MOTIVATING AGENTS TO ACQUIRE INFORMATION * Charles Angelucci † June 2015 Abstract It is difficult to motivate advisors to acquire information through standard performance-contracts when outcomes are uncertain and/or only observed in the distant future. Decision-makers may however exploit advisors’ desire to influence the outcome. I build a model in which a decision-maker and two advisors care about the decision’s con- sequences, which depend on an unknown parameter. To reduce uncer- tainty, the advisors can produce information of low or high accuracy; this information becomes public but the decision-maker cannot assess its accuracy with certainty, and monetary transfers are unavailable. I show that, when the decision-maker can commit to a decision-rule, he should behave as if the provided information is more accurate than it actually is, and thus grant excessive importance to the produced information. Keywords: Information acquisition, experts, communication, advice. JEL Classification Numbers: D00, D20, D80, D89 P16. * I am deeply indebted to Philippe Aghion, Emmanuel Farhi, Oliver Hart, Patrick Rey, and Jean Tirole for their invaluable guidance and support. I am also grateful to Nava Ashraf, Daisuke Hirata, and Andrei Shleifer for insightful comments. This paper has also benefited from discussions with Julia Cag´ e, Wouter Dessein, Alfredo Di Tillio, Matthew Gentzkow, Renato Gomes, Ben Hebert, Christian Hellwig, Emir Kamenica, Nenad Kos, Thomas Le Barbanchon, Lucas Maestri, Eric Mengus, Simone Meraglia, Marco Ottaviani, Nicola Persico, Alexandra Roulet, Alp Simsek, Kathryn Spier, David Sraer, Eric Van den Steen, and audiences at Harvard, HBS, Yale SOM, Chicago Booth, Columbia, Kellogg, WUSTL, Toulouse, Bocconi, CUNEF, HEC, Rochester, Ecole Polytechnique, Universit´ e Paris Dauphine and the Petralia Workshop. Opinions and errors are mine. † Columbia Business School. Email: [email protected]1
57
Embed
MOTIVATING AGENTS TO ACQUIRE INFORMATION€¦ · MOTIVATING AGENTS TO ACQUIRE INFORMATION Charles Angelucciy June 2015 Abstract It is di cult to motivate advisors to acquire information
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
MOTIVATING AGENTS TO ACQUIREINFORMATION∗
Charles Angelucci†
June 2015
AbstractIt is difficult to motivate advisors to acquire information through
standard performance-contracts when outcomes are uncertain and/oronly observed in the distant future. Decision-makers may howeverexploit advisors’ desire to influence the outcome. I build a model inwhich a decision-maker and two advisors care about the decision’s con-sequences, which depend on an unknown parameter. To reduce uncer-tainty, the advisors can produce information of low or high accuracy;this information becomes public but the decision-maker cannot assess itsaccuracy with certainty, and monetary transfers are unavailable. I showthat, when the decision-maker can commit to a decision-rule, he shouldbehave as if the provided information is more accurate than it actuallyis, and thus grant excessive importance to the produced information.
Keywords: Information acquisition, experts, communication, advice.
∗I am deeply indebted to Philippe Aghion, Emmanuel Farhi, Oliver Hart, Patrick Rey,and Jean Tirole for their invaluable guidance and support. I am also grateful to NavaAshraf, Daisuke Hirata, and Andrei Shleifer for insightful comments. This paper has alsobenefited from discussions with Julia Cage, Wouter Dessein, Alfredo Di Tillio, MatthewGentzkow, Renato Gomes, Ben Hebert, Christian Hellwig, Emir Kamenica, Nenad Kos,Thomas Le Barbanchon, Lucas Maestri, Eric Mengus, Simone Meraglia, Marco Ottaviani,Nicola Persico, Alexandra Roulet, Alp Simsek, Kathryn Spier, David Sraer, Eric Van denSteen, and audiences at Harvard, HBS, Yale SOM, Chicago Booth, Columbia, Kellogg,WUSTL, Toulouse, Bocconi, CUNEF, HEC, Rochester, Ecole Polytechnique, UniversiteParis Dauphine and the Petralia Workshop. Opinions and errors are mine.†Columbia Business School. Email: [email protected]
1
1 Introduction
Information acquisition is central to decisions made under uncertainty, yet
it often relies on intermediaries. For instance, regulatory agencies benefit from
internal staff expertise; judges rely on prosecutors to collect evidence; and
managers ask subordinates to estimate projects’ returns. However, delegating
information acquisition responsibilities triggers a natural concern about the
quality of the evidence upon which the decision will be made.
In many environments, performance-based contracts offer a powerful means
ensuring the provision of accurate information from experts (e.g., portfolio
managers). In other environments, performance is difficult to measure and
so alternative levers must be found. This is particularly true when the ap-
propriateness of decisions becomes apparent only in the distant future (e.g.,
regulatory actions), if ever (e.g., judicial sanctions).1 In addition, specifying
the desired quality of advice in a contract—to reward the information itself
rather than the outcomes—is often prohibitively costly or unenforceable in a
court of law, so that not only is performance difficult to assess, but also mon-
etary transfers are a rather coarse instrument.2 However, in many cases, a
decision-maker may be capable of motivating information-acquisition by ex-
ploiting her advisors’ desire to influence the outcome.3 This alternative channel
1The appropriateness of a past decision may also be hard to measure in absence of clearcounterfactuals (e.g., pollution standards).
2In many environments, monetary transfers contingent on performance are banned al-together (e.g., regulatory hearings, expert witnesses in courts of law, etc). See Szalay [2005]and Alonso and Matouschek [2008] for discussions of this matter. In other settings—e.g.,scientific journals—the decision-maker has only a limited budget (if any) to allocate.
3An advisor’s desire to influence the outcome may be due to his reputational concerns,or simply because the decision affects him directly.
2
of influence is the object of this paper.
I develop a model whose main insight is that the decision-maker should
make its decision exceedingly sensitive to the provided information. In par-
ticular, in certain informational environments, the decision-maker should act
as if the information provided is more accurate than it actually is. This, I
show, is the optimal way of structuring the final decision in a way that makes
it privately interesting for her advisors to acquire costly but accurate informa-
tion. I argue that this result squares well with the way organizations motivate
information acquisition and make decisions in practice.
