This paper develops a unified structural framework that identifies multiple, potentially-confounded belief-updating biases by using elicited prior and posterior beta distributions, and tests it in a laboratory urn experiment; at the population level base-rate neglect persists but overinference vanishes once richer bias channels are included, while individual-level analysis reveals widespread heterogeneity with sequence-related (hot-hand / gambler) and motivated-belief biases as primary drivers .
速读内容
Experimental setup & elicitation tool
Design: 88 subjects × 30 tasks; each task draws one urn from 99 (1–99% red) and elicits a prior (after 1–3 draws) and posterior (after additional 3/5/7 draws) as parametric beta distributions using a two-slider graphical interface .
Incentives: binarized/quadratic-style scoring to truthfully elicit mean and variance; 15 tasks (“dollar urns”) attach monetary stakes to induce preference-based motives .
Theoretical model and identification strategy
Core: prior ~ Beta(a0,b0) and signal ~ Binomial → Bayesian posterior Beta(a0+k, b0+n−k); biases introduced as distortions of likelihood/prior (exponents γ, δ) and extended via asymmetric weights (α, β), confirmation term ρ·c, sequence dummies, and precision multiplier ν to separately identify over/under-inference, base-rate biases, confirmation, motivated asymmetries, sequence effects, and precision biases .
Empirical strategy: compare a baseline (γ,δ-like) regression to a complete specification (Eqs.14–16) that adds preference and sequence interactions and a variance equation to disentangle channels .
Population-level findings
Baseline (parsimonious) summary: apparent overinference (overweighting signal) and base-rate neglect at the aggregate level .
Complete model: evidence for overinference disappears at the population level once richer bias channels are included; base-rate neglect remains the robust aggregate distortion .
Individual heterogeneity & bias importance
Individual estimates: all tested bias types appear among subjects to some extent; the complete model classifies many more individuals and reduces the “No Bias” group dramatically .
Quantitative importance: gambler’s fallacy and optimism produce the largest expected-value deviations, while hot-hand is the most prevalent bias and exerts its main effect via variance (overprecision) rather than mean shift .
Co-occurrence structure (bias networks)
Biases cluster: hot-hand and gambler’s fallacy are negatively correlated and act as hubs separating two clusters — (1) prior-based / against-signal biases linked to gambler’s fallacy and base-rate overuse, and (2) “jumping-to-conclusions” cluster (overprecision, optimism, overinference) linked to hot-hand — implying recurring patterns rather than independent biases .
Robustness & reduced-form link
Results robust to simulated measurement error in variance (SIMEX-style) and Bonferroni multiple-testing corrections; a Grether-style binary-state reduced-form recovers the baseline pattern (overinference + base-rate neglect) when continuous beliefs are collapsed to binary odds .