URL (Slide and paper): ux.nu/BYfFi
Persistence and Snap Decision Making: Inefficient Decisions by a Reputation-Concerned Expert Tomoya Tajika IER, Hitotsubashi University
June 22, 2018 AMES @ Sogang University
1 / 26
URL (Slide and paper): ux.nu/BYfFi
Introduction
• Behaving consistently is widely observed and valued phenomenon. • E.g., • Politicians who behave inconsistently are criticized. • Economists who predict economic trends that oppose their previous ones would not be trusted.
• Consistent behavior may lead to inefficient decisions.
• Consider a politician who decides to invest a public project, which is likely to make a profit. • After a while, he finds that the project is likely to make a loss. • Behaving consistency prevents him from withdrawing from the investment ○ inefficient decisions
2 / 26
URL (Slide and paper): ux.nu/BYfFi
• Given that persons tend to behave consistently even it is inefficient, why they do not wait further information? ∵ Waiting information enables better choice without behaving inconsistently. ○ This study shows that “reputation concern” forces a person to “behave consistently” and make a such “snap decision.”
3 / 26
URL (Slide and paper): ux.nu/BYfFi
Study overview
• There is unknown state ∈ {x , y }. • Player; a reputation-concerned expert and an evaluator. • The expert has two opportunities to recommend a choice from {x , y }. • If the last recommendation matches the realized state, the expert is (monetarily) rewarded. • In each opportunity, he receives signal ∈ {x , y }, whose accuracy is his ability. • Accuracy grows in the second opportunity. • An evaluator assesses the expert’s ability by observing the recommendations and realized state.
4 / 26
URL (Slide and paper): ux.nu/BYfFi
KEY High-ability expert is less likely to receive inconsistent signal. ○ The expert has an incentive to pretend to have high-ability by behaving consistently. • This study shows that
5 / 26
• Truthful recommendation is not an equilibrium. • Consistent recommendation is an equilibrium.
• Furthermore, we also consider the case that the expert can remain silent at the first opportunity.
○ Breaking a silence can be observed in an equilibrium.
KEY Assessment differs whether the expert remains silence or not.
• Remain silence: the expert assessed by the second signal. ○ assessment becomes severe. • Break silence: it leads to consistent recommendation ○ the expert assessed by the first signal. ○ assessment becomes mild (c.f. self-handicapping).
○
Incentive to break silence arises.
URL (Slide and paper): ux.nu/BYfFi
Literature
6 / 26
• In social psychology, many studies find the evidences and discuss consistent behavior (Cialdini, 2006). • Falk and Zimmerman (2016) is the closest study. The differences are • knowledge/set of abilities • observation on the realized state
• Ferreira and Rezende (2007): information disclosure as a commitment for persistent behavior • Reputational herding and anti-herding: Levy (2004), Sabourian and Sibert (2009). • Belief persistence: Rabin and Schrag (1999). • Self-handicapping: Tirole (2002).
URL (Slide and paper): ux.nu/BYfFi
7 / 26
Setup
• Players: expert and evaluator • The state of the world ω ∈ {x , y } • Expert recommends a choice from R1 = R2 = {x , y }. • Expert has two chances to recommend. • If the last recommendation matches to the realized state, expert is rewarded K . • Before each chance, expert receives a signal in S1 = S2 = {x , y } with accuracy θt = Prob(st = ω), t ∈ {1, 2} . • Assume
θ2 1−θ2
1 = (1 + α) 1−θ . Equiv., θ2 =
θ
1
(1+α)θ1 . 1+αθ1
URL (Slide and paper): ux.nu/BYfFi
8 / 26
• θ1 ∈ D ⊆ [1/2, 1] is called ability. The expert knows θ1 . • Assume |D | ⩾ 2, inf D ∈ D and sup D ∈ D . • Let f : D → R++ be the density of θ1 (if D is discrete, f (θ) = Prob(θ1 = θ)). • By observing recommendations and state, Evaluator computes the belief about θ1 . • The expert’s vNM utility is I (r2 = ω)K + Eβ [θ1 | r1 , r2 , ω], risk neutral.
URL (Slide and paper): ux.nu/BYfFi
1st opportunity Expert observes signal s 1 ∈ S1
Expert recommends r1 ∈ R 1
9 / 26
2nd opportunity
Expert observes signal s 2 ∈ S2
Expert recommends r2 ∈ R 2
Evaluator observes state ω∈X
Evaluator computes reputation Eβ [θ1 | r1 , r2 , ω]
Figure: Timeline of the model
DEFINITION 1. Recommendation r is a mapping such that r = (r1 , r2 ), r1 : S1 → R1 and r2 : R1 × S1 × S2 → R2 .
URL (Slide and paper): ux.nu/BYfFi
Truthful recommendation
10 / 26
DEFINITION 2. Recommendation r is truthful if r1 (s1 ) = s1 and r2 (r1 , s1 , s2 ) = s2 . ¯ > 0 such PROPOSITION 1. For each K > 0, there exists α ¯ that for each α < α, the truthful recommendation is not a PBE.