A decision-maker interacts with one or two agents before taking a (contin-
uous) action. All parties are interested in the decision’s consequences, which
depend on some unknown state of the world.4 In fact, for most of the analysis,
players are assumed to agree on which decision is optimal when they have
access to the same information. To reduce uncertainty, agents privately ac-
quire one of two informationally-ranked signals about the state of the world,
with the more accurate signal being also the costliest one. The decision-maker
wishes to motivate agents to acquire the more accurate signal, but is unable
to infer with certainty which signal was acquired by looking at its realization.5
Monetary transfers contingent on the information itself are ruled out, and thus
the choice of the final action is the only means by which the decision-maker
can motivate the agents.6 The decision-maker is assumed capable of commit-
4See for instance Prendergast [2007], [2008] on the intrinsic motivation of bureaucrats.5The signals share a common support.6In the spirit of Grossman and Hart [1986] and Hart and Moore [1990], I posit that a
contract specifying monetary transfers as a function of the information provided cannot bedeposited in a court of law.
3
ting to a decision rule, which specifies the action to implement as a function of
the information gathered by the agents. This provides a good approximation
of the settings in which decision-makers can restrain their behavior through
tiary rules in the judicial system).7 Finally, the informational environment I
consider is such that higher realizations of either signal lead to correspondingly
higher actions preferred by the decision-maker and the agents, that is, each
player’s schedule of desired actions is upward sloping.
I first look at the version of the model with one agent. In this environment
I show that, to induce the acquisition of the more accurate signal, it is optimal
for the decision-maker to implement a higher (lower) action than desired when
observing a realization that is such that the agent would desire a lower (higher)
action under the less accurate signal. Such a distortion hurts the agent (and
the decision-maker) under the more accurate signal, but relatively more so
under the less accurate signal, thereby motivating information acquisition.
I then restrict attention to environments in which signals induce sched-
ules of desired actions that are rotations one of another around the action that
would be implemented under the prior belief (i.e., the status quo). In this envi-
ronment, I show that it is optimal for the decision-maker to implement higher
actions than desired when observing high realizations—realizations that lead
to desired actions higher than the status-quo under either signal, and lower
actions than desired when observing low realizations—i.e., realizations that
lead to desired actions smaller than the status-quo under either signal. In
7Moreover, judges’ decisions are potentially subject to review by higher courts.
4
other words, it is optimal for the decision-maker to act as if the information
provided is more accurate than it actually is. Introducing an upward (down-
ward) distortion when observing a high (low) realization motivates the agent
to acquire the signal of high accuracy and have his posterior beliefs be less
anchored on the status-quo.
It is common practice for organizations to commit to ex-post inefficient
rules to foster information acquisition by its members (e.g., through the del-
egation of decision rights, the choice of default-actions, etc).8 One approach
close in spirit to the one advanced in this paper is for organizations to put ex-
cessive weight on the information produced by their own staff. Choosing and
committing to a weight to put on certain types of information (e.g., estimations
and forecasts) is routinely done by administrations which, for democratic and
efficiency reasons, must commit to transparent decision-making procedures
(e.g., US Sentencing Guidelines). An alternative approach consists of dis-
missing already available relevant information (e.g., reputation, past history,
information coming from the outside). For instance, the non-admissibility of
certain relevant pieces of evidence in the inquisitorial system of law (e.g., crim-
inal records) gives incentives to the prosecutor to find better information.9 10
More generally, it has been documented by organizational sociologists that
decision-makers within firms ignore relevant information or have poor “orga-
nizational memory”, (see for instance Feldman and March [1981] and Walsh
8See for instance Aghion and Tirole [1997], Dessein [2002], and Szalay [2005] on delega-tion, and Li [2001] on the choice of default-actions.
9See Lester et al. [2012] on a related matter.10Similarly, the editor of a scientific journal may acquire a reputation for not taking
into account the authors’ publishing records, so that the referee’s input becomes even morepivotal.
5
and Ungson [1991]). My paper provides one possible efficiency rationale: by
disregarding available information, firms motivate employees by making their
(better) information pivotal.
If to motivate information acquisition by one agent involves making the
final decision excessively sensitive to the information he provides, it is not
clear a priori how this intuition survives when listening to several agents. To
investigate this issue, I extend the baseline model by incorporating a second
agent. To motivate each agent, I show that it is optimal for the decision-maker
to once again acts as if the jointly provided information is more accurate than
it actually is.
The rest of the paper is structured as follows. A short review of the liter-
ature concludes the introduction. Section 2 presents the setup of the model.
Section 3 characterizes the optimal decision rule for the baseline model with
one agent (Section 3.1); it then turns to the analysis with multiple agents
(Section 3.2). Section 4 provides a discussion, and Section 5 concludes.
Related Literature
This paper contributes to a recent strand of literature that incorporates
information acquisition in principal-agent models without monetary transfers
or, more generally, with limited or costly transfers (see, for instance, Aghion
and Tirole [1997], Dewatripont and Tirole [1999], Szalay [2005], Gerardi and
6
Yariv [2008a], Krishna and Morgan [2008], and Che and Kartik [2009].11,12
Unlike Aghion and Tirole [1997] and Che and Kartik [2009], I assume that
the decision-maker can commit to an action to implement as a function of the
information provided. Szalay [2005] considers a set-up with, on the one hand, a
single agent who acquires costly information and, on the other hand, a decision-
maker unable to verify the validity of the acquired information. Szalay shows
that to induce information acquisition, it is optimal to delegate decision rights
to the agent while eliminating from the decision set those actions that the latter
would implement when uninformed. My paper is complementary to Szalay’s
in that it considers instances in which the decision-maker is able to verify the
acquired information. In the single-agent case, I show that the decision-maker
retains authority, possibly implements all actions, and makes the final decision
exceedingly sensitive to the agent’s information. In particular, the decision-
maker introduces small distortions when the information provided confirms
the prior, and large ones otherwise. Within this literature, to the best of
my knowledge, this paper is the first to investigate the issue of motivating
information acquisition by multiple agents in environments of nontransferable
utility.13
This paper is also closely related to the literature on experts and their
11See also Austen-Smith [1994], Dur and Swank [2005], Che and Kartik [2009], Argen-ziano et al. [2013] and Di Pei [2013] for models with information acquisition but lack ofcommitment.
