Available online at www.sciencedirect.

com

ScienceDirect
Journal of Economic Theory 199 (2022) 105188
www.elsevier.com/locate/jet

Non-Bayesian updating in a social learning

experiment ✩
Roberta De Filippis a,b , Antonio Guarino b,∗ , Philippe Jehiel b,c ,
Toru Kitagawa b
a European Banking Authority, France
b Department of Economics, University College London (UCL), United Kingdom of Great Britain and Northern Ireland
c Paris School of Economics, France

Received 20 September 2019; final version received 7 December 2020; accepted 1 January 2021
Available online 26 January 2021

Abstract
In our laboratory experiment, subjects, in sequence, have to predict the value of a good. The second
subject in the sequence makes his prediction twice: first (“first belief”), after he observes his predecessor’s
prediction; second (“posterior belief”), after he observes his private signal. We find that the second subjects
weigh their signal as a Bayesian agent would do when the signal confirms their first belief; they overweight
the signal when it contradicts their first belief. This way of updating, incompatible with Bayesianism, can
be explained by the Likelihood Ratio Test Updating (LRTU) model, a generalization of the Maximum
Likelihood Updating rule. It is at odds with another family of updating, the Full Bayesian Updating. In
another experiment, we directly test the LRTU model and find support for it.
© 2021 Published by Elsevier Inc.

JEL classification: D01; D81; D90

Keywords: Ambiguous belief updating; Multiple priors; Social learning; Experiment

✩
We thank Marco Cipriani, Olivier Compte, Pietro Ortoleva and participants in conferences and seminars at many
institutions for useful comments and suggestions. We thank Hugo Freeman, Andrea Giacometti, Huilei Kang, Justin
Franco Lam, Jeff Rowley and Teresa Steininger for excellent research assistance. We gratefully acknowledge the financial
support from the ERC (grant numbers 210772, 742816, and 715940) and the ESRC through the ESRC Centre for
Microdata Methods and Practice (CeMMAP) (grant number RES-589-28-0001). Any views xpressed are solely those
of the author(s) and so cannot be taken to represent those of the European Banking Authority (EBA) or to state EBA
policy.
* Corresponding author.
E-mail addresses: [email protected] (R. De Filippis), [email protected] (A. Guarino),
[email protected] (P. Jehiel), [email protected] (T. Kitagawa).

https://doi.org/10.1016/j.jet.2021.105188
0022-0531/© 2021 Published by Elsevier Inc.
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

1. Introduction

Suppose you are contemplating the possibility of investing in a new project. Since it seems, a
priori, equally likely that it succeeds or fails, you ask for the opinion of an independent advisor.
After collecting some information on the project, he evaluates the probability of success to be
70%. On the basis of this recommendation only, clearly your belief on the probability of suc-
ceeding depends on how much you trust your advisor’s ability. If you fully trust him, you may
agree with him and evaluate the probability of success to be 70% as well. If you do not think
he has done a good job, or you suspect he is not so talented as you originally thought, you may
even completely discard his view and keep your prior belief of a 50% probability of success.
For the sake of the example, let us assume you trust him, although not completely, and assess
that the project will succeed with probability 65%. You now receive further information on the
project, independently of your advisor’s information. The information is of the same quality as
that received by your advisor, but is negative, that is, if you had to base your evaluation on this
information only, you would update your prior belief to a value lower than 50% (actually, 30%,
since the information is of the same quality). How would you now use this information to make
inferences on the quality of the project? Would you change your mind on your advisor’s ability?
And how would you revise your 65% belief?
One way of reasoning is that the negative private information you receive (contradicting
the advisor’s view) makes it more doubtful that he used his information correctly. You should,
therefore, discount the advisor’s evaluation even more than before and mainly rely on your in-
formation. Since you now trust him less, your belief based on his advice only would be less than
65%. Moreover, you have now negative information, which pushes your belief further down. For
instance, if you completely lost trust in your advisor’s ability, you would now totally discard his
view and on the basis of his advice only you would evaluate the probability of success to be 50%;
considering also the negative content of your information, you would evaluate the probability of
success to be 30%.
While this reasoning seems intuitive, one has to be careful in applying it, at least if one wants
to respect standard Bayesian updating.1 Indeed, a Bayesian agent, once expressed his belief of
65%, would simply update it by considering the probability of the received information condi-
tional on the project being a success or a failure. To the extent that the information received by
the advisor and yourself are independent conditional on the profitability of the project, the 65%
probability summarizes all the relevant information in order to update the belief. Given the values
we have used in our example, a Bayesian agent would never give a valuation of 30%; rather, he
would update the 65% belief to 44.3%.2
Decision problems like the one we have just, informally, described are very common in many
applications. Traders in financial markets typically try to infer information, say about asset fun-
damental values, from the order flow and also use their own private information. Companies
make investment decisions using consultants and their own internal profitability studies. Editors
of scientific journals make their editorial decisions on the basis of referee reports and of their
own reading of the papers.

1 We presented this work in many seminars and conferences, and played this game with the audience. It occurred
frequently that some professional economists gave an answer in agreement with this reasoning (and with our theory, as
we shall see) and were not convinced they were violating Bayesianism until they saw the proof.
2 This percentage is the solution to the equation x 65 30
100−x = 35 70 (where we have expressed the updating through the
x = 0.3(0.65)
likelihood ratio), or, using the familiar Bayesian updating formula, to: 100 0.3(0.65)+0.7(0.35) .

2
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

In this work we aim to study human behavior in a controlled experiment in which subjects
have to make similar types of decisions. Our purpose is to analyze how well human subjects’ be-
havior conforms to Bayesian updating when they have to make inferences from a private signal
and from the decision of another human subject, and, if appropriate, to understand the determi-
nants of deviations from it. To be specific, we ask subjects to predict whether a good is worth 0 or
100 units, two events that are, a priori, equally likely. A first subject receives a noisy symmetric
binary signal about the true value realization: either a “good signal”, which is more likely if the
value is 100; or a “bad signal”, which is more likely if the value is 0. After receiving his signal,
the subject is asked to state his belief on the value being 100. We then ask a second subject to
make the same type of prediction based on the observation of the first subject’s decision only.
Finally, we provide the second subject with another, conditionally independent, signal about the
value of the good and ask him to make a new prediction.
The main result of our study is that when at time 2 subjects observe their private signal, they
update their belief in an “asymmetric way.” When the signal is in agreement with their first belief
(e.g., when they first state a belief higher than 50% and then receive a signal indicating that the
more likely value is 100), they weigh the signal as a Bayesian agent would do. When, instead,
they receive a signal contradicting their first belief, they put considerably more weight on it (i.e.,
as if they had observed more than one signal, or as if the signal had a higher precision than it
actually has).
This asymmetric updating is incompatible with standard Bayesianism, even allowing for sub-
jective precisions of the signals, in the vein of Savage (1954), as long as the two pieces of
information are independent conditional on the value of the good. Given that the conditional
independence follows from principles of logic, we conclude that there is a need to go beyond the
Bayes/Savage paradigm. Moreover, as we will demonstrate, the asymmetric updating: i) cannot
be explained by known psychological biases such as base-rate neglect or confirmation bias; ii) is
not due to risk preferences; iii) is inherently related to social learning (indeed, it does not occur
in control, individual decision making, treatments).
Another main contribution of the paper is to provide a simple explanation for such asymmetric
updating, by building on the literature on ambiguity and multiple priors (and, in particular, on
the branch devoted to belief updating). We propose that, like in models with multiple priors,
the subject at time 2 may entertain several possible theories about time 1 subject’s “rationality”
(where a subject is considered “rational” if he chooses an action higher than 50 when he receives
a good signal and lower than 50 when he receives a bad signal; a “noise” subject, in contrast,
chooses any action between 0 and 100 independently of whether he receives a good or bad
signal).3 Moreover, each time he has to make a decision, he selects the theory that maximizes the
likelihood of the realized observations. Based on this selected theory about the rationality of the
predecessor, subject 2 updates his belief in a standard Bayesian fashion (possibly using subjective
representations of the precision of the signals). We label this mode of updating Likelihood Ratio
Test Updating model, a generalization of the Maximum Likelihood Updating model.
Intuitively, this explains the asymmetry we observe for the following reason. Imagine a subject
observing the predecessor taking an action greater than 50 (i.e., an action that presumably comes
from a good signal, indicating the value is 100). Suppose he considers that the event is most likely
under the prior that the predecessor is rational and, therefore, chooses his own action (his “first

3 As we will discuss in the next section, we use this minimal requirement in our definition of rationality, since it is
enough to infer the signal from the subject’s action, which is the only thing that matters.

3
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

belief”) accordingly. After he observes a private signal confirming his first belief (that the value
is more likely to be 100), the subject remains confident that the predecessor was rational, that is,
sticks to the same prior on the predecessor’s rationality. He updates on that prior belief and so the
weight he puts on the signal seems identical to that of a Bayesian agent. Consider now the case
in which he receives a signal contradicting his first belief (i.e., a bad signal, indicating that the
more likely value is 0). In such a case he now deems it an unlikely event that the predecessor was
rational. In other words, he selects another prior belief on the predecessor’s rationality, giving a
much higher probability to his predecessor being noise. Once he has selected this new prior on
the predecessor’s rationality, he updates on the basis of the signal realization. This time it will
look like he puts much more weight on the signal, since the signal first has made him change the
prior on the rationality of the predecessor (becoming more distrusting) and then update on the
basis of that prior. It is important to remark that it is as if the subject put different weights on the
signal depending on its realization, not that he consciously uses different weights.
If instead of considering the Likelihood Ratio Test Updating model one were to consider the
Full Bayesian Updating model, a popular alternative model of updating with multiple priors, we
establish that the agent would behave as if putting a lower (rather than higher) weight on his own
signal when it is contradicting relative to when it is confirming. Thus, whereas the Likelihood
Ratio Test Updating model can explain our experimental finding, the Full Bayesian Updating
model cannot.
More formally, we offer statistical and experimental evidence in support of this theory. We
perform a careful econometric analysis and find evidence in support of this theory against the
Bayesian (Savage) approach and the Full Bayesian Updating model. As for experimental ev-
idence, in a treatment we directly test how human subjects update on the rationality of the
predecessor. Again, we find that the updating is at odds with the Savage approach and the Full
Bayesian Updating model and in line with our model.
Our findings are of direct relevance for theory purposes. In particular, to the best of our knowl-
edge, we are the first to use multiple priors to model beliefs about rationality of others and provide
experimental evidence in its support. We show that ambiguity is relevant not just when it is about
the composition of an urn (which has inspired an enormous literature, following Ellsberg’s fa-
mous experiments), but also in very common and economically relevant decision problems. In
our analysis, we also contribute to the debate on how to update multiple priors (an area in which
there is very little theoretical consensus) and propose a novel identification strategy to distinguish
between different theories of updating.
It should be clear that our results have implications for applied work. In applications, the
Savage approach has been widely used. For instance, in behavioral finance, researchers model
overconfidence by allowing subjective precisions of the signals (an agent considers his own sig-
nal more informative than the others’). While this is left for future research, our finding suggests
the need to allow for multiple priors, which may unveil important phenomena, e.g., in the way
economic agents react to good and bad news depending on the prevailing current belief, market
conditions or state of the economy (recession or boom).
The paper is organized as follows. Section 2 describes the model, the Perfect Bayesian Equi-
librium predictions, as well as the predictions of the Likelihood Ratio Test Updating model and
of the Full Bayesian Updating model. Section 3 presents the experiment. Section 4 contains
the results of the main treatments. Section 5 illustrates the control treatments. Sections 6 and 7
present econometric and experimental evidence in support of the Likelihood Ratio Test Updating
rule. Section 8 discusses the related literature and Section 9 concludes. The Appendix contains
additional material and the instructions.

