Efficient Coding and Risky Choice

EFFICIENT CODING AND RISKY CHOICE∗
CARY FRYDMAN AND LAWRENCE J. JIN
Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

We experimentally test a theory of risky choice in which the perception of a
lottery payoff is noisy due to information processing constraints in the brain. We
model perception using the principle of efficient coding, which implies that per-
ception is most accurate for those payoffs that occur most frequently. Across two
preregistered laboratory experiments, we manipulate the distribution from which
payoffs in the choice set are drawn. In our first experiment, we find that risk tak-
ing is more sensitive to payoffs that are presented more frequently. In a follow-up
task, we incentivize subjects to classify which of two symbolic numbers is larger.
Subjects exhibit higher accuracy and faster response times for numbers they have
observed more frequently. In our second experiment, we manipulate the payoff
distribution so that efficient coding modulates the strength of valuation biases. As
we experimentally increase the frequency of large payoffs, we find that subjects
perceive the upside of a risky lottery more accurately and take greater risk. To-
gether, our experimental results suggest that risk taking depends systematically
on the payoff distribution to which the decision maker’s perceptual system has re-
cently adapted. More broadly, our findings highlight the importance of imprecise
and efficient coding in economic decision making. JEL Codes: G02, G41.
∗ We are grateful to Andrei Shleifer (the editor), four anonymous referees,

Nicholas Barberis, Nicola Gennaioli, Katrin Gödker, Alex Imas, Shimon Kogan,
John O’Doherty, Stavros Panageas, Antonio Rangel, Elke Weber, Michael Wood-
ford, and seminar participants at Caltech, the Chinese University of Hong Kong,
Harvard University, the London School of Economics, the National University
of Singapore, the Ohio State University, Tsinghua University, the University of
Hong Kong, the University of Mannheim, the University of New South Wales, the
University of Notre Dame, the University of Technology Sydney, the University
of Utah, the University of California, San Diego, the University of Pennsylva-
nia, the University of Southern California, the University of Texas at Dallas, the
University of Warwick, the University of Zurich, Washington University in St.
Louis, Yale University, the Behavioral Economics Annual Meeting, the Chicago
Booth Conference in Behavioral Finance and Decision Making, the Kentucky Fi-
nance Conference, the LA Finance Day Conference, the NBER Behavioral Finance
Meeting, the Society for Neuroeconomics Conference, and the Sloan-Nomis Work-
shop on the Cognitive Foundations of Economic Behavior for helpful comments.
Frydman acknowledges financial support from the National Science Foundation
(Grant #1749824); Jin acknowledges financial support from the Linde Institute at
Caltech.

C The Author(s) 2021. Published by Oxford University Press on behalf of the Presi-
dent and Fellows of Harvard College. All rights reserved. For Permissions, please email:
journals.permissions@oup.com
The Quarterly Journal of Economics (2022), 161–213. https://doi.org/10.1093/qje/qjab031.
Advance Access publication on August 19, 2021.
161
162 THE QUARTERLY JOURNAL OF ECONOMICS
I. INTRODUCTION
In nearly all economic models of risky choice, the decision
maker (DM) is assumed to make a choice based on a precise

representation of available lotteries. Yet a large literature in nu-
merical cognition finds that humans perceive numerical quanti-
ties with noise, even when the quantities are clearly presented
to the DM through Arabic numerals (see Dehaene 2011 for a re-
view). This basic premise leads to the hypothesis, recently pro-
posed by Khaw, Li, and Woodford (2021) (KLW), that risky choice
will also be based on a noisy representation of available lotter-
ies. As KLW show theoretically, noisy perception of lottery payoffs
can provide a microfoundation for small-stakes risk aversion and
stochastic choice.
The idea that perceptual noise drives risk aversion has a vari-
ety of important but untested implications. For instance, if percep-
tual noise systematically varies across environments, so should
the DM’s appetite for risk (Woodford 2012a, 2012b). This implica-
tion is particularly relevant because there is evidence that noise
in perception of sensory stimuli—such as light or sound—changes
optimally with the environment. Specifically, a core principle from
neuroscience called efficient coding states that the brain should
allocate resources so that perception is relatively more precise for
those stimuli that are expected to occur relatively more frequently
(Barlow 1961; Laughlin 1981).1 This principle explains the tem-
porary “blindness” that we experience when moving from a dark
room to a brightly lit one, because resources have not yet been ad-
justed for precisely perceiving objects in the new bright environ-
ment. If the principle of efficient coding also governs choice under
risk, then the DM’s perception of a lottery payoff—and hence her
appetite for risk—will vary with the environment.
In this article, we design and conduct two preregistered exper-
iments to test the hypothesis that efficient coding operates during
risky choice. In each experiment, we measure how the demand for
a risky lottery varies as we change the payoff distribution to which
a subject has recently adapted. To guide our experimental design,
we build a theoretical framework that combines principles from
two existing models. First, the foundation of our framework is the
1. For experimental evidence consistent with efficient coding in sensory per-

ception, see Girshick, Landy, and Simoncelli (2011), Wei and Stocker (2015,
2017), Heng, Woodford, and Polanı́a (2020), and Payzan-LeNestour and Wood-
ford (forthcoming). See also the evidence from Polanı́a, Woodford, and Ruff (2019)
on efficient coding in choice between food items.
EFFICIENT CODING AND RISKY CHOICE 163
KLW model, which assumes that the DM observes noisy signals

of lottery payoffs and subsequently forms optimal estimates of
these payoffs through Bayesian inference. Second, we rely on
the efficient coding model from Heng, Woodford, and Polanı́a

