Using Randomized Experiments to Learn About Market Competition


This page in:

A fundamental question in development economics is whether markets are competitive. That is, do firms set prices equal to marginal cost? The extent of competition determines the impact of public policy (e.g., how much does a tax or subsidy pass-through to consumers) and so understanding it is crucial for policy design. More generally, the extent of competition affects welfare. Markups over marginal cost can have benefits, allowing firms to recoup the cost of building a new plant, or of research and development, but can also be nefarious, arising when firms collude to fix prices. I discuss here several recent examples that illustrate different approaches for measuring the extent of market competition using experiments.

The Many-Markets Reduced Form Approach

One approach is to start with a set of separate markets, and then randomly vary the amount of competition in some of those markets. The key challenge in this approach is identifying what constitutes a market and having them spatially distinct enough from one another that competition is limited across markets. This has been the approach of several firm experiments that aim to see whether firms receiving training or loans benefit at the expense of their untreated competitors in the same markets. In these examples, markets are different rural marketplaces in Kenya, and different retail shopping clusters in China respectively. Since retail services are non-tradable, the retail industry is well suited for the many-markets approach.

If competition itself can be randomly varied at the market-level, then the impacts of competition can be measured in a reduced form equation by comparing outcomes in markets with more or less competition. This approach has been used to measure the effects of competition on consumer welfare. In the Dominican Republic, an experiment by Mattis Busso and Sebastian Galiani randomized the entry of grocery stores into different districts. The authors don’t convince firms to randomize their market entry strategy---which to my knowledge has been done yet---but rather partner with the government to change the number of firms in a district that will accept a special debit card distributed to households eligible to participate in an anti-poverty program. In partnership with the government, they select a random number of firms (i.e., 0,1,2,3) to recruit to join the program in each district, experimentally varying competition among grocery stores serving beneficiaries. They find that entry leads to reductions in price ranging from 2 to 6 percent and an improvement in service quality reported by customers. Prices dropped more in areas where the number of entrants was larger.

These reduced from results confirm the insight from industrial organization theory that more competitors can lower prices, but leave other questions unanswered. Why were prices higher in the first place? Were grocery stores conspiring to raise prices in an illegal cartel that was broken by the new entrants, or did the initially high prices emerge because the markets could sustain only a few entrants, between whom competition was imperfect but non-cooperative?

Combining Industrial Organization Theory with Market-Level and Individual-Level Randomized Variation

To identify collusion, an economic model is required to interpret the experimental results. Berry and Haile (2014), building on Bresnahan (1982), describe an approach, the basic idea of which is to posit different models of how firms set prices and then test between these models. Given demand and supply curves, do prices look as if they are being generated by many firms, or a small number of firms? If the prices appear to be generated by many firms, it is unlikely that firms are colluding to jointly maximize profits.

Another study in rural Kenyan markets by Lauren Bergquist and Michael Dinnerstein implements this approach using market-level randomization. These authors assume markets are independent across space and four-week periods and then randomize cost and demand subsidies to maize traders across markets to identify demand and supply curves. Their estimates imply that maize traders act consistently with joint profit maximization and earn median markups of 39 percent on maize sold to consumers, suggesting that there may be a cartel with monopoly power in this setting.

This market level approach relies on the definition of market boundaries, and the assumption that there is no interference between treatment and control markets---that is, outcomes in control markets are unaffected by the treatment. Outside of the retail context, one may not wish to make this assumption. For instance, in the context of an export crop produced in single region, and where traders are itinerant across the geography, it is not clear exactly how many compete for a single farmers’ supply. Defining market boundaries for instance at the village level to implement market-level randomization and test for monopsony power would be arbitrary, as traders are likely to compete across those boundaries. Online markets are another example in which delineating market boundaries is challenging.

Recognizing this, Lorenzo Casaburi and I use individual-level randomization across firms within a market to measure the extent of competition while avoiding the problem of defining market boundaries. It is well known that cocoa farmers receive only a small share of the surplus from chocolate sales, but there is uncertainty as to why. One hypothesis is that the traders that buy at the farm-gate collude to hold down prices. To evaluate this hypothesis, we provide subsidies randomly to a subset of traders buying raw cocoa (cacao) from farmers in Sierra Leone, generating experimental variation in their marginal revenue. We find that treatment traders pay slightly higher prices and secure much more quantity compared to control traders.

Since these treatment and control traders compete to buy from farmers in the same villages, one cannot interpret these average treatment and control differences as average treatment effects in the usual way of experimental analysis, which assumes no interference. Nonetheless, viewed through the lens of a model in which treatment and control traders compete with one another, the average difference in prices paid by treatment and control traders recovers the degree of differentiation among traders, which summarizes a trader’s ability to buy while paying farmers a lower price than competitors. The results suggest that traders are relatively homogeneous, with limited ability to price differently than one another, pointing to a competitive market. Combined with treatment and control differences in quantity, we also infer the elasticity of the supply curve traders face.

Armed with these parameter estimates, it is possible to infer the number of traders in the market without defining market boundaries, and test whether prices appear to be set by one trader (as under a cartel) or many. The experimental results are combined with an estimate of the pass-through rate to farmers of a change in the wholesale price affecting all traders in the market, not just those receiving the experimental treatment. The world price is used as an instrument for the wholesale price, yielding an estimate of the pass-through that is quite high. Note, the maize study also uses a pass-through rate for inference, but estimates it using market-level randomized variation, which, unlike the instrumental variable approach, requires defining market boundaries. The number of traders implied by the pass-through rate of the cocoa price, the trader differentiation rate, and the supply elasticity is 40% higher and significantly different from the average number of traders counted operating in a village, illustrating that markets cannot easily be delineated by village boundaries, in this case because farmers can sell their cocoa to traders outside their village.

Another general consideration in studying competition is non-price competition, that is, the idea that firms compete on other variables than price, such as trade credit, supply assurance, or customer service. The cocoa study deals with by allowing traders to be differentiated, and by identifying this differentiation with individual-level randomized variation. Further, the study measures treatment and control differences in trade credit provision and combines them with treatment and control differences in price to create a difference in “effective price” that summarizes both price and non-price competition. A similar approach could be taken with other measures.

Together, the cocoa and maize studies illustrate how different experimental approaches combined with industrial organization theory can teach us about competition, and how the appropriate approach for an industry depends on whether one is willing to delineate market boundaries as part of an experimental design. If researchers continue to deploy both market-level and individual-level approaches in diverse industries, together we can build more intuition about the extent of competition, helping to guide the design public policy. For instance, the cocoa and maize studies suggest that collusion can be present in different stages in the value chain. In the case of cocoa in Sierra Leone, there is no evidence of monopsony power among traders buying from farmers, but in the market for maize in Kenya, there is evidence that traders have monopoly power over consumers. These insights have implications for agricultural policy and antitrust enforcement in both settings.


Join the Conversation