Multi-Attribute Auctions with Unobserved Heterogeneity in Supplier Qualities and Buyers Tastes

Multi-attribute auctions are widely used in industry procurement. Recently many Internet-based auction markets adopted multi-attribute structure. Multi-attribute auctions differ from standard low price auctions in that (a) a buyer chooses a winner on the basis of several indicators instead of taking only a price quote into account; (b) the exact weighting scheme that is used to determine the winner is not announced. Often the exact set of characteristics included in the score is left unspecified. These features create natural obstacles to the analysis of multi-attribute auctions since they imply that auctioneers' tastes as well as some of the relevant bidder characteristics may not be observed in the data. The structure of the market is such that exactly the same set of suppliers is considered by a negligibly small set of buyers. Therefore, in contrast to the markets with differentiated products we cannot assume that the probability of winning given the choice set is observed in the data. This paper shows that the multi-attribute auction model with unobserved bidder characteristic and unobserved auctioneer tastes is identified from the data. An important identification step deals with recovering bidders' quality rankings. This step essentially proposes an algorithm to recover bidder classification into unobserved groups. Once quality classification is constructed unobserved bidders tastes are recovered by exploiting the variation in a bidder probability of winning in response to price variation. We show that are model is capable of generating price variation required for identification. We apply our methodology to the data for on-line multi-attribute procurement auctions for services and estimate the distribution of buyers' tastes, sellers' unobserved characteristics and the distribution of sellers' private costs.