(2020) (HWP) to endogenize the conditional distribution of noisy
signals—which is called the “efficient code.” As in KLW, our
framework generates stochastic choice between a risky lottery
and a certain option; crucially, by adding the efficient coding
mechanism from HWP, we can also assess how the environment
modulates the probability of choosing the risky lottery.
To build intuition, consider a DM who chooses between a
binary risky lottery and a certain option. Furthermore, suppose
the DM is in a low-volatility environment where the upside of the
risky lottery is drawn from a narrow range between $15 and $25.
Efficient coding implies that the brain will allocate its limited
resources across this narrow range, allowing the DM to easily
distinguish between payoffs in the range [$15, $25]. Suppose now
the volatility increases, so that the upside of the risky lottery is
drawn from a wider range between $5 and $35. Efficient coding
then predicts that resources will be partially reallocated away
from the narrow range and toward the extremes of the new range.
As the perceptual system must now “cover more ground” with its
limited resources, the DM finds it more difficult to distinguish
between payoffs in the range [$15, $25].
The shift in perceptual resources immediately leads to a
testable prediction about choice. In the low-volatility environ-
ment, if we increase the upside of the risky lottery from, say,
$20 to $21, the DM will find it easy to distinguish between the
two payoffs, and therefore she can easily perceive the increase in
the attractiveness of the risky lottery. As a result, the increase in
the risky lottery’s upside payoff will have a large impact on the
likelihood that the DM accepts the risky lottery. Conversely, in
the high-volatility environment, perceptual resources are spread
across a wider range, and therefore the DM will have greater dif-
ficulty distinguishing between $20 and $21. As such, the same $1
increase in the risky lottery payoff will have a smaller impact on
the DM’s likelihood of accepting the risky lottery, compared with
that in the low-volatility environment.
More generally, the efficient coding model of HWP pre-
dicts that perception—and hence behavior—is more sensitive
to changes in payoff values when the dispersion of potential
payoffs is smaller.2 This prediction is inconsistent with most

standard economic models of risky choice in which valuation
is nonstochastic and independent of context. At the same time,
the prediction is shared by a broad class of theories, including

the prominent decision-by-sampling model from cognitive science
(Stewart, Chater, and Brown 2006), theories of normalization from
neuroscience (Carandini and Heeger 2012; Rangel and Clithero
2012; Louie, Glimcher, and Webb 2015), and alternative specifica-
tions of efficient coding (Wei and Stocker 2015; KLW).
In our first experiment, we test the above prediction by in-
centivizing subjects to make a series of decisions between a risky
lottery and a certain option. We manipulate the range of payoffs
across a high-volatility condition and a low-volatility condition,
and crucially, we include a set of 30 “common trials” presented
in both conditions. These common trials allow us to cleanly com-
pare behavior across conditions and identify the effect of the prior
distribution. Another important component of our design is that
our tests do not depend on whether the DM’s objective is to maxi-
mize the precision of her payoff estimate (as in models of sensory
perception) or to maximize her expected financial gain (as in mod-
els of economic decision making) (Rustichini et al. 2017; Ma and
Woodford 2020).3 Thus, the data we produce can be used to simul-
taneously test different specifications of efficient coding in choice
under risk.
The results from our first experiment provide strong evidence
that efficient coding influences the demand for risky lotteries.
We find that in the low-volatility condition, a $1 increase in the
payoff of the certain option is associated with an 18.6% increase
in the frequency of choosing the certain option, compared with a
smaller increase of 13.7% in the high-volatility condition. These
estimates are based on the same choice sets from each condition,
and the effect is significant between and within subjects. We also
find that subjects execute decisions significantly faster in the low-
2. This prediction is derived under the assumption that coding resources do

not change with the volatility of the environment. Later in the article, we discuss
implications for the case when coding resources can change with the volatility of
the environment.
3. As emphasized in Ma and Woodford (2020), there are differences in how
resource constraints are imposed across different models of efficient coding. Al-
though we use a specific constraint levied by HWP, the main prediction we test is
qualitatively similar to other models of efficient coding in sensory perception that
assume different constraints, such as Wei and Stocker (2015).
volatility condition, and therefore our results cannot be driven

by an alternative hypothesis where subjects in the low-volatility
condition choose to process information for a longer period of time.
As an additional test of the mechanism, we present each sub-

ject with a “perceptual choice” task following the risky choice
task. In this second task, subjects still need to perceive numer-
ical quantities, but they do not need to perceive any probabilities
or integrate them with payoffs. We incentivize subjects to classify
whether a two-digit number displayed on each trial is above or
below a reference number. We find that even in this simpler en-
vironment, classification accuracy depends strongly on the distri-
bution of numbers to which the subject has adapted. Subjects are
significantly more accurate and they respond faster if the number
they are classifying was presented more frequently in the recent
past.
When viewed through the lens of efficient coding, the results
from our first experiment indicate that noisier perception gener-
ates noisier choice. At the same time, valuation remains largely
unbiased, in the sense that the average perception of a payoff is
approximately the same as the true payoff. The “bias toward the
prior” effect, which is commonly associated with Bayesian mod-
els, does not arise in our model when the prior is uniform—as
it is in our first experiment. For other priors, however, efficient
coding can give rise to strong biases in valuation. As emphasized
by Woodford (2012a, 2012b), efficient coding can theoretically gen-
erate a value function that exhibits several features from prospect
theory, including reference dependence and diminishing sensitiv-
ity (Kahneman and Tversky 1979). Moreover, these features will
fluctuate over time as an optimal response to changes in the en-
vironment. For example, an environment with a decreasing dis-
tribution of payoffs will induce the familiar value function with
diminishing sensitivity as payoffs become larger; an environment
with an increasing distribution of payoffs will generate dimin-
ishing sensitivity as payoffs become smaller. Both types of value
functions are a consequence of the perceptual system’s limited ca-
pacity to discriminate between payoffs that occur infrequently
(Robson 2001; Rayo and Becker 2007; Netzer 2009; Payzan-
LeNestour and Woodford forthcoming).
In our second experiment, we test whether risk taking is
greater when the DM has adapted to an increasing distribution
compared to a decreasing distribution. Across two experimental
conditions, we manipulate the shape of the payoff distribution
while holding its range constant. When analyzing identical choice
sets across the two conditions, we find evidence consistent with

a systematic bias in valuation. As predicted by efficient coding,
a subject exhibits greater risk taking when she is adapted to the
increasing distribution, compared to the decreasing distribution.