4
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

2. The theoretical framework

2.1. The basic model

In our economy there is a good that can take two values, V ∈ {0, 100}. The two values are
equally likely. There are two agents assumed to be risk neutral who make a decision in sequence.
The decision consists in choosing a number in the interval [0, 100]. Each agent t (t = 1, 2)
receives a symmetric binary signal st ∈ {0, 1} whose objective probability distribution is:

Pr(st = 1 | V = 100) = Pr(st = 0 | V = 0) = q ≡ 0.7.

Conditional on the value of the good, the signals are identically and independently distributed
over time, with precision 0.7. Since the signal st = 1 increases the probability that the value is
100, we will also refer to it as the good signal, and to st = 0 as the bad signal.
At time 1, agent 1 observes the signal s1 and takes an action a1 . At time 2, agent 2 observes
a1 and takes a first action a21 . He then observes the private signal s2 and takes a second action a22 .
j
The agent’s payoff from each choice j = 1, 2 is equal to 100 − 0.01(V − at )2 .4

2.1.1. Theoretical predictions of the basic model

Given the quadratic payoff function, both agents optimally state their belief (expressed as
j j
a percentage) that the value is 100. For given information It , the agent chooses at to max- 
j 2 j j∗ j
imize his expected payoff E[100 − 0.01(V − at ) |It ], that is, he chooses at = E V |It .
In the Perfect Bayesian Equilibrium (PBE), agent 1 chooses 70 upon observing s1 = 1 and
30 upon observing s1 = 0. After observing a1∗ = 30 or 70, agent 2 chooses as his first
action a21∗ = a1∗ . After observing the private signal s2 , the agent updates his belief and
chooses a22∗ = E (V |a1 = 70, s2 = 1) = 84.48, a22∗ = E (V |a1 = 30, s2 = 0) = 15. 52 or a22∗ =
E (V |a1 = 30, s2 = 1) = E (V |a1 = 70, s2 = 0) = 50. In the first two cases, the signal points in
the same direction as the predecessor’s action, therefore, we refer to them saying that the signal
is “confirming;” in the last two cases, instead, the signal points in the opposite direction to the
predecessor’s action, and we say that the signal is “contradicting.” Note that when the signal
is contradicting, the agent’s posterior belief is equal to the prior belief, since the two pieces of
information the agent receives “cancel out.”

2.2. The extended model

j
In the experiment, we can expect to see actions a1 and a2 different from those computed in
the PBE above for a variety of reasons. We summarize these reasons by letting agents attach
subjective precisions to the signals, possibly different from the objective ones, in the spirit of
Savage. We also wish to consider the possibility that agent 2 may “distrust” agent 1 by letting
him believe that agent 1 did not make a correct inference from his signal.
Specifically, we let agents have subjective views about the signal precisions: q1S ∈ (0.5, 1]
and q2tS ∈ (0.5, 1], where q1S stands for the subjective precision attached by agent 1 to the signal
at time 1 and q2tS stands for the subjective precision attached by agent 2 to signals at times

4 The superscript j is immaterial when t = 1 since agent 1 acts only once.

5
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

t = 1, 2.5 We also let agent 2 have subjective beliefs on the “rationality” of agent 1, that is, to
think that agent 1 may be of two types, rational (τr ) or noise (τn ). The subjective probability
that the subject is noise is denoted by Pr(τ = τn ) ≡ θ and the subjective probability that he is
rational by Pr(τ = τr ) = 1 − θ (later on we will also use μ ≡ 1 − θ ). The two types of agents
behave as follows. A rational agent always chooses an action greater than 50 upon observing a
good signal and an action lower than 50 upon observing a bad signal. Specifically, we let f (a1 )
denote the subjective distribution of actions on (50, 100] as perceived by agent 2 when agent 1 is
rational and receives a signal s1 = 1, and by symmetry we let f (100 − a1 ) be the corresponding
subjective distribution of actions on [0, 50) when agent 1 receives a signal s1 = 0.6 A noise
agent, by contrast, chooses in a random way, according to a distribution g(a1 ) symmetric around
50 independently of the observed signal.7 We are defining an agent at time 1 as “rational” as long
as he updates in the correct direction, since the only thing that agent 2 has to learn from agent 1 is,
indeed, the signal realization (given that the objective precisions of the signals are known and do
not have to be learned), and this is revealed under the minimal requirement that agent 1 updates
in the right direction. We assume that, after observing a1 , agent 2 thinks that agent 1 is rational
with probability μ(a1 ). This probability can be derived from the prior subjective probability θ
and the subjective distributions of actions at time 1 by a rational or a noise agent 1.8 Formally,
for any a1 ∈ (50, 100], we have that

(1 − θ )f (a1 )
μ (a1 ) = .
(1 − θ )f (a1 ) + 2g(a1 )θ
A similar expression can be derived for the case of a1 ∈ [0, 50).
j
Given the subjective precisions and rationality beliefs and given his information It , the agent
j j j
chooses
 at to maximize his expected payoff E[100 − 0.01(V − at )2 |It ]: his optimal action is
j
E V |It .

2.2.1. Theoretical predictions of the extended model

The predictions of the extended model are qualitatively similar to the basic model’s. Since
q1S > 0.5, agent 1 updates in the “correct direction” although not necessarily as in the PBE.
Given μ(a1 ) and q21S , agent 2 computes his expected value of the good, which he expresses by
stating a21 . In particular, a21∗ is such that
⎧
⎪ μ(a1 )q21S +(1−μ1 (a1 ))g(a1 )
a21∗ Pr (V = 100|a1 ) ⎨ μ(a1 )(1−q21S )+(1−μ(a1 ))g(a1 )
if a1 > 50,
= = (1)
100 − a21∗ Pr (V = 0|a1 ) ⎪
⎩ μ(a1 )(1−q21S )+(1−μ(a1 ))g(a1 )
if a1 < 50,
μ(a1 )q21S +(1−μ(a1 ))g(a1 )

where we are using the law of total probabilities and interpret Pr(·) as agent’s subjective proba-
bility.

5 The superscript S stands for “subjective.” Note that we let agent 2 assign a different precision to signal 1 (that he
does not observe) and to signal 2. Also note that we maintain the assumption of symmetric signals.
6 Symmetry is not required for our arguments; it is only to simplify the exposition.
7 We could have instead assumed that the non-rational agent always chooses an action in the wrong direction. This
would not alter our results.
8 The subjective distribution of actions by a rational agent 1 is derived from the distribution of q 1S .
2

6
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Once he has stated this belief, he then updates it, using his subjective precision q22S . A key
aspect that we wish to highlight is how the second action a22 of agent 2 relates to his first action
a21 and to the signal s2 he observes. Simple application of Bayes’s law yields:
Pr (V = 100|a1 , s2 ) Pr (s2 |V = 100, a1 ) Pr (V = 100|a1 )
= . (2)
Pr (V = 0|a1 , s2 ) Pr (s2 |V = 0, a1 ) Pr (V = 0|a1 )
Given the conditional independence of the signals, the expression simplifies to
Pr (V = 100|a1 , s2 ) Pr (s2 |V = 100) Pr (V = 100|a1 )
= , (3)
Pr (V = 0|a1 , s2 ) Pr (s2 |V = 0) Pr (V = 0|a1 )
that is, to
 2s2 −1
a22∗ q22S a21∗
= . (4)
100 − a22∗ 1 − q22S 100 − a21∗
Our main focus is on the weight agent 2 attaches to his own signal and how this depends on
the signal realization s2 and on agent 1’s action a1 . To this end, let us define this weight as the
value α22 that satisfies the following equality:
α22 ·(2s2 −1)
a22∗ q a21∗
= . (5)
100 − a22∗ 1−q 100 − a21∗
How α22 compares to 1 indicates, for each (s2 , a1 ), whether agent 2 attaches more (α22 > 1) or
less (α22 < 1) weight to his signal s2 relative to the objective precision of s2 (q ≡ 0.7). From (4)
and (5), α22 of the Bayesian agent is given by
q22S
ln
1−q22S
2
α2,B (s2 , a1 ; q22S ) ≡   . (6)
q
ln 1−q

From this expression, it is clear that while α2,B2 (s , a ; q 2S ) may differ from 1 whenever q 2S = q,
2 1 2 2
it is always the case that α2,B (s2 , a1 ; q2 ) is the same, irrespective of a1 and of s2 . In particular,
2 2S

we have proven the following result:

Proposition 1 (Symmetric Updating). In the PBE of the extended model, agent 2 attaches the
 
2 (s , a ; q 2S ) = ln q22S q
same weight α2,B 2 1 2 2S / ln 1−q to his private signal s2 for either real-
1−q2
ization 0 or 1 and for any a1 .

The signal realization s2 affects how agent 2 updates his belief (this can be seen in the expo-
nent 2s2 − 1 in (4)), but the weight attached to it as defined in (6) does not depend on the signal
realization. In particular, it is the same whether agent 2 observes a signal that contradicts agent
1’s action (e.g., s2 = 0 after a1 > 50) or a signal that confirms agent 1’s action (e.g., s2 = 1 after
a1 > 50). The fundamental reason for this “symmetric” updating is that, conditional on the value
of the good, the signal at t = 2 is independent of the signal at t = 1 (by experimental design) as
well as of the rationality of agent 1, which appears to derive from principles of logic.
So far we have considered risk neutral agents. The property of symmetric updating is, how-
ever, more general. Suppose agents have the following utility

7
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

⎛  ⎞
j 1+η
|V − at |
u(V , at ; η) = 100 ⎝1 − ⎠,
j
(7)
100

where η ∈ [0, ∞) is a risk preference parameter. For η = 1, u(V , at ; η) coincides with the mone-
tary payoff, that is, this is the case of risk neutrality. When η > 1 the agent is risk averse, whereas
a value η < 1 implies risk loving preferences.
By following the same steps as above, we obtain the relationship between the first and second
optimal actions of agent 2, a generalization of (4):
 η−1 (2s2 −1)
a22∗ q22S a21∗
= . (8)
1 − a22∗ 1 − q22S 1 − a21∗
In this extended model that allows for different risk preferences, by using (4) and (8), we obtain

q22S
ln
1−q22S
2
α2,B (s2 , a1 ; q22S , η) = η−1   . (9)
q
ln 1−q

We have therefore proven the following proposition:

Proposition 2 (Symmetric Updating for Different Risk Preferences). Suppose agent 2 has prefer-
ences represented by (7). In the PBE of the extended model with risk preferences, he attaches the
 
2 (s , a ; q 2S , η) ≡ η−1 ln q22S q
same weight α2,B 2 1 2 2S / ln 1−q to his private signal s2 for either
1−q2
realization 0 or 1 and for any a1 .

