Disambiguating Information Interventions: Recovering Beliefs and Ambiguity Attitudes from Virtual Twins

Elicited probabilities conflate latent beliefs, ambiguity attitudes, and response er- ror. We develop a measurement method that identifies these components from noisy subjective-probability data using consequential bets on singleton and union events. Ambiguity weighting is parameterized by two indices (a,b), capturing insensitivity and elevation. A hierarchical Bayesian model places latent beliefs on the simplex, al- lows respondent-level (ai,bi), and treats elicited probabilities as noisy measurements. We implement the method in a randomized information experiment in the Dutch LISS panel, where respondents make incentivized forecasts about a demographically matched virtual twin’s GP utilization. The treatment effect on raw probabilities is imprecise and indistinguishable from zero; model-recovered beliefs show meaningful updating, improving predictive accuracy by 0.026–0.032 Brier points and lowering elevation-based ambiguity aversion, with a smaller and less robust change in likelihood insensitivity. Treatment effects on elicited probabilities need not identify treatment effects on beliefs.