DIVISION OF THE HUMANITIES AND SOCIAL SCIENCES
CALIFORNIA INSTITUTE OF TECHNOLOGY
PASADENA, CALIFORNIA 91125
Cursed Sequential Equilibrium
Meng
-Jhang Fong
California Institute of Technology
Po
-Hsuan Lin
California Institute of
Technology
Thomas R. Palfrey
California Institute of Technology
SOCIAL SCIENCE WORKING PAPER 1465
April
2023
Cursed Sequential Equilibrium
*
Meng-Jhang Fong
Po-Hsuan Lin
Thomas R. Palfrey
§
April 11, 2023
Abstract
This paper develops a framework to extend the strategic form analysis of cursed
equilibrium (CE) developed by Eyster and Rabin (2005) to multi-stage games. The
approach uses behavioral strategies rather than normal form mixed strategies, and
imposes sequential rationality. We define cursed sequential equilibrium (CSE) and
compare it to sequential equilibrium and standard normal-form CE. We provide a
general characterization of CSE and establish its properties. We apply CSE to five
applications in economics and political science. These applications illustrate a wide
range of differences between CSE and Bayesian Nash equilibrium or CE: in signaling
games; games with preplay communication; reputation building; sequential voting;
and the dirty faces game where higher order beliefs play a key role. A common theme
in several of these applications is showing how and why CSE implies systematically
different behavior than Bayesian Nash equilibrium in dynamic games of incomplete
information with private values, while CE coincides with Bayesian Nash equilibrium
for such games.
JEL Classification Numbers: C72, D83
Keywords: Multi-stage Games, Private Information, Cursed Equilibrium, Learning
*
Grants from the National Science Foundation (SES-0617820) and the Gordon and Betty Moore Foun-
dation (1158) supported this research. We are grateful to Shengwu Li and Shani Cohen for recent corre-
spondence that helped to clarify the differences between the CSE and SCE approaches to the generalization
of cursed equilibrium for dynamic games. We thank participants of the Caltech Theory Seminar and Colin
Camerer for comments and also thank Matthew Rabin for earlier discussions on the subject during his visit
at Caltech as a Moore Distinguished Scholar.
Division of the Humanities and Social Sciences, California Institute of Technology, Pasadena, CA 91125
USA. mjfong@caltech.edu
Division of the Humanities and Social Sciences, California Institute of Technology, Pasadena, CA 91125
USA. plin@caltech.edu
§
Corresponding Author: Division of the Humanities and Social Sciences, California Institute of Technol-
ogy, Pasadena, California 91125 USA. trp@hss.caltech.edu. Fax: +16263958967 Phone: +16263954088
1 Introduction
Cursed equilibrium (CE) proposed by Eyster and Rabin (2005) is a leading behavioral equi-
librium concept that was developed to explain the “winner’s curse” and related anomalies
in applied game theory. The basic idea behind CE is that individuals do not fully take
account of the dependence of other players’ strategic actions on private information. Cursed
behavior of this sort has been detected in a variety of contexts. Capen et al. (1971) first
noted that in oil-lease auctions, “the winner tends to be the bidder who most overestimates
the reserves potential” (Capen et al. (1971), p. 641). Since then, this observation of overbid-
ding relative to the Bayesian equilibrium benchmark, which can result in large losses for the
winning bidder, has been widely documented in laboratory auction experiments (Bazerman
and Samuelson, 1983; Kagel and Levin, 1986; Kagel et al., 1989; Forsythe et al., 1989; Dyer
et al., 1989; Lind and Plott, 1991; Kagel and Levin, 2009; Ivanov et al., 2010; Camerer et al.,
2016). In addition, the neglect of the connection between the opponents’ actions and private
information is also found in non-auction environments, such as bilateral bargaining games
(Samuelson and Bazerman, 1985; Holt and Sherman, 1994; Carrillo and Palfrey, 2009, 2011),
zero-sum betting games with asymmetric information (Rogers et al., 2009; Søvik, 2009), and
voting and jury decisions (Guarnaschelli et al., 2000).
While CE provides a tractable alternative to Bayesian Nash equilibrium and can explain
some anomalous behavior in games with a winner’s-curse structure, a significant limitation is
that it is only developed as a strategic form concept for simultaneous-move Bayesian games.
Thus, when applying the standard CE to dynamic games, the CE analysis is carried out on
the strategic form representation of the game, implying that CE cannot distinguish behavior
across dynamic games that differ in their timing of moves but have the same strategic form.
That is, players are assumed to choose type-dependent contingent strategies simultaneously
and
not
update their beliefs as the history of play unfolds. A further limitation implied
by the strategic form approach is that CE and standard Bayesian Nash equilibrium make
identical predictions in games with a private-values information structure (Eyster and Rabin
(2005), Proposition 2). In this paper we extend the CE in a simple and natural way to
multi-stage games of incomplete information. We call the new equilibrium concept
Cursed
Sequential Equilibrium
(CSE).
In Section 2, we present the framework and our extension of cursed equilibrium to dy-
namic games. We consider the framework of multi-stage games with observed actions,
introduced by Fudenberg and Tirole (1991b), where players’ private information is repre-
sented by types, with the assumption that the set of available actions is independent of their
types at each public history. Our new solution concept is in the same spirit of the cursed
1
equilibrium—in our model, at each stage, players will (partially) neglect the dependence of
the other players’ behavioral strategies on their types, by placing some weight on the incor-
rect belief that all types adopt the average
behavioral
strategy. Specifically, at each public
history, this corresponds to the average distribution of actions
given the current belief about
others’ types at that stage
. Therefore, as players update their beliefs about others’ private
information via Bayes’ rule, but with incorrect beliefs about the other players’ behavioral
strategies, in later stages this can lead them to have
incorrect
beliefs about the other players’
average distribution of actions.
Following Eyster and Rabin (2005)’s notion of cursedness, we parameterize the model by
a single parameter
χ
∈
[0
,
1] which captures the degree of cursedness and define fully cursed
(
χ
= 1) CSE analogously to fully cursed (
χ
= 1) CE. Recall that in a fully cursed (
χ
= 1) CE,
each type of each player chooses a best reply to expected (cursed) equilibrium distribution
of other players’ actions, averaged over the type-conditional strategies of the other players,
with this average distribution calculated using the prior belief on types. Loosely speaking,
a player best responds to the average CE strategy of the others. In a
χ
-CE, players are
only partially cursed, in the sense that each player best responds to a
χ
-weighted linear
combination of the
average
χ
-CE strategy of the others and the
true
(type-dependent)
χ
-CE
strategy of the others.
The extension of this definition to multi-stage games with observed actions is different
from
χ
-CE in two essential ways: (1) the game is analyzed with behavioral strategies; and
(2) we impose sequential rationality and Bayesian updating. In a fully cursed (
χ
= 1) CSE,
(1) implies at every stage
t
and each public history at
t
, each type of each player
i
chooses a
best reply to the expected (cursed) equilibrium distribution of other players’ stage-
t
actions,
averaged over the type-conditional stage-
t
behavioral strategies of other players, with this
average distribution calculated using
i
’s current belief about types at stage
t
. That is, player
i
best responds to the average stage-
t
CSE strategy of others. Moreover, (2) requires that
each player’s belief at each public history is derived by Bayes’ rule wherever possible, and
best replies are with respect to the continuation values computed by using the fully cursed
beliefs about the behavioral strategies of the other players in current and future stages.