Intuitively, the effect of risk aversion is to lower the weight attached to the signal, as it makes
actions closer to 50 more attractive (compared to the case of risk neutrality). This, however, does
not affect the conclusion about the symmetry in updating, since it holds no matter what the signal
realization is.
While we have established that the symmetric updating result holds for a broad class of exten-
sions that adhere to the principle of Bayesian updating, it may be worth mentioning an incorrect
intuition that would suggest otherwise. After observing a contradicting signal (e.g., s2 = 0 after
a1 > 50), an agent would seem to be right in updating down his belief on the rationality of the
predecessor (the belief previously used when stating a21∗ ) and, as a result, in inferring more from
his own private signal, compared to the case of a confirming signal (e.g., s2 = 1 after a1 > 50).
Yet, with Bayesian updating, while it is true that agent 2 does update on the rationality of the pre-
decessor after observing his private signal, his posterior belief on the value of the good follows
expression (4), as results from the statistical property that, conditional on the value of the good,
the signal s2 is independent of the action a1 . In other words, the intuition is incorrect because it
implicitly amounts to incorrectly applying Bayes’s law.9

9 To see this, note that a Bayesian agent updates on the rationality of the predecessor after observing the action a and
1
receiving the signal s2 as follows: let μ2 (a1 , s2 ) = Pr(τr |a1 , s2 ),

μ2 (a1 , s2 ) k∈{0,100} Pr(τr |a1 , s2 , V = k) Pr(V = k|a1 , s2 )
=  = (10)
1 − μ (a1 , s2 ) k∈{0,100} Pr(τn |a1 , s2 , V = k) Pr(V = k|a1 , s2 )
2

8
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

In summary, we have shown that even enriching the model with risk aversion, subjective
precisions and subjective expectations on the rationality of the predecessor, the Bayesian model
dictates that the weight put on the signal does not depend on the signal realization. As we shall
see, this basic implication of Bayesian updating is not consistent with our experimental data.

2.3. Updating multiple priors

In the Introduction we mentioned the (incorrect) intuition that observing a signal contradicting
the first belief makes an agent update down on the predecessor’s rationality and put more weight
on his own signal. While this intuition is incorrect in the realm of Bayesianism, it is, however,
compatible with a model of updating in which an economic agent has multiple priors on the
predecessor’s rationality. In such a model, the own signal serves two purposes: it makes the
agent select the prior on the predecessor’s rationality, and, once this is done, update on the first
belief. We first review this method of updating referred to as the Likelihood Ratio Test Updating
rule, and then discuss an alternative popular method of updating referred to as Full Bayesian
Updating, which will be shown to imply markedly different predictions.

2.3.1. Likelihood ratio test updating

Compared to the previous Extended Model, we now introduce one modification: we allow
agent 2 to have multiple priors about θ = Pr(τ = τn ). Whereas a Bayesian agent has a unique
prior θ , an agent
 at time 2 has now ambiguous beliefs on θ , that is, multiple priors belonging to
the set θ∗ , θ ∗ ⊆ [0, 1], θ∗ < θ ∗ .
To update his belief upon observing an event E, first the agent selects one of the priors in the
set. If he is sufficiently confident that the event could occur conditional on the predecessor being
rational, he will pick up the lowest prior θ∗ , in the complementary case, he will pick up θ ∗ . In
other words,
Pr(E|τ = τr )
if ≥ c, then θ = θ∗ , and (11)
Pr(E|τ = τn )
Pr(E|τ = τr )
if < c, then θ = θ ∗ ,
Pr(E|τ = τn )
where c ∈ [0, ∞).
Pr(E|τ =τr )
We refer to this model of updating based on the likelihood ratio Pr(E|τ =τn ) as Likelihood Ratio
Test Updating (LRTU) rule. It can be seen as a simple generalization of the Maximum Likelihood
Updating (MLU) model, in which the time 2 subject estimates θ to be the value in θ∗ , θ ∗ that
maximizes the likelihood of observing the event E. Indeed, since
Pr(E) = Pr(E|τ = τr ) Pr(τ = τr ) + Pr(E|τ = τn ) Pr(τ = τn ),
that is,
Pr(E) = Pr(E|τ = τr )(1 − θ ) + Pr(E|τ = τn )θ ,


k∈{0,100} Pr(τr |a1 , V = k) Pr(V = k|a1 , s2 )
 .
k∈{0,100} Pr(τn |a1 , V = k) Pr(V = k|a1 , s2 )

If instead agent 2 after observing a1 and s2 first updated on the rationality of agent 1 and then plugged μ2 (a1 , s2 ) into (1)
to form a belief on V as in (4), he would violate Bayesian updating (as he would make the mistake of using the posterior
belief—after observing a1 and s2 —about the rationality of agent 1 as if it were the belief after observing a1 only).

9
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

according to the MLU rule, the agent chooses either θ∗ or θ ∗ , depending on whether the event is
more likely conditional on the predecessor being rational or noise. That is,
Pr(E|τ = τr )
if ≥ 1, then θ = θ∗ , and (12)
Pr(E|τ = τn )
Pr(E|τ = τr )
if < 1, then θ = θ ∗ .
Pr(E|τ = τn )
The LRTU model generalizes the MLU model to take into account that agents may need stronger
or weaker evidence in favor of one type in order to select a specific prior.10
Let us now describe the LRTU in the context of our model. In our model subject 2 makes
a decision twice, first after observing the event E ≡ {a1 } and then after observing the event
E ≡ {a1 , s2 }.11 Thus, there are two events in which agent 2 updates his belief about the rationality
of agent 1. Note also that after observing {a1 , s2 } the agent also uses the signal realization s2 to
update on the first belief.
From now on we consider the case in which a1 ∈ (50, 100]; the other case is analogous. In
Pr(E|τ =τr ) f (a1 )
this case, for E ≡ {a1 }, Pr(E|τ =τn ) = 2g(a1 ) ; hence,
 f (a1 )
θ∗ if ≥ c,
θ (a1 ) =
1 2g(a1 )
θ∗ if f (a1 )
2g(a1 ) < c.
Suppose now that agent 2 receives a confirming signal, s2 = 1, then E = {a1 , s2 = 1} and

Pr(E|τ = τr ) [q21S q22S + (1 − q21S )(1 − q22S )]f (a1 )

= .
Pr(E|τ = τn ) g(a1 )
If, instead, agent 2 receives a contradicting signal, s2 = 0, then E = {a1 , s2 = 0} and

Pr(E|τ = τr ) [q21S (1 − q22S ) + (1 − q21S )q22S ]f (a1 )

= .
Pr(E|τ = τn ) g(a1 )

10 This is equivalent to assuming that the subject acts as if he received another signal ϕ about the predecessor’s type
(and uncorrelated with the event). In this case, he would choose the prior to maximize the following probability:

Pr(E, ϕ) = Pr(E, ϕ|τ = τr ) Pr(τ = τr ) + Pr(E, ϕ|τ = τn ) Pr(τ = τn ).

That is, he would select θ = θ∗ (or θ = θ ∗ ) if the following inequality is (or is not) satisfied:

Pr(E, ϕ|τ = τr )
≥ 1,
Pr(E, ϕ|τ = τn )
that is,

Pr(E|τ = τr ) Pr(ϕ|τ = τr )
≥ 1,
Pr(E|τ = τn ) Pr(ϕ|τ = τn )
or

Pr(E|τ = τr ) Pr(ϕ|τ = τn )
≥ .
Pr(E|τ = τn ) Pr(ϕ|τ = τr )
=τn )
By setting Pr(ϕ|τ
Pr(ϕ|τ =τr ) ≡ c, one obtains the LRTU model.
11 Since a is a continuous variable, Pr({a }|T = t ) should be read as a conditional density function.
1 1 r

10
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Note that [q21S q22S + (1 − q21S )(1 − q22S )] > 12 > [q21S (1 − q22S ) + (1 − q21S )q22S ]: a confirming
signal makes it more likely that agent 1 is rational; conversely, a contradicting signal makes it
less likely.12
Therefore, we have the following proposition:

f (a1 )
Proposition 3 (Belief on the Predecessor’s Rationality with LRTU Rule). Suppose 2g(a1 ) ≥ c so
that θ 1 (a1 ) = θ∗ . Then
θ 2 (a1 , s2 = 1) = θ∗
and

θ∗ if [q21S (1 − q22S ) + (1 − q21S )q22S ] fg(a
(a1 )
≥ c,
θ (a1 , s2 = 0) =
2 1)
θ ∗ if [q21S (1 − q22S ) + (1 − q21S )q22S ] fg(a
(a1 )
1)
< c.

Suppose f (a1 )
2g(a1 ) < c so that θ 1 (a1 ) = θ ∗ . Then

θ 2 (a1 , s2 = 0) = θ ∗
and

θ∗ if [q21S q22S + (1 − q21S )(1 − q22S )] fg(a
(a1 )
≥ c,
θ (a1 , s2 = 1) =
2 1)
θ ∗ if [q21S q22S + (1 − q21S )(1 − q22S )] fg(a
(a1 )
1)
< c.

Suppose, as it will prove more relevant to interpret our data, that agent 2 initially trusts that
agent 1 is more likely to be rational and thus chooses θ∗ . Then Proposition 3 says that agent
2 must a fortiori do so when receiving a confirming signal, since the likelihood that agent 1 is
rational increases. This is not necessarily true after a contradicting signal, since the likelihood
that agent 1 is rational decreases. Thus, when agent 2 observes a confirming signal, he keeps
the same belief θ∗ about agent 1 being a noise agent, and, therefore, his two choices (one before
and one after his own signal) are made as if the agent 2 held a unique prior θ∗ , that is, as if he
were Bayesian. In contrast, when a contradicting signal is observed and agent 2 switches from
θ∗ to θ ∗ , his two choices are made as follows: first, his belief on the predecessor’s rationality is
computed as
(1 − θ∗ )f (a1 )
μ (a1 ) = ,
(1 − θ∗ )f (a1 ) + 2θ∗ g(a1 )
1 such that
and he makes the first action a2,L
1
a2,L μ (a1 ) q21S + (1 − μ (a1 ))g(a1 )
= ;
100 − a2,L
1 μ (a1 ) (1 − q21S ) + (1 − μ (a1 ))g(a1 )
then, after observing the signal s2 = 0, his belief on the predecessor’s rationality is

12 A simple proof of the inequality proceeds as follows. Observe that q S → q S q S + (1 − q S )(1 − q S ) is a strictly
1 1 2 1 2
increasing function for all q2S ∈ (0.5, 1] and thus in the considered range, q1S q2S + (1 − q1S )(1 − q2S ) is strictly larger than
q2S 12 + (1 − 12 )(1 − q2S ) = 12 .
Similarly, q1S → q1S (1 − q2S ) + (1 − q1S )q2S is strictly decreasing for all q2S ∈ (0.5, 1] and thus q1S (1 − q2S ) + (1 −
q1 )q2S < 12 .
S

11
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

(1 − θ ∗ )f (a1 )
μ2 (a1 , s2 = 0) = ,
(1 − θ ∗ )f (a1 ) + 2θ ∗ g(a1 )
2 such that
and he makes the second action a2,L
2
 2s2 −1
a2,L q22S a21∗∗
= , (13)
100 − a2,L
2 1 − q22S 100 − a21∗∗
where
1
a21∗∗ μ2 (a1 , s2 = 0) q21S + (1 − μ2 (a1 , s2 = 0))g(a1 ) a2,L
= < .
100 − a21∗∗ μ2 (a1 , s2 = 0) (1 − q21S ) + (1 − μ2 (a1 , s2 = 0))g(a1 ) 100 − a2,L
1

