Bayesian Elicitation Mark Whitmeyer∗ January 28, 2020 Abstract How can a receiver design an information structure in order to elicit information from a sender? We study how a decision-maker can acquire more information from an agent by reducing her own ability to observe what the agent transmits. Intuitively, when the two parties’ preferences are not perfectly aligned, this garbling relaxes the sender’s concern that the receiver will use her information to the sender’s disadvantage. We characterize the optimal information structure for the receiver. The main result is that under broad conditions, the receiver can do just as well as if she could commit to a rule mapping the sender’s message to actions: information design is just as good as full commitment. Similarly, we show that these conditions guarantee that ex ante information acquisition always benets the receiver, even though this learning might actually lower the receiver’s expected payo in the absence of garbling. We illustrate these eects in a range of economically relevant examples. arXiv:1902.00976v2 [econ.TH] 26 Jan 2020 Keywords: Costly Signaling, Cheap Talk, Information Design, Bayesian Persuasion JEL Classications: C72; D82; D83 ∗Department of Economics, University of Texas at Austin Email: [email protected]. Thanks to V. Bhaskar, Gleb Domnenko, William Fuchs, Rosemary Hopcroft, Vasudha Jain, Meg Meyer, Vasiliki Skreta, Max Stinchcombe, Yiman Sun, Tymon Tatur, Alex Teytel- boym, Thomas Wiseman, and Joseph Whitmeyer. Thanks also to various seminar and conference audiences. This paper is a shorter version of the job market paper of the same name. 1 Contents 1 Introduction2 2 The Model 14 3 Opacity = Commitment 20 4 The Structure of Optimal Transparency and the Rewards to Ex Ante Learning 28 5 Examples 34 6 Discussion 44 A Section 3 Proofs 48 B Section 4 Proofs 53 C Section 5 (Examples) Derivations 54 1 Introduction In that open eld If you do not come too close, if you do not come too close, On a summer midnight, you can hear the music... T.S. Eliot East Coker There is a decision-maker faced with choosing an action in an uncertain world. She does not have direct access to information about the state of the world, but there is a second person who does. The second person (or sender) observes the state (equivalently, his type) before sending a message1, which the decision-maker (or receiver) observes before taking an 1We term the sender’s action a message in order to distinguish it from the receiver’s action. In some settings, like for instance cheap talk games, the moniker is tting. In others, like for instance the Spence (1978) [31] model of signaling through education attainment, where the sender chooses his level of education, labeling his action 2 action. This scenario is called a communication game (or a signaling game), and the class of such games includes cheap talk games, like that studied in Crawford and Sobel (1982) [10]; and costly signaling games, like that studied in Spence (1978) [31]. In nearly all of the literature, and for much of this paper, the message chosen by the sender does not enter into the receiver’s utility function; hence, the sender’s message has only instrumental value to the receiver. Consequently, the receiver’s only concern is infor- mation transmission: she2 values only the information content of the message, and so nat- urally her welfare increases as the sender’s messages become more informative. However, there are a number of potential frictions that could impede this information transmission. First, the sender’s and the receiver’s preferences over the action taken may be imperfectly aligned. Second, the messages may be costly to the sender, which costs aect the messages that are chosen in equilibrium. As a result, less than full information may be transmitted at equilibrium, and the frictions may even be so severe that no information is transmitted. Now, suppose that the receiver is not forced to observe the sender’s message directly, but may commit ex ante to observe a noisy signal of the message instead. That is, suppose the receiver may choose the degree of transparency, or information structure, in the game. Can less than full transparency help the receiver? Moreover, if less than full transparency is benecial, then what does the (receiver-)optimal degree of transparency look like, and how can the receiver solve the problem of designing the optimal information structure? If this were a decision problem in which the information were exogenous, i.e. if there were no sender and instead the message sent to the receiver about the state were from nature, then the answer to the rst question would clearly be no. Namely, if the message were exogenous, then the receiver would always (at least weakly) prefer to observe the message itself rather than some noisy signal. Conversely, here, the message is not exogenous, but is instead an equilibrium choice of the sender. Crucially, the sender is aware of the degree of transparency, which thus aects the message, or distribution over messages, that he sends at equilibrium. There is an important trade-o present in the receiver’s choice of information structure. a message is less appropriate. 2Throughout, to simplify the language, we arbitrarily impose that the receiver is female and the sender is male. 3 By choosing a more informative signal of the message, the receiver obtains more information for any xed strategy vector of the sender types. On the other hand, the vector of strategies chosen by the sender types, and hence the information, is an endogenous choice of the sender that he makes cognizant of the information structure. Thus, less transparency may beget a more informative vector of strategies at equilibrium. The optimal degree of transparency arises as a consequence of these trade-os, and it may strictly benet the receiver to choose a less informative signal of the sender’s message. In short, less than full transparency can strictly benet the receiver. The following example illustrates this observation. Sender and receiver are employees of a rm: the receiver is part of upper management, the Chief Operations Ocer (COO), say; and the sender is a (local) branch manager. The COO is contemplating whether to close the branch (action C) or not (action O). The branch is either viable (state G) or not (state B) and the COO would like to close the branch if the state is B and keep the branch open otherwise. Explicitly, the COO’s payos are uR(O, G) = uR(C, B) = 1, and uR(O, B) = uR(C, G) = 0 The COO is unable to observe directly the viability of the branch. Instead, the branch manager is “on the ground" and observes the state. The branch manager must make an investment decision, whether to buy new equipment, say; and all else equal, the branch manager would prefer to invest (message I) if and only if the state is G. On the other hand the branch manager would prefer the branch be kept open no matter the state; and, crucially, the payos are such that the branch manager would rather choose the incorrect investment decision for the state and have the branch be kept open than choose the correct investment level and have the branch be closed. The state-dependent utilities for the branch manager for each message, action combination are (I,O) (I,C) (N,C) (N,O) G 3 1 0 2 B 2 0 1 3 Both COO and branch manager share the common prior 0 ∶= Pr (Θ = G) = 2/3: the branch is more likely to be viable than not. 4 Figure 1: Whether to Close the Branch, Full Transparency The full transparency scenario is depicted in Figure1. There, no equilibria exist in which any information is transmitted. The only equilibria are pooling–those in which the branch manager chooses the same message no matter the state–and so the COO’s posterior is the same as her prior. The logic behind this is simple: there can be no equilibrium in which both I and N are sent in such a way that the COO strictly prefers to take a dierent action after each message. In such a circumstance, the local manager who knows the state is B always prefers to deviate to the message that is followed by the branch being kept open (O). The COO’s payo is 2/3. Suppose we introduce a neutral third party, a middle manager, say, who oversees the branch manager. The COO no longer observes the message of the branch manager; instead, the middle manager witnesses the message before communicating to the COO. Because the middle manager is neutral, we may model him as a signal, a stochastic map ∶ M → Δ(X) where X is some (nite) set of signal realizations. For any message m sent by the branch manager, the middle manager sends signal realization x to the COO with probability (x|m). One possible signal is one that is completely uninformative and involves just one signal realization, the statement (everything is ) f ine–no matter what message the sender chooses, the middle manager says that “everything is ne." Formally, the set of signal realizations is X = {f ine}, and (f ine|I) = (f ine|N ) = 1 We term this signal no transparency, and this scenario is depicted in Figure2. As in 5 Figure 2: Whether to Close the Branch, No Transparency the full transparency case, no information is transmitted at equilibrium, but now the branch managers separate and send dierent messages. In particular, the branch manager in the low state does not invest, since he knows that the receiver will not be able to identify him from this choice due to the garbling by the signal.