A
χ
-CSE, for
χ <
1, is then defined in analogously to
χ
-CE, except for using a
χ
-weighted
linear combination of the
average
χ
-CSE
behavioral strategies
of others and the
true
(type-
dependent)
χ
-CSE
behavioral strategies
of others. Thus, similar to the fully cursed CE, in
a fully cursed (
χ
= 1) CSE, each player believes other players’ actions at each history are
independent
of their private information. On the other hand,
χ
= 0 corresponds to the
standard sequential equilibrium where players have correct perceptions about other players’
2
behavioral strategies and are able to make correct Bayesian inferences.
1
After defining the equilibrium concept, in Section 3 we explore some general properties of
the model. We first prove the existence of a cursed sequential equilibrium in Proposition 1.
Intuitively speaking, CSE mirrors the standard sequential equilibrium. The only difference is
that players have incorrect beliefs about the other players’ behavioral strategies at each stage
since they fail to fully account for the correlation between others’ actions and types at every
history. We prove in Proposition 2 that the set of CSE is upper hemi-continuous with respect
to
χ
. Consequently, every limit point of a sequence of
χ
-CSE points as
χ
converges to 0 is a
sequential equilibrium. This result bridges our behavioral solution concept with the standard
equilibrium theory. Finally, we also show in Proposition 4 that
χ
-CSE is equivalent to
χ
-CE
for one-stage games, demonstrating the connection between the two behavioral solutions.
In multi-stage games, cursed beliefs about behavioral strategies will distort the evolution
of a player’s beliefs about the other players’ types. As shown in Proposition 3, a direct
consequence of the distortion is that in
χ
-CSE players tend to update their beliefs about
others’ types too passively. That is, there is some persistence in beliefs in the sense that
at each stage
t
, each
χ
-cursed player’s belief about any type profile is at least
χ
times the
belief about that type profile at stage
t
−
1. Among other things, this implies that if the
prior
belief about the types is full support and
χ >
0, the full support property will persist
at all histories, and players will (possibly incorrectly) believe every profile of others’ types is
possible at every history.
This dampened updating property plays an important role in our framework. Not only
does it contribute to the difference between CSE and the standard CE through the updating
process, but it also implies additional restrictions on off-path beliefs. The effect of dampened
updating is starkly illustrated in the pooling equilibria of signaling games where every type
of sender behaves the same everywhere. In this case, Proposition 5 shows if an assessment
associated with a pooling equilibrium is a
χ
-CSE, then it also a
χ
′
-CSE for all
χ
′
≤
χ
, but it
is not necessarily a pooling equilibrium for all
χ
′
> χ
. This contrasts with one of the main
results about CE, that if a pooling equilibrium is a
χ
-CE for some
χ
, then it is a
χ
′
-CE for
all
χ
′
∈
[0
,
1] (Eyster and Rabin (2005), Proposition 3).
This suggests that perhaps the dampened updating property is an equilibrium selection
device that eliminates some pooling equilibrium, but actually this is
not
a general property.
As we demonstrate later, the
χ
-CE and
χ
-CSE sets can be non-overlapping, which we il-
1
For the off-path histories, similar to the idea of Kreps and Wilson (1982), we impose the
χ
-consistency
requirement (see Definition 2) so the assessment is approachable by a sequence of totally mixed behavioral
strategies. The only difference is that players’ beliefs are incorrectly updated by assuming others play the
χ
-cursed behavioral strategies. Hence, in our approach if
χ
= 0, a CSE is a sequential equilibrium.
3
lustrate with a variety of applications. The intuition is that in CSE, players generally do
not have correct beliefs about the opponents’ average behavioral strategies. The pooling
equilibrium is just a special case where players have correct beliefs.
In Section 4 we explore the implications of cursed sequential equilibrium with five ap-
plications in economics and political science. Section 4.1 analyzes the
χ
-CSE of signaling
games. Besides studying the theoretical properties of pooling
χ
-CSE, we also analyze two
simple signaling games that were studied in a laboratory experiment (Brandts and Holt,
1993). We show how varying the degree of cursedness can change the set of
χ
-CSE in these
two signaling games in ways that are consistent with the reported experimental findings.
Next, we turn to the exploration of how sequentially cursed reasoning can influence strategic
communication. To this end, we analyze the
χ
-CSE for a public goods game with communi-
cation (Palfrey and Rosenthal, 1991; Palfrey et al., 2017) in Section 4.2, finding that
χ
-CSE
predicts there will be less effective communication when players are more cursed.
Next, in Section 4.3 we apply
χ
-CSE to the centipede game studied experimentally by
McKelvey and Palfrey (1992) where one of the players believes the other player might be
an “altruistic” player who always passes. This is a simple reputation-building game, where
selfish types can gain by imitating altruistic types in early stages of the game. The public
goods application and the centipede game are both private-values environments, so these
two applications clearly demonstrate how CSE departs from CE and the Bayesian Nash
equilibrium, and shows the interplay between sequentially cursed reasoning and the learning
of types in private-value models.
In strategic voting applications, conditioning on “pivotality”—the event where your vote
determines the final outcome—plays a crucial role in understanding equilibrium voting be-
havior. To illustrate how cursedness distorts the pivotal reasoning, in Section 4.4 we study
the three-voter two-stage agenda voting game introduced by Ordeshook and Palfrey (1988).
Since this is a private value game, the predictions of the
χ
-CE and the Bayesian Nash equilib-
rium coincide for all
χ
. That is, cursed equilibrium predicts no matter how cursed the voters
are, they are able to correctly perform pivotal reasoning. On the contrary, our CSE predicts
that cursedness will make the voters less likely to vote strategically. This is consistent with
the empirical evidence about the prevalence of sincere voting over sequential agendas when
inexperienced voters have incomplete information about other voters’ preferences (Levine
and Plott, 1977; Plott and Levine, 1978; Eckel and Holt, 1989).
Finally, in Section 4.5 we study the relationship between cursedness and epistemic rea-
soning by considering the two-person dirty faces game previously studied by Weber (2001)
and Bayer and Chan (2007). In this game,
χ
-CSE predicts cursed players are, to some extent,
4
playing a “coordination” game where they coordinate on a specific learning speed about their
face types. Therefore, from the perspective of CSE, the non-equilibrium behavior observed
in experiments can be interpreted as possibly due to a coordination failure resulting from
cognitive limitations.
The cursed sequential equilibrium extends the concept of cursed equilibrium from static
Bayesian games to multi-stage games with observed actions. This generalization preserves
the spirit of the original cursed equilibrium in a simple and tractable way, and provides
additional insights about the effect of cursedness in dynamic games. A contemporaneous
working paper by Cohen and Li (2023) is closely related to our paper. That paper adopts an
approach based on the coarsening of information sets to define sequential cursed equilibrium
(SCE) for extensive form games with perfect recall. The SCE model captures a different
kind of cursedness
2
that arises if a player neglects the dependence of other players’ unob-
served (i.e., either future or simultaneous) actions on the history of play in the game, which
is different from the dependence of other players’ actions on their type (as in CE and CSE).
