Title: When social influence promotes the wisdom of crowds

Abstract: Questions regarding whether, and if so, under what conditions, groups exhibit ''crowd wisdom'' have spurred numerous studies in many disciplines, including management and organizational science, psychology, sociology, complex systems, and computer science. Substantial effort in previous research on these questions has focused on investigating the role of social influence in promoting the wisdom of the crowd or, conversely, leading the crowd astray. Specifically, many previous studies have sought to infer the importance of social influence network attributes (such as influence centralization) to explain the accuracy of collective estimates. In this paper, we argue that this approach is limited and can lead to inconsistent conclusions. Based on our theoretical analysis, numerical simulation, and reanalysis of four previously published experiments (which included a total of 4,002 human participants, organized in 131 independent groups), we demonstrate that the wisdom of crowds in estimation tasks depends on the interaction between the following two factors: (i) centralization of the social influence network, and (ii) the features of the estimation context---i.e., the distribution of the initial (pre-influence) estimates. Specifically, we find that centralized influence is desirable in situations where a crowd is predisposed to overestimation bias and/or have a high likelihood of committing egregious errors. By adopting a framework that integrates both the structure of social influence and the estimation context, we bring the previously conflicting results under one theoretical framework and clarify the effects of influence centralization on the quality of crowd wisdom.