Incentivizing Information Design

A principal hires an agent to acquire costly information that will influence the decision of a third party. While the realized piece of information is observable and contractible, the experimental process is not. Assuming a general family of information cost functions (inclusive of Shannon’s mutual information), we show that the first best is achievable when the agent has limited liability or when he is risk averse, in contrast to standard moral hazard models. However, when the agent is risk averse and has limited liability, efficiency losses arise generically. Specifically, we show that the principal obtains his first best outcome if and only if she intends to implement a ”symmetric” experiment, i.e. one in which the cost of generating each piece of evidence is the same. On the other hand, ”asymmetric” experiments that are relatively uninformative with high probability but occasionally produce conclusive evidence will bear large agency costs.