In the terminology of Eyster and Rabin (2005) (p. 1665), the cursedness is with respect to
endogenous information, i.e., what players observe about the path of play. The idea is to
treat the unobserved actions of other players in response to different histories (endogenous
information) similarly to how cursed equilibrium treats players’ types. A two-parameter
model of partial cursedness is developed, and a series of examples demonstrate that for
plausible parameter values, the model is consistent with some experimental findings related
to the failure of subjects to fully take account of unobserved hypothetical events, whereas
behavior is “more rational” if subjects make decisions after directly observing such events.
At a more conceptual level, our paper is related to several other behavioral solution concepts
developed for dynamic games, such as agent quantal response equilibrium (AQRE) (McK-
elvey and Palfrey, 1998), dynamic cognitive hierarchy theory (DCH) (Lin and Palfrey, 2022;
Lin, 2022), and the analogy-based expectation equilibrium (ABEE) (Jehiel, 2005; Jehiel and
Koessler, 2008), all of which modify the requirements of sequential equilibrium in different
ways than cursed sequential equilibrium.
2 The Model
Since CSE is a solution concept for dynamic games of incomplete information, in this pa-
per we will focus on the framework of multistage games with observed actions (Fudenberg
and Tirole, 1991b). Section 2.1 defines the formal structure of multi-stage games with ob-
2
We illustrate some implications of these differences in the application to signaling games in Section 4.1.
For a more detailed discussion of the differences between CSE and SCE, see Fong et al. (2023)
5
served actions, followed by Section 2.2, where the
χ
-cursed sequential equilibrium is formally
developed.
2.1 Multi-Stage Games with Observed Actions
Let
N
=
{
1
,...,n
}
be a finite set of players. Each player
i
∈
N
has a
type
θ
i
drawn from
a finite set Θ
i
. Let
θ
∈
Θ
≡ ×
n
i
=1
Θ
i
be the type profile and
θ
−
i
∈
Θ
−
i
≡ ×
j
̸
=
i
Θ
j
be the
type profile without player
i
. All players share a common (full support) prior distribution
F
(
·
) : Θ
→
(0
,
1). Therefore, for every player
i
, the belief of other players’ types conditional
on his own type is
F
(
θ
−
i
|
θ
i
) =
F
(
θ
−
i
,θ
i
)
P
θ
′
−
i
∈
Θ
−
i
F
(
θ
′
−
i
,θ
i
)
.
At the beginning of the game, players observe their own types, but not the other players’
types. That is, each player’s type is his own private information.
The game is played in stages
t
= 1
,
2
,...,T
. In each stage, players simultaneously choose
actions, which will be revealed at the end of the stage. The feasible set of actions can vary
with histories, so games with alternating moves are also included. Let
H
t
−
1
be the set of
all possible histories at stage
t
, where
H
0
=
{
h
∅
}
and
H
T
is the set of terminal histories.
Let
H
=
∪
T
t
=0
H
t
be the set of all possible histories of the game, and
H\H
T
be the set of
non-terminal histories.
For every player
i
, the available information at stage
t
is in Θ
i
×H
t
−
1
. Therefore, player
i
’s
information sets can be specified as
I
i
∈Q
i
=
{
(
h,θ
) :
h
∈H\H
T
,θ
i
∈
Θ
i
}
. That is, a type
θ
i
player
i
’s information set at the public history
h
t
can be defined as
S
θ
−
i
∈
Θ
−
i
(
θ
i
,θ
−
i
,h
t
).
With a slight abuse of notation, it will be denoted as (
θ
i
,h
t
). For the sake of simplicity, we
assume that, at each history, the feasible set of actions for every player is independent of
their type and use
A
i
(
h
t
−
1
) to denote the feasible set of actions for player
i
at history
h
t
−
1
.
Let
A
i
=
×
h
∈H\H
T
A
i
(
h
) denote player
i
’s feasible actions in all histories of the game and
A
=
A
1
×···×
A
n
. In addition, we assume
A
i
is finite for all
i
∈
N
and
|
A
i
(
h
)
|≥
1 for all
i
∈
N
and any
h
∈H\H
T
.
A behavioral strategy for player
i
is a function
σ
i
:
Q
i
→
∆(
A
i
) satisfying
σ
i
(
h
t
−
1
,θ
i
)
∈
∆(
A
i
(
h
t
−
1
)). Furthermore, we use
σ
i
(
a
t
i
|
h
t
−
1
,θ
i
) to denote the probability player
i
chooses
a
t
i
∈
A
i
(
h
t
−
1
). We use
a
t
= (
a
t
1
,...,a
t
n
)
∈ ×
n
i
=1
A
i
(
h
t
−
1
)
≡
A
(
h
t
−
1
) to denote the action
profile at stage
t
and
a
t
−
i
to denote the action profile at stage
t
without player
i
. If
a
t
is
the action profile realized at stage
t
, then
h
t
= (
h
t
−
1
,a
t
). Finally, each player
i
has a payoff
function
u
i
:
H
T
×
Θ
→
R
,
and we let
u
= (
u
1
,...,u
n
) be the profile of payoff functions. A
multi-stage game with observed actions, Γ, is defined by the tuple Γ =
⟨
T,A,N,
H
,
Θ
,
F
,u
⟩
.
6
2.2 Cursed Sequential Equilibrium
In a multi-stage game with observed actions, a solution is defined by an “assessment,” which
consists of a (behavioral) strategy profile
σ
, and a belief system
μ
. Since action profiles will
be revealed to all players at the end of each stage, the belief system specifies, for each player,
a conditional distribution over the set of type profiles conditional on each history. Consider
an assessment (
μ,σ
). Following the spirit of the cursed equilibrium, for player
i
at stage
t
,
we define the
average behavioral strategy profile of the other players
as:
̄
σ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
i
) =
X
θ
−
i
∈
Θ
−
i
μ
i
(
θ
−
i
|
h
t
−
1
,θ
i
)
σ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
−
i
)
for any
i
∈
N
,
θ
i
∈
Θ
i
and
h
t
−
1
∈H
t
−
1
.
In CSE, players have incorrect perceptions about other players’ behavioral strategies.
Instead of thinking they are using
σ
−
i
, a
χ
-cursed
3
type
θ
i
player
i
would believe the other
players are using a
χ
-weighted average of the average behavioral strategy and the true
behavioral strategy:
4
σ
χ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
−
i
,θ
i
) =
χ
̄
σ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
i
) + (1
−
χ
)
σ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
−
i
)
.
The beliefs of player
i
about
θ
−
i
are updated in the
χ
-CSE via Bayes’ rule, whenever
possible, assuming other players are using the
χ
-cursed behavioral strategy rather than the
true behavioral strategy. We call this updating rule the
χ
-cursed Bayes’ rule
. Specifically, an
assessment satisfies the
χ
-cursed Bayes’ rule if the belief system is derived from the Bayes’
rule while perceiving others are using
σ
χ
−
i
rather than
σ
−
i
.
Definition 1.
