Signalling

• Principal is informed
tries to signal type through contract o er or through ex-ante choice of action

• Spence, M. (1973)
Workers are better informed about their ability
They choose education level as a signal of ability
This works if education is more costly for low ability types

Signalling games: Dynamic games of incomplete (asymmetric) information

• Incomplete information (asymmetric)

modeled using a move of Nature ( ctitious player)
objective probabilities on types of a player

• A Signalling game is a ”sender-receiver” game of incomplete information in two stages

First stage principal is privately informed about the type, sends a message to the agent
Agent acts accordingly in second stage

• A Perfect Bayesian Equilibrium consists of

a strategy of each player
AND beliefs of each player (about other players’ types, actions...)
SUCH THAT beliefs are consistent with strategies and prior beliefs + Bayes rule
AND strategies are optimal at every point in the game given beliefs (sequential rationality)

Beer Quiche Game

• Player 1 (Principal) is privately informed about his characteristic

He is either Tough (T) or Weak (W) and to signal his type he chooses between beer (B) or quiche
(Q)

• Player 2 (Agent) does not know the type but he observes the action of player 1
If he observes beer vs quiche he may deduce that its the tough or weak type
Based on his beliefs he chooses whether to ght (f) or retreat (r)

• Set of types of player 1: Θ = {W , T }

• Sets of possible actions:

A1 = {B , Q } and A2 = {f , r }

• Prior Probabilities of each type t denoted by p(t )

• Player 2’s (posterior) beliefs denoted by µ(t |a1), i.e. belief that player 1 is type t given that action a1
is observed
Example: µ(W|Q) would mean the belief that player 1 is weak after observing him having ordered
quiche
• We denote best response functions of player 1 by β1(a1|t ) and of player 2 by β2(a2|a1)
conditional probabilities of actions a1 and a2
Example: β1(q|T), probability that player 1 would choose quiche given that her type is tough
‣ 1-β1(b|T), prob that he would choose beer even though her type is tough
β2(f|q), probability that player 2 would ght after he observed player one choosing quiche

• Top node is nature, who assigns the

type
• Probability of weak is 0.1
• Second level is player 1 choosing
beer or quiche and nally player 2
reacting conditional on his
observations
• Dotted lines show same information
set: nodes that player 2 can’t
distinguish between as he doesn’t
know the type

Payo s player 1
• “Dominance” e ect
No ght is preferred to ghting
Shifts payo up by 2
• “Conditional” e ect
Depending on type pl1 prefers beer or quiche -> tough type prefers beer
Matching right type gives extra payo of +1

Payo s player 2
• “Dominance” e ect
Prefers to face a weak player 1
Shifts payo up +1
• “Conditional” e ect
Matching action of ghting or retreating to weak or strong player
Matching right action to right type gives extra payo of +1

PBE: Strategies and beliefs

Strategies
• Player 2’s pure strategies after observing each action set: (f , f ), (f , r ), (r , f ), (r , r )
(R,f) for example would be to ght after observing quiche and retreat after observing beer
‣ Dynamic game: a di erent action for each past action of other

• Player 1’s pure strategies (action for each type): (B , B ), (B , Q ), (Q , B ), (Q , Q ).

(B, Q) would be to choose beer if weak type and to choose quiche of tough type
‣ Incomplete information game: a di erent action for each player’s type

• Mixed strategies as pairs of conditional probabilities:

Player 1: (β1(B|T ), β1(B|W ))
Player 2: (β2(f|B ), β2(f|Q )).
Beliefs
• Theres no need for beliefs of player 1 as everything is known about player 2

Player 2
• We already know prior beliefs
prob that type is weak p(w) and prob that type is tough p(t)

• We need to de ne Player 2’s posterior beliefs on player 1’s type µ(t |B ) and µ(t |Q )
These are de ned by prior beliefs p(t ) combined with player 1’s equilibrium mixed strategy
(β1(B |W ), β1(B |T )) via Bayes rule:

P2(W1B)
—————-
P2B

Probability that type is week and beer has been chosen

Posterior belief of player two that the type
=
is weak after having observed beer Probability that beer has been chosen

• by chain rule of conditional probability P2(W1B) also equals P2(B|W)xP(w), which is the prob that beer has been
chosen given the type is week times the prior probabilty that the type is weak

• Denominator is total prob that beer has been chosen whcih could happen after 2 contingencies
(β1(B |T )p(T), whcih is the probability that the tough type chooses beer times probabilty that the
type is tough
(β1(B |W)p(W), prob that weak type chooses beer times prob that type is weak

• whenever we can apply bayes rule we should

• These are the strategies and beliefs for a Perfect Bayesian equilibrium, they must also be pro t
maximizing

Let’s start analysing the game:

Equilibria:

• with signalling games there are multiple equilibria

• We focus on pure strategies and categorise into two classes of equilibria:

Separating Equilibria:
‣ di erent types choose di erent actions
Pooling equilibria:
‣ both types choose the same action
‣ Both types go for quiche for example
•
Separating equilibria

• remember player 1 does not want to ght so any signal that could induce a ght would not be
chosen

Scenario 1
• Suppose T type plays B and W type plays Q, s1 = (B , Q )
Means strategy of player 1 is to play beer if type is tough and quiche if type is weak

• Given s1, Bayes rule = player 2’s posterior beliefs must be µ(T |B ) = 1, µ(W |Q ) = 1
Prob that type is tough after beer is 1, prob that weak after quiche also 1

‣ Best response of player 2: choose r if B and choose f if Q, s2 = (r , f )

‣ Best response of player 1: no incentive to deviate for T type but the W type can deviate to
B and get -1 instead of -2, hence s1 not rational and new strategy will be (B, B)

‣ Hence not an Perfect Bayesian Equilibrium as Sequential Rationality of W type violated

Scenario 2

• T type plays Q and W type plays B, s1 = (Q, B )

• Bayes Rule = player 2’s beliefs µ(T |Q ) = 1, µ(W |B ) = 1

• Best response of player 2: choose f if B and choose r if Q, s2 = (f , r )

• Best response of player 1: no incentive to deviate for the T type but the W type would prefer to
deviate to Q

‣ So, not a PBE as Sequential Rationality of W type violated

NO SEPERATING EQUILIBRIUM IN THIS GAME

POOLING EQUILIBRIUM

• Suppose both play Beer S1= (B,B)

• Bayes Rule = µ(T |B ) = 0.9 and µ(W |B ) = 0.1, can be shown using formula
Posterior beliefs will be equal to prior beliefs

• Seq. Rationality of 2 following B is to play r given µ

Payo (f|B) = .9 x (-1) + .1 x (1)
Payo (r|B) = 0
‣ Will retreat

• Seq. Rationality of 2 following Q is to play ?? (No beliefs assigned by Bayes rule since o the
equilibrium path)
• Sequential Rationality of Player 1, W type implies that player 2 must play f when he observes Q
Player 2 must ght after Q
To support this, the corresponding beliefs of player 2 must satisfy: µ(W |Q ) ≥ 1/2

‣ This is a pooling PBE:

β1(B |T ) = β1(B |W ) = 1,
β2(r|B ) = β2(f |Q ) = 1,
µ(W|B ) = 0.1 µ(W|Q ) ≥ 1/2

• other pooling equilibrium:

where both types choose Quiche
This is a pooling PBE: β1(Q |T ) = β1(Q |W ) = 1, β2(r |Q ) = 1, µ(W|Q ) = 0.1 SR of type T of 1
implies that β2(f |B ) = 1 so µ(W |B ) ≥ 1/2
‣ Continuum of PBE pooling

Formal De nition

• a Perfect Bayesian Equilibrium is a collection of strategies and beliefs (β1, β2, µ) that satis es

1) Sequential Rationality:
Strategies maximize expected payo s given beliefs:

2) Bayes Rule:
for any a1 such that β1(a1|t ) > 0,

A doubt
• In second pooling equilibrium with (Q, Q ) beliefs feel “strange”

• Suppose o the equilibrium is played meaning player 1 chooses B

player 2 sees B , then notice that the W type is getting a maximum of -1 while in equilibrium he
gets 0
T type however can gain from B, since he can get 0 if 2 plays r .
So player 2 should put 0 probability on type W after B.

• But if µ(W |B ) = 0, the player 2 plays r after B and the equilibrium collapses as T chooses B... (recall
 the requirement µ(W |B ) ≥ 1/2)
‣ So intuitively this should not be an equilibrium!
‣ intuitive criterion

Intuitive Criterion

• A player’s non-equilibrium action is equilibrium-dominated if any of its possible outcomes leads to

a STRICTLY LOWER payo than the equilibrium payo

• If, under some ongoing equilibrium, a non-equilibrium action is observed which is equilibrium-
dominated for some types but not others, then beliefs must put zero probability on the former set of
types

In second pooling equilibrium:

• Player 1 W type playing B is equilibrium dominated, while player 1 T type playing B is not.
Thus, by IC we have µ(W |B ) = 0 and second pooling equilibrium collapses.

First pooling equilibrium :

• When o path action A1= Q is taken:
Q is equilibrium dominated for type T since he gets 0 in equilibrium, and his best payo with Q
is -1
Q is not equilibrium dominated for type W
‣ Hence reasonable beliefs must give 0 probability to T playing Q
‣ Thus, µ(W |Q ) = 1, which is consistent with 2 playing f
So this pooling equilibrium satis es IC.

Finally

Unique pure strategy equilibrium:

Pooling on B , player 2 plays r when he observes B and f when he observes Q . Beliefs µ(T |B ) = .9 and
µ(T |Q ) = 0.

Satis es IC