12For models in which performance-contracts are available, see Demski and Sappington[1987] and Gromb and Martimort [2007].
13The presence of multiple experts has however been considered in other environ-ments (though most often without information acquisition). See for instance Milgrom andRoberts [1986], Krishna and Morgan [2001], Battaglini [2004], Ambrus and Takahashi [2008],Gentzkow and Kamenica [2011], Gul and Pesendorfer [2012], and Bhattacharya and Mukher-jee [2013].
7
career concerns, which typically posits that experts have no intrinsic interest
in the decision being made, but wish to signal that they are of high ability (see
Osband [1989], Scharfstein and Stein [1990], Holmstrom [1999], Ottaviani and
Sorensen [2006], and Bar-Isaac [2008]). In many of these papers, it is also the
case that the expert is delegated decision rights (see Demski and Sappington
[1987], Prendergast and Stole [1996], Milbourn et al. [2001], Prat [2005], Dai
et al. [2006], Malcomson [2009], and Fox and Van Weelden [2012]). To focus
on the implications of the agent’s interest in the outcome of the decision, I
ignore signalling issues. Also, in my environment, the delegation of decision
rights is never desirable.
The fact that it may be optimal for the decision-maker to exaggerate the
accuracy of the information provided is reminiscent of the optimality of hyper-
powered incentives schemes present in models of procurement that incorporate
information acquisition (see for instance Lewis and Sappington [1997] and Sza-
lay [2009]).14 In addition to showing that the spirit of this finding is present
also in models of expertise with nontransferable utility, I investigate whether it
survives in presence of multiple agents. More generally this paper contributes
to the mechanism design literature that explicitly considers information ac-
quisition (see Kim [1995], Persico [2000], Bergemann and Valimaki [2002], and
Shi [2012]).
This paper is also related to the literature on persuasion and disclosure
(see Grossman [1981], Milgrom [1981], Milgrom and Roberts [1986], Glazer and
14In a setting in which the agent has private information regarding the accuracy of apublic signal about its costs of production, Chu and Sappington [2010] show that it may beoptimal to induce production levels above or below the first-best.
8
Rubinstein [2004], Glazer and Rubinstein [2006], and Kamenica and Gentzkow
[2011]).15,16 One insight delivered by the model is that inducing an agent to
disclose information limits the decision-maker’s ability to motivate information
acquisition.
This paper touches upon similar issues as those found in the literature on
committees, most notably the issue of keeping several agents sufficiently pivotal
that they would engage in information acquisition (see Gilligan and Krehbiel
[1987], Persico [2004], Gerardi and Yariv [2008a], Gershkov and Szentes [2009],
and Hao and Suen [2009b] for a survey).
It also relates to papers investigating whether it is preferable for a decision
maker to take advice from agents who share her preferences or beliefs (e.g.,
see Dur and Swank [2005], Gerardi and Yariv [2008a] and Che and Kartik
[2009]).17 In line with these papers, I find that when agents need to acquire
information, some heterogeneity may be desirable.
Although the environment I consider is such that delegating decision rights
is never optimal, some of the techniques and concepts used—in particular the
construction of incentive-compatible mechanisms—are also present in the liter-
ature on delegation (see for instance Holmstrom [1977], Melumad and Shibano
[1991], Alonso and Matouschek [2008], Goltsman et al. [2009], Armstrong and
Vickers [2010], and Nocke and Whinston [2013]), which also assumes commit-
15See Milgrom [2008] for a survey.16The case of unverifiable information has also received much attention, starting with the
seminal paper by Crawford and Sobel [1982] and with more recent contributions (Dessein[2002], Battaglini [2004], Martimort and Semenov [2008], Kartik [2009], and Che et al.[2013],). See Sobel [2010] for a survey of the literature on Cheap Talk.
17For a model investigating this issue, but without information acquisition, see Landieret al. [2009].
9
ment.18 Furthermore, some aspects of my analysis—and, in particular, the role
played by the likelihood ratio—are reminiscent of the early analyses of moral
hazard (see Mirrlees [1976], Holmstrom [1979], Holmstrom [1982], Grossman
and Hart [1983],and Mirrlees [1999]), except that in my environment the moral
hazard problem cannot be solved with monetary transfers. Finally, I build
upon recent advances in comparative statics techniques to model the infor-
mation acquisition process (in particular Persico [2000] and Athey and Levin
[2001]) and use an information criterion introduced by Johnson and Myatt
[2006].19
2 The Setting
I consider a game with a decision-maker, DM , and up to two agents, A1
and A2. DM has payoff function u (a, ω), and Ai, for i = 1, 2, has payoff
function u (a, ω)−C (ei). Payoff functions depend on DM ’s action a ∈ R and
the unobservable state of the world ω ∈ Ω ⊆ R. The function C (ei) represents
the cost to Ai of exerting effort ei, and is explained in Section 2.2.
The state of the world ω is a random variable W with cumulative distribu-
tion function FW (ω) on support Ω. W is taken to have full support, i.e., the
associated density function satisfies fW (ω) > 0 for any ω ∈ Ω.
Finally, monetary transfers are assumed unfeasible and thus do not enter
payoffs. The importance of this assumption is discussed in Section 4.
18See Van den Steen [2008] for a set-up in which it is the potential difference of opinionsbetween players that paves the way to the delegation of decision-rights.
19See Shi [2012] for a model on auctions also focusing on this informational environment.
10
2.1 Payoff Functions
I restrict my attention to gross payoff functions u (a, ω) that are continuous
in both arguments, strictly concave in a, and whose second-order derivative
with respect to a is independent of both ω and a.20 Common functional forms
that satisfy these conditions include, for instance, − (ω − a)2 and ωa − a2
2.
Moreover, let
a(ω) := argmaxa∈R
u (a, ω)
denote the players’ ideal action when W = ω. I assume the payoff function is
such that a(ω) is strictly increasing in ω. Players wish to match higher states
with higher actions.