2 (s , a ; q 1S , q 2S ) be weight α 2 attached to s as
Similarly to the Bayesian model, let α2,L 2 1 2 2 2 2
defined by
2
a2,L α22 ·(2s2 −1) 1
a2,L
q
= ,
100 − a2,L
2 1−q 100 − a2,L
1

where agent 2’s action is now obtained using the LRTU rule. When θ (a1 ) = θ∗ and the signal
 
q22S q
is confirming (s2 = 1), it is immediate that α2,L
2 (s = 1, a ; q 1S , q 2S ) = ln
2 1 2 2 2S / ln 1−q , 1−q2
like in the Bayesian case (since in this case there is no change of prior belief θ∗ ). When, instead,
in response to a contradicting signal, agent 2’s belief on the predecessor’s rationality switches
from θ∗ to θ ∗ we obtain
⎡ ⎤ ⎛ 1 ⎞
q22S a2,L a21∗∗
ln ln − ln
⎢ 1−q2 2S ⎥ ⎜ 1
100−a2,L 100−a2 1∗∗ ⎟
2
α2,L (s2 = 0, a1 ; q21S , q22S ) = ⎢
⎣   ⎥
⎦ ·⎜
⎝1+ ⎟
⎠
q q22S
ln 1−q ln
1−q22S

q22S q
> ln / ln = α2,L
2
(s2 = 1, a1 ; q21S , q22S ).
1 − q22S 1−q
The greater distrust of agent 1’s rationality after a contradicting signal is equivalent to a larger
weight attached by agent 2 to his own signal. Agent 2’s weights α2,L 2 (s , a ; q 1S , q 2S ) are there-
2 1 2 2
fore no longer necessarily equal for the confirming and contradicting signal cases.
A similar analysis holds for the case in which 2g(af (a1 )
1)
< c, so that θ 1 (a1 ) = θ ∗ . Agent 2 can
revise up the trust about the predecessor’s rationality in response to a confirming signal s2 = 1,
while remaining distrustful if the signal is contradicting. We summarize these results in the next
proposition:

Proposition 4 (Asymmetric Updating with LRTU Rule). Under the LRTU updating rule, the
weight that agent 2 attaches to his signal satisfies the following conditions:
f (a1 )
(i) if 2g(a 1)
≥ c so that θ 1 (a1 ) = θ∗ ,

q22S q
2
α2,L (s2 = 0, a1 ; q21S , q22S ) ≥ α2,L
2
(s2 = 1, a1 ; q21S , q22S ) = ln / ln ,
1 − q2 2S 1 − q

where the inequality is strict if and only if [q21S (1 − q22S ) + (1 − q21S )q22S ] fg(a
(a1 )
1)
< c;

12
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

(ii) if f (a1 )
2g(a1 ) < c so that θ 1 (a1 ) = θ ∗ ,

q22S q
2
α2,L (s2 = 1, a1 ; q21S , q22S ) ≥ α2,L
2
(s2 = 0, a1 ; q21S , q22S ) = ln / ln ,
1 − q22S 1−q

where the inequality is strict if and only if [q21S q22S + (1 − q21S )(1 − q22S )] fg(a
(a1 )
1)
≥ c.

We now illustrate our results on the LRTU rule with a simple example in which agent 2 first
f (a1 )
trusts the predecessor’s rationality (i.e., 2g(a 1)
≥ c).
 
An Example Supposes that agent 2 has multiple priors θ∗ , θ ∗ = [0, 1] on the predecessor’s
type. He observes a1 = 70 and then the signal s2 = 0. Let us assume the threshold is c = 1, so
that the model is equivalent to the MLU model. Suppose he has expectations on the rational and
Pr(a1 =70|τ =τr )
noise types’ actions at time 1 such that Pr(a 1 =70|τ =τn )
≥ 1. In this case, the agent selects the prior
θ∗ = 0. The agent is confident about the predecessor’s rationality, and, therefore, chooses a2,L 1 =

70. After receiving the signal s2 = 0, the agent now reassesses the predecessor’s rationality.
The probability of observing an action greater than 50 and a negative signal conditional on the
Pr(a1 =70,s2 =0|τ =τr )
predecessor being rational is now lower. If, in particular, Pr(a 1 =70,s2 =0|τ =τn )
< 1, then the agent
∗
chooses θ = 1. Being now confident that the predecessor was a noise type, the agent considers
a1 = 70 completely uninformative, which would imply a belief of 0.5 on V = 100. On top of
this, the subject has observed a bad signal: by applying Bayes’s rule to a belief of 0.5, the subject
obtains a posterior belief of 0.3 on the value being 100 and, as a result, chooses a2,L 2 = 30.

In terms of our previous analysis, this is equivalent to a subject overweighting the signal, with
70
(1−0.7)2 100
α22 = 2, since 30 = 100 70

70
. A similar analysis applies to the case in which
(1−0.7) 100 +0.72 1− 100
2

Pr(a1 =70|τ =τr )

the subject observes a signal s2 = 1. It is easy to see that if Pr(a1 =70|τ =τn ) ≥ 1, then a fortiori
Pr(a1 =70,s2 =1|τ =τr )
Pr(a1 =70,s2 =1|τ =τn ) ≥ 1 (as per Proposition 3). Therefore, in this case the subject sticks to the
prior θ∗ = 0. Since the agent is still confident that the predecessor was rational, he does not
change his first belief on V = 100, which remains 0.7. Since the agent has observed a good
signal, by applying Bayes’s rule to a belief of 0.7, he obtains a posterior belief of 0.84 on the
2 = 84. This is equivalent to an agent weighing the
value being 100 and, as a result, chooses a2,L
signal as a Bayesian agent would do, with α22 = 1.

2.3.2. Full Bayesian updating

The LRTU (MLU) is only one family of updating rules for multiple priors. Another, well
known, family goes under the name of Full Bayesian Updating (FBU) (for a discussion, see
Gilboa and Marinacci, 2013). We now show that this updating rule does not produce the same
asymmetry we illustrated for LRTU (MLU), and is not compatible with the intuition mentioned
in the Introduction.
Here we consider the FBU model with maxmin preferences, axiomatized by Pires (2002).
From now on, our statements about the FBU model should be read as statements about the FBU
rule with maxmin preferences. In the Appendix, we will consider FBU with Hurwitz preferences
as an alternative to the maxmin preferences (and we will show that this model is not compatible
with the asymmetry explained before either).
Let us consider, as usual, the case in which a1 ∈ (50, 100].
 Suppose,
 similarly to the previous
section, that an agent at time 2 starts with a set of priors θ∗ , θ ∗ on the proportion of noise-type

13
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

agents at time 1.
 Upon observing the predecessor’s action a1 , the agent applies Bayes’s rule to
each prior θ in θ∗ , θ ∗ and obtains the following likelihood ratio:

p21 (θ, q21S ) Pr(V = 100|a1 ; θ, q21S ) (1 − θ )q21S f (a1 ) + g(a1 )θ

≡ = . (14)
1 − p21 (θ, q21S ) Pr(V = 0|a1 ; θ, q21S ) (1 − θ )(1 − q21S )f (a1 ) + g(a1 )θ
According
 to the FBU, the range  of θ spans a range of beliefs on the value of the good being
100: p21 (θ ∗ , q21S ), p21 (θ∗ , q21S
 . 
)
For any given prior θ in θ∗ , θ ∗ , after observing a signal s2 the agent updates his likelihood
ratio as follows:
 2s2 −1
p22 (θ, q21S , q22S ) (1 − θ )q21S f (a1 ) + g(a1 )θ q22S
= .
1 − p22 (θ, q21S , q22S ) (1 − θ )(1 − q21S )f (a1 ) + g(a1 )θ 1 − q22S

 2Again, the range of θ spans a range of beliefs on the value of the good being 100:
p2 (θ ∗ , q21S , q22S ), p22 (θ∗ , q21S , q22S ) .  
A maxmin expected utility agent with a set of beliefs p, p̄ chooses the optimal action
amax min such that
 
amax min = arg max min   Ep 100 − 0.01(V − a) ,
2
a
p∈ p,p̄

that is,
⎧
⎨ 100p, if p > 12 ,
amax min = 50, if p ≤ 12 and p̄ ≥ 12 ,
⎩
100p̄, if p̄ < 12 .

Therefore, for the case a1 ≥ 50, since the lower bound of the beliefs satisfies p = p21 (θ ∗ , q21S ) ≥
∗
2 , the agent’s first action is based on the most pessimistic prior, θ = θ :
1

1
a2,F = 100p21 (θ ∗ , q21S ).
This action equals to that of the Bayesian model with θ = θ ∗ . The agent acts as if he distrusted
the predecessor’s rationality (a pessimistic agent maximizes the minimum payoff he can obtain,
and the minimum payoff coincides with the predecessor’s action being the least informative).
Similarly, the second action is
⎧
⎨ 100p22 (θ ∗ , q21S , q22S ), if p22 (θ ∗ , q21S , q22S ) > 12 ,
a2,F =
2
50, ∗
if p2 (θ , q2 , q22S ) ≤ 12 , and p22 (θ∗ , q21S , q22S ) ≥ 12 ,
2 1S
⎩ 2 1S 2S
100p2 (θ∗ , q2 , q2 ), if p22 (θ∗ , q21S , q22S ) < 12 .
After receiving a confirming signal, the agent keeps choosing according to the Bayesian model
with the same prior θ = θ ∗ . Instead, after receiving a contradicting signal he chooses as in the
Bayesian model but with θ ≤ θ ∗ , since now the worst-case scenario is that in which the prede-
cessor’s action was the most informative (if p22 (θ∗ , q21S , q22S ) ≤ 12 he chooses as in the Bayesian
model with θ = θ∗ ).
It is useful to rewrite these conclusions in terms of weight on the private signal, and contrast
them to those we obtained for the LRTU model (as shown in Proposition 4).

14
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

2 (s , a ; q 1S , q 2S ) be the weight attached to s as defined by

Let α2,F 2 1 2 2 2

2
a2,F α22 ·(2s2 −1) 1
a2,F
q
= ,
100 − a2,F
2 1−q 100 − a2,F
1

where agent 2’s action is now obtained using the FBU rule. When agent 2 receives a confirm-
 
2 (s = 1, a ; q 1S , q 2S ) = ln q22S q
ing signal, α2,F 2 1 2 2 2S / ln 1−q since the two actions by agent 2
1−q2
(before and after receiving the signal) are based on the same prior belief θ = θ ∗ (note though
that this is not necessarily the same prior belief as in the LRTU model). When, instead, agent 2
receives a contradicting signal s2 = 0, α2,F 2 (s = 0, a ; q 1S , q 2S ) is smaller than in the Bayesian
2 1 2 2
case:
⎡ ⎤ ⎛ 2 ⎞
q22S a2,F a22∗∗
ln ln − ln
⎢ 1−q22S ⎥ ⎜ 100−a2,F2 100−a22∗∗ ⎟
2
α2,F (s2 = 0, a1 ; q21S , q22S ) = ⎢
⎣   ⎥ · ⎜1 −
⎦ ⎝
⎟
⎠
q q22S
ln 1−q ln
1−q22S

q22S q
≤ ln / ln = α2,F2
(s2 = 1, a1 ; q21S , q22S ),
1 − q2 2S 1 − q

where a22∗∗ is the second action of agent 2 he would take if he updated a2,F 1 in the Bayesian
2
way (i.e, the action identical to a2,L obtained in the expression (13) above). Here, the inequality
2 ≥ a 2∗∗ and is strict when p 2 (θ ∗ , q 1S , q 2S ) < 1 . Hence, the FBU model yields
follows from a2,F 2 2 2 2 2
an asymmetric way of updating, as summarized in the next proposition.

Proposition 5 (Asymmetric Updating with FBU Rule). Under the FBU rule, the weight agent 2
attaches to his private signal depends on its realization and satisfies

q22S q
ln / ln = α2,F
2
(s2 = 1, a1 ; q21S , q22S ) ≥ α2,F
2
(s2 = 0, a1 ; q21S , q22S ),
1 − q2 2S 1−q

where the inequality is strict if and only if p22 (θ ∗ , q21S , q22S ) < 12 .