Intuitively, when the subject is adapted to an increasing distri-
bution, perceptual resources are allocated toward large payoffs;
this resource allocation enables the DM to accurately perceive
the large and attractive upside of the risky lottery, and thus she
chooses to take risk. In contrast, when the subject is adapted to
a decreasing distribution, she has difficulty recognizing the large
upside of the risky lottery and is unwilling to take risk. Our results
are consistent with those from a recent perceptual choice experi-
ment by Payzan-LeNestour and Woodford (forthcoming), who find
that “outlier” stimuli are perceived less accurately than frequently
occurring stimuli. More generally, we provide novel evidence con-
sistent with the hypothesis that diminishing sensitivity to payoffs
arises in part from an optimal allocation of perceptual resources.
We emphasize that the existing evidence of efficient coding,
which comes almost exclusively from data on sensory perception,
in no way implies that the same mechanisms are deployed dur-
ing decision making under risk. Indeed, it is plausible that sen-
sory perception is governed by efficient coding while a different
decision system is activated when a subject is presented with a
decision concerning monetary risk. What we test in this article is
precisely the hypothesis that efficient coding and noisy perception
are also active in higher-level decision systems that govern risky
choice. Our experimental evidence therefore supports a nascent
theoretical agenda on the implications of imprecise and efficient
coding for economic behavior (Woodford 2012a, 2012b; Steiner and
Stewart 2016; Gabaix and Laibson 2017; Natenzon 2019; Wood-
ford 2020; KLW; Enke and Graeber 2021).
At a broader level, our results contribute to a growing lit-
erature that builds cognitive and perceptual foundations for the
psychological assumptions in behavioral economics. For instance,
several behavioral models of financial markets demonstrate that
prospect theory preferences can explain puzzling facts such as the
high equity premium of the aggregate stock market (see Barberis
2018 for a review). Our data support the proposition that efficient
coding provides a normative foundation for the value function as-
sumed in prospect theory (Woodford 2012a, 2012b).4
4. For alternative approaches to endogenizing the value function in prospect

theory, see Friedman (1989) and Denrell (2015).
The rest of the article is organized as follows. In Section II, we

present a theory of efficient coding that guides our experimental
design. Sections III and IV provide experimental tests of the model
by studying how the payoff distribution affects choice. Section V

provides additional discussions, and Section VI concludes.
II. THE MODEL

In this section, we present a theory of efficient coding and
risky choice that integrates two existing theoretical models. The
foundation of our theory is the KLW model of noisy perception of
lottery payoffs. In their baseline model, KLW assume a particular
form of noisy coding of lottery payoffs and a specific prior distribu-
tion, which they use to derive novel implications for risky choice.
We build on KLW by integrating their model with the efficient
coding model of HWP. By combining these two models, the theory
is able to generate predictions about how noisy coding—and the
probability of risky choice—systematically changes for any payoff
distribution to which the DM has adapted.5
II.A. Choice Environment

The DM faces a choice set that contains two options: a certain
option and a risky lottery. The certain option, denoted as (C, 1),
pays C > 0 dollars with certainty. The risky lottery, denoted as (X,
p; 0, 1 − p), pays X > 0 dollars with probability p and zero dollars
with probability 1 − p. The DM’s task is to choose between these
two options.
Under expected utility theory, a DM with utility U(·) and
no background wealth chooses the risky lottery over the certain
option if and only if
(1) p · U (X) + (1 − p) · U (0) > U (C).
Conditional on X, C, and p, the DM’s choice is nonstochastic.

We now present the KLW framework of noisy coding, which
departs from expected utility theory by assuming that perceptions
of X and C are noisy. The noisy coding assumption is motivated
5. KLW also provide an extension to their baseline model in which the precision
of noisy coding can flexibly change with the volatility of a particular (log-normal)
prior distribution. Our framework further generalizes this flexibility by deriving
optimal coding rules for any prior distribution.
by the literature in sensory perception, where a common finding

is that when an identical stimulus is presented on different occa-
sions (e.g., a fixed number of dots presented across different trials
of an experiment), experimental subjects judge the stimulus dif-

ferently across the different occasions. Before observing the choice
set that contains X and C, the DM holds prior beliefs about X and
C, denoted by f(X, C). We further assume the DM believes that
X and C are drawn independently: f(X, C) = f(X)f(C). Upon ob-
serving the choice set, the DM’s perceptual system spontaneously
generates a noisy signal, Rx , of X, and a noisy signal, Rc , of C.6
Each signal is drawn from a distinct conditional distribution; Rx
is randomly drawn from f(Rx |X) and Rc is randomly drawn from
f(Rc |C). In the language of Bayesian inference, these conditional
distributions are the likelihood functions, which we define in Sec-
tion II.B.
Given the prior beliefs and noisy signals, the DM follows
Bayes’ rule to generate posterior distributions about X and C,
which we denote by f(X|Rx ) and f(C|Rc ). As in KLW, we assume
that the DM has linear utility.7 Under this assumption, the op-
timal decision rule depends only on the conditional means of the
posterior distributions, E[ X̃|Rx ] and E[C̃|Rc ].8 The DM chooses the
risky lottery if and only if the perceived expected value of the risky
lottery exceeds that of the certain option, which occurs under the
following condition: p · E[ X̃|Rx ] > E[C̃|Rc ].
It is worth noting that the encoding process described above—
the process that maps X and C to Rx and Rc —is conditional on the
values of X and C, which we assume are perfectly observable to
the econometrician but not to the DM. In other words, even after
the DM is presented with a choice set, she still faces uncertainty
about the payoff values of X and C. As such, Bayesian inference
takes place at the level of a single choice set, and it characterizes
6. In the choice environment we study here, our interpretation is that the