(
μ,σ
)
satisfies
χ
-cursed Bayes’ rule if the following rule is applied to update
the posterior beliefs whenever
P
θ
′
−
i
∈
Θ
−
i
μ
i
(
θ
′
−
i
|
h
t
−
1
,θ
i
)
σ
χ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
′
−
i
,θ
i
)
>
0
:
μ
i
(
θ
−
i
|
h
t
,θ
i
) =
μ
i
(
θ
−
i
|
h
t
−
1
,θ
i
)
σ
χ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
−
i
,θ
i
)
P
θ
′
−
i
∈
Θ
−
i
μ
i
(
θ
′
−
i
|
h
t
−
1
,θ
i
)
σ
χ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
′
−
i
,θ
i
)
.
Let Σ
0
be the set of totally mixed behavioral strategy profiles, and let Ψ
χ
be the set of
assessments (
μ,σ
) such that
σ
∈
Σ
0
and
μ
is derived from
σ
using
χ
-cursed Bayes’ rule.
5
3
We assume throughout the paper that all players are equally cursed, so there is no
i
subscript on
χ
. The
framework is easily extended to allow for heterogeneous degrees of cursedness.
4
If
χ
= 0, players have correct beliefs about the other players’ behavioral strategies at every stage.
5
In the following, we will use
μ
χ
(
·
) to denote the belief system derived under
χ
-cursed Bayes’ Rule. Also,
note that both
σ
χ
−
i
and
μ
χ
are induced by
σ
; that is,
σ
χ
−
i
(
·
) =
σ
χ
−
i
[
σ
](
·
) and
μ
χ
(
·
) =
μ
χ
[
σ
](
·
). For the ease
of exposition, we drop [
σ
] when it does not cause confusion.
7
Lemma 1 below shows that another interpretation of the
χ
-cursed Bayes’ rule is that players
have correct perceptions about
σ
−
i
but are unable to make perfect Bayesian inference when
updating beliefs. From this perspective, player
i
’s cursed belief is simply a linear combination
of player
i
’s cursed belief at the beginning of that stage (with
χ
weight) and the Bayesian
posterior belief (with 1
−
χ
weight). Because
σ
is totally mixed, there are no off-path histories.
Lemma 1.
For any
(
μ,σ
)
∈
Ψ
χ
,
i
∈
N
,
h
t
= (
h
t
−
1
,a
t
)
∈H\H
T
and
θ
∈
Θ
,
μ
i
(
θ
−
i
|
h
t
,θ
i
) =
χμ
i
(
θ
−
i
|
h
t
−
1
,θ
i
) + (1
−
χ
)
"
μ
i
(
θ
−
i
|
h
t
−
1
,θ
i
)
σ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
−
i
)
P
θ
′
−
i
μ
i
(
θ
′
−
i
|
h
t
−
1
,θ
i
)
σ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
′
−
i
)
#
Proof.
See Appendix A.
This is analogous to Lemma 1 of Eyster and Rabin (2005). Another insight provided
by Lemma 1 is that even if player types are independently drawn, i.e.,
F
(
θ
) = Π
n
i
=1
F
i
(
θ
i
),
players’ cursed beliefs about other players’ types are generally
not
independent across players.
That is, in general,
μ
i
(
θ
−
i
|
h
t
,θ
i
)
̸
= Π
j
̸
=
i
μ
ij
(
θ
j
|
h
t
,θ
i
)
.
The belief system will preserve the
independence only when the players are either fully rational (
χ
= 0) or fully cursed (
χ
= 1).
Finally, we place a consistency restriction, analogous to consistent assessments in sequen-
tial equilibrium, on how
χ
-cursed beliefs are updated off the equilibrium path, i.e., when
X
θ
′
−
i
∈
Θ
−
i
μ
i
(
θ
′
−
i
|
h
t
−
1
,θ
i
)
σ
χ
−
i
(
a
t
−
i
|
h
t
−
1
,θ
′
−
i
,θ
i
) = 0
.
An assessment satisfies
χ
-consistency if it is in the closure of Ψ
χ
.
Definition 2.
(
μ,σ
)
satisfies
χ
-consistency if there is a sequence of assessments
{
(
μ
k
,σ
k
)
}⊆
Ψ
χ
such that
lim
k
→∞
(
μ
k
,σ
k
) = (
μ,σ
)
.
For any
i
∈
N
,
χ
∈
[0
,
1],
σ
, and
θ
∈
Θ, let
ρ
χ
i
(
h
T
|
h
t
,θ,σ
χ
−
i
,σ
i
) be player
i
’s perceived
conditional realization probability of terminal history
h
T
∈H
T
at history
h
t
∈H\H
T
if the
type profile is
θ
and player
i
uses the behavioral strategy
σ
i
whereas perceives other players’
using the cursed behavioral strategy
σ
χ
−
i
. At every non-terminal history
h
t
, a
χ
-cursed player
in
χ
-CSE will use
χ
-cursed Bayes’ rule (Definition 1) to derive the posterior belief about the
other players’ types. Accordingly, a type
θ
i
player
i
’s conditional expected payoff at history
h
t
is given by:
E
u
i
(
σ
|
h
t
,θ
i
) =
X
θ
−
i
∈
Θ
−
i
X
h
T
∈H
T
μ
i
(
θ
−
i
|
h
t
,θ
i
)
ρ
χ
i
(
h
T
|
h
t
,θ,σ
χ
−
i
,σ
i
)
u
i
(
h
T
,θ
i
,θ
−
i
)
.
8
Definition 3.
An assessment
(
μ
∗
,σ
∗
)
is a
χ
-cursed sequential equilibrium if it satisfies
χ
-
consistency and
σ
∗
i
(
h
t
,θ
i
)
maximizes
E
u
i
(
σ
∗
|
h
t
,θ
i
)
for all
i
,
θ
i
,
h
t
∈H\H
T
.
3 General Properties of
χ
-CSE
In this section, we characterize some general theoretical properties of
χ
-CSE. The first result
is the existence of the
χ
-CSE. The definition of
χ
-CSE mirrors the definition of the sequential
equilibrium by Kreps and Wilson (1982)—the only difference is that players in
χ
-CSE update
their beliefs by
χ
-cursed Bayes’ rule and best respond to
χ
-cursed (behavioral) strategies.
Therefore, one can prove the existence of
χ
-CSE in a similar way as in the standard argument
of the existence of sequential equilibrium.
Proposition 1.
For any
χ
∈
[0
,
1]
and any finite multi-stage game with observed actions,
there is at least one
χ
-CSE.
Proof.
We briefly sketch the proof here, and the details can be found in Appendix A.
Fix any
χ
∈
[0
,
1]. For any
i
∈
N
and any information set
I
i
= (
h
t
−
1
,θ
i
), player
i
has
to choose every action
a
t
i
∈
A
i
(
h
t
−
1
) with probability at least
ε
. Since there are no off-path
histories, the belief system is uniquely pinned down by
χ
-cursed Bayes’ rule and a
χ
-CSE
exists in this
ε
-constrained game. We denote this
χ
-CSE as (
μ
ε
,σ
ε
). By compactness, there
is a converging sub-sequence of assessments such that (
μ
ε
,σ
ε
)
→
(
μ
∗
,σ
∗
) as
ε
→
0, which is
a
χ
-CSE, as desired.