2.2 Effort and Information Acquisition
Ai privately chooses his effort level ei ∈ E, which translates into the acqui-
sition of signal Xei that he chooses from a set of signals Xee∈E. A signal Xe
is distributed as GXe|W (x|ω). All signals share common support X ⊂ R, and
x denotes a typical realization. Let GXe (x) = EW[GXe|W (x|ω)
]represent the
marginal distribution of signal Xe, with associated density function gXe(x).
Agents have access to the same set of signals Xee∈E, and realizations are
conditionally independent.
Throughout the paper, I maintain the assumption that effort levels (i.e.,
signals) are privately chosen (i.e., unobservable to the other players), but that
20Assuming the second-order derivative with respect to a is independent of both ω and aimplies that a player’s expected payoff under a given belief about ω is a translation of thisplayer’s expected payoff under any other belief about ω.
11
signal realizations are publicly observable.
If effort levels (i.e., signals) were observable, the players’ desired action
when observing Xe1 = x and Xe2 = x′ would be:
a (x, x′ | Xe1 , Xe2) := argmaxa∈R
∫Ω
u (a, ω) dFW |Xe1 ,Xe2 (ω | x, x′) ,
where updating occurs according to Bayes’ rule. suppose that signals are such
that the players’ ideal action is, all else equal, nondecreasing in each signal’s
realization, regardless of which signals are acquired.21
For most of the analysis, I suppose E = e, e, and let X = Xe andX = Xe.
Moreover, I assume signalX to be more informative than signal X, in the sense
that:
∫X
∫X
∫Ωu(a(x, x′ |X,Xe2 , ω
)dFW |X,Xe2
(ω | x, x′
)dgX (x) dGXe2
(x′)>
∫X
∫X
∫Ωu(a(x, x′ | X,Xe2 , ω
)dFW |X,Xe2
(ω | x, x′
)dgX (x) dGXe2
(x′),
∀Xe2 .22 Moreover, I assume C (e) = 0 < C (e) = c. Acquiring the more
informative signal implies a higher disutilty to the agents.
To ensure an interior solution, I will be requiring that the signals X andX
are such that the highest value the likelihood ratiogX(x)
gX(x)takes over the support
X is lower than some threshold, where this threshold is problem-specific.23
21For most applications of my analysis, it is enough to restrict attention to signalsthat induce conditional densities for W that satisfy the monotone likelihood ratio property(MLRP). Recall that the conditional density function fW |X has the MLRP if for x′ > x,
fW |X (ω | x′) /fW |X (ω | x) is nondecreasing in ω.22Because agents have access to identical signals X andX, it is also the case that, holding
Xe1 constant, DM ’s expected payoff is higher when Xe2 =X than when Xe2 = X.23This assumption is made to focus on the most interesting case in which DM is unable
12
Finally, note that because of the common support assumption, DM will be
unable to infer with certainty which signal was acquired when observing its
realization.24
2.2.1 Rotations
I sometimes focus on payoff functions and signals that induce schedules of
desired actions that are rotations one of another. To ease exposition, suppose
there is one agent only. The intuition for the case with two agents is provided
in Section 3.2. The players’ schedule of desired actions induced by signal
X, denoted a (x) = a(x |X
)is a rotation of the schedule of desired actions
induced by signal X, denoted a (x) = ai (x | X), around some realization x if:
(a (x)− a (x)) (x− x) > 0 ∀x 6= x.
In these environments, I refer to the realizations higher than x as high real-
izations. The desired action when observing a high realization is higher under
signalX than under signal X. Conversely, I refer to the realizations lower than
x as low realizations. The desired action when observing a low realization is
lower under signal X than under signal X.
Many payoff functions and signals lead to schedules of desired actions that
to (almost) implement the first-best by committing to take extreme actions when observ-ing certain realizations x highly informative of the effort level. The intuition for this isreminiscent Mirrlees [1999]).
24Take the case of an agent collecting data and carrying out a statistical analysis (say,computing a correlation coefficient). Although the decision-maker can replicate the statis-tical analysis if given access to the data, she may lack the time or resources to verify theactual representativeness of the sample, and cannot infer it perfectly from observing thestatistical output.
13
can be depicted as rotations one of another. Suppose, for instance, that payoff
functions are such that the players’ desired actions are equal to the expected
value of W (e.g., payoffs are of the form − (ω − a)2). Then, an environment in
which Ω = R, W ∼ N (0, 1), X = ρW +√
1− ρ2ε, and X = ρW +√
1− ρ2ε,
with 0 < ρ < ρ < 1, leads to schedules a (x) = ρx and a(x) = ρx, which are
rotations one of another around x = 0. Moreover, one can compute that the
likelihood ratiogX(x)
gX(x)is constant and equal to 1.
Similarly, an environment in which Ω = X = R+, with fW (ω) = 1λe−
ωλ ,
fW |X(ω |X = x
)= 1
λxe−
ωλx , and fW |X (ω | X = x) = 1
λxe−
ωλx , where 1 < λ <
λ < λ, yields schedules a (x) = λx and a (x) = λx, which are are rotations one
of another around x = 0.
Further examples include environments in which W ∼ U [0, 1], X = W + ε,
and X = W + ε, with ε ∼ N(
0, 1α
)and ε ∼ N
(0, 1
α
), and with X (ω) =
[ω − ν, ω + ν].25
Decision Rules and Timing Most of the analysis is carried out assuming
that DM can credibly commit to a deterministic decision rule a (x1, x2), which
is a mapping from signals into actions.26 The timing is as follows. First, DM
announces a decision rule. Second, agents privately choose their signals. Third,
the realizations of the signals are observed by all and, fourth, DM implements
the action specified in the decision rule.
25As a last example, one can also consider an environment in which W ∼ N(
12 ,
1β
)on
Ω = [0, 1], with X = W and X ∼ U [0, 1].26In the environment with public information, the framework is one of moral hazard with
binary actions. Using more general communications schemes—e.g., attempting to elicit theagents’effort levels—would not improve DM ’s payoff.
