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.