Let Φ(
χ
) be the correspondence that maps
χ
∈
[0
,
1] to the set of
χ
-CSE. Proposition 1
guarantees Φ(
χ
) is non-empty for any
χ
∈
[0
,
1]. Because
χ
-cursed Bayes’ rule changes con-
tinuously in
χ
, we can further prove in Proposition 2 that Φ(
χ
) is an upper hemi-continuous
correspondence.
Proposition 2.
Φ(
χ
)
is upper hemi-continuous with respect to
χ
.
Proof.
The proof follows a standard argument. See Appendix A for details.
As shown in Corollary 1, a direct consequence of upper hemi-continuity is that every
limit point of a sequence of
χ
-CSE when
χ
→
0 is a sequential equilibrium. This result
bridges our behavioral equilibrium concept with standard equilibrium theory.
Corollary 1.
Every limit point of a sequence of
χ
-CSE with
χ
converging to 0 is a sequential
equilibrium.
9
Proof.
By Proposition 2, we know Φ(
χ
) is upper hemi-continuous at 0. Consider of a se-
quence of
χ
-CSE. As
χ
→
0, the limit point remains a CSE, which is a sequential equilibrium
at
χ
= 0. This completes the proof.
Finally, by a similar argument to Kreps and Wilson (1982), for any
χ
∈
[0
,
1],
χ
-CSE is
also upper hemi-continuous with respect to payoffs. In other words, our
χ
-CSE preserves
the continuity property of sequential equilibrium.
The next result is the characterization of a necessary condition for
χ
-CSE. As seen
from Lemma 1, players update their beliefs more passively in
χ
-CSE than in the stan-
dard equilibrium—they put
χ
-weight on their beliefs formed in previous stage. To formalize
this, we define the
χ
-dampened updating property
in Definition 4. An assessment satisfies
this property if at
any
non-terminal history, the belief puts at least
χ
weight on the belief
in previous stage—both on and off the equilibrium path. In Proposition 3, we show that
χ
-consistency implies the
χ
-dampened updating property.
Definition 4.
An assessment
(
μ,σ
)
satisfies the
χ
-dampened updating property if for any
i
∈
N
,
θ
∈
Θ
and
h
t
= (
h
t
−
1
,a
t
)
∈H\H
T
,
μ
i
(
θ
−
i
|
h
t
,θ
i
)
≥
χμ
i
(
θ
−
i
|
h
t
−
1
,θ
i
)
.
Proposition 3.
χ
-consistency implies
χ
-dampened updating for any
χ
∈
[0
,
1]
.
Proof.
See Appendix A.
It follows that if assessment (
μ,σ
) satisfies the
χ
-dampened updating property, then for
any player
i
, any history
h
t
and any type profile
θ
, player
i
’s belief about
θ
−
i
is bounded by
χμ
i
(
θ
−
i
|
h
t
−
1
,θ
i
)
≤
μ
i
(
θ
−
i
|
h
t
,θ
i
)
≤
1
−
χ
X
θ
′
−
i
̸
=
θ
−
i
μ
i
(
θ
′
−
i
|
h
t
−
1
,θ
i
)
.
One can see from this condition that when
χ
increases, the feasible range of
μ
i
(
θ
−
i
|
h
t
,θ
i
)
shrinks, and the restriction on the belief system becomes more stringent. Moreover, if the
history
h
t
is an off-path history of (
μ,σ
), then this condition characterizes the feasible set of
off-path beliefs, which shrinks as
χ
increases.
An important implication of this observation is that Φ(
χ
) is not lower hemi-continuous
with respect to
χ
. The intuition is that for some
χ
-CSE that contains off-path histories, the
off-path beliefs to support the equilibrium might not be
χ
-consistent for sufficiently large
χ
.
10
In this case, the
χ
-CSE is not attainable by a sequence of
χ
k
-CSE where
χ
k
converges to
χ
from above, causing the lack of lower hemi-continuity.
6
Lastly, another implication of
χ
-dampened updating property is that for each player
i
,
history
h
t
and type profile
θ
, the belief
μ
i
(
θ
−
i
|
h
t
,θ
i
) has a lower bound that is
independent
of the strategy profile. The lower bound is characterized in Corollary 2. This result implies
that when
χ >
0,
F
(
θ
−
i
|
θ
i
)
>
0 implies
μ
i
(
θ
−
i
|
h
t
,θ
i
)
>
0 for all
h
t
, so that if prior beliefs are
bounded away from zero, beliefs are always bounded away from 0 as well. In other words,
when
χ >
0, because of the
χ
-dampened updating, beliefs will always have full support even
if at off-path histories.
Corollary 2.
For any
χ
-consistent assessment
(
μ,σ
)
,
i
∈
N
,
θ
∈
Θ
and
h
t
∈H\H
T
,
μ
i
(
θ
−
i
|
h
t
,θ
i
)
≥
χ
t
F
(
θ
−
i
|
θ
i
)
Proof.
See Appendix A.
If the game has only one stage, then the dampened updating property has no effect, in
which case
χ
-CSE and
χ
-CE are equivalent solution concepts. This is formally stated and
proved in Proposition 4.
Proposition 4.
For any one-stage game and for any
χ
,
χ
-CSE and
χ
-CE are equivalent.
Proof.
For any one-stage game, the only public history is the initial history
h
∅
. Thus, in any
χ
-CSE, for each player
i
∈
N
and type profile
θ
∈
Θ, player
i
’s belief about other players’
types at this history is
μ
i
(
θ
−
i
|
h
∅
,θ
i
) =
F
(
θ
−
i
|
θ
i
)
.
Since the game has only one stage, the outcome is simply
a
1
= (
a
1
1
,...,a
1
n
), the action profile
at stage 1. Moreover, given any behavioral strategy profile
σ
, player
i
believes
a
1
will be the
outcome with probability
σ
i
(
a
1
i
|
h
∅
,θ
i
)
×
χ
̄
σ
−
i
(
a
1
−
i
|
h
∅
,θ
i
) + (1
−
χ
)
σ
−
i
(
a
1
−
i
|
h
∅
,θ
−
i
)
.
Therefore, if
σ
is the behavioral strategy profile of a
χ
-CSE in an one-stage game, then for
6
An example is provided in Section 4.1 (see Footnote 7).
11
each player
i
, type
θ
i
∈
Θ
i
and each
a
1
i
∈
A
i
(
h
∅
) such that
σ
i
(
a
1
i
|
h
∅
,θ
i
)
>
0,
a
1
i
∈
argmax
a
1
′
i
∈
A
i
(
h
∅
)
X
θ
−
i
∈
Θ
−
i
F
(
θ
−
i
|
θ
i
)
×
X
a
1
−
i
∈
A
−
i
(
h
∅
)
χ
̄
σ
−
i
(
a
1
−
i
|
h
∅
,θ
i
) + (1
−
χ
)
σ
−
i
(
a
1
−
i
|
h
∅
,θ
−
i
)
u
i
(
a
1
′
i
,a
1
−
i
,θ
i
,θ
−
i
)
,
which coincides with the maximization problem of
χ
-CE. This completes the proof.
