Comparisons of learning models in repeated games have been a central preoccupation of experimental and behavioral economics over the last decade. Much of this work begins with pooled estimation of the model(s) under scrutiny. I show that in the presence of parameter heterogeneity, pooled estimation can produce a severe bias that tends to unduly favor reinforcement learning relative to belief learning. This occurs when comparisons are based on goodness of fit and when comparisons are based on the relative importance of the two kinds of learning in hybrid structural models. Even misspecified random parameter estimators can greatly reduce the bias relative to pooled estimation.

The double auction trading institution (DA) has been highly efficient across diverse marginal-cost market structures, whether human subjects or "zero-intelligence" robots populated those markets. Accordingly, many researchers suspect that DA performance transcends market structure and agent strategy. But we show that (1) large avoidable costs undermine the efficiency and stability of human subject DAs, and (2) these low human efficiencies are simultaneously well above zero-intelligence efficiencies. Our results dramatically illustrate the potential havoc wrought by highly competitive institutions when they must cope with nonconvex technologies.

The double auction trading institution has been highly efficient across diverse marginal-cost market structures, whether human subjects or 'zero-intelligence' robots populated those markets. Accordingly, many researchers suspect that double auction performance transcends market structure and agent strategy. But the authors show that large avoidable costs undermine the efficiency and stability of human subject double auctions and these low human efficiencies are simultaneously well above zero-intelligence efficiencies. Their results dramatically illustrate the potential havoc wrought by highly competitive institutions when they must cope with nonconvex technologies.