Bayesian Persuasion under Bias Management

A principal delegates choice to an agent whose decision depends on both beliefs and tastes. The principal can steer the delegated decision using two costly instruments: (i) an information policy that determines a Bayes–plausible distribution of posteriors, and (ii) a bias-management policy that shifts the agent’s effective taste. We study a binary-state, two-action, convex hull of two benchmark tastes specialization with posterior-separable information costs. The analysis admits an inner–outer decomposition: optimal bias management is bang–bang (either no intervention or the minimal intervention needed to flip the agent’s action), while the optimal information policy is characterized by concavification of an endogenous posterior value function that already incorporates optimal management and information costs. This structure clarifies how information acquisition and bias management interact; they can be complements, substitutes, or both depending on the primitives of the model. Information changes which posteriors are realized and hence where management is used; management reshapes the curvature and kinks of the posterior value function and hence the marginal value of information. The model delivers regime classifications for pooling vs. informativeness and for management at different posteriors within informative signals, and highlights how comparative statics can be monotone or non-monotone depending on how concavification contact points move with costs.

💡 Research Summary

The paper develops a unified theoretical framework for a principal who delegates a binary‑state, binary‑action decision to an agent whose choice depends on both beliefs and an intrinsic “taste” that may be misaligned with the principal’s objective. The principal can influence the outcome through two costly levers: (1) an information policy that selects a Bayes‑plausible distribution of posterior beliefs, and (2) a bias‑management policy that shifts the agent’s effective taste within the convex hull of two benchmark tastes. Information acquisition incurs a posterior‑separable convex cost k_P·∫κ(p)τ(dp), while bias management costs k_V·q, where q∈