Averaging versus sampling in collaborative judgment

How does a dyad combine information from different members in order to arrive at a consensus judgment? One suggestion is that groups combine information in a Bayes optimal fashion: the group calculates a weighted average of individuals' estimates, with the weightings being proportional to the quality of the information each individual possesses. Alternatively, the dyad may seek to identify which member's estimate is the best, and return that as a joint judgment. These models were tested by asking members of a dyad to make private estimates of a continuous quantity (the direction of movement of a coherent motion stimulus), and to then make a joint judgment. Joint judgments were more accurate than individual judgments, but were often in the neighborhood of one of the individual judgments. Fitting of two candidate models suggested that dyads formed a combined probability distribution across possible responses; that this distribution was bimodal by virtue of the heavy-tailed nature of the individual distributions; and that dyads sampled their joint judgment from the combined distribution. In both models, relative accuracy on the current trial, as well as historical accuracy, were both predictors of whose individual estimate was given more weight in the joint judgment. Rather than relying on an averaging process to reconcile estimates to reach consensus, the results highlight the role of more categorical resolution in continuous domains of judgment, and that joint judgments are well explained by sampling from a bimodal combined distribution.