Vertigo: Comparing structural models of imperfect behavior in experimental games

Abstract

We introduce the game of Vertigo to study learning in experimental games with one-sided incomplete information. Our models allow players to make errors when choosing their actions. We compare six models where the players are modeled as sophisticated (taking errors in action into account when constructing strategies) or unsophisticated on one dimension, and employ Bayes' rule, a faster updating rule, or no updating at all on the second. Using a fully Bayesian structural econometric approach, we find that unsophisticated models perform better than sophisticated models, and models with no (or slower) updating perform better than models with faster updating.