From the proof of Proposition 4, one can see that in one-stage games players have
correct
perceptions about the average strategy of others. Therefore, the maximization problem
of
χ
-CSE coincides with the problem of
χ
-CE. For general multi-stage games, because of
the
χ
-dampened updating property, players will update beliefs incorrectly and thus their
perceptions about other players’ future moves can also be distorted.
4 Applications
In this section, we will explore
χ
-CSE in five applications of multi-stage games with observed
actions, in order to illustrate the range of effects it can have and to show how it is different
from the
χ
-CE and sequential equilibrium.
Our first application is the sender-receiver signaling game, which is practically the sim-
plest possible multi-stage game. From our analysis, we will see both the theoretical and
empirical implications of our
χ
-CSE.
4.1 Pooling Equilibria in Signaling Games
We first make a general observation about pooling equilibria in multi-stage games. Player
j
follows a
pooling strategy
if for every non-terminal history,
h
t
, all types of player
j
take
the same action
a
t
+1
j
∈
A
j
(
h
t
). Conceptually, since every type of player
j
takes the same
action, players other than
j
cannot make any inference about
j
’s type from
j
’s actions. A
pooling
χ
-CSE
is a
χ
-CSE where every player follows a pooling strategy. Hence, every player
has correct beliefs about any other player’s future move because every type of every player
chooses the same action.
Since in any pooling
χ
-CSE, players can correctly anticipate other players’ future moves
no matter how cursed they are, one may naturally conjecture that a pooling
χ
-CSE is also a
χ
′
-CSE for any
χ
′
∈
[0
,
1]. As shown by Eyster and Rabin (2005), this is true for one-stage
12
Bayesian games: if a pooling strategy profile is a
χ
-cursed equilibrium, then it is also a
χ
′
-cursed equilibrium for any
χ
′
∈
[0
,
1]. Surprisingly, this result does not extend to multi-
stage games. Proposition 5 shows if a pooling behavioral strategy profile is a
χ
-CSE, then it
remains a
χ
′
-CSE only for
χ
′
≤
χ
, which is a weaker result than Eyster and Rabin (2005).
This result is driven by the
χ
-dampened updating property which restricts the set of
off-path beliefs. As discussed above, when
χ
gets larger, the set of feasible off-path beliefs
shrinks, eliminating some pooling
χ
-CSE.
Proposition 5.
A pooling
χ
-CSE is a
χ
′
-CSE for
χ
′
≤
χ
.
Proof.
See Appendix B.
The proof strategy is similar to the one in Eyster and Rabin (2005) Proposition 3. Given
a
χ
-CSE behavioral strategy profile, we can separate the histories into on-path and off-path
histories. For on-path histories in a pooling equilibrium, since all types of players make the
same decisions, players cannot make any inference about other players’ types. Therefore,
for on-path histories, their beliefs are the prior beliefs, which are independent of
χ
. On the
other hand, for off-path histories, as shown in Proposition 3, a necessary condition for
χ
-CSE
is that the belief system has to satisfy the
χ
-dampened updating property. When
χ
gets
larger, this requirement becomes more stringent, and hence some pooling
χ
-CSE may break
down.
Example 1 is a signaling game where the sender has only two types and two messages, and
the receiver has only two actions. This example demonstrates the implication of Proposition
5 and shows the lack of lower hemi-continuity; i.e., it is possible for a pooling behavioral
strategy profile to be a
χ
-CSE, but not a
χ
′
-CSE for
χ
′
> χ
. We will also use this example to
illustrate how the notion of cursedness in sequential cursed equilibrium proposed by Cohen
and Li (2023) departs from our CSE.
Example 1.
The sender has two possible types drawn from the set Θ =
{
θ
1
,θ
2
}
with
Pr(
θ
1
) = 1
/
4. The receiver does not have any private information. After the sender’s type
is drawn, the sender observes his type and decides to send a message
m
∈ {
A,B
}
, or any
mixture between the two. After that, the receiver decides between action
a
∈{
L,R
}
or any
mixture between the two, and the game ends. The game tree is illustrated in Figure 1.
If we solve for the
χ
-CE of the game (or the sequential equilibria), we find that there
are two pooling equilibria for every value of
χ
. In the first pooling
χ
-CE, both sender types
choose
A
; the receiver chooses
L
in response to
A
and
R
at the off-path history
B
. In
the second pooling
χ
-CE, both sender types pool at
B
and the receiver chooses
R
at both
13
2
,
2
L
−
1
,
4
R
A
4
,
−
1
L
1
,
0
R
B
θ
1
[
1
4
]
2
,
1
L
−
1
,
0
R
A
4
,
−
2
L
1
,
0
R
B
θ
2
[
3
4
]
Nature
1
1
2
2
Figure 1: Game Tree for Example 1
histories. By Proposition 3 of Eyster and Rabin (2005), these two equilibria are in fact
pooling
χ
-CE for all
χ
∈
[0
,
1]. The intuition is that in a pooling
χ
-CE, players are not
able to make any inference about other players’ types from their actions because the average
normal form strategy is the same as the type-conditional normal form strategy. Therefore,
their beliefs are independent of
χ
, and hence a pooling
χ
-CE will still be an equilibrium for
any
χ
∈
[0
,
1].
However, as summarized in Claim 1 below, the
χ
-CSE imposes
stronger
restrictions than
χ
-CE in this example, in the sense that when
χ
is sufficiently large, the second pooling equi-
librium cannot be supported as a
χ
-CSE. The key reason is that when the game is analyzed
in its normal form, the
χ
-dampened updating property shown in Proposition 3 does not have
any bite, allowing both pooling equilibria to be supported as a
χ
-CE for any value of
χ
. Yet,
in the
χ
-CSE analysis, the additional restriction of
χ
-dampened updating property eliminates
some extreme off-path beliefs, and hence, eliminates the second pooling
χ
-CSE equilibrium
for sufficiently large
χ
. For simplicity, we use a four-tuple [(
m
(
θ
1
)
,m
(
θ
2
)); (
a
(
A
)
,a
(
B
))] to
denote a behavioral strategy profile.
Claim 1.
In this example, there are two pure pooling
χ
-CSE, which are:
1.
[(
A,A
); (
L,R
)]
is a pooling
χ
-CSE for any
χ
∈
[0
,
1]
.
2.
[(
B,B
); (
R,R
)]
with
μ
2
(
θ
1
|
A
)
∈
1
3
,
1
−
3
4
χ
is a pooling
χ
-CSE if and only if
χ
≤
8
/
9
.
Proof.
See Appendix B.
From previous discussion, we know in general, the sets of
χ
-CSE and
χ
-CE are non-
overlapping because of the nature of sequential distortion of beliefs in
χ
-CSE. Yet, a pooling
14
χ
-CSE is an exception. In a pooling
χ
-CSE, players can correctly anticipate others’ future
moves, so a pooling
χ
-CSE will mechanically be a pooling
χ
-CE. In cases such as this, we
can find that
χ
-CSE is a
refinement
of
χ
-CE.
7
Remark.
