Quantal response and nonequilibrium beliefs explain overbidding in maximum-value auctions

We report new experimental data on a simple common value auction to investigate the extent to which bidding can be explained by logit QRE, in combination with different models about bidder beliefs: cursed equilibrium, level-k, and cognitive hierarchy. There is a close correspondence between the predicted bidding patterns in those models and the distribution of observed bids. The pattern of median bids in the data consists of a combination of overbidding with low signals, and near-value-bidding with higher signals. Logit QRE with heterogeneous bidders approximates this pattern. Combining QRE with any of the other models of belief formation leads to a closer match with the data. All the estimated models predict only small treatment effects across different versions of the game, consistent with the data. We also reanalyze an earlier dataset for the same game (Ivanov et al., 2010), which exhibited much more overbidding, and reach similar qualitative conclusions. •We report an experiment on a simple common value auction to investigate the extent to which bidding can be explained by quantal response equilibrium, in combination with different assumptions about the structure of bidder beliefs—the cursed equilibrium model and models that posit levels of strategic sophistication.•Using a structural estimation approach, we find a close correspondence between the theoretical predictions of those models and experimental behavior. The basic pattern of median bids in the data consists of a combination of overbidding for low signals, and value-bidding for higher signals.•The logit QRE model with heterogeneous bidders fits this pattern reasonably well. Combining quantal response with either cursed beliefs (CE-QRE) or a level-k of strategic sophistication (LK-QRE, CH-QRE) leads to a close match with the data. All these variations on quantal response models predict minimal differences of median bidding behavior across different versions of the game, consistent with the experimental findings.•Finally, we reanalyze data from an earlier experiment on the same auction by Ivanov, Levin and Niederle (2010). While their data exhibit much more variance compared with ours, we still find that these models fit their data reasonably well, even in the presence of extreme overbidding.