Qualitative interview suggest that sellers often posted prices that reflected strategic price offers, fishing for higher prices, much in the way that one would typically make offers in more traditional, in-person negotiations. In fact, sellers often posted prices that were not only higher than their local market price, but even in excess of what was being paid in hubs or super hubs. As a result, ask prices were on average substantially above bid prices.Buyers’ average bid prices, on the other hand, track hub market prices very well. Figure B.8 plots these values over time for maize, and Figure B.9 provides a box-and-whisker plot of bid and ask prices within each season, in which we can see that the median bid price is typically at or below the 25th percentile of ask prices Nonetheless, about 7,300 tons of grain were successfully transacted, worth about ✩2.3 million USD. 22% of treated traders and 2% of treated households successfully traded on the platform. Figure B.10 shows the cumulative sales over the platform during the duration of the study. We now turn the impacts of the platform. We first explore the effects on market integration and trade flows. Figure 2 presents impacts of the platform on several outcomes: whether any trade is occurring between sub-counties, the number of traders engaged in trade between sub-counties, the volume of trade flowing between sub-counties, and price dispersion between markets. The first three of these outcomes is drawn from our panel survey of traders, in which we asked detailed questions about their trading behavior at the sub-county level finding. The last is drawn from our market-level price surveys and is therefore at the market dyad level. Within these dyads, we analyze the experiment using indicator variables for dyads in which both markets are treated and dyads in which one market is treated,dutch buckets using no-treatment dyads as the control.

Figure 2 presents Fan regressions of each outcome on the distance between the pair, estimated separately for our three treatment groups. Distance is measured as the road distance of the shortest route connecting the two. Before examining treatment effects, we first note some important patterns observed among our control-only pairs. In the upper left panel, we see that while the probability of any trade is high for nearby sub-counties, this diminishes rapidly with distance. The probability of any trade occurring between the sub-counties is close to zero beyond 200km distance. Consistent with this, the number of traders finding and total trade volumes finding also falls quickly with distance.These increasing trade costs with distance lead to notably higher price dispersion between markets located at further distances, as shown in the bottom right panel. What is the effect of introducing a mobile clearinghouse? In Figure 2, we see increases in the probability of any trade occurring, the number of traders engaged in trade between sub-counties, and the volume of trade flowing between sub-counties. We also see a drop in price dispersion.Again, we see increases in the probability of trade occurring finding, increases in the number of traders operating between sub-counties finding; increases in trade volumes finding, and reductions in price dispersion finding.Returning to Figure 2, we also note a striking pattern by distance. Treatment effects are strongly concentrated among nearby markets. In fact, we see almost no treatment effect beyond 200km, the point at which the probability of trade drops close to zero. The one exception to this is price dispersion, which continues to drop in our treated market pairs beyond 200km, albeit at a slower rate. This may be due either to treatment effects on large, long-distance traders not included in our sample as resident in study markets. Alternatively, we may observe transitive convergence finding or to transshipment finding. Table 2 explores this further, estimating Equation 1 separately for market pairs above and below the median distance observed in our sample finding.

We again see that treatment effects for all outcomes are concentrated in nearby sub-counties and markets, with significant increases in the probability of trade occurring finding, significant increases in the number of traders operating between sub-counties, significant increases in trade volumes finding, and significant reductions in price dispersion. In contrast, we see almost no effect of the platform on direct trading outcomes finding for markets that are at above median distance, though the point estimates suggest that price convergence results may persist along a farther distance, albeit at a smaller rate. The platform is therefore quite successful in generating additional short-distance trade. However, it falls to live up to the often-touted promise of such online marketplaces to directly connect remotely-located farmers and markets with urban consumers finding. While perhaps initially surprising, this pattern of large effects over the shortest distances is consistent with Figure 1, which shows that very little of the existing price dispersion at short distances is explained by transport costs. Given the ubiquity of mobile phones even in the control group, it is likely that the very large price gaps necessary to motivate long-distance trade are already arbitraged away, meaning that the more marginal improvements in information revealed by our system typically only exceed the pecuniary of trade over shorter distances.We have already seen in the previous section that the platform encouraged greater intermediary activity in treated sub-counties and markets. Here, we explore in greater detail trader take-up of the platform and effects on their businesses. Table 4 presents trader take-up results. We see that, by the endline survey, 91% of treated traders report having heard of Kudu, while only 32% of control traders have heard of the platform. Therefore, while Kudu was not restricted to be operational only in treated areas, we do see a significant and large difference in awareness of the platform generated by our encouragement design. We also observe a 42 percentage point increase in the likelihood of receiving any price information via SMS finding. However, in terms of knowledge of prices, we do not see a substantial treatment effect. We ask traders to report their best guess of the current market price in their local market, their hub market, and their superhub market, which we then compare the the actual price as measured by our market surveys. We call the absolute value of the gap the “error” in price knowledge. Although traders’ knowledge of nearby local and hub markets is slightly better than their knowledge of superhub prices, we see no differences between treatment and control traders in knowledge for any market type.