noisy signals are generated unconsciously; they are not the outcome of deliberate
and conscious information acquisition in the sense of Stigler (1961). For more
discussion on this point, see Ma and Woodford (2020).
7. Although we focus on how imperfect perception affects risky choice, our
model does not preclude the more traditional source of risk aversion that operates
through diminishing marginal utility of wealth. In Section III.E, we extend the
model by allowing for both imperfect perception and intrinsic risk aversion.
8. We use the notations X̃ and C̃ to emphasize that when the DM forms the
posterior means E[ X̃|Rx ] and E[C̃|Rc ], she does not directly observe X and C; she
treats these payoffs as random variables.
how the DM’s prior belief shifts after observing a noisy signal of
the true payoff. As we show next, the distribution of the noisy
signal drives the main predictions of our model.

II.B. Likelihood Function
1. Deriving the Optimal Likelihood Functions. We depart from
KLW by allowing the DM to choose the optimal likelihood func-
tions according to the prior distribution. In this case, the DM still
encodes each payoff with noise, but the noise distribution can be
optimized to meet a specific performance objective. We rely on
the efficient coding model of HWP to endogenize the likelihood
functions f(Rx |X) and f(Rc |C). We also retain the assumption from
KLW that the DM encodes X and C independently; this is a natu-
ral assumption given the DM’s prior belief that X and C are drawn
independently. In addition, we assume the DM’s performance ob-
jective is to maximize expected financial gain, defined as
Expected financial gain

≡ pX · Prob(choose the risky lottery|X, C) · f (X) f (C) · dXdC

+ C · Prob(choose the certain option|X, C) · f (X) f (C) · dXdC.
(2)
We focus on the performance objective in equation (2) because

it is commonly used in economic settings—although we discuss
alternative objectives below.
When maximizing expected financial gain, the DM faces an
information-processing constraint. Specifically, HWP assume that
the DM encodes X through a finite number of n “neurons,” where
the output state of each neuron takes the value 0 or 1. The output
states of these n neurons are assumed to be mutually independent,
and each neuron takes the value 1 with probability θ (X) and 0
with the remaining probability 1 − θ (X). The encoded value of X
is therefore represented by an output vector of zeros and ones,
with length n. Given that the neurons are mutually independent,
a sufficient statistic for the output vector is the sum across the n
output values, which we denote by Rx . Thus, the noisy signal Rx
can take on integer values from 0 to n. The likelihood function of
X can then be defined by

n
(3) f (Rx |X) = (θ (X)) Rx (1 − θ (X))n−Rx .
Rx

We do not interpret the encoding process described above as a
literal description of brain function; rather, we share with HWP
the view that equation (3) is useful in analyzing how a system
with limited resources may transmit information through a set of
n binary signals. As such, n can be thought of as representing an
individual-specific capacity constraint: as n goes to infinity, the
random variable Rnx converges almost surely to its mean θ (X), and
therefore the amount of noise in perceiving X is reduced to zero.
The DM encodes C using a process that is identical to that for
encoding X. Each neuron takes the value 1 with probability θ (C)
and 0 with the remaining probability 1 − θ (C). The sum across
the n output values is denoted as Rc . The likelihood function of C
is defined by

n
(4) f (Rc |C) = (θ (C)) Rc (1 − θ (C))n−Rc .
Rc
Equations (3) and (4) show that the likelihood functions are
driven by the coding rules, θ (X) and θ (C), which map X and C
into the probability that a neuron emits a value of 1. Intuitively,
if the DM is particularly concerned about perceiving values of X
within a given range, then a good coding rule, θ (X), should be very
sensitive to X over that range.
When n is large, and when pX and C are i.i.d., HWP show
that the coding rules that maximize expected financial gain are
given by9
⎡ ⎛ X
⎞ ⎤2
2
⎢ ⎜π f (x) dx ⎟⎥ 3
⎢ ⎜ ⎟⎥
(5) θ (X) = ⎢sin ⎜ −∞ ⎟⎥
⎣ ⎝2 ∞ 2 ⎠⎦
f (x) 3 dx
−∞
9. The coding rules described in this section are derived when n, which param-
eterizes the capacity constraint, is sufficiently large. Appendix 7 of HWP shows
that for any finite n that is greater than or equal to 10, the coding rules remain
approximately optimal. When illustrating the model’s implications in Section II,
we set n to 10.
and
⎡ ⎛ C
⎞⎤2
2
⎢ ⎜π f (c) dc ⎟⎥
3

⎢ ⎜ ⎟⎥
(6) θ (C) = ⎢sin ⎜ −∞ ⎟⎥ .
⎣ ⎝2 ∞ 2 ⎠⎦
f (c) dc
3
−∞
Importantly, we see that the coding rules depend explicitly on

the prior distributions f(X) and f(C), which gives the model its
inherent context dependence.
As in HWP, we also consider two alternative performance
objectives besides maximizing expected financial gain, which we
use to assess the robustness of the model’s predictions. The first
alternative objective is to maximize mutual information between
X and its noisy signal, Rx ,
(7) max I(X, Rx ),

θ(X)
where the mutual information I(X, Rx ) is defined as the difference

between the marginal entropy of Rx and the entropy of Rx con-
ditional on X. Similarly, the DM is assumed to maximize mutual
information between C and Rc . The second alternative objective
is to maximize the probability of an accurate choice, given by
Probaccurate

≡ Prob(Rx > Rc |θ (X) > θ (C)) · 1θ(X)>θ(C) · f (X) f (C) · dXdC

+ Prob(Rx < Rc |θ (X) < θ (C)) · 1θ(X)<θ(C) · f (X) f (C) · dXdC.
(8)
Interestingly, there exists a class of priors for which the three

performance objectives in equations (2), (7), and (8) all lead to the
same optimal coding rules (as n goes to infinity). Specifically, in
Online Appendix A, we show that when the following two condi-
tions
(i) pX and C are i.i.d.