14
Finally, for simplicity, participation constraints at the agents’ level are
ignored. This is a natural assumption when agents are directly influenced by
the decision being made, so that their “participation” is not a choice. A brief
discussion is devoted to this issue in Section 4.
3 The Model
3.1 The Single-Agent Case
We begin by analyzing the single-agent case by assuming that DM does
not observe which signal is acquired but does observe its realization. Since we
here treat the case with one agent only, we simplify the notation by writing
DM and A’s preferred actions under signal X and signal X as, respectively,
a(x) and a(x).
The decision-maker’s problem. To focus on the case of interest, I assume
that A’s disutility c when acquiring signal X is high enough that the former
has insufficient private incentives to acquireX if DM implements her preferred
action under signal X for ∀x. Formally, I assume that
∫X
∫Ωu (a (x) , ω) dFW |X (ω | x) dGX (x)−c <
∫X
∫Ωu (a (x) , ω) dFW |X (ω | x) dGX (x) .
If the inequality were reversed, then, rather trivially, it would always be op-
timal for DM to induce A to acquire signal X and systematically implement
a (x) for ∀x. The immediate consequence of this assumption is that DM must
necessarily introduce distortions in the decision rule to induce the acquisition
15
of signalX, so that it is unclear a priori whether such a policy is desirable. In
the following, again to focus on the case of interest, assume that it is optimal
for the decision-maker to induce the agent to acquire signal X.27
Suppose DM wishes A to acquire signal X. She then designs the schedule
of decisions a (x) to maximize her expected payoff:
∫X
∫Ωu (a (x) , ω) dFW |X (ω | x) dGX (x) , (1)
subject to inducing A to acquire signal X rather than signal X:
∫X
∫Ωu (a (x) , ω) dFW |X (ω | x) dGX (x)−c ≥
∫X
∫Ωu (a (x) , ω) dFW |X (ω | x) dGX (x) .
(2)
Inequality (2) tells us that DM must design the decision rule a (x) so that it is
a sufficiently better match with the posterior beliefs induced by signalX than
with those induced by signal X. The existence of a solution is guaranteed by
the strict concavity of payoff functions and the unboundedness of A.28
I solve this problem by setting up a Lagrangian, where γ is the multi-
27It will be so whenever the informational gain outweighs the payoff loss associated tothe distortions.
28To see this, consider two realizations x′ and x′′, such that x < x′ < x′′, and such thatgX(x′)
gX(x′) > 1 , and set a (x) = aA(x) + η for ∀x ∈ [x′, x′′], where η > 0, and, say, a (x) = aA(x)
for all other realizations. Then, observe that, ∀x ∈ [x′, x′′], we have that:
limη→∞
[gX(x)
∫Ω
v (aA (x) + η, ω) dFW |X (ω | x)− gX(x)
∫Ω
v (aA (x) + η, ω) dFW |X (ω | x)
]= +∞,
since (i)∫
Ωv (a, ω) dFW |X(ω | x) and
∫Ωv (a, ω) dFW |X(ω | x) are strictly concave
(and with identical second-order derivatives), (ii)gX(x)
gX(x) > 1, and (iii) since, by def-
inition of x > x, aA(x) > aA(x) for ∀x ∈ [x′, x′′]. This, in turn, implies that∫ x′′
x′ gX(x)∫
Ωv (+∞, ω) dFW |X (ω | x) dx −
∫ x′′
x′ gX(x)∫
Ωv (∞, ω) dFW |X (ω | x) dx = +∞,
and thus that inequality (2) holds under this hypothetical solution.
16
plier associated with (2). The decision-maker then chooses a (x) to maximize
pointwise: ∫X
∫Ωu (a (x) , ω) dFW |X (ω | x) dGX (x) (3)
+γ
∫X
(gX (x)
∫Ωu (a (x) , ω) dFW |X (ω | x)− c− gX (x)
∫Ωu (a (x) , ω) dFW |X (ω | x)
)dx.
The assumptions made regarding the payoff function u (a, ω) (see Section 2)
ensure that this optimization problem is convex ∀x ∈ X , i.e., that the first-
order condition is necessary and sufficient ∀x.
In what follows, let K = x ∈ X | a(x) = a(x), that is, let K be the
set of realizations such that the players’ schedules of desired actions under
signals X and X intersect. Similarly, let K+ = x ∈ X | a(x) > a(x) and
K− = x ∈ X | a(x) < a(x).
Proposition 1 Suppose the decision-maker wishes the agent to acquire signal
X. The optimal decision-rule specifies an action a∗ (x) such that:
1. a∗ (x) = a (x) if x ∈ K,
2. a∗ (x) > a (x) if x ∈ K+, and
3. a∗ (x) < a (x) if x ∈ K−.
Moreover, when x /∈ K, the distortion | a∗ (x) − a (x) | is increasing in |
a(x)− a(x) | andgX(x)
gX(x).
Proof. See Appendix 1.
Proposition 1 states that if A has insufficient private incentives to gather
signal X, then DM has no choice but to introduce an upward (downward)
17
distortion when the realization x is such that A’s desired action is higher
(lower) in case he has exerted high effort (i.e., acquired X) than in case he
has shirked (i.e., acquired X). Even though introducing such distortions hurts
both DM and A in case the latter has exerted high effort, it hurts relatively
more so A in case he has shirked, because the implemented action is then
very distant from what he desires. Specifically, because payoffs are concave,
departing from the ex-post optimal action a (x) in the opposite direction than
a (x) involves a second-order payoff loss to DM and A in equilibrium, but a
first-order payoff loss to A in case he has shirked, which increases A’s incentives
to exert high effort. Building on this logic, the greater the distance between
what A desires in and out of equilibrium is (i.e., the greater | a(x)− a(x) | is),
the more it pays to introduce a distortion, and thus the greater the distance
between a∗ (x) and a (x) is. Similarly, the greater the likelihood ratiogX(x)
gX(x), the
more the distortion hurts A relatively more under signal X than under signal
X, and thus the greater the distortion. As in standard moral hazard problems,
DM behaves as if interpreting a high likelihood ratiogX(x)
gX(x)as indicative of A
having shirked, and conversely for low likelihood ratios. Finally, when the
realization is such that A desires the same action in and out of equilibrium,
introducing a distortion does not help incentivize information acquisition.