This may be because knowledge in our control is already quite high, as demonstrated by the relatively small error size. We do, however, see strong treatment effects in terms of self-reported impacts on negotiations, both with farmers from whom traders buy and with buyers to whom they sell. Treated traders are more likely to report that they are aware of farmers and buyers receiving price information via SMS. They are also more likely to state that this information changed how they negotiated with their trading partner. Finally, in terms of Kudu take-up, 80% of treated traders used Kudu finding, while 22% successfully completed a deal on the platform. In comparison, only 12% of control traders tried Kudu, and only 3% successfully completed a transaction. This appears to come mainly from a reduction in trader markups finding; point estimates suggest a reduction of about 8%, though this effect is measured with imprecision and is not significant finding. Volumes traded appear to increase, perhaps sizably, though again, this point estimate is not significant finding. Columns 4-5 present effects on the price at which traders sell maize, while Columns 6-7 present treatment effects on the price at which traders purchase maize. Similar to results presented in Table 3, we see no significant effects on the level of prices finding. However, looking at heterogeneity by relative deficit and surplus areas finding, we see that in relative deficit areas finding, treatment results in trader sale prices that are significantly lower finding. Conversely, in areas that are relative deficit finding, we observe that treated traders sell at higher prices. We see similar, albeit slightly muted, effects for the trader purchase price in Column 7. We will return later to discuss the relative magnitudes of the sale vs. purchase price treatment effects. Finally, we explore treatment effects based on baseline heterogeneity. Figure 4 presents treatment effects on profits, markups, and trade volumes based on their baseline levels finding. The top panels plot Fan regression estimates of the outcome at endline on baseline levels separately by treatment and control, while the bottom panel presents the difference finding, along with the 90% and 95% confidence intervals. Density in the baseline measure is presented in red. While the negative effects on profits and positive effects on volumes traded appear fairly consistent across their baseline distribution, we do interestingly see that markups are higher among treated traders at the low end of the baseline markup distribution and lower among treated traders at the high end of the baseline markup distribution. These estimates therefore suggest that the introduction of the mobile marketplaces appears to lead to convergence in markups, helping low markup traders and harming high markup traders. This is consistent with the idea that trading in the absence of our intervention is skill- and human capital-intensive,grow bucket in which case those with stronger trading networks and superior information can reliably reap greater profits.

Our intervention, by reducing the cost of information to a symmetric low level appears to have removed a substantial portion of the heterogeneity in trader markups. We have seen thus far that the introduction of a mobile clearinghouse platform induces greater market integration and lowers intermediaries’ profits. These results are often seen as stepping stones along a causal chain ending in the ultimate goal of improving the welfare of smallholder farmers. We turn now to effects of the platform on farmers. First, we explore measures of awareness and take-up of the platform among farmers. Table 6 presents these results. We see that 52% percent of farmers in treated sub-counties have heard of Kudu, compared to only 12% in control sub-counties. Similarly, 55% of treated farmers have received price information via some form of an SMS-based platform, compared to only 16% of control farmers. Next, we explore impacts on price knowledge. We first note that overall price knowledge is lower among farmers than among traders, with error rates that are 42-84% higher than observed among traders. This is consistent with the presence of information asymmetries between farmers and traders. Similar to those of traders, farmers’ error rates grow with distance. In contrast to traders, however, we do find some suggestive evidence that information to farmers improves their knowledge of prices; errors rates are smaller among treated farmers than among control farmers, albeit not quite significantly so finding. 23% of farmers in treated sub-counties report using price information received via SMS when negotiating prices in the past year, compared to only 1% in control sub-counties. Turning to take-up of Kudu itself, we see that 26% of treated farmers have ever used Kudu finding, compared to 2% of control farmers. However, the success rate for study farmers managing to transact on Kudu is relatively low; only 2% of treated farmers completed a transaction on Kudu. Though this rate is significantly higher than the 0% observed in the control group, this low adoption of Kudu is notable, and suggests that farmers either see low benefits or high barriers to adoption of the platform. We will explore determinants of adoption below. Columns 2, 4, 6, and 8 of Table 7 present results. The coefficient on β2 helps to characterize likely adopters. Who is likely to adopt Kudu depends on who faces the largest benefits to adoption, relative to the cost. One might predict that smaller, poorer farmers who currently receive low prices would have the most to gain from access to a new platform on which to sell their crops. However, one could also imagine that the platform would have a hard time finding a buyer for farmers who sell relatively small surpluses, and therefore that it may be the larger, wealthier farmers who are best positioned to use Kudu. Consistent with Allen finding, our evidence suggests the latter interpretation. We see that take up is significantly higher among farmers with higher total revenues and higher prices received, as evidenced by the statistically significant correlation between the propensity score and these outcomes.