(9)
and (ii) pX and C are uniformly distributed
are satisfied, the coding rules under all three objectives reduce to
2
π X − Xl
θ (X) = sin

(10)
2 Xu − Xl
and
2
π C − Cl
(11) θ (C) = sin ,
2 Cu − Cl
where Xl , Xu , Cl , and Cu are the bounds of the uniform distribu-

tions, with Xl < Xu and Cl < Cu . This result is useful because it
enables tests of efficient coding which are robust to changing the
assumption that the DM maximizes expected financial gain. In
our main risky choice experiment, we build a design that satisfies
the conditions in expression (9).
2. Properties of the Likelihood Function. Here we illustrate
how the likelihood function depends explicitly on the DM’s prior
beliefs. For now, we suppose that the DM’s prior belief about X is a
uniform distribution between Xl and Xu , and that her prior belief
about C is a uniform distribution between Cl and Cu . In keeping
with the conditions in expression (9), we further set Cl = p · Xl and
Cu = p · Xu , so that pX and C are identically distributed. Under
these assumptions, the likelihood functions of X and C are given
by
2 Rx
n π X − Xl
f (Rx |X) = sin
Rx 2 Xu − Xl
2 n−Rx
π X − Xl
× 1 − sin ,
2 Xu − Xl
2 Rc
n π C − Cl
f (Rc |C) = sin
Rc 2 Cu − Cl
2 n−Rc
π C − Cl
(12) × 1 − sin .
2 Cu − Cl
The expressions in equation (12) show that the likelihood func-

tions depend directly on the parameters of the prior distributions,
Xl , Xu , Cl , and Cu . This dependence of the likelihood function on

the prior is a signature characteristic of efficient coding.
Figure I illustrates the malleability of the likelihood function.
Panel A presents two different prior distributions over X, one

with high volatility and the other with low volatility. In the
high-volatility environment, X is distributed uniformly over a
wide range (Xl = 8 and Xu = 32). In the low-volatility environ-
ment, X is distributed uniformly over a narrow range (Xl = 16
and Xu = 24). Equation (10) implies that these two distributions
induce different coding rules θ (X). Panel B shows that the coding
rule is steeper for the low-volatility distribution, compared
with the high-volatility distribution. Recall that the coding rule
gives the “success probability” of the binomial distribution in
equation (3). Thus, a steeper coding rule implies that the success
probability is more sensitive to changes in X. Panel C plots the
implied likelihood function for two values, X = 18 and X = 22, and
for each prior distribution. In the low-volatility environment, a
payoff of X = 18 generates a very different distribution of signals
f(Rx |X) compared to a payoff of X = 22. Thus, as X increases
from 18 to 22, the DM’s perceptual system can easily detect this
change. In the high-volatility distribution, however, X = 18 and
X = 22 generate distributions of signals that overlap extensively.
The more extensive overlap of the likelihood functions in the
high-volatility environment leads to less discriminability between
the two payoffs, compared to the low-volatility environment. As
we show in the next section, this difference in discriminability
has a direct effect on risky choice.
II.C. Value Function and Implications for Choice

Given the prior and likelihood functions defined above, the
DM proceeds by using Bayesian inference to compute a posterior
distribution of each payoff in the choice set. Under the assumption
of linear utility, the DM’s optimal decision rule depends only on
the conditional means of the posterior distributions. Specifically,
the posterior means of X and C, conditional on Rx and Rc , are
given by
Xu
Xu f (Rx |X) f (X)XdX
X
(13) E[ X̃|Rx ] ≡ f (X|Rx )XdX = l X
u
Xl
f (Rx |X) f (X)dX
Xl
(A)
0.2
0.15

0.1
0.05
10 15 20 25 30
(B)
1
0.8
0.6
0.4
0.2
0
10 15 20 25 30
(C)
0.5
0.4
0.3
0.2
0.1
0
0 5 10
FIGURE I
Prior Distributions, Coding Rules, and the Optimal Likelihood Functions
Panel A plots two uniform prior distributions for X, one with low volatility (Xl =
16 and Xu = 24) and the other with high volatility (Xl = 8 and Xu = 32). Panel B
plots the coding rule θ (X), defined in equation (10), for both volatility environments.
Panel C plots the implied likelihood function f(Rx |X), defined in equation (12), for
two values, X = 18 and X = 22, and for each of the two prior distributions. The
capacity constraint parameter n is set to 10.
and
Cu
Cu f (Rc |C) f (C)CdC

Cl
(14) E[C̃|Rc ] ≡ f (C|Rc )CdC = Cu
,
Cl
f (Rc |C) f (C)dC
Cl
where f(X) and f(C) are the DM’s prior beliefs about X and C, and
the likelihood functions f(Rx |X) and f(Rc |C) are from equation (12).
Importantly, equation (13) shows that the DM’s estimate of X
is a random variable, and the randomness comes from Rx . There-
fore, the DM faces a distribution of perceived values for each X.
We now characterize the mean and standard deviation of this
distribution. Specifically, we define the value function, v(X), by

n
(15) v(X) = f (Rx |X) · E[ X̃|Rx ].
Rx =0
That is, v(X) represents the subjective valuation of X averaged

across different values of Rx . Moreover, we define the standard
deviation for the subjective valuation, σ (X), by
⎡ ⎤ 12

n
σ (X) = ⎣ f (Rx |X)(E[ X̃|Rx ]) − v 2 (X)⎦ .
2
(16)
Rx =0
Equations (15) and (16) indicate that the curvature of the value
function and the randomness in subjective valuation are jointly
determined by the DM’s prior beliefs and the implied likelihood
functions.
In keeping with the running example from the previous sec-
tion, Figure II, Panel A plots, for both the high- and low-volatility
environments, the average subjective valuation v(X), as well as
its one standard deviation bounds v(X) ± σ (X).
The figure shows that randomness in utility, σ (X), is substan-
tially higher in the high-volatility environment. This is driven by
the greater overlap of likelihood functions in the high-volatility
environment, compared with the low-volatility environment. Be-
cause subjective valuation is noisier in the high-volatility envi-
ronment, the model predicts that choices will also be noisier and
hence less sensitive to a given change in payoff values.
(A) (B)
0.2 0.4
0.3