URL (Slide and paper): ux.nu/BYfFi
11 / 26
Detail for Proposition 1. Let θTruth be the ex post expectation of θ1 . r r ω 1 2
∫
θTruth = θTruth = ∫ xxx yyy = θTruth θTruth xxy yyx
∫
θ2 θ2 f (θ)d θ θθ2 f (θ)d θ
∫
=∫
(1−θ)2
θ 1+αθ f (θ)d θ = ∫ (1−θ)2 f (θ)d θ 1+αθ
∫
= θTruth = θTruth = θTruth = ∫ θTruth xyx xyy yxx yxy
θ3 f (θ)d θ 1+αθ θ2 f (θ)d θ 1+αθ
θ2 (1−θ) f (θ)d θ 1+αθ θ(1−θ) f (θ)d θ 1+αθ
URL (Slide and paper): ux.nu/BYfFi
12 / 26
Detail for Proposition 1.—Cont’d • Suppose that Expert receives s1 = x in the 1st opp. and s2 = y in the 2nd opp. pxy = Prob(ω = x | s1 = x , s2 = y ). • To the TR become PBE, the following is necessary
Truth Truth Truth Truth pxy (θxxx + K ) + (1 − pxy )θxxy ⩽ pxy θxyx + (1 − pxy )(θxyy +K)
|
{z
}
utility of r1 = r2 = x
⇐⇒
1
1+α
=
pxy
1 − pxy
|
⩽
{z
utility of r1 = x , r2 = y
θTruth − θTruth +K xyy xxy Truth − θTruth + K θxxx xyx
}
URL (Slide and paper): ux.nu/BYfFi
Consistent recommendation
13 / 26
DEFINITION 3. Recommendation r is consistent if r1 (s1 ) = s1 and r2 (r1 , s1 , s2 ) = s1 . • Assume that if r1 = r2 , the evaluator believes that r2 = s2 . ¯ > 0 such PROPOSITION 2. For each K > 0, there exists α ¯ that for each α < α, a consistent recommendation is a PBE.
URL (Slide and paper): ux.nu/BYfFi
14 / 26
Detail for Proposition 2. Let θCons be the ex post expectation of θ1 . r r ω 1 2
θCons xxx
=
θCons yyy
∫
= ∫ ∫
θ2 f (θ)d θ θf (θ)d θ
=
m2 μ
θ(1 − θ)f (θ)d θ μ − m2 θCons = θCons = ∫ = xxy yyx 1− μ (1 − θ)f (θ)d θ
θCons = θCons = θCons = θCons xyx xyy yxx yxy ∫ θ2 (1−θ) ∫ f (θ)d θ θ2 (1 − θ2 )f (θ)d θ 1+αθ = ∫ = ∫ θ(1−θ) θ(1 − θ2 )f (θ)d θ f (θ)d θ 1+αθ
URL (Slide and paper): ux.nu/BYfFi
Detail for Proposition 2.—Cont’d • Suppose that Expert receives s1 = x in the 1st opp. and s2 = y in the 2nd opp. • To the CR become PBE, the following is necessary (and sufficient) 1
1+α
=
pxy
1 − pxy
⩾
θCons − θCons +K xyy xxy Cons − θCons + K θxxx xyx
15 / 26
URL (Slide and paper): ux.nu/BYfFi
16 / 26
Example
Assume D = {1/2, 3/4, 1}, f (1/2) = a , f (1) = b ,
f (3/4) = 1 − a − b . Let V (α) := W (α) =
Truth −θTruth +K θxyy xxy Truth −θTruth +K θxxx xyx
.
Cons −θCons +K θxyy xxy
Cons −θCons +K θxxx xyx
and
1.0
0.9
0.8
W(α)
0.7
V(α)
0.6
1 1+α
0.5 0
0.2
0.4
0.6
0.8
1.0
URL (Slide and paper): ux.nu/BYfFi
Right of silence
17 / 26
• Introducing a right of silence in the first opportunity. ○ R1 = {x , y , ∅}, R2 = {x , y }.
• ∅ means keeping silence.
DEFINITION 4. Recommendation r is the waiting strategy if r1 (s1 ) = ∅ and r2 (r1 , s1 , s2 ) = s2 . • Waiting strategy enables to conceal receiving inconsistent signals. • (Off-path-belief) Assume if r1 (s1 ) ̸= ∅, the evaluator believes that r1 (s1 ) = r2 (r1 , s1 , s2 ) = s1 , but not update about θ1 .
