Updating ambiguous beliefs in a social learning experiment

Abstract

We present a social learning experiment in which subjects predict the value of a good in sequence. We elicit each subject's belief twice: …rst ("…rst belief"), after he observes his predecessors' prediction; second, after he also observes a private signal. Our main result is that subjects update on their signal asymmetrically. They weigh the private signal as a Bayesian agent when it con…rms their …rst belief and overweight it when it contradicts their …rst belief. This way of updating, incompatible with Bayesianism, can be explained by ambiguous beliefs (multiple priors on the predecessor's rationality) and a generalization of the Maximum Likelihood Updating rule. Our experiment allows for a better understanding of the overweighting of private information documented in previous studies.

Social information use and social information waste

“…… expert advice utilization was 48 percent, while subjects put 52 percent of weight on their own initial estimate. This implies that on average, given own and expert's experience levels, subjects were egocentric towards their own opinions and discounted advice" (p. [16][17][18][19][20][21][22] Novaes Tump et al (2018), one unique experiment. "… individuals were relatively reluctant to incorporate social information and instead used suboptimal switching thresholds" (p. 7)…”

“…These two principles imply that the average random participant should give equal weight to her opinion and to that of a random participant from the same group [18]. This basic principle can be formalized in various ways, the most common being Bayesian updating rules [19][20][21][22][23][24] or the averaging heuristic [18,25]. This point of view is not universally shared.…”

“…In most of the 'bandit' and 'arrowhead' tasks, participants get private feedback on their actions, which they take into account in a near-optimal way, contrasting with their poor use of social information [33,55]. Likewise, participants in advice-taking tasks use new evidence efficiently when it is not social [18,24]. Self-confidence is a reliable predictor of egocentric discounting [31]: indeed, as Hawthorne-Madell & Goodman [26] show, it is rational (in the authors' specific sense) for a self-confident agent to discount divergent opinions.…”

Information choice in a social learning experiment

“…Relative to the symmetric design of our paper, in both of these earlier experiments, the two forms of information are treated asymmetrically. De Filippis et al (2016) also modify the Anderson and Holt setup but by adding a belief elicitation stage after subjects observe the decisions of others and once again after they have received their own private signal. This design enables them to assess the extent to which subjects update their beliefs, conditional on each type of information received.…”

Information Redundancy Neglect versus Overconfidence: A Social Learning Experiment

“…Such a tendency is documented in various studies (e.g. Nöth and Weber, 2003;Çelen and Kariv, 2004;Goeree et al, 2007;and De Filippis et al, 2016), and it is typically referred to as overcondence, since subjects seem to trust their own information (or own ability to learn from it) more than their predecessors' information (or ability to learn from it).…”