To summarize, to motivate the acquisition of accurate information, DM
has no choice but to make the final decision excessively sensitive to the agent’s
information. If the agent was to (unobservably) deviate and acquire signal X,
the final decision would be an overwhelmingly poor match with the posterior
beliefs induced by that signal.
18
In the next two corollaries, we impose more structure on the informational
environment.
Corollary 1 If the schedule a(x) is a rotation of the schedule a(x), then the
optimal decision-rule a∗ (x) is also a rotation such that:
(a∗ (x)− a (x)) (x− x) > 0 ∀x 6= x.
Moreover, if the schedule a(x) is a steeper rotation of the schedule a(x), and
the likelihood ratiogX(x)
gX(x)is non-decreasing in x as | x − x | increases, then
a∗ (x) is a steeper rotation of a(x) around x.
Proof. See Appendix 1.
Corollary 1 states that if A has insufficient private incentives to gather
signalX, and the schedule of desired actions induced by signalX is a rotation
of that induced by signal X, then DM designs a schedule of decisions which is
itself akin to a rotation around the realization x. In particular, the optimal de-
cision rule involves implementing actions higher than desired when observing
high realizations, and actions lower than desired when observing low realiza-
tions. The reason as to why introducing such distortions induces information
acquisition is as follows. Take a given high realization x (the logic is symmetric
for low realizations), and recall that high observations are those for which both
the decision-maker and the agent desire a higher action under signal X than
under signal X. By implementing an even higher action than desired under
signalX, the decision-maker hurts the agent under either signal, but relatively
19
more so under signal X. This is because players suffer increasingly more as
the implemented action becomes more distant from their ideal action. (Note
that the ex post optimal course of action a(x) is by construction farther away
from the agent’s desired action under signal X than that under X. However,
if the cost c is high enough, this is insufficient to motivate the acquisition of
accurate information.)
Finally, if the schedule a(x) is a steeper rotation of the schedule a(x),
and the likelihood ratiogX(x)
gX(x)is non-decreasing in x as | x− x | increases, the
magnitude of these distortions—that is, the distance between a∗(x) and a(x)—
becomes larger the more extreme the realization x is (i.e., the farther away x is
from x). It is optimal to do so because, in relative terms, realizations become
increasingly more likely under signal X than under signal X as x becomes
more distant from x, so that distortions become increasingly more costly to
the agent under signal X (in expectation).
Proposition 2 At the optimum, the decision-maker behaves as if signal X is
more informative than it is actually is if:
1. u (a, ω) is such that aX(x) = EW |X=x [ω],
2. the schedule a(x) is a steeper rotation of the schedule a(x) around x, and
3. x is such that EW |X=x [ω] = EW |X=x [ω] = EW [ω].
Proof. From Proposition 1, we know the optimal decision-rule is such that
a∗(x) > a(x) when x > x, and conversely when x < x. Moreover, if aX(x) =
20
EW |X=x [ω], then preferences must be of the form:
l (ω) + ωa− a2
2,
where l (ω) : Ω −→ R. When preferences are of this form, signalX is more in-
formative than signal X if and only if VX
[EW |X=x [ω]
]> VX
[EW |X=x [ω]
].
Because a∗(x) is a steeper rotation of a(x) around x, it follows that
VX [a∗(x)] > VX
[EW |X=x [ω]
]= VX [a(x)].
When preferences are such that players wish to match the action with the
expected value of W , and when signalsX and X are such that it is optimal to
put more weight on x when it is generated by X (i.e., when a(x) is a rotation
of a(x)), DM optimally puts excessive weight on the realization x to motivate
information acquisition. The resulting schedule of implemented actions is itself
a rotation of a(x), as if the provided information was generated by a signal
more informative than X.
3.1.1 An Explicit Solution
To gain further insights I now specify preferences and distributions so as
to compute explicit solutions. Suppose u (a, ω) = − (ω − a)2, W ∼ N (0, 1),
and let the agent’s signals be constructed as:
X = ρW +√
1− ρ2ε,
and
X = ρW +√
1− ρ2ε,
21
where 0 < ρ < ρ < 1, and where ε ∼ N (0, 1) and is independent of W .
One may compute that X | W = ω ∼ N (ρω, 1− ρ2) and W | X = x ∼
N (ρx, 1− ρ2), for X ∈X,X
. In addition, we have that a(x) = ρx and
a(x) = ρx. Finally, one may also verify that X ∼ N (0, 1) and X ∼ N (0, 1),
implying that the likelihood ratio is constant and equal to 1, thereby making
the analysis particularly amenable to comparative statics.29
Solving the optimization problem presented in Subsection 3.1 leads to the
following proposition.30
Proposition 3 If c ≤ 2ρ(ρ− ρ
), then a∗ (x) = ρx. Otherwise, the optimal
decision-rule is given by:
a∗ (x) =
(c
2(ρ− ρ
))x,where c
2(ρ−ρ)> ρ.
Proof. See Appendix 1.
Proposition 3 tells us that the decision-maker must introduce a distortion
in the decision rule whenever the cost c of acquiring signal X is sufficiently
high. In addition, these distortions are increasing in the cost c and decreasing
in the informational superiority of signal X compared to signal X, measured
by ρ− ρ. To see the latter comparative static, recall that the agent also cares
about the accuracy of the match between the action and the state of the world.
29However, we have that cov(W,X
)= ρ > cov (W,X) = ρ
30In solving for the solution, I relax the assumption, made otherwise in the paper, thatc is large enough that the incentive compatibility constraint always binds.
22
Corollary A threshold c exists such that it is strictly optimal to induce the
agent to acquire signal X if and only if c < c.
Proof. The decision-maker’s expected payoff when inducing the agent to
acquire signal X (and c > 2ρ(ρ− ρ
)) is equal to:
−1 + 2ρ
(c
2(ρ− ρ
))−( c
2(ρ− ρ
))2
,
while the decision-maker’s expected payoff when inducing the agent to instead
acquire signal X is equal to:
−1− ρ2 + 2ρ2. (4)
The statement in the corollary immediately follows from equating both ex-
pected payoffs and rearranging for c.
