The Strategy Aggregation Effect in Judgment

Abstract

Predicting a criterion that is probabilistically related to pieces of information, or cues, is a paradigmatic judgment task that has been investigated both, in research trying to identify the individual judgment and decision making strategies people use, and in the wisdom-of-crowds literature where the focus is on how aggregation can improve accuracy. I combine these two lines of research to investigate how the performance of individual and aggregated linear strategies are affected by different environmental and group aggregation factors and how performance differences between individual and aggregated linear strategies can be understood in a unified framework. I show that constrained linear strategies are more accurate for individual judgments, but when these judgments are averaged, an unconstrained linear strategy is more accurate. This strategy aggregation effect can be understood by analyzing a decomposition of the mean squared error into bias, variance, and covariance. Because of their lower bias but higher variance, unconstrained linear strategies perform worse for individual judgments, but better for averaged judgments where aggregation minimizes variance. In simulations with artificial and real environments, I further show that this aggregation effect does not occur if there are correlations between individual judgments. Here, constrained linear strategies always outperform an unconstrained linear strategy, because the larger covariance component of the unconstrained linear strategy outweighs its lower bias. I end with real-world implications of the results for cognitive strategies and decision environments in group and organizational settings.

Beyond collective intelligence: Collective adaptation

“…The field of dynamic decision making, a subdiscipline of cognitive science, has focused on these types of tasks, showing for example that individuals and groups with shorter memories and noisier copying strategies can be more successful over time in changing environments (Gonzalez et al, 2015;Lejarraga et al, 2014). Other important distinctions include the way rewards are split among the members of the collective (Ostrom, 2010), distribution of relevant information (Simsek & Buckmann, 2015), predictability (Olsson, 2021), and speed of environmental change (Aoki & Feldman, 2014).…”

“…For example, the fitness of different cooperative strategies (Boyd & Lorberbaum, 1987) and social learning strategies (Rendell et al, 2010) is highly dependent on the distribution of competing strategies in the population. Inference strategies that produce high variance in individual solutions perform better when used in groups than when individuals use them alone (Olsson, 2021). Diversity of solutions in general promotes collective performance (Brush et al, 2018;Page, 2008).…”