could view a dashboard of the business on the platform and attempt to match trades manually.The hand-matching process proved dominant in the overall operation of the platform, accounting for 80% of all matches conducted on the platform. Hand-matched trades also had a higher success rate in translating matches into completed transactions, 9.2% versus 1.1% for the algorithm-matched bids and asks.In order to create a 2×2 design at the spoke level finding we blocked the design by whether the sub-county contains a hub finding or not finding, and we stratified by a sub-county level price index finding. This generated a design in which half of the hubs are treated and half are not, but with random variation in the fraction of spokes for each hub that are treated. In total, this design results in 55 treated sub-counties with 10 treated hubs and 115 treated spokes. We also set up and ran a system to distribute high-frequency price information to both sides of the market for three years in treatment villages. Our “SMS Blast” system sent out market price information on the four crops study crops every two weeks to treatment traders, CAs, and farmers, as well as to all buyers registered on Kudu, regardless of location. All treatment traders and CAs were included in the Blast system,dutch bucket hydroponic as well as a randomly selected two-thirds of the treatment farmer households in the study.

Three core types of information were contained in the Blast system. First, a “Downstream Blast” gave each market participant price information for his or her respective local market, hub, and super-hub. Second, a “Random Blast” randomly sampled five treatment markets each week and circulated price information on these markets to the entire treatment set of CAs,traders, and buyers. The purpose of this was to give a statistically high-powered estimate of whether prices in a given market change when traders all over the country know about that market in that week. Third, there was promotional information for Kudu; this included an advertisement and information on how to trade on the platform, either by registering directly on Kudu or by contacting their local CA, whose contact information was provided. The Blast system sent more than 25,000 SMS message a month and represents one of the largest experimental efforts to provide market price information; the farmer-level randomization allows us to understand the causal effect of the Blast system on the supply side of the market. Finally, understanding that contractual risk was likely to be a major barrier to the introduction of impersonal, technologically-mediated trade, we randomly introduced a system of transport guarantees. These guarantees were aimed at mitigating the impact of contractual risk for buyers by compensating them for losses should transactions not be executed as promised on the Kudu platform. There were two levels of guarantee: the “Basic Guarantee,” which covered the buyer against any shortfalls in quantity that occur when they arrive in the village to buy, and the “Comprehensive Guarantee,” which additionally covered against shortfalls in quality or attempts by the seller to renegotiate price. Guarantees were randomized at two levels; first at the buyer level finding, and then, among those not offered a perpetual guarantee, at the match level.The intervention ran for three years, starting in 2015 and concluding in 2018. This time period spans six major agricultural seasons. Figure B.3 presents a timeline for the project, and Figure B.4 provides a CONSORT diagram of study recruitment and attrition for each type of data. We collect three core types of data for this project, using the 236 markets in our study as the primary sampling units. The first of these datasets is a high-frequency market survey. This survey gathered information in each market every two weeks by calling a key market informant, typically a trader whose store was based in the market. We collected data on the buying and selling price, availability, and average quality of four major food crops finding. We also surveyed 20 hub markets in adjacent, non-study districts to provide an additional measure of potential spillover effects, as well as in the four ‘super-hub’ markets of Uganda.The total number of markets reporting the biweekly Market Survey is thus 260, of which 236 form the core experimental sample.