Essentially, in the FBU model the agent updates like a Bayesian with a prior belief θ ∗ when
receiving a confirming signal, while he puts a lower weight on his own signal when the signal is
contradicting. This is in stark contrast with the LRTU model (case i of Proposition 4), in which
the agent updates like a Bayesian with prior belief θ∗ when receiving a confirming signal, while
he puts a higher weight on his own signal when it is contradicting.
The essential observation from this analysis, which we will use for the interpretation of our
experimental data, is that neither the Bayesian model nor the FBU model can explain that the
weight on the own signal is greater upon receiving a contradicting signal than upon receiving
a confirming signal, whereas the LRTU model can. For theoretical purposes, one may further
observe that there is a similarity between the weights in the FBU model and in the LRTU model
when the agent originally distrusts the predecessor’s rationality (case ii of Proposition 4). Indeed,
for both models the weight on the contradicting signal is lower than on the confirming signal.
But note the difference in this case too. Take the case in which q22S = q. In the FBU model, the
agent updates like a Bayesian (he puts a weight 1 on the own signal) for the confirming case,

15
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

while he puts a weight lower than 1 for the contradicting case. In the case ii of Proposition 4 for
the LRTU rule, instead, the agent, after distrusting the predecessor, puts a weight higher than 1
on a confirming signal and equal to 1 on a contradicting signal.
We now illustrate the working of FBU in our model with a simple example.
 
An Example Supposes that agent 2 has multiple priors θ∗ , θ ∗ = [0, 1] on the predecessor’s
type. He observes a1 = 70. The agent updates his belief on the value of the good using each
prior θ ∈ [0, 1]. This means that his posterior beliefs on V = 100 lie in [0.5, 0.7]. Therefore, he
1 = 50, the action that maximizes the minimum payoff he can obtain. After receiving
chooses a2,F
the contradicting signal s2 = 0, the subject updates his set of beliefs to [0.3, 0.5]. This implies
2 = 50, which is equivalent to α 2 = 0. After receiving the signal s = 1,
that again he chooses a2,F 2 2
instead, the agent updates his set of beliefs to [0.7, 0.84]. He will then maximize his minimum
payoff by choosing a2,F2 = 70, which is equivalent to α 2 = 1. Note that in this example this
2
updating rule implies no updating at all (rather than overweighting the signal) after receiving
a contradicting signal, and updating as a Bayesian after observing a confirming signal. This
asymmetric way of updating is sharply different from that in the previous example about the
LRTU rule.

3. The experiment and the experimental design

3.1. The experiment

We ran the experiment in the Experimental Laboratory for Finance and Economics (ELFE) at
the Department of Economics at University College London (UCL) in 2009, 2010, 2011, 2014,
2018 and 2019.13 The subject pool mainly consisted of undergraduate students in all disciplines
at UCL. They had no previous experience with this experiment. In total, we recruited 287 stu-
dents. Each subject participated in one session only.
The sessions started with written instructions given to all subjects. We explained to partici-
pants that they were all receiving the same instructions. Subjects could ask clarifying questions,
which we answered privately.
Here we describe the baseline treatment (SL1). In the next section, we will explain the exper-
imental design. We ran five sessions for this treatment. In each session we used 10 participants.
The procedures were the following:

1. Each session consisted of fifteen rounds. At the beginning of each round, the computer pro-
gram randomly chose the value of a good. The value was equal to 0 or 100 with the same
probability, independently of previous realizations.
2. Participants were not told the value of the good. They knew, however, that they would receive
information about the value, in the form of a symmetric binary signal. If the value was equal
to 100, a participant would receive a “green ball” with probability 0.7 and a “red ball” with
probability 0.3; if the value was equal to 0, the probabilities were inverted. That is, the green
signal corresponded to the good signal, st = 1, and the red signal to the bad signal, st = 0.

13 ELFE is the new name of the laboratory, formerly known as ELSE Laboratory. The SL treatments (described below)
were conducted with a built-on-purpose program. The IDM and GC treatments were programmed and conducted with
the software z-Tree (Fischbacher, 2007) in 2014 and 2019.

16
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

3. A subject was randomly chosen to make a decision. He received a signal and chose a number
between 0 and 100, up to two decimal places. The other subjects observed the decision made
by the first subject on their screen. The identity of the subject was not revealed.
4. In a second period, another subject was randomly chosen and asked to choose a number
between 0 and 100, having observed the first subject’s choice only.
5. After he had made that choice, he received a signal (another “ball draw” from the same urn
with replacement) and had to make a second decision. This time, therefore, the decision was
based on the observation of the predecessor’s action and of the private signal.
6. The experiment then continued with a third, fourth,..., tenth period, until all 10 subjects had
acted.14
7. At the end of the round, after all 10 subjects had made their decisions, subjects observed
a feedback screen, in which they observed the value of the good and their own payoff for
that round. The payoffs were computed as 100 − 0.01(V − at )2 of a fictitious experimental
currency called “lira.” After participants had observed their payoffs and clicked on an OK
button, the software moved to the next round.

Note that essentially we asked subjects to state their beliefs. To elicit the beliefs, we used a
quadratic scoring function, a quite standard elicitation method. In the instructions, we followed
Nyarko and Schotter (2002) and explained to subjects that to maximize the amount of money
they could expect to gain, it was in their interest to state their true belief.15
To compute the final payment, we randomly chose (with equal chance) one round among the
first five, one among rounds 6 − 10 and one among the last five rounds. For each of these rounds
we then chose either decision 1 or decision 2 with equal chance (with the exception of subject 1,
who was paid according to the only decision he made in the round). We summed up the payoffs
obtained in these decisions and then converted the sum into pounds at the exchange rate of 100
liras for £7. Moreover, we paid a participation fee of £5. Subjects were paid in cash, in private,
at the end of the experiment. On average, in this treatment subjects earned £21 for a 2 hour
experiment.

3.2. Experimental design

This section describes the main features of each treatment, as summarized in Table 3.1.
Social Learning (SL). In addition to the social learning treatment (SL1) just described, we
ran a second treatment (SL2) which only differed from the first because the signal had a preci-
sion which was randomly drawn in the interval [0.7, 0.71] (instead of having a precision always
exactly equal to 0.7). Each subject observed not only the signal (the “ball color”) but also the
exact precision of his own signal.16 A third treatment (SL3) was identical to SL2, with the ex-
ception that instead of having sequences of 10 subjects, we had sequences of 4 subjects. Given

14 The experiment was designed to address many research questions. The data collected on periods beyond 2 are not
relevant for this paper’s research question. They are analysed in the related study by Angrisani et al. (2021).
15 This explanation helps the subjects, since they do not have to solve the maximization problem by themselves (and
to which extent they are able to do so is not the aim of this paper). For a discussion of methodological issues related to
elicitation methods, see Schotter and Trevino (2014).
16 Drawing the precision from the tiny interval [0.7, 0.71], instead of having the simpler set up with fixed precision equal
to 0.7, was only due to a research question motivated by the theory of Guarino and Jehiel (2013), where the precision
is, indeed, supposed to differ agent by agent; this research question, however, is not the object of this paper (since it
becomes relevant only for periods beyond the second).

17
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Table 3.1
Treatments’ main features.
Treatments Signal Precision Sequence Subjects in a group Groups Participants Rounds
SL1 0.7 10 10 5 50 15
SL2 [0.7,0.71] 10 10 5 49 15
SL3 [0.7,0.71] 4 4 5 20 30
IDM1 0.7 1 or 2 - - 36 30
IDM2 0.7 2 - - 32 15
GC 0.7 2 10 5 56 15
SL: Social Learning; IDM: Individual Decision Making; GC: Guess the Color. Note that in SL2 there are 49 subjects
since one session was run with 9 participants rather than 10 due to a last minute unavailability of one subject.

the smaller number of subjects, each round lasted less time, obviously; for this reason, we de-
cided to run 30 rounds per session, rather than 15. The results we obtained for times 1 and 2 for
these three treatments are not statistically different (as we show in the next section and in the Ap-
pendix). For the purposes of this paper, we consider the three treatments as just one experimental
condition. We will refer to it as the SL treatment.
Individual Decision Making (IDM). In the social learning treatments subjects make deci-
sions after observing private signals and the actions of others. We may expect departures from
the PBE even independently of the social learning aspect if subjects do not update in a Bayesian
fashion. To control for this, we ran two treatments in which subjects observed more than one
signal and made more than one decision. In a first treatment (IDM1), a subject received a signal
and had to make a choice in the interval [0, 100] (as subject 1 in the SL treatments). Then, with a
50% probability, he received another signal and had to make a second decision (like the second
decision of subject 2 in the SL treatments).17 In a second treatment (IDM2) we wanted to take
care of another concern. In social learning, subject 2 observes a signal (subject 1’s action) about
another signal (subject 1’s ball color): could it be that the behavior we observe is due to this sort
of compound lottery? In the IDM2 treatment, the computer drew a first signal. The subject did
not observe it; rather, he observed a computer’s “announcement” about this signal. Such an an-
nouncement coincides with the true signal with probability 0.75 and is equal to the other signal
with probability 0.25 (it is important to remark that these probabilities are known to the subject).
After observing this announcement, the subject had to make a choice in the interval [0, 100].
Then, he received another signal and had to make a second decision (similarly to the SL treat-
ments). The announcement is essentially a signal with precision 0.6 (= 0.7 ∗ 0.75 + 0.3 ∗ 0.25).
Following the same logic of our basic model, in the PBE the first action is 0.6 or 0.4 (for a good
or bad signal, respectively). The second action is 0.22 after a bad announcement and bad signal;
0.39 after a good announcement and a bad signal; 0.61 after a bad announcement and a good
signal; 0.78 after a good announcement and good signal. Allowing for risk aversion and subjec-
tive precisions can change the specific predictions but does not alter the result for the symmetric
updating, as we proved in Proposition 2.
Guess the Color (GC). In the social learning treatments, subject 2 observes subject 1’s action,
can infer his signal, and, presumably, on the basis of this, state the value of the good. We also
ran a treatment in which we ask subject 2 to report his belief on the signal observed by his

17 We did not ask subjects to make a second decision in all rounds so that the first decision was made without knowing
whether there would be a second decision; with this method, the subject was in a condition, we believe, very similar to
that of subject 1 in the SL treatments.

18
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

predecessor. Specifically, subject 1 had exactly the same task as in all other treatments, that is,
to state his belief in the interval [0, 100] after observing a signal. Subject 2 observed this choice
and was asked to choose a number in [0, 100], expressing the probability that the first subject had
observed a good signal (green ball). Then he received a signal and had to make the same choice
again. The elicitation method was identical to the other treatments. We ran 5 sessions for this
treatment, using the same protocol as in Treatment SL and with a sequence of only two subjects.
In each session there were 10 or 12 subjects and in each round they were randomly matched in
pairs: one would be the first in the sequence and one the second. 18
Finally, to study the robustness of our results to different specifications, we also ran a treatment
in which 44 subjects received signals of stochastic precision in (0.5, 1]. The precision was only
known to the subject receiving the signal. We used sequences of two subjects (like in the GC
treatment). In the interest of space, we discuss this treatment in the Appendix.