This game is useful for illustrating some of the differences between the notions
of “cursedness” in
χ
-CSE and the sequential cursed equilibrium ((
χ
S
,ψ
S
)-SCE) proposed
by Cohen and Li (2023). The first distinction is that the
χ
and
χ
S
parameters capture
substantively different sources of distortion in a player’s beliefs about the other players’
strategies. In
χ
-CSE, the degree of cursedness,
χ
, captures how much a player neglects the
dependence of the other players’ behavioral strategies on those players’ (exogenous)
private
information
, i.e, types, drawn by nature, and as a result, mistakenly treats different types
as behaving the same with probability
χ
. In contrast, in (
χ
S
,ψ
S
)-SCE, the cursedness
parameter,
χ
S
, captures how much a player neglects the dependence of the other players’
strategies on future moves of the others, or current moves that are unobserved because of
simultaneous play. Thus, it is a neglect related to endogenous information. If player
i
observes
a previous move by some other player
j
, then player
i
correctly accounts for the
dependence of player
j
’s chosen action on player
j
’s private type, as would be the case in
χ
-CSE only at the boundary where
χ
= 0.
In the context of pooling equilibria in sender-receiver signaling games, if
χ
S
= 1, then
in SCE the sender believes the receiver will respond the same way both on and off the
equilibrium path. This distorts how the sender perceives the receiver’s future action in
response to an off-equilibrium path message. In
χ
-CSE, cursedness does not hinder the
sender from correctly perceiving the receiver’s strategy since the receiver only has one type.
Take the strategy profile [(
A,A
); (
L,R
)] for example, which is a pooling
χ
-CSE equilibrium
for all
χ
∈
[0
,
1]. However, with (
χ
S
,ψ
S
)-SCE, a sender misperceives that the receiver, upon
receiving the off-path message
B
, will, with probability
χ
S
, take the same action (
L
) as
when receiving the on-path message
A
. If
χ
S
is sufficiently high, the sender will deviate to
send
B
, which implies that [(
A,A
); (
L,R
)] cannot be supported as an equilibrium when
χ
S
is sufficiently large (
χ
S
>
1
/
3). The distortion induced by
χ
S
also creates an additional SCE
if
χ
S
is sufficiently large: [(
B,B
); (
L,R
)]. To see this, if
χ
S
= 1, then a sender incorrectly
believes that the receiver will continue to choose
R
if the sender deviates to
A
, rather than
switching to
L
, and hence
B
is optimal for both sender types. However, [(
B,B
); (
L,R
)] is
not
a
χ
-CSE equilibrium for any
χ
∈
[0
,
1], or a
χ
-CE in the sense of Eyster and Rabin
7
Note that the
χ
-CSE correspondence Φ(
χ
) is not lower hemi-continuous with respect to
χ
. To see
this, we consider a sequence of
{
χ
k
}
where
χ
k
=
8
9
+
1
9
k
for
k
≥
1. From the analysis of Claim 1, we
know [(
B,B
); (
R,R
)]
̸∈
Φ(
χ
k
) for any
k
≥
1. However, in the limit where
χ
k
→
8
/
9, [(
B,B
); (
R,R
)] with
μ
2
(
θ
1
|
A
) = 1
/
3 is indeed a CSE. That is, [(
B,B
); (
R,R
)] is not approachable by this sequence of
χ
k
-CSE.
15
(2005), or a sequential equilibrium.
In the two possible pooling equilibria analyzed in the last paragraph, the second SCE
parameter,
ψ
S
, does not have any effect, but the role of
ψ
S
can be illustrated in the context of
the [(
B,B
); (
R,R
)] sequential equilibrium. This second SCE parameter,
ψ
S
, is introduced to
accommodate a player’s possible failure to fully account for the informational content from
observed
events. The larger (1
−
ψ
S
) is, the greater extent a player neglects the informational
content of observed actions. Although the parameter
ψ
S
has a similar flavor to 1
−
χ
in
χ
-
CSE, it is different in a number of ways. In particular this parameter only has an effect
via its interaction with
χ
S
and thus does not independently arise. In the two parameter
model, the overall degree of cursedness is captured by the product,
χ
S
(1
−
ψ
S
), and thus
any cursedness effect of
ψ
S
is shut down when
χ
S
= 0. For instance, under our
χ
-CSE, the
strategy profile [(
B,B
); (
R,R
)] can only be supported as an equilibrium when
χ
is sufficiently
small. However, [(
B,B
); (
R,R
)] can be supported as a (
χ
S
,ψ
S
)-SCE even when (1
−
ψ
S
) = 1
as long as
χ
S
is sufficiently small. In fact, when
χ
S
= 0, a (
χ
S
,ψ
S
)-SCE is equivalent to
sequential equilibrium regardless of the value of
ψ
S
.
45
,
30
30
,
30
C
15
,
0
0
,
0
D
30
,
15
50
,
35
E
I
30
,
90
45
,
90
C
0
,
15
15
,
15
D
45
,
15
100
,
30
E
S
θ
1
[
1
2
]
30
,
30
30
,
30
C
0
,
45
30
,
45
D
30
,
15
30
,
0
E
I
45
,
0
45
,
0
C
15
,
30
0
,
30
D
30
,
15
0
,
15
E
S
θ
2
[
1
2
]
BH 3
BH 4
Nature
1
1
2
2
Figure 2: Game Tree for
BH 3
and
BH 4
in Brandts and Holt (1993)
Example 2.
Here we analyze two signaling games that were studied experimentally by
Brandts and Holt (1993) (
BH 3
and
BH 4
) and show that
χ
-CSE can help explain some
of their findings. In both Game
BH 3
and Game
BH 4
, the sender has two possible types
{
θ
1
,θ
2
}
which are equally likely. There are two messages
m
∈{
I,S
}
available to the sender.
8
After seeing the message, the receiver chooses an action from
a
∈{
C,D,E
}
. The game tree
and payoffs for both games are summarized in Figure 2.
8
I
stands for “
I
ntuitive” and
S
stands for “
S
equential but not intuitive”, corresponding to the two pooling
sequential equilibria of the two games.
16
In both games, there are two pooling sequential equilibria. In the first equilibrium, both
sender types send message
I
, and the receiver will choose
C
in response to
I
and choose
D
in response to
S
. In the second equilibrium, both sender types send message
S
, and the
receiver will choose
D
in response to
I
while choose
C
in response to
S
. Both are sequential
equilibria, in both games, but only the first equilibrium where the sender sends
I
satisfies
the intuitive criterion proposed by Cho and Kreps (1987).
Since the equilibrium structure is similar in both games, the sequential equilibrium and
the intuitive criterion predict the behavior should be the same in both games. However, this
prediction is strikingly rejected by the data. Brandts and Holt (1993) report that in the
later rounds of the experiment, almost all type
θ
1
senders send
I
in Game
BH 3
(97 %), and
yet all type
θ
1
senders send
S
in Game
BH 4
(100%). In contrast, type
θ
2
senders behave
similarly in both games—46
.
2% and 44
.
1% of type
θ
2
senders send
I
in Games
BH 3
and
BH 4
, respectively. Qualitatively speaking, the empirical pattern reported by Brandts and
Holt (1993) is that
sender type
θ
1
is more likely to send
I
in Game BH 3 than Game BH 4
while sender type
θ
2
’s behavior is insensitive to the change of games.