URL (Slide and paper): ux.nu/BYfFi
Pooling equilibrium
18 / 26
LEMMA 1. Suppose that sup D = 1 and α is sufficiently small. Suppose also that the evaluator does not update her belief about whether the expert remains silent or not. Then, the waiting strategy fails to be a PBE. • (Intuition) The evaluator’s assessment about θ1 is based on • θ2 = 1+αθ 1 if r1 = ∅ 1 • θ1 if r1 ̸= ∅ ○ Since θ2 > θ1 , the assessment becomes severe if r1 = ∅. (c.f. Self-handicapping). (1+α)θ
URL (Slide and paper): ux.nu/BYfFi
• (Re)consider consistent recommendation, that is r1 (s1 ) ̸= ∅. This is referred to as a snap decision. • (Off-path-belief) Assume if r1 (s1 ) = ∅, the evaluator does not update about θ. LEMMA 2. Suppose that inf D = 1/2 and K > 3. Suppose also that the evaluator does not update her belief about whether the expert ¯ remains silent or not. Then, there exists α ¯ the consistent such that for each α < α, recommendation fails to be a PBE.
19 / 26
URL (Slide and paper): ux.nu/BYfFi
20 / 26
PROPOSITION 3.
(i) The waiting strategy is a PBE. The out-of-equilibrium belief places probability 1 on θ = inf D . (ii) Suppose that α is sufficiently small. Then, the consistent recommendation is a PBE. The out-of-equilibrium belief places probability 1 on θ = inf D .
The out-of-equilibrium belief is natural in the following sense.
(i) In the equilibrium, h (θ) is minimized at θ = inf D . This type is most likely to break silence. (ii) In the equilibrium, h (θ) is maximized at θ = inf D . This type is most likely to keep silence.
URL (Slide and paper): ux.nu/BYfFi
21 / 26
Keeping or breaking silence as a signaling
• Let r1 : S1 × D → R1 . • Let β be the belief system. • Let h (θ : r , β) = EU (r1 = ∅ | s1 , θ) − EU (r1 = s1 | s1 , θ). LEMMA 3. With given r , h (θ : r , β) is a concave quadratic function of θ. ○ (With sufficiently small α), Possible separation is • If θ ∈ / [θ∗ , θ∗ ]: r1 = s1 and r2 = r1 when r1 = s1 . • If θ ∈ [θ∗ , θ∗ ]: r1 = ∅ and r2 = s2 . • The evaluator believes that if r1 = ∅, the expert employs the waiting strategy and if r1 ̸= ∅, the expert employs the consistent recommendation.
URL (Slide and paper): ux.nu/BYfFi
h (θ : r , μ)
r1 = s1 1 2
22 / 26
r1 = s1
r1 = ∅
θ∗
θ∗
1
θ
Figure: A tripartite equilibrium
h (θ : r , μ)
θ∗ =
1 r 2 1
=∅
r1 = s1
θ∗
1
r1 = s1 θ
1 2
h (θ : r , μ)
θ∗
Figure: Bipartite equilibria
θ ∗ r1 = ∅ θ = 1
URL (Slide and paper): ux.nu/BYfFi
23 / 26
LEMMA 4. There is no bipartite separation equilibrium associated with ability such that θ∗ = 1. • Intuition: When θ = 1, the expert only concerns about his reputation. The benefit of breaking silence is only reputation. The advantage vanishes if θ = 1 expert keeps silent. • Implication: In each separating equilibrium, sufficiently high ability expert keeps silence and make a snap decision. ○ Precision of decision is not monotone in ability. • Separating equilibrium may not exist. PROPOSITION 4. Suppose that K > 3. Then, there exists ¯ such that for each α < α, ¯ no separating α equilibrium is associated with the expert’s ability.
URL (Slide and paper): ux.nu/BYfFi
Extension 1: Knowledge about θ1
24 / 26
• Suppose that the expert does not know his ability θ1 but receive signal τ about θ1
COROLLARY 1. Suppose that the expert receives a signal about his ability. Then, for each K > 0, there ¯ > 0 such that for each α < α, ¯ the exists α truthful recommendation is not a PBE. COROLLARY 2. Suppose that the expert receives a signal about his ability. Then, for each K > 0, there ¯ > 0 such that for each α < α, ¯ a exists α consistent recommendation is a PBE. PROPOSITION 5. Suppose that the expert receives no information about his ability in the interim stage. Then, if K > 0, for a sufficiently small α, the waiting strategy is a PBE.
URL (Slide and paper): ux.nu/BYfFi
Extension 2: Different α
25 / 26
• Assume that α depends on θ. • Assume also that
θα ′ (θ) α(θ)
∈ (−1, 0).
COROLLARY 3. Suppose that the expert receives a signal about his ability. Then, for each K > 0, there ¯ > 0 such that for each function α exists α ¯ a consistent that satisfies sup α(θ) < α, recommendation is a PBE.
URL (Slide and paper): ux.nu/BYfFi
Conclusion
26 / 26
• We study the incentive of behaving consistently. • Recommending different choices hurts the expert’s reputation, which forces her to behave consistently. • Right of silence may not be a remedy for the consistent behavior. • Breaking a silence can work as a self-handicapping, which is enabled by the consistent behavior.