4. Results of the SL treatments

Our main interest is to understand how human subjects weigh their private signal relative
to the public information contained in the observation of the predecessor’s action. To this aim,
we focus on subjects’ second decisions at time 2, when both private and public information are
available. Before doing so, however, let us very briefly discuss the decisions of subjects at time
1 (who have only observed a private signal) and the first decisions of subjects at time 2 (based
on the observation of their predecessor’s choice only). The interested reader can find a detailed
analysis in the Online Appendix
In summary,
1) at time 1, the median action is in line with Bayesian updating (30 and 70), for both signal
realizations (bad and good). There is a lot of heterogeneity, but rarely subjects update in the
wrong direction (i.e., choosing a1 < 50 after receiving s1 = 1), as shown in Fig. 4.1.
2) at time 2, the distribution of the first actions is different from that at time 1, with higher
frequencies in the middle of the support (see Fig. 4.2). This reflects the fact that subjects seem
to “discount” the information contained in the predecessor’s action, to some extent. Indeed, a
deeper analysis shows that the discounting occurs when the observed action is close to 50 (a1 ∈
[50, 66.7) or, symmetrically, a1 ∈ (33.3, 50]) or not too extreme (a1 ∈ [66.7, 83.4) or, symmet-
rically, a1 ∈ (16.6, 33.3]). For more extreme values (a1 > 83.4 or a1 < 16.6), subjects at time 2
trust more the predecessor’s action in terms of reflecting the correct signal realization.19

4.1. How do subjects weigh their signal relative to their predecessor’s action?

Let us now study how subjects update their beliefs upon receiving the signal at time 2, that
is, let us focus on the second action at time 2. We will refer to the first action that subjects take
at time 2 as their “first belief” and to the second as their “posterior belief.” Fig. 4.3 shows the
frequency of posterior beliefs conditional on whether the subject received a signal confirming his
first belief (i.e., s2 = 1 after an action a21 > 50 or s2 = 0 after an action a21 < 50) or contradicting

18 The total number of participants was 56 since in two sessions we recruited 10 subjects and in three sessions we
recruited 12 subjects (which made the probability of a subject being matched with the same other participant even
lower). The exchange rate for this treatment was 100 liras for £6.
19 We report here the specific intervals chosen in our analysis for illustration. Different partitioning of the interval does
not alter the results.

19
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Fig. 4.1. Distribution of actions at time 1. The top (bottom) panel refers to actions upon receiving s1 = 1 (s1 = 0).

Fig. 4.2. Distribution of first actions at time 2 (the top panel refers to a1 > 50, the middle to a1 = 50 and the bottom to
a1 < 50).

20
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Fig. 4.3. Distribution of a22 for confirming (top panel) or contradicting (bottom panel) s2 .

it (i.e., s2 = 1 after an action a21 < 50 or s2 = 0 after an action a21 > 50).20 The figure is obtained
(and all the following analysis is conducted) after transforming an action a21 into 100 − a21 and
the corresponding signal s1 into 1 − s1 whenever a1 < 50.
If subjects acted as in the PBE of the basic model (with objective precision, risk neutrality and
no doubt about the rationality of their predecessor), in the case of confirming signal we would
observe the entire distribution concentrated on 84. The empirical distribution shows much more
heterogeneity. Nevertheless, the median action as well as the mode is indeed close to this predic-
tion. For the contradicting signal, the picture is rather different. Whereas in the PBE we would
observe the entire distribution concentrated on 50, the empirical distribution looks very asym-
metric around 50, with more than 70% of the mass below 50. To quantify these observations,
we compute the weight that the subject puts on his signal by using the notation of Section 2 (to
alleviate notation, we omit the arguments of α22 ):

20 In this analysis we exclude the cases in which the action at time 1 was a = 50, since observing a 50 is uniformative.
1
We do study the case in which a subject at time 2 observed an informative action at time 1 and chose a21 = 50; in this
case we distinguish whether the action observed at time 1 confirmed or contradicted the realization of the signal s2 . Note
that an alternative definition of confirming and contradicting signal would be in reference to a1 rather than to a21 . This
would not affect our results, since the difference is in one observation only (in which a21 > 50 and a1 < 50).

21
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Table 4.1
Distribution of weights on the own signal in the SL treatment.
1st Quartile Median 3rd Quartile
α22 0.69 1.16 2.03

α22 (upon observing confirming signal) 0.55 0.96 1.34

α22 (upon observing contradicting signal) 1.00 1.70 2.72
The table shows the quartiles of the distribution of the weight on the own signal for the second action at time 2 in the SL
treatment. The data refer to all cases in which the first action at time 2 was different from 50.

2 a1 2 a21
a22 0.7α2 100
2
(1 − 0.7)α2 100
= s1 +(1−s1 ) .
100 2 a1 2 a21 2 a21 2 a21
0.7α2 100
2
+ (1 − 0.7)α2 1 − 100 (1 − 0.7)α2 100 + 0.7α2 1 − 100

(15)
In this formulation α22 can be thought of as expressing subjective precisions (like in expression
(6)) or as reflecting both subjective precisions and risk preferences (like in expression (9)). We
prefer to think of α22 as expressing subjective beliefs only, but this is a matter of interpretation.21
In this interpretation, for α22 = 1, expression (15) gives the Bayesian updating formula, and
so α22 = 1 is the weight that a Bayesian agent would put on the signal. A value higher (lower)
than 1 indicates that the subject overweights (underweights) the signal. For instance, for α22 = 2,
the expression is equivalent to Bayesian updating after receiving two conditionally independent
signals and can, therefore, be interpreted as the action of a Bayesian agent acting upon receiving
two signals (with the same realization). A subject that does not put any weight on the signal
(α22 = 0) does not update at all upon observing it (a22 = 50), whereas a subject who puts an
infinite weight on it chooses an extreme action (a22 = 0 or a22 = 100), as if he were convinced
that the signal fully reveals the value of the good. Finally, a negative value of α22 indicates that
the subject misreads the signal, e.g., interpreting a good signal as a bad one.
Table 4.1 reports the results.22 While in the case of a confirming signal the median subject
puts only a slightly lower weight on the signal than a Bayesian agent would do, in the case of a
contradicting signal the weight is considerably higher, 1.70.23 The different weight is observed
also for the first and third quartiles. Essentially, subjects update in an asymmetric way, depending
on whether the signal confirms or not their first beliefs: contradicting signals are overweighted
with respect to Bayesian updating.24
To sharpen our understanding of the behavior in the SL treatment, we now look at how the
weight on the signal changes with the first belief. Table 4.2 reports the quartiles for α22 for three

21 In the laboratory, subjects may choose different actions for many different reasons, including errors in updating
beliefs or implementing the actions, risk preferences, subjective beliefs. As we said, we find subjective precisions a
simple approach to capture all these departures, in the spirit of Savage.
22 The value of α 2 is undetermined when a 1 = 100, therefore we exclude these cases. When a 2 = 100, we approximate
2 2 2
it to 99.99.
23 We ran a Mann-Whitney U test (Wilcoxon rank-sum test) on the median weight for the confirming and contradicting
signal; we can reject the null hypothesis that their distribution is the same (p-value = 0.000002).
24 As we said, our results do not change if we define the signal as contradicting or confirming with respect to the action
a1 rather than with respect to the first belief a21 , since the difference is for one observation only. Moreover, we cannot
reject the hypothesis that the results, both for confirming and contradicting signals, are the same for the three treatments
SL1, SL2 and SL3. (see the Appendix for details).

22
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Table 4.2
Distribution of weights for second actions at time 2 in the SL treatment.
1st Quartile Median 3rd Quartile
α22 (upon observing confirming signal)
Conditional on 50 < a21 ≤ 66.7 0.66 0.97 1.16
Conditional on 66.7 < a21 ≤ 83.4 0.18 0.91 1.55
Conditional on a21 > 83.4 0.54 2.10 3.74

α22 (upon observing contradicting signal)

Conditional on 50 < a21 ≤ 66.7 0.55 0.96 1.60
Conditional on 66.7 < a21 ≤ 83.4 1.03 1.68 2.11
Conditional on a21 > 83.4 2.72 3.34 4.23
The table shows the quartiles of the distribution of weights for second actions at time 2, conditional on different values
of the first belief.

different classes of a21 . As one can immediately observe, the asymmetry occurs for the last two
classes, but not for the first.
As we mentioned above, the median subject chose an action a21 > 66.7 mainly when he ob-
served an action at time 1 greater than the theoretical Bayesian decision. These are cases in
which the subject “trusted” the predecessor. These are also the cases in which subjects update in
an asymmetric way.
This result is in line with the LRTU rule, as shown in case (i) of Proposition 4. Essentially,
subjects at time 2 who trust the predecessor’s rationality, stop trusting it after a contradicting
signal. This asymmetric updating is not in agreement with the Bayesian model, even considering
the extended one, and cannot be explained by the FBU rule. In Section 6, we will perform an
econometric comparison of the different models of updating, taking into account the heterogene-
ity of actions, and we will show that indeed the LRTU model is the best fit of the data. Before
we do that, let us make some important observations and (in the next section) discuss the results
of our control treatments.
First, our result cannot be explained by and does not fall into categories of psychological
biases sometimes invoked in decision making under uncertainty such as the base rate neglect or
the confirmatory bias. Base rate neglect in our experiment would mean neglecting the first belief
once the new piece of information (the private signal) is received. With such a bias, we should
expect that the median choice of subjects first observing an action a1 > 50 and then a signal
s2 = 1 be equal to that at time 1 after observing a signal s1 = 1, which is not the case (this would
be equivalent to α22 equal to 0, whereas it is slightly greater than 1). As for the confirmatory bias,
if subjects had the tendency to discard new information in disagreement with their original view,
and only accept information confirming their original opinion (the definition of confirmatory
bias), they would ignore (i.e., not update upon receiving) a contradicting signal, in sharp contrast
with our results. Note that had we inverted the order in which information is presented to subjects
in the experiment (i.e., first the private signal and then the predecessor’s action) we would not
have been able to rule out this possibility.25

25 It is also worth mentioning that whereas in the social learning literature, as in much psychological literature, re-
searchers have talked about “overconfidence” (see Section 7), in other experiments subjects show “underconfidence.” In
particular, in experiments on decision making with naive advice, it has been observed that “when given a choice between
getting advice or the information upon which the advice is based, subjects tend to opt for the advice, indicating a kind of

23
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Second, if one thinks that the only inference subjects had to make from the predecessor’s
action was the predecessor’s signal realization (and not the precision, since it was known), it is
even more surprising that subjects simply did not choose 50 after a contradicting signal, since
the fact that a good and a bad piece of information “cancel out” does not require sophisticated
understanding of Bayes’s rule.26

5. Controls: the IDM treatments

One may wonder whether the asymmetric updating result is due to the social learning aspect
of our experiment or, instead, is just the way human subjects update upon receiving two con-
secutive signals (possibly corresponding to a psychological bias other than base rate neglect or
confirmation bias as mentioned above). To tackle this issue, as we explained, we ran two, indi-
vidual decision making, control treatments. Table 5.1 shows the results of the IDM1 treatment,
in which subjects receive two consecutive signals of the same precision 0.7. As one can see, the
asymmetry and the overweight of the contradicting signal are not present in this treatment: the
median weight is equal to 1 for the contradicting signal and a bit higher for the confirming signal
(it should be observed, though, that the order for the first quartile is reversed). Table 5.2 shows
the results of the IDM2 treatment. Remember that in such a treatment we want to mimic the fact
that in social learning subject 2 only observes “a signal (subject 1’s action) about a signal (subject
1’s ball color).” In the IDM2 treatment a subject first observes a computer announcement about
the signal (color of the ball) and then directly observes the color of a second ball drawn from the
same urn. Again, there is no asymmetry in updating. Subjects slightly overweigh the signal they
observe directly, but this slight overweight does not depend on whether the signal confirms or
contradicts the announcement (we cannot reject the hypothesis that the two median weights are
the same; Mann-Whitney U test, p-value = 0.27).

