The paper presents a new meta data set covering 13 experiments on the social learning games by Bikhchandani, Hirshleifer, and Welch (1992). The large amount of data makes it possible to estimate the empirically optimal action for a large variety of decision situations and ask about the economic significance of suboptimal play. For example, one can ask how much of the possible payoffs the players earn in situations where it is empirically optimal that they follow others and contradict their own information. The answer is 53% on average across all experiments - only slightly more than what they would earn by choosing at random. The players' own information carries much more weight in the choices than the information conveyed by other players' choices: the average player contradicts her own signal only if the empirical odds ratio of the own signal being wrong, conditional on all available information, is larger than 2:1, rather than 1:1 as would be implied by rational expectations. A regression analysis formulates a straightforward test of rational expectations, which rejects, and confirms that the reluctance to follow others generates a large part of the observed variance in payoffs, adding to the variance that is due to situational differences.
Social learning information cascades failure of rational expectations meta analysis