Research

From Individual Information to Group Beliefs: Memory, Aggregation, and Stereotype Persistence

(joint work with Zhuokun Liu)

We study why experiences that contradict stereotypes may have a limited influence on group beliefs. We distinguish two cognitive frictions: imperfect memory of individual performance and imperfect aggregation of individual assessments into group assessments. In two online experiments, subjects graded students on a carknowledge quiz and reported their predicted and recalled performance, individually and by gender. We randomly vary access to grading records to manipulate reliance on memory. When women outperform men, hiding records makes group predictions more male-favoring. Yet making records available does not ensure that group assessments reflect all individual information equally. Even with visible records, directly reported group predictions and recall imply more male-favoring gender gaps than those aggregated from individual assessments. This wedge suggests information reflected in individual assessments may not be aggregated accurately in group assessments. To facilitate aggregation, we test an intervention that first elicits all individual predictions and displays them when subjects make group predictions. When within-gender performance variation is low, this intervention offsets 67% of the effect of hiding records on the gender gap from directly reported group predictions.

Correlation Neglect in Information Demand and Belief Formation

This paper experimentally investigates how correlation neglect distorts both the ex-ante valuation of information and ex-post belief formation. Motivated by a stylized model, I jointly elicit subjects’ information valuations and posterior beliefs across randomized environments featuring independent, positively correlated, or negatively correlated signals. Asymmetry is documented in how cognitive belief biases affect information demand. Driven by correlation neglect, subjects fail to price in redundancy, overvaluing positively correlated information by 6.5% relative to Bayesian-equivalent independent information. However, when faced with negatively correlated information, this behavioral mapping breaks down. Although subjects continue to exhibit correlation neglect in their ex-post belief updating, they do not discount the negatively correlated signals ex ante. Ultimately, this paper demonstrates that the link between cognitive biases and information valuation is asymmetric and dependent on the direction of the correlation.