Table 5.1
Distribution of weights on the own signal in the IDM1 treatment.
1st Quartile Median 3rd Quartile
α22 0.64 1.08 2.06

α22 (upon observing confirming signal) 0.64 1.30 2.48

α22 (upon observing contradicting signal) 0.94 1.00 1.74
The table shows the quartiles of the distribution of the weight on the own signal for the action at time 2 in the IDM1
treatment. The data refer to all cases in which the action at time 1 was different from 50.

From these two control treatments, we can conclude that the asymmetric updating we observe
in the SL treatment is not just a deviation from Bayesian updating after receiving two pieces of
information; it is, instead, inherent to the social learning aspect of our study.

underconfidence in their decision making abilities [...]” (Schotter, 2003). Our result is again not explained by this type of
bias.
26 One could observe that if a subject chose, e.g., a 1 = 84 and then, after receiving a bad signal, a 2 = 50, the cor-
2 2
responding α22 would be 2, which is compatible with the overweight we documented. It must be noticed, though, that
if we exclude the cases in which a22 = 50, the asymmetry remains and is actually even stronger (see Table B.3 in the
Appendix).

24
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Table 5.2
Distribution of weights on the own signal in the IDM2 treatment.
1st Quartile Median 3rd Quartile
α22 0.76 1.24 2.03

α22 (upon observing confirming signal) 0.71 1.30 2.11

α22 (upon observing contradicting signal) 0.81 1.24 1.73
The table shows the quartiles of the distribution of the weight on the own signal for the action at time 2 in the IDM2
treatment. The data refer to all cases in which the action at time 1 was different from 50.

6. Econometric analysis

We now estimate the three models of updating presented in Section 2, the Bayesian Updating
(BU) model, the LRTU model and the FBU model, with the data of the SL treatments. We will
then perform a formal statistical comparison among them to quantify the evidence in favor of
one model against the others.
As we know from the descriptive statistics, there is a lot of heterogeneity in the actions. We
also know from the description of the three models that subjective precisions play a big role. To
take this into account we use the data from the IDM1 treatment, in which we can infer subjective
signal precisions directly from the chosen actions. Inferring some features of a model directly
from the data of an auxiliary treatment in lieu of imposing extra parametric assumptions is, to
the best of our knowledge, another novel contribution of our work to the experimental literature.
For the BU model, we estimate the parameter θ ; for the LRTU model we estimate the param-
eters (θ∗ , θ ∗ , c); finally, for the FBU model we estimate the parameters (θ∗ , θ ∗ ). We obtain the
distribution of actions f and g as described in Section 2 directly from the subjects’ actions at
time 1.
We proceed to the estimation of the parameters by the generalized method of moments
(GMM). In each of our models, the heterogeneity in the subjective precision of signals (q21S , q22S )
induces a distribution of weights of updating of agent 2 for any fixed value of the parameters. The
estimation strategy consists in finding the parameter values such that the distribution of weights
predicted by a model is closest to the actual distribution of weights. With maximum likelihood,
we would need to specify a parametric distribution for (q21S , q22S ) to quantify fitness of the distri-
bution of predicted weights of updating to the observed one. In our IDM experiments, however,
we do observe the empirical distribution of (q21S , q22S ). We make use of this empirical distri-
bution to construct a nonparametric distribution of predicted weights of updating, and match it
with the observed one by GMM. That is, our GMM estimation procedure does not rely on any
parametric assumptions. We have a gain in terms of robustness of the estimates, with a potential
sacrifice in terms of efficiency relative to a correctly specified parametric maximum likelihood
procedure.
As we did in our descriptive analysis, we focus on the three quartiles of the empirical distri-
bution of the updating weights for a) the first action at time 2; b) the second action at time 2,
conditional upon receiving a confirming signal; c) the second action at time 2, conditional upon
receiving a contradicting signal. In each of the three cases, we consider the distribution of weights
conditional on a subject at time 2 having observed a1i belonging to either B1 = (50, 66.7], or
B2 = (66.7, 83.4] or B3 = (83.4, 100]. We then match such distributions of weights predicted

25
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Table 6.1
Parameter estimates.
Model θ θ∗ θ∗ c
BU 0.30
(0.052)
LRTU 0 0.30 [1.65, 1.71]
(0.016) (0.045) (0.071)
FBU (and maxmin) 0.30 0.30
(0.070) (0.067)
The table shows the parameter estimates of the three models. The standard errors in parenthesis are computed by non-
parametric bootstrap with 1000 bootstrap samples. The standard error for c refers to the lower bound of the set of
minimizers.

by the model with the empirical ones at each of the three quartiles reported in Table 4.2. The
estimate will, therefore, result from 27 moment conditions.27
In the Appendix we describe all the other details of the econometric models and the estimation
procedure. Here we only discuss the estimation results, reported in Table 6.1 (non-parametric
bootstrapped standard errors in parenthesis).28
It is useful to start our description with the estimates for the LRTU model. The lower bound
estimate of c shows that, to “trust” a predecessor’s action, a subject needs the likelihood ratio
to be greater than a threshold equal to 1.65, that is, he requires stronger evidence of rationality
than what assumed in the MLU model (in which c = 1). When this threshold is reached, the
subject considers the observed action as fully rational (since the estimated lower bound for the
proportion of a noise type is θ∗ = 0). When, instead, the threshold is not reached, he attaches a
probability of 0.30 to the action coming from a noise type.
It is interesting to note one implication of these parameter estimates for subjects’ behavior.
After receiving a confirming signal, subjects keep trusting an action a1 ∈ (83.4, 100] (which they
trust even before receiving the signal). After receiving a contradicting signal, in 68.9% of the
cases they stop trusting such an action.
The BU model is a special case of the LRTU model in which the support of prior beliefs
collapses to a singleton, θ∗ = θ ∗ = θ . We estimated θ for the BU model and obtained an estimate
of 0.30 (standard error: 0.052), which coincides with the estimate of θ ∗ in the LRTU model.
The crucial question is whether the LRTU model provides a better explanation for the observed
behavior than the BU model. A simple comparison of the minimized GMM objective functions
for the two models would not be an appropriate way of measuring their relative fitness, since one
model allows for more degrees of freedom (has more parameters) than the other. We develop a
model comparison hypothesis test between BU and LRTU models (see Appendix). The p-value
of the test is 0.005, that is, we can reject the null hypothesis that the BU model with parameter
value θ = 0.3 is the true data generating process and consider our evidence in support of the
LRTU model. The LRTU model fits the data significantly better than the BU model after we
have properly taken into account the gain of over-parametrization.
As for FBU model, the estimates for θ∗ and θ ∗ are both identical to 0.30. That is, under
the FBU specification, the estimates indicate that subject 2 does not have multiple priors for
the predecessor’s rationality. Furthermore, if we compare this result to that for the BU model,

27 For the BU model, as observed above, given agent 2’s first action, his second action only depends on q 2S . For this
2
reason, the identifying information for θ is only based on the first action at time 2 (i.e., 9 moment conditions).
28 The GMM objective function does not have a unique minimizer for the parameter c: c ∈ [1.65, 1.71]. Table 6.1 hence
reports the set of minimizers for c. Nevertheless, the other parameters have the same estimate for any c ∈ [1.65, 1.71].

26
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

we note that the FBU model’s estimates coincide with the estimated BU, despite the fact that
FBU model has an extra degree of freedom. In other words, adding multiple priors and assuming
FBU does not improve the fitness of the more restrictive BU model. Combined with the model
comparison result of LRTU versus BU model, we conclude that the fitness of the LRTU model
outperforms the FBU model. In the Appendix, we also estimate the FBU model with the Hurwitz
criterion (Hurwicz, 1951) instead of the maxmin criterion, and again find that the FBU model
with Hurwitz criterion fits the data significantly worse than the LRTU model.
We would like to emphasize that our approach does not impose any parametric restriction
on the subjective precisions: the evidence in favor of the LRTU model does not depend on any
distributional specification for heterogeneous beliefs.

7. Results of the GC treatment

We also tested the hypothesis that subjects change their mind about the predecessors’ ratio-
nality in a Bayesian fashion (or not) directly, through the GC treatment. Recall that in such a
treatment, subject 2 is asked to state his belief that the predecessor observed the good signal
(the green ball) first after observing the predecessor’s action only and, then, after receiving the
private signal too. The stated belief is equivalent to the stated belief that the first decision maker
is rational, according to the definition used in Section 2.
Our main interest is in the beliefs after the second subject has also received the private signal.
Fig. 7.1 reports the distributions of these beliefs, distinguishing the cases of contradicting and
j
confirming signals.29 We denote these beliefs about the predecessor’s signal by xt , to avoid
j
confusion with the beliefs about the value, previously denoted by at . Analogously to the analysis
of Section 4, the figure is obtained after transforming an action x21 < 50 into 100 − x21 and the
corresponding signal s1 into 1 −s1 . For each case, we plot two frequencies: the left bar in each bin
shows the empirical frequency of x22 ; the right bar, instead, shows the frequency of the theoretical
beliefs (denoted by x̂22 ) that subjects would have reported, had they updated in a Bayesian way
(with signal precisions equal to the objective precisions).30
The empirical and theoretical frequencies are remarkably different in the contradicting signal
case, whereas they are remarkably similar in the confirming signal case. For the contradicting
case, the theoretical belief distribution first-order stochastically dominates the empirical belief
distribution: subjects’ beliefs about their predecessors’ signals (equivalent to their predecessors’
rationality, according to the definitions of Section 2) shift more drastically than what Bayesian
updating predicts. Note, in particular, that the mode of the empirical distribution is 50, that is,
after a contradicting signal, subjects give equal chance to the predecessor having received a green
or red ball.
For confirming signals, the theoretical, Bayesian, belief distribution matches the empirical
belief distribution remarkably well, indicating that subjects update their beliefs in line with
Bayesian updating. We perform two-sample Kolmogorov-Smirnov tests and reject the null hy-
pothesis of equivalence of the distributions (p-value < 0.01) for the contradicting signal case,
whereas we cannot reject the null for the confirming signal case (p-value = 0.13).

29 The signals are defined as confirming or contradicting with respect to a . There are no relevant differences when the
1
definition of confirming or contradicting is with respect to x21 .
30 We also simulated the updated beliefs using subjective precisions, elicited from the IDM treatment (as in the econo-
metrics analysis). The results are almost identical and presented in the Appendix (Fig. D.3).

27
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Fig. 7.1. Distribution of second decisions (beliefs on the predecessor’s signal) at time 2. The top panel refers to confirming
signals; the bottom to contradicting signals. In each panel, we report the frequency of the actual decisions (left bars) and
of the simulated (using objective precisions), theoretical ones (right bars).

In the Appendix we quantify the departure from Bayesian updating in the GC treatment, by
computing a subject’s weight on the signal with a similar method to that of Section 4. In the case
of a confirming signal the median weight coincides with the Bayesian one, whereas in the case
of a contradicting signal it is considerably higher, 2.86.
We also simulated the actions (beliefs on the value) implied by the beliefs about the predeces-
sor’s signal and obtained distributions remarkably similar to the empirical distributions reported
in Fig. 4.3 of Section 4 (see Fig. D.5 in Appendix).
Overall, the results of this treatment show how subjects change their view about the predeces-
sors’ rationality in a direct way. The changes are not in line with Bayesian updating, but in line
with the LRTU, the theoretical model we proposed to explain the asymmetric updating observed
in the SL treatments.