0.15
0.2
0.1
0.1
0.05
0
10 20 30 2 4 6 8
8
30
25
6
20
15 4
10
2
10 20 30 2 4 6 8
FIGURE II
Prior Distributions and Value Functions
Panel A: the upper graph plots two uniform prior distributions for X, one with
low volatility (Xl = 16 and Xu = 24) and the other with high volatility (Xl = 8
and Xu = 32). The lower graph plots the subjective valuations implied by efficient
coding, v(X), and their one standard deviation bounds v(X) ± σ (X). Panel B: the
upper graph plots two prior distributions, one increasing and one decreasing. The
increasing distribution is characterized by

i , h, l = l, if Xl X Xm
i
f X; Xl , Xu, Xm ,
h, if Xm < X Xu
i
7 1
where Xl = 2, Xu = 8, Xm
i = 4.5, h =
25 , and l = 125 . The decreasing distribution
is characterized by

h, if Xl X Xm
d
f X; Xl , Xu, Xm
d , h, l =
d < XX
,
l, if Xm u
where Xl = 2, Xu = 8, Xm d = 5.5, h = 7 , and l = 1 . The lower graph plots the

25 125
subjective valuations implied by efficient coding, v(X), and their one standard
deviation bounds v(X) ± σ (X). For both panels, the capacity constraint parameter
n is set to 10. In the lower graph of each panel, the dotted line is the 45-degree
line.
To formalize this prediction, we compute the probability of

choosing the risky lottery—which we refer to from now on as
the “probability of risk taking.” Recall that conditional on X
and C, the noisy signals Rx and Rc are drawn from the like-

lihood functions f(Rx |X) and f(Rc |C). For a given realization of
(Rx , Rc ), the DM then chooses between the risky lottery and the
certain option by comparing the posterior means of X and C in
equations (13) and (14). As a result, when fixing X, C, and the
prior distributions, we compute the probability of risk taking as
follows:
Prob(risk taking|X, C)

n
n

= 1 p·E[ X̃|Rx ]>E[C̃|Rc ] · f (Rx |X) · f (Rc |C)
Rx =0 Rc =0
n
n
1
(17) + 1 p·E[ X̃|Rx ]=E[C̃|Rc ] · f (Rx |X) · f (Rc |C) .
2
Rx =0 Rc =0
Equation (17) says that the DM chooses the risky lottery over the
certain option when p · E[ X̃|Rx ] > E[C̃|Rc ], and the DM randomly
chooses between the two options when p · E[ X̃|Rx ] = E[C̃|Rc ].
Figure III plots, for both the high- and low-volatility envi-
ronments, the probability of risk taking against the difference in
expected values between the two options, namely, pX − C. Natu-
rally, a higher value of pX − C increases the attractiveness of the
risky lottery and hence increases the probability of risk taking.
Note that for an expected-utility maximizer with no background
wealth, the probability of risk taking should be a step function of
pX − C with a single step at pU −1 ( U (C)−(1− p
p)U (0)
) − C. However,
Figure III shows that under noisy coding, the probability of risk
taking has an S-shaped relationship with pX − C. More impor-
tant, under efficient coding, the slope of this function is negatively
related to the volatility of the stimulus distribution (for those val-
ues of pX − C that do not deliver an extreme probability near 0
or 1). Thus, for a given increase in X, the probability of choosing
the risky lottery increases more in the low-volatility environment.
This heightened sensitivity in the low-volatility environment can
be traced back to the property illustrated in Figure I, Panel C: a
given increase in X leads to a larger difference in the distribution
of noisy signals in the low-volatility environment, compared with
the high-volatility environment.

FIGURE III
Model-Predicted Probability of Choosing the Risky Lottery
The graph plots the probability of risk taking, defined in equation (17), for the
two volatility environments (low volatility: Xl = 16, Xu = 24, Cl = 8, and Cu =
12; high volatility: Xl = 8, Xu = 32, Cl = 4, and Cu = 16). The prior distributions
for X and C are uniform. The probability p that the risky lottery pays X dollars
is set to 0.5. The capacity constraint parameter n is set to 10. For each volatility
environment, we draw X uniformly from [16, 24] and C uniformly from [8, 12]; that
is, we draw the payoffs from the common support of the low-volatility distribution
and the high-volatility distribution. We then compute, for a given X and C, the
probability of risk taking. Finally, we aggregate these probabilities for each level
of pX − C.
II.D. Increasing and Decreasing Payoff Distributions

