Learning under ambiguity: An experimental investigation

Abstract

We investigate learning in ambiguous situations where subjects bet on a winning event whose probability depends on an unknown proportion of winning chips in an urn. Varying the number of draws prior to choice allows us to “scan” ambiguity attitudes across differing amounts of information. By separately eliciting posterior beliefs in addition to matching probabilities, we disentangle the impact of learning on ambiguity attitude from its impact on beliefs, including divergences from Bayesian update. Both “raw data” and smooth ambiguity model-based analyses show that learning affects ambiguity attitude in the direction of ambiguity neutrality. Moreover, at small sample sizes, the impact of these changes on preferences is comparable to that of the divergence from Bayesian update.