Rather intuitively, it is optimal for the decision-maker to induce the acqui-
sition of signalX only if its cost is not too high and/or the associated reduction
in uncertainty is sufficiently large.
3.2 The Two-Agent Model
We have learnt from the single-agent case that, to motivate information
acquisition, the agent has to be made excessively influential, in the sense that
the final decision must be made excessively sensitive to the information pro-
vided. In this section we investigate how this intuition survives in the presence
of multiple agents. For the sake of exposition, I assume that all three players
23
share (gross) payoff function u (a, ω).
Suppose now that there are two agents, agent 1 and agent 2, where
each may either acquire less accurate signal X, at zero cost, or more ac-
curate signal X, at cost c. Recall that realizations are conditionally in-
dependent. To save on notation, since players have identical gross pay-
offs, in the following, let aX1,X2 (x1, x2) = a (x1, x2 | X1, X2). For simplicity,
I restrict the analysis with two agents to environments whereby player i’s
schedule of desired actions induced by signal X is a steeper rotation of the
schedule induced by signal X around some realization xi (xj), for ∀Xej and
∀xj. Moreover, the realization xi (xj) is such that EW |X,Xej [ω | xi (xj) , xj] =
EW |X,Xej [ω | xi (xj) , xj] = EW |Xej [ω | xj]. Finally, I suppose the likelihood ra-
tiogX(x)
gX(x)is nondecreasing in x as x becomes distant from x, where x is defined
as EW |X [ω | x] = EW |X [ω | x] = a0.
In this section, we focus on the case of interest, that is, we look at the case
in which the decision-maker induces each agent to acquire his more accurate
signal. Extending the analysis to the other possible cases is immediate.
The decision-maker now chooses schedule a (x1, x2) to maximize pointwise
her expected payoff:
∫X
∫X
∫Ω
u (a(x1, x2), ω) dFW |X,X(ω | x1, x2)dGX(x)dGX(x), (5)
subject to inducing agent 1 to acquire signal X (with associated multiplier λ):
∫X
∫X
∫Ωu (a(x1, x2), ω) dFW |X,X(x1, x2)dGX(x)dGX(x)− c ≥ (6)
24
∫X
∫X
∫Ωu (a(x1, x2), ω) dFW |X,X(x1, x2)dGX(x)dGX(x),
and subject to induce agent 2 to acquire signal X (with associated multiplier
γ): ∫X
∫X
∫Ωu (a(x1, x2), ω) dFW |X,X(x1, x2)dGX(x)dGX(x)− c ≥ (7)
∫X
∫X
∫Ωu (a(x1, x2), ω) dFW |X,X(x1, x2)dGX(x)dGX(x).
The decision-maker needs to design a schedule a(x1, x2) which is on average
a sufficiently better match with the posterior beliefs jointly induced by the
pair of signalsX1 = X,X2 = X
than with those jointly induced by both
the pairsX1 = X,X2 = X
and
X1 = X,X2 = X
.31
Recall from the single-agent case that the signs of the distortions depended
on whether the realization x was “high” or “low”, that is, depended on whether
the desired action when observing x under signal X was higher or lower than
that under signal X. These considerations are still going to be useful in this
analysis with two agents, but some slight clarifications are first warranted. The
notion of “low” and “high” realization/posterior belief must now be extended
to account for the fact that posterior beliefs are now jointly induced by the
signals X1 and X2. In particular, a distinction must be made whether we are
at the stage of the game in which realizations are known, or prior to this.
Take agent 1 for instance. His posterior beliefs after having observed
a given realization x2, and anticipating that agent 2 chose X, are still in-
creasing in the realization x1 of the chosen signal. In addition, as before,
31Much like the single-agent case, a solution to this problem exists. In addition, by virtueof the restrictions imposed on payoff functions, this optimization problem is convex.
25
his posterior beliefs are still increasing more steeply with x1 under signal
X than under signal X. However, since the higher the realization x2 is,
the higher is agent 1’s desired action for any realization and chosen sig-
nal, we have that the ex-post rotation-point EW |X[ω | X = x2
]is increasing
in x2, i.e., dx1(x2)dx2
> 0, where x1(x2) is defined as the realization x1 such
that EW |X,X[ω |X = x1 (x2) ,X = x2
]= EW |X,X
[ω | X = x1 (x2) ,X = x2
]=
Eω|X[W |X = x2
].32
However, prior to knowing the realization of agent 2’s signal, the relevant
rotation-point is still EW [ω], because agent 1 expects agent 2’s signal to con-
firm his prior on average.33 This last remark is of importance since the decision
rule is designed under the prior belief. In the following, let x denote as before
the rotation-points computed under the prior belief.
Before proceeding to the characterization of the optimal decision rule, I
provide an example to gain further intuition.
Example. I extend the framework provided in Section 4.1.1 to the case with
two agents. Suppose W ∼ N (0, 1), and let the two signals at the disposal of
agent i = 1, 2 be constructed as:
X = ρW +√
1− ρ2εi
and
X = ρW +√
1− ρ2εi,
32Not surprisingly, by an identical reasoning, we also have dx2(x1)dx1
> 0.33Again, by an identical reasoning, it is also the case that x2 = EW [ω].
26
where 0 < ρ < ρ < 1, and where εi ∼ N(0, 1) and is independent of W .
Suppose further that all three players have payoff function u (a, ω) =
− (ω − a)2. We may then compute that each player’s desired action under
signals X1 =X and X2 =X, when observing x1 and x2, is equal to:
aX,X (x1, x2) =
(1− ρ2
)ρ
1− ρ4 (x1 + x2) , (8)
while that under signals X1 = X and X2 = X, when observing x1 and x1, is
equal to:
aX,X (x1, x2) =
(1− ρ2
)ρx1 +
(1− ρ2
)ρx2
1− ρ2ρ2(9)
and that under signals X1 = X and X2 = X, when observing x1 and x2, is
equal to:
aX,X (x1, x2) =
(1− ρ2
)ρx2 +
(1− ρ2
)ρx1
1− ρ2ρ2(10)
From these expressions, it follows that agent 1’s “rotation” point, conditional
on having observed x2, and anticipating that agent 2 has acquired X, is:
x1(x2) =
(ρ
1− ρ
)(ρ2 − ρ2
ρ− ρ
)x2
while agent 2’s “rotation-point” conditional on having observed x1, and antic-
ipating that agent 1 acquired X is:
x2(x1) =
(ρ
1− ρ
)(ρ2 − ρ2
ρ− ρ
)x1
where(
ρ1−ρ
)(ρ2−ρ2
ρ−ρ
)< 1.