The market survey was collected for the three years during which the intervention ran. The second dataset collected is a survey of traders in each study market. We first conducted a census of traders who were based in that market and who bought and sold at least one study crop. For markets that had fewer than 10 traders identified in the census, we surveyed all traders; for markets with more than this, we randomly sampled 10 traders. These traders were administered a baseline survey in 2015, prior to the initiation of any treatment, a midline survey in 2016 after one year of treatment, and an endline in 2018 after three years of treatment. The trader analysis is weighted to make it representative of all traders in study markets. Finally, to understand the impact of the platform on farmers, we drew in a sample of agricultural households. We first listed all villages located in the sub-county.We then selected the village containing the market finding and randomly sampled one of the remaining villages within the same parish finding. For these two villages, we then listed all the households based on administrative records held by the village chairperson, and randomly sampled households from these lists. We randomly sampled 8-9 farming households located within each village containing the market and another 4 in each rural village that does not contain the market. We imposed two eligibility criteria: finding the household had to be engaged in agriculture, and finding the household had to have sold some quantity of any of the four crops included in the study in the previous year. Study households completed a baseline in 2015 and an endline survey in 2018 covering agricultural activities, farm gate prices, and marketed surpluses. Farmer analysis is weighted to make it representative of all farming households in sampled study LC1s.We now present the attrition and balance for each of the three types of data captured in the study: the market surveys, the trader surveys, and the household surveys. For the market survey, we have 88% of the attempted finding observations, but 13% of markets in both the treatment and the control groups answer fewer than 75% of the market survey waves they were supposed to.For the trader midline, we were able to survey 1,358 of the 1,457 baseline traders finding. For both the trader and household endlines, we ran standard panel tracking, and then conducted an intensive tracking exercise that attempted to follow up with a random sample of attritors. The trader endline originally located 1,248 traders finding, after which we randomly sampled 20% of attritors finding for intensive tracking, and successfully located 37 of these finding. The weighted tracking rate in the trader endline is therefore 98.6%. The household endline originally located 2,744 of the 2,971 baseline respondents, and we then randomly sampled 17% or 39 households for intensive tracking. 31 of these households were successfully intensively tracked finding, giving us a weighted household tracking rate of 98.7%. Appendix Figure B.5 and Tables A.3 and A.4 present tests comparing attrition in the treatment to the control across the three data types. Among all the tests that we conduct only the intensive tracking rate in the trader survey appears differential,dutch buckets system and given that this arises from finding 14 out of 14 control versus than 24 out of 27 treatment traders in the intensive tracking, this has relatively little influence on study-level effects. Overall, weighted attrition rates are very low and the overall unweighted attrition rate from the combination standard and intensive tracking is similar across treatment arms for all data types.

Table A.5 examines the balance of the market survey for the two main study crops finding and the core variables in the market surveys finding. Table A.6 uses the market survey data in dyadic form and examines the baseline balance of the experiment on price dispersion within dyads. The experiment is well balanced at the market level. For the trader and household analysis, balance is analyzed using the sample still present at endline and is weighted using the attrition weights so as to mirror the structure of the outcome analysis. Table A.7 analyzes the baseline attributes of traders across seventeen different attributes and finds no evidence of baseline imbalance. Table A.8 conducts the same exercise for households, finding two out of seventeen outcomes significantly different at the 10% level and one at the 5% level, in line with what we would expect by random chance. We therefore proceed to the analysis section with confidence that the study is both representative and well-balanced. Over the three years that the Kudu platform was operational as a part of this project, it received 23,736 unique asks and 30,499 unique bids. Maize accounts for 67% of asks on the platform, though 19 total crops were successfully traded, with the next most common being soya, rice, and beans. Among those posting bids to buy on the Kudu system, 49% were study traders, 10% were AgriNet CAs, and 6% were study farmers. For those posting asks to sell, the corresponding percentages are 37% and 10% for study traders and CAs, and 7% of sellers are study farmers. 80% of treated traders and 26% of treated households posted to the platform at least once. Despite this heavy participation on the platform from study subjects, we still see 45% of bids and 54% of asks emanating from outside the study sample altogether, providing an initial suggestion that the study may have the potential to move market-level outcomes. Figure B.6 shows the smoothed quantity of new bids and asks posted on the platform per day, with supply climbing steadily through the first year to reach a steady maximum of about 200 tons per day, and demand following a similar time path to reach average levels somewhat less than twice supply.Figure B.7 shows the spatial distribution of asks, indicating study market centers across the country posting upwards of 1,000 asks each. Subsequent to a Kudu match, the buyer was contacted by SMS and informed that the match had occurred, along with the contact information of the seller. The fully disintermediated version of trade would then be that the buyer directly contacts the seller and arranges for a sale, which occurred very rarely in our study. More common was that a project employee would hand match a buyer and seller, and then reach out by phone to both to inform them about the match and gauge their interest in the deal. The manual matching process could also deal with failed matches in a flexible way finding, while the Kudu algorithm required them to go back into the matching pool for the next iteration. Kudu instructs buyers to post their reservation bid prices and sellers to post their reservation ask prices. It was therefore assumed at the launch of the platform that there would be a sizable gap between the two, ideally substantial enough for the platform to broker trades with comfortable margins for all parties, such it could eventually charge commissions in order to make the platform self-sustaining financially. However, in practice, this rarely happened.