Our results so far show that payoff volatility affects the dis-
persion in perceived valuation, σ (X). At the same time, the average
subjective valuation v(X) largely coincides with X when the prior
distribution is uniform. Under other priors, however, efficient cod-
ing can induce strong biases in valuation, in the sense that v(X)
can differ significantly from X (Woodford 2012a, 2012b; KLW). To
illustrate how these strong biases arise, we consider a different
environment in which the payoffs of X and C are drawn either
from an increasing distribution or from a decreasing distribution.
We specify the increasing distribution by

l, if Xl X Xmi
(18) f X; Xl , Xu, Xm
i
, h, l = .
i
h, if Xm < X Xu
We specify the decreasing distribution by

h, if Xl X Xm
d
(19) f X; Xl , Xu, Xm, h, l =
d
.
l, if Xm < X Xu
d

Figure II, Panel B plots, for the increasing and decreasing
prior distributions, the average subjective valuation v(X), as well
as its one standard deviation bounds v(X) ± σ (X).10 For the in-
creasing distribution, small values of X occur with low frequency,
and the DM therefore allocates little coding resources toward
these infrequent and small values. The lack of coding resources
dedicated to small values of X gives rise to a positive perceptual
bias toward the mean of the distribution: v(X) > X when X is small.
Conversely, for the decreasing distribution, the DM allocates lit-
tle coding resources toward large values of X. As a result, the DM
is insensitive to high values of X and exhibits a negative percep-
tual bias toward the mean of the distribution: v(X) < X when X
is large. We test implications of these value functions after pre-
senting our main experiment. Specifically, for large values of X,
we experimentally test whether demand for risk taking is higher
in the increasing condition, compared to the decreasing condition.
III. EXPERIMENT 1: VOLATILITY MANIPULATION

In this section, we provide our main experimental test of the
model by manipulating the volatility of payoffs. We preregistered
the experiment and recruited 150 students from the University of
Southern California to participate in the laboratory; see Online
Appendix B for the preregistration document. Each subject first
completed a risky choice task and then a perceptual choice task.
This ordering was chosen to minimize any fatigue effects in the
risky choice task, which is our main task of interest. Subjects were
paid a $7 participation fee in addition to earnings from each task.
III.A. Design of the Risky Choice Task

On each trial, subjects choose between the risky lottery (X,
p; 0, 1 − p) and the certain option (C, 1). The probability p is
10. When the prior distribution is increasing or decreasing, conditions (9) for
the equivalence of coding rules no longer hold. Figure II, Panel B presents the
subjective valuations based on the coding rule that maximizes the DM’s expected
financial gain. The results are quantitatively similar if the subjective valuations
are instead based on the coding rule that maximizes mutual information.

FIGURE IV
Design of the Risky Choice Task in Experiment 1
The task consists of two blocks of trials: one block contains trials from the high-
volatility condition, and the other block contains trials from the low-volatility
condition. The order of the two blocks is randomized across subjects. Each block
begins with 30 “adaptation trials,” followed by 270 “test trials.” Among the 270
test trials, we designate 30 “common trials” that are identical across both volatil-
ity conditions and are the basis of our main tests. On each trial, subjects have
unlimited time to decide which option they prefer.
fixed at 0.5 for all trials. The values of X and C are drawn in-
dependently, and we manipulate the distribution of each payoff
across two volatility conditions. In the high-volatility condition, X
is drawn uniformly from [8, 32], and C is drawn uniformly from
[4, 16]. In the low-volatility condition, X is drawn uniformly from
[16, 24], and C is drawn uniformly from [8, 12].
We choose these design parameters for two reasons. First, be-
cause our goal is to isolate the effect of volatility, we keep the mean
of each payoff distribution constant across conditions. The mean
of X is fixed at 20, and the mean of C is fixed at 10. Second, our
parameter values satisfy conditions (9): the distributions of X and
C are independent, and pX and C are identically and uniformly
distributed. These conditions imply that the efficient coding rule
is robust to changing the performance objective from maximiz-
ing expected financial gain to maximizing mutual information or
maximizing the probability of an accurate choice. Thus, our design
is optimized to test generic predictions of efficient coding.
Figure IV shows a schematic of the task design. Each sub-
ject goes through both the high- and low-volatility conditions; the
order of the two conditions is randomized across subjects. Each

condition contains 300 trials, which are broken into two phases:
an initial “adaptation” phase with 30 trials, and a subsequent
“test” phase with 270 trials. The adaptation phase is intended to

allow subjects to adapt to the condition-specific payoff distribu-
tion. The test phase contains the trials that we are interested in
analyzing.
To generate a clean test of efficient coding, we want to com-
pare decisions on the same choice sets across the two volatility
conditions. One constraint we face when designing these “com-
mon trials” in the test phase is that the lottery payoffs must fall
in the support of the distribution of both conditions. Our goal is
to maximize the number of common trials that satisfy this con-
straint, while staying faithful to the statistical properties of each
payoff distribution.
To do so, we first note that the support of the low-volatility dis-
tribution is a subset of the support of the high-volatility distribu-
tion; therefore, payoffs on common trials must fall in the support
of the low-volatility distribution. Specifically, with 19 probability, a
pair (X, C) drawn from the high-volatility distribution falls in the
support of the low-volatility distribution. As such, in each condi-
tion, we designate 30 of the 270 trials in the test phase as common
trials. These common trials are identical across conditions, and we
generate them by drawing 30 pairs of (X, C) over approximately
equally spaced grid points of the low-volatility distribution; see
Online Appendix, Table D.1 for exact values. In each condition,
the location of a common trial is randomized at the subject level
across the 270 possible test trial locations.
We draw the remaining 240 test trials in the low-volatility
condition from the low-volatility distribution. For the remain-
ing 240 test trials in the high-volatility distribution, we draw (X,
C) uniformly from the high-volatility distribution, but critically,
we “redraw” the pair (X, C) if it falls in the support of the low-
volatility distribution—because this part of the high-volatility dis-
tribution is already covered by the common trials. Therefore, for
each volatility condition, the distribution of payoffs across all tri-
als accurately reflects the appropriate population distribution. In
summary, the common trials simultaneously serve two purposes:
they allow for a clean comparison of behavior across conditions,
and they also reinforce the prior on subsequent trials.
Subjects are not explicitly informed about the payoff distribu-
tions from which X and C are drawn. We believe that such a design
is more natural than telling subjects the payoff distributions they

will experience. In particular, if the experimenter explicitly gives
information about the distribution of payoffs, subjects may arti-
ficially direct their attention to this information, which in turn

could generate an experimenter demand effect. Furthermore, our
design enables us to test for learning effects, which are important
when conducting our within-subject analyses.
One of the 600 trials was randomly selected for payment, and
the subject was paid according to their choice on this randomly
selected trial. The average earning for the risky choice task was
$10.14. Online Appendix B provides the instructions that were
given to subjects before the experiment.
III.B. Results from the Risky Choice Task