Figure 2 depicts the classification of realizations as to whether they are
27
“high” or “low” once realizations are known. The lines x1(x2) and x−12 (x2)
represent, respectively, agent 1 and agent 2’s rotation-points given the other
player’s realization. For a given realization x2, if x1 > x−12 (x2) then x2 is low.
Similarly, for a given realization x2, x1 is high if and only if x1 > x1(x2). The
areas filled in grey (A and B) are those for which both agents’ realizations
are either low or high. These are all the combinations of x1 and x2 for which
both agents’ information is congruent. In particular, each agent’s realization
is “high” (“low”) if and only if both realizations are positive (negative) and
not too different one from another. The area above the grey surfaces, i.e., area
C, represents all the combinations of x1 and x2 such that x1 is high and x2 is
low. Finally, the area below the grey surfaces, i.e., area D, represents all the
combinations of x1 and x2 such that x1 is low and x2 is high.
[Insert Figure 2 here]
Returning to the general framework, it therefore follows from our dis-
cussion that when x1 and x2 are both high then we must have that
aX,X (x1, x2) , aX,X (x1, x2) < aX,X (x1, x2). Further, when x1 and x2 are both
low, then aX,X (x1, x2) < aX,X (x1, x2) , aX,X (x1, x2). If instead x1 is high and
x2 is low then aX,X (x1, x2) < aX,X (x1, x2) < aX,X (x1, x2), and similarly when
x1 is low and x2 is high.
Proposition 4 Suppose the decision-maker wishes each agent to acquire his
more accurate signal. If (8) and (9) bind, then the optimal decision rule is
To see this, suppose first that x1, x2 > x, implying that x2(x1) > x1(x2). It
then follows that x1− x1(x2) > x2− x2(x1) since x1−x2 > 0 > x1(x2)− x2(x1).
Suppose now that x1 > x > x2. We then have that x2(x1) > 0 > x1(x2), from
which it again follows that | x1 − x1(x2) |>| xx − x2(x1) |.35 Finally, the
inequality x1 − x1(x2) > x2 − x2(x1) implies (22) by virtue of (i) the fact that
the derivative of aX1,X2 (x1, x2) with respect to x1 is independent of x2, and
that with respect to x2 is independent of x1, (ii) the fact that gross preferences
are identical, and (iii) the fact that the signals at the disposal of A1 and A2
have identical conditional densities.
Let us interpret I as the marginal cost of moving away from a =
aX,X (x1, x2), and −(II+III) as the marginal benefit. Suppose to begin with
that λ ≥ γ. The function II is negative at a = aX,X(x1, x2) since x1 is high.
Conversely, because x2 is low, it follows that the function III is positive. The
function II+III is however negative at a = aX,X (x1, x2) because of the fact
that:
1. payoff functions being concave, inequality (22) implies that |∫Ω∂u(a,ω)∂a
dFW |X,X(ω | x1, x2) |> |∫
Ω∂u(a,ω)∂a
dFW |X,X(ω | x1, x2) | at
35It could also be the case that agent 1’s posterior belief is high while agent 2’s is low forx2 < x1 < 0. However, this would violate the condition that | x1 |>| x2 |.
52
a = aX,X (x1, x2),
2.gX(x1)
gX(x1)>
gX(x1)
gX(x1)since | x1 − x |>| x2 − x |, and
3. λ ≥ γ.
Finally, since I is equal to zero at aX,X (x1, x2), we have that the first-order
derivative at a = aX,X (x1, x2) is positive, and thus that it is strictly optimal
to set a∗ (x1, x2) > aX,X (x1, x2).
The argument so far relied on the assumption that λ ≥ γ. If the reverse
is true, it is clear that everything is as before as long as γ is not too large.
If for a given x1, γ is sufficiently large that II + III > 0 then it is optimal
to introduce a downward distortion. There thus exists a nonnegative function
r (λ, γ), such that it is optimal to set a∗ (x1, x2) > aX,X (x1, x2) if and only if
| x1− x |> r (λ, γ) | x2− x |, where it is immediate that ∂r(λ,γ)∂λ
< 0, ∂r(λ,γ)∂γ
> 0,
and where r (λ, γ) = 1 if γ = λ.
The proof for the case in which the agent 1’s posterior belief is high while
agent 2’s is low, but x2 is more extreme than x1 and x2 < x1 < 0 is symmetric
to that solved above in which 0 < x2 < x1: it is then optimal to introduce
a downward distortion (if and only if λ is not too large). Finally, the proof
for the case in which agent 1’s posterior belief is low while agent 2’s is high is
symmetric and thus left out.
To conclude, note that because (i) agents have identical payoff functions
and access to an identical set of signals, and (ii) both agents’ incentive com-
patibility constraint is binding, it must be the case that the function a∗ (x1, x2)
is such that λ = γ at the optimum, so that r (λ, γ) = 1. To see this, suppose
53
λ > γ. Then, the left-hand side of both agents’ incentive-compatibility con-
straint is equal one to another. However, the right-hand side cannot be equal
one to another: the right-hand side of (9) is lower than that of (10), which is
a contradiction with the assumption that both constraints are binding.
54
a(x)
a(x)
a∗(x)
X
R
Figure 1: The Optimal Solution when α = 1
55
X
X
x1(x2)
x−12 (x2)
A
B
C
D
Figure 2: The Classification of the Signals’ Realizations
56
aA(x)
aA(x)
aDM(x)
a∗(x)
X
R
Figure 3: The Optimal Decision Rule when α = 1 and b > 0