8. Discussion and related literature

To conclude, we want to explain how our study relates to different strands of literature.
First, the idea that one uses the data to select the prior from a set of priors is well and long
established in statistics. It dates back to the Type-I I maximum likelihood of Good (1965), in
which new observations are used to estimate a prior for an unknown parameter (see, e.g., Berger,
1985). Selecting the prior is at the root of the debate in Bayesian statistics and, in particular, in
empirical Bayes procedures (see, also Cox and Hinkley, 1979; Good, 1983; Berger and Berliner,
1986). In this methodology, the set of priors from which one prior is estimated is invariant to the
arrival of new information, similarly to our LRTU rule.
Second, in economic decision theory, MLU was first introduced in the literature on ambigu-
ity aversion. It is, indeed, one of the two main families of updating with ambiguous beliefs, the
other being FBU (see Gilboa and Marinacci, 2013 for a survey). The literature has mostly con-

28
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

sidered one-shot decision problems, noting important conceptual difficulties when dealing with
the arrival of new information and belief updating.31 In their axiomatization of the MLU model,
Gilboa and Schmeidler (1993) do not consider a multi-period problem. In their MLU framework
an agent only updates once, therefore the problem of how to update once new information arrives
is not immediately relevant. Nevertheless, in their analysis, implicitly the choice of the prior is
once and for all. This would be equivalent, in our experiment, to the subject having to stick to the
prior he has selected after observing the predecessor’s action only. Our LRTU model generalizes
the MLU rule since it lets the agent require particularly strong evidence to choose a prior. More
importantly, the set of priors from which the agent chooses does not shrink to a singleton as,
implicitly, in Gilboa and Schmeidler (1993).32
Third, more recently, a model of updating based on hypothesis testing has been reproposed
in the economics literature in a contribution by Ortoleva (2012) who axiomatizes this model to
deal with the important question of belief updating after observing unexpected events.33 In his
Hypothesis Testing model, the decision maker has a prior over possible priors (referred to as
theories to avoid confusion). Initially, the theory with the highest prior probability is selected.
When a new event occurs, if the likelihood of that event is lower than a specific threshold, the
prior over priors is updated on the basis of the likelihood of the event, and the theory with the
highest posterior probability is selected. Ortoleva (2012) highlights the case in which paradigm
shift would occur essentially after unforeseen events (which amounts to setting the threshold
probability to 0), but the general version of his theory is very similar in spirit to our approach,
here applied to the modeling of an agent’s belief about the rationality of another agent. Our
parameter c can be viewed as playing a role similar to that of the prior over priors in Ortoleva’s
model (since in both set ups, an agent can be a priori biased in favor of a particular theory —
a bias we actually estimate). Our evidence can also be read as supportive of Ortoleva (2012)’s
approach.
Finally, our experiment is related to the experimental social learning literature. Compared to
the canonical experiments (following Anderson and Holt, 1997), we made two important proce-
dural changes. First, in previous experiments subjects made a decision in a binary action space,
whereas we ask subjects to choose in a very rich space which practically replicates the contin-
uum. This allows us to elicit their beliefs, rather than just observing whether they prefer one
action to the another.34 Second, in previous experiments subjects made one decision only, af-

31 There is also an important experimental literature devoted to estimating ambiguity aversion in one-shot decision
problems, see, e.g., Ahn et al. (2014) and the papers discussed therein.
32 Gilboa and Marinacci (2013) describe the MLU and FBU models as two extremes: one in which only one prior
is used and one in which all are. Perhaps, our model can be seen somehow in between these two extremes. In the
LRTU model, the subject does pick one prior, but this does not eliminate the other priors forever, since the subject
can pick another prior after new information arrives. A model in which the agent picks different priors every time
new information arrives exhibits a form of time inconsistency. In such model preferences are not stable, which may be
problematic from a normative view point (similar objections apply to Epstein and Schneider, 2007). For a theoretical
investigation of dynamically consistent updating of ambiguous beliefs see Hanany and Klibanoff (2007). Nevertheless,
from a descriptive viewpoint, the model that best fits the data lets the subjects choose the prior every time (from a set
that we estimate). As we emphasized already before, we view this as being very much in the tradition of the statistics
literature dating back to the Type-I I maximum likelihood of Good (1965).
33 In another recent contribution, Weinstein (2017) also proposes that agents may be non Bayesian when something
unexpected happens and the agent decides for a “paradigm shift.”.
34 Within the discrete action space experiments, exceptions to the binary action space are the financial market experi-
ments of Cipriani and Guarino (2005, 2009) where subjects can choose to buy, to sell or not to trade. In the interesting
experimental design of Çelen and Kariv (2004), subjects choose a cut off value in a continuous signal space: depending

29
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

ter observing both the predecessors and the private signal. In our experiment, instead, subject 2
states his belief twice (before and after the private signal), which allows us to identify the weight
on the signal in a neat way. Many previous papers (e.g., Nöth and Weber, 2003; Çelen and Kariv,
2004; Goeree et al., 2007) have reported that subjects overweigh their own signal. Our results
are clearly in line with this finding. In those experiments, though, the behavior we document
and explain could not be observed, exactly because each subject only made one decision and in
a binary action space. When subjects had a signal in agreement with the previous actions, they
typically followed it and chose the same action. This decision is essentially uninformative for the
experimenter on how subjects update on their private information. In fact, on the basis of previ-
ous results, one could have thought that overweighing private information is a general feature of
human subjects’ updating in this type of experiments. Our work shows that this is not the case,
since it only happens when the private information contradicts the first belief.

9. Conclusion

We studied the behavior of human subjects in an experiment in which they report their belief
about the value of a good, first after observing a predecessor’s evaluation and then upon receiving
a private signal. We found an important deviation from Bayesian updating: news contradicting the
previous belief are overweighted compared to news that confirm it. This asymmetric updating is
inherent to how people learn from others (it is not a psychological bias in updating) and cannot be
explained by subjective beliefs in the vein of Savage (despite the flexibility of this approach). We
explain this asymmetry by using a model of updating multiple priors on the rationality of others.
We see our work as a step in understanding how to update multiple priors (something very little
explored) and how to model beliefs about others’ rationality. In future work, this approach could
be combined with models of bounded rationality, such as level-k, QRE, ABEE. We suspect our
findings are relevant in the field, given that rarely there are data on the probability distribution of
others’ rationality. The way subjects update, as we have observed it in our study, may also have
many interesting implications for social and economic applications.35 For instance, in a financial
market, after a period of boom (recession), the arrival of new, negative (positive) information
may trigger a stronger reaction by traders and so higher price volatility. More empirical work to
explore such applications would be, we believe, very valuable.

Appendix A. Supplementary material

Supplementary material related to this article can be found online at https://doi.org/10.1016/

j.jet.2021.105188.

References

Ahn, D., Choi, S., Gale, D., Kariv, S., 2014. Estimating ambiguity aversion in a portfolio choice experiment. Quant.
Econ. 5 (2), 195–223.
Anderson, L., Holt, C., 1997. Information cascades in the laboratory. Am. Econ. Rev. 87 (5), 847–862.

on the realization of the signal, one of the two actions is implemented (as in a Becker et al. (1964) mechanism). That
design allows the authors to distinguish herd behavior from informational cascades.
35 In a recent study, Giustinelli and Pavoni (2017) use survey measures of ambiguous beliefs in the context of high
school track choice.

30
R. De Filippis, A. Guarino, P. Jehiel et al. Journal of Economic Theory 199 (2022) 105188

Angrisani, M., Guarino, A., Jehiel, P., Kitagawa, T., 2021. Information redundancy neglect versus overconfidence: a
social learning experiment. Am. Econ. J. Microecon. Forthcoming.
Becker, G., DeGroot, M., Marschak, J., 1964. Measuring utility by a single-response sequential method. Behav. Sci. 9
(3), 226–232.
Berger, J., 1985. Statistical Decision Theory and Bayesian Analysis. Springer.
Berger, J., Berliner, L.M., 1986. Robust Bayes and empirical Bayes analysis with ε-contaminated priors. Ann. Stat. 14
(2), 461–486.
Çelen, B., Kariv, S., 2004. Distinguishing informational cascades from herd behavior in the laboratory. Am. Econ. Rev. 94
(3), 484–498.
Cipriani, M., Guarino, A., 2005. Herd behavior in a laboratory financial market. Am. Econ. Rev. 95 (5), 1427–1443.
Cipriani, M., Guarino, A., 2009. Herd behavior in financial markets: an experiment with financial market professionals.
J. Eur. Econ. Assoc. 7 (1), 206–233.
Cox, D., Hinkley, D., 1979. Theoretical Statistics. CRC Press.
Epstein, L.G., Schneider, M., 2007. Learning under ambiguity. Rev. Econ. Stud. 74 (4), 1275–1303.
Fischbacher, U., 2007. Z-tree: Zurich toolbox for ready-made economic experiments. Exp. Econ. 10 (2), 171–178.
Gilboa, I., Marinacci, M., 2013. Ambiguity and the Bayesian paradigm. In: Acemoglu, D., Arellano, M., Dekel, E. (Eds.),
Advances in Economics and Econometrics, vol. 1. Cambridge University Press, pp. 179–242.
Gilboa, I., Schmeidler, D., 1993. Updating ambiguous beliefs. J. Econ. Theory 59 (1), 33–49.
Giustinelli, P., Pavoni, N., 2017. The evolution of awareness and belief ambiguity in the process of high school track
choice. Rev. Econ. Dyn. 25, 93–120.
Goeree, J., Palfrey, T., Rogers, B., McKelvey, R., 2007. Self-correcting information cascades. Rev. Econ. Stud. 74 (3),
733–762.
Good, I., 1965. The Estimation of Probabilities: An Essay on Modern Bayesian Methods. The MIT Press.
Good, I., 1983. Good Thinking: The Foundations of Probability and Its Applications. University of Minnesota Press.
Guarino, A., Jehiel, P., 2013. Social learning with coarse inference. Am. Econ. J. Microecon. 5 (1), 147-74.
Hanany, E., Klibanoff, P., 2007. Updating preferences with multiple priors. Theor. Econ. 2 (3), 261–298.
Hurwicz, L., 1951. The Generalised Bayes Minimax Principle: A Criterion for Decision Making Under Uncertainty.
Cowles Commision Discussion Paper Statistics, 355.
Nöth, M., Weber, M., 2003. Information aggregation with random ordering: cascades and overconfidence. Econ. J. 113
(484), 166–189.
Nyarko, Y., Schotter, A., 2002. An experimental study of belief learning using elicited beliefs. Econometrica 70 (3),
971–1005.
Ortoleva, P., 2012. Modelling the change of paradigm: non-Bayesian reactions to unexpected news. Am. Econ. Rev. 102
(6), 2410–2436.
Pires, C., 2002. A rule for updating ambiguous beliefs. Theory Decis. 53 (2), 137–152.
Savage, L., 1954. The Foundations of Statistics. Wiley, New York.
Schotter, A., 2003. Decision making with naive advice. Am. Econ. Rev. Pap. Proc. 93 (2), 196–201.
Schotter, A., Trevino, I., 2014. Belief elicitation in the lab. Annu. Rev. Econ. 6, 103–128.
Weinstein, J., 2017. Bayesian Inference Tempered by Classical Hypothesis Testing. Mimeo.

31