To explain this finding, Brandts and Holt (1993) propose a descriptive story based on
naive receivers
. A naive receiver will think both sender types are equally likely, regardless of
which message is observed. This naive reasoning will lead the receiver to choose
C
in both
games. Given this naive response, a type
θ
1
sender has an incentive to send
I
in Game
BH
3
and choose
S
in Game
BH 4
. (Brandts and Holt (1993), p. 284 – 285)
In fact, their story of naive reasoning echoes the logic of
χ
-CSE. When the receiver is
fully cursed (or naive), he will ignore the correlation between the sender’s action and type,
causing him to not update the belief about the sender’s type. Proposition 6 characterizes
the set of
χ
-CSE of both games. Following the notation in Example 1, we use a four-tuple
[(
m
(
θ
1
)
,m
(
θ
2
)); (
a
(
I
)
,a
(
S
))] to denote a behavioral strategy profile.
Proposition 6.
The set of
χ
-CSE of Game BH 3 and BH 4 are characterized as below.
In Game BH 3, there are three pure
χ
-CSE:
1.
[(
I,I
); (
C,D
)]
is a pooling
χ
-CSE if and only if
χ
≤
4
/
7
.
2.
[(
S,S
); (
D,C
)]
is a pooling
χ
-CSE if and only if
χ
≤
2
/
3
.
3.
[(
I,S
); (
C,C
)]
is a separating
χ
-CSE if and only if
χ
≥
4
/
7
.
In Game BH 4, there are three pure
χ
-CSE:
1.
[(
I,I
); (
C,D
)]
is a pooling
χ
-CSE if and only if
χ
≤
4
/
7
.
17
2.
[(
S,S
); (
D,C
)]
is a pooling
χ
-CSE if and only if
χ
≤
2
/
3
.
3.
[(
S,S
); (
C,C
)]
is a pooling
χ
-CSE for any
χ
∈
[0
,
1]
.
Proof.
See Appendix B.
As noted earlier for Example 1, by Proposition 3 of Eyster and Rabin (2005), pooling
equilibria (1) and (2) in games BH 3 and BH 4 survive as
χ
-CE for
all
χ
∈
[0
,
1]. Hence,
Proposition 6 implies that
χ
-CSE refines the
χ
-CE pooling equilibria for larger values of
χ
.
Moreover,
χ
-CSE actually eliminates
all
pooling equilibria in BH 3 if
χ >
2
/
3. Proposition
6 also suggests that for any
χ
∈
[0
,
1], sender type
θ
2
will behave similarly in both games,
which is qualitatively consistent with the empirical pattern. In addition,
χ
-CSE predicts that
a highly cursed (
χ >
2
/
3) type
θ
1
sender will send different messages in different games—
highly cursed type
θ
1
senders will send
I
and
S
in Games
BH 3
and
BH 4
, respectively. This
is consistent with the empirical data.
4.2 A Public Goods Game with Communication
Our second application is a threshold public goods game with private information and pre-
play communication, variations of which have been studied in laboratory experiments (Pal-
frey and Rosenthal, 1991; Palfrey et al., 2017). Here we consider the “unanimity” case where
there are
N
players and the threshold is also
N
.
Each player
i
has a private cost parameter
c
i
, which is independently drawn from a
uniform distribution on [0
,K
] where
K >
1. After each player’s
c
i
is drawn, each player ob-
serves their own cost, but not the others’ costs. Therefore,
c
i
is player
i
’s private information
and corresponds to
θ
i
in the general formulation.
9
The game consists of two stages. After
the profile of cost parameters is drawn, the game will proceed to stage 1 where each player
simultaneously broadcasts a public message
m
i
∈ {
0
,
1
}
without any cost or commitment.
After all players observe the message profile from this first stage, the game proceeds to stage
2 which is a unanimity threshold public goods game. Player
i
has to pay the cost
c
i
if he
contributes, but the public good will be provided only if all players contribute. The public
good is worth a unit of payoff for every player. Thus, if the public good is provided, each
player’s payoff will be 1
−
c
i
.
If there is no communication stage, the unique Bayesian Nash equilibrium is that no
player contributes, which is also the unique
χ
-CE for any
χ
∈
[0
,
1]. In contrast, with the
communication stage, there exists an efficient sequential equilibrium where each player
i
9
This application has a continuum of types. The framework of analysis developed for finite types is
applied in the obvious way.
18
sends
m
i
= 1 if and only if
c
i
≤
1 and contributes if and only if all players send 1 in the
first stage.
10
Since this is a private value game, the standard cursed equilibrium has no bite,
and this efficient sequential equilibrium is also a
χ
-CE for all values of
χ
, by Proposition 2
of Eyster and Rabin (2005). In the following, we demonstrate that the prediction of
χ
-CSE
is different from CE (and sequential equilibrium).
To analyze the
χ
-CSE, consider a collection of “cutoff” costs,
{
C
χ
c
,C
χ
0
,C
χ
1
,...,C
χ
N
}
. In
the communication stage, each player communicates the message
m
i
= 1 if and only if
c
i
≤
C
χ
c
. In the second stage, if there are exactly 0
≤
k
≤
N
players sending
m
i
= 1 in the
first stage, then such a player would contribute in the second stage if and only if
c
i
≤
C
χ
k
.
A
χ
-CSE is a collection of these cost cutoffs such that the associated strategies are a
χ
-CSE
for the public goods game with communication. The most efficient sequential equilibrium
identified above for
χ
= 0 corresponds to cutoffs with
C
0
0
=
C
0
1
=
···
=
C
0
N
−
1
= 0 and
C
0
c
=
C
0
N
= 1.
There are in fact multiple equilibria in this game with communication. In order to
demonstrate how the cursed belief can distort players’ behavior, here we will focus on the
χ
-CSE that is similar to the most efficient sequential equilibrium identified above, where
C
χ
0
=
C
χ
1
=
···
=
C
χ
N
−
1
= 0 and
C
χ
c
=
C
χ
N
. The resulting
χ
-CSE is given in Proposition 7.
Proposition 7.
In the public goods game with communication, there is a
χ
-CSE where
1.
C
χ
0
=
C
χ
1
=
···
=
C
χ
N
−
1
= 0
, and
2. there is a unique
C
∗
(
N,K,χ
)
≤
1
s.t.
C
χ
c
=
C
χ
N
=
C
∗
(
N,K,χ
)
that solves:
C
∗
(
N,K,χ
)
−
χ
C
∗
(
N,K,χ
)
K
N
−
1
= 1
−
χ.
Proof.
See Appendix B.
To provide some intuition, we sketch the proof by analyzing the two-person game, where
the
χ
-CSE is characterized by four cutoffs
{
C
χ
c
,C
χ
0
,C
χ
1
,C
χ
2
}
, with
C
χ
0
=
C
χ
1
= 0 and
C
χ
c
=
C
χ
2
. If players use the strategy that they would send message 1 if and only if the cost is
less than
C
χ
c
, then by Lemma 1, at the history where both players send 1, player
i
’s cursed
10
One can think of the first stage as a poll, where players are asked the following question: “Are you
willing to contribute if everyone else says they are willing to contribute?”. The message
m
i
= 1 corresponds
to a “yes” answer and the message
m
i
= 0 corresponds to a “no” answer.
19