Generality, repetition, and the role of descriptive learning models

Previous studies of learning in certain classes of games appear to lead to inconsistent conclusions. Studies that focus on the observed sequential changes in behavior support models that imply high-action inertia and allow for different parameters in different games. Studies that use simulation-based analysis and focus on the prediction of behavior in new games support models that imply little-action inertia, and demonstrate the value of models that assume general parameters over certain classes of games. We show that this apparent inconsistency emerges even when analyzing a large data set with a single model. We then show that the inconsistency between the two analyses can be a product of the tendency by subjects to repeat past choices.