Our experiment investigates managers' self-predictions of their subsequent performance and, based upon, their choice of a collaborator. Our results show that managers' self-predictions are not biased anymore after they are informed about the performance of a reference group. In spite of this, most managers do not rationally choose a collaborator given their beliefs. In a second treatment, superiors (who are assumed to be at a higher hierarchy level than the managers) obtain various information, e.g. about managers' self-predictions, and have to predict the managers' performances. Our data show that superiors adapt their predictions into the direction of the managers' self-predictions, although not completely. Particularly, superiors think that their managers' self-predictions are biased if they are lower than the average performance of the reference group. Based upon their predictions, superiors have to select a collaborator for their managers. We find that superiors' collaborator choices do not significantly differ from the managers' choices. This proves due to excellent information processing by both, managers and superiors, which on the whole leads to very similar predictions of managers' subsequent performance.