We produce a large data set that contains 90,000 total ob-
servations across all subjects and conditions. As part of our pre-
registered data exclusion rule, we drop one subject who chose the
certain option on all trials in the first condition. We are then left
with 89,400 trials across the first and second experimental condi-
tions, of which we analyze only the 8,940 common trials.
1. Frequency of Risk Taking. We begin our analysis with
between-subjects tests, which we construct using only trials from
the first condition. We find that on average, subjects choose the
risky option on 52.3% of trials (std. err.: 2.3%) with an average
response time of 2.0 seconds (std. err.: 0.085 seconds). Table I pro-
vides results from a set of mixed-effects linear regressions, which
account for heterogeneity across subjects in average levels of risk
taking and in sensitivity to X and C.11 Column (1) shows that risk
taking increases significantly in X and decreases significantly in
C. The coefficients of interest are those on the interaction terms:
the coefficient on X × high is negative (p-value = .038) and the
coefficient on C × high is positive (p-value = .056). This is our first
piece of evidence consistent with efficient coding, namely, that a
11. We have also estimated an analogous set of mixed-effects logistic regres-

sions; however, with random effects on X, C, and the intercept, we find that these
mixed-effects logistic regressions do not converge to numerically stable estimates.
As an alternative, we estimate a set of logistic regressions without random effects,
in which we pool all subjects. Online Appendix Table D.2 shows that the results
from the logistic regressions are consistent with those presented in Table I. Our
preferred specification is the mixed-effects linear regression because it accounts
for heterogeneity across subjects.
TABLE I
FREQUENCY OF CHOOSING THE RISKY LOTTERY IN EXPERIMENT 1 (VOLATILITY MANIPULATION)
Between-subjects tests Within-subject tests
Unrestricted Restricted Restricted Restricted sample Restricted sample Restricted sample

sample sample sample (w/out trials 301–450) —low-volatility first —high-volatility first
Dependent variable: (w/out trials 301–450) (w/out trials 301–450)
“Choose risky lottery” (1) (2) (3) (4) (5) (6)
high 0.023 0.001 − 0.023 − 0.061 0.016 − 0.164

(0.191) (0.200) (0.097) (0.113) (0.168) (0.139)
X 0.066∗∗∗ 0.074∗∗∗ 0.062∗∗∗ 0.064∗∗∗ 0.073∗∗∗ 0.052∗∗∗
(0.006) (0.006) (0.004) (0.004) (0.006) (0.006)
C − 0.167∗∗∗ − 0.186∗∗∗ − 0.163∗∗∗ − 0.170∗∗∗ − 0.186∗∗∗ − 0.155∗∗∗
(0.013) (0.012) (0.009) (0.009) (0.012) (0.014)
X × high − 0.017∗∗ − 0.023∗∗∗ − 0.006 − 0.009∗∗ − 0.015∗∗∗ − 0.002
(0.008) (0.008) (0.003) (0.004) (0.006) (0.005)
C × high 0.033∗ 0.049∗∗∗ 0.013∗ 0.025∗∗∗ 0.030∗∗∗ 0.019
(0.017) (0.017) (0.008) (0.009) (0.011) (0.013)
Constant 0.776∗∗∗ 0.787∗∗∗ 0.826∗∗∗ 0.850∗∗∗ 0.801∗∗∗ 0.943∗∗∗
(0.166) (0.179) (0.106) (0.123) (0.180) (0.135)
EFFICIENT CODING AND RISKY CHOICE
Observations 4,470 4,170 8,257 6,411 3,125 3,286
Notes. The table reports results from mixed-effects linear regressions in which the dependent variable takes the value of 1 if the subject chooses the risky lottery, and 0 otherwise.
The dummy variable, high, takes the value of 1 if the trial belongs to the high-volatility condition, and 0 if it belongs to the low-volatility condition. Only data from common trials are
included. There are random effects on the independent variables X, C, and the intercept. Standard errors of the fixed-effect estimates are clustered at the subject level and reported
in parentheses. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.
183

$1 increase in X and a $1 decrease in C each leads to a greater

increase in the frequency of choosing the risky lottery in the low-
volatility condition.
Before conducting further analyses, we take one additional

step to clean the data. We exclude the trials on which a sub-
ject exhibits an excessively fast response time of less than 0.5
seconds, which constitutes 7.6% of the data; the remaining data
are referred to as the “restricted sample.” This exclusion was not
preregistered, but unsurprisingly these fast decisions are not re-
sponsive to the underlying payoff values, and thus we employ this
exclusion in all subsequent analyses. Table I, column (2) shows
that after excluding fast decisions, the conclusion remains largely
the same: the coefficient on X × high is negative (p-value = .003)
and the coefficient on C × high is positive (p-value = .003).
To see a graphical representation of the difference in be-
havior across conditions, Figure V, Panel A uses the data from
Table I, column (2) to plot the empirical frequency of risk taking
as a function of the difference in expected values between the two
options.
The figure shows a striking difference across conditions: a $1
increase in pX − C leads to a greater increase in the frequency
of choosing the risky lottery in the low-volatility condition, com-
pared with the high-volatility condition. We emphasize that this
difference is based on the same exact set of 30 common trials in
each condition. Figure V, Panel B presents the data in a different
manner, without imposing the assumption that the frequency of
risk taking is a function of pX − C. Each point represents one of
the 30 common trials. The x-axis measures the frequency of risk
taking in the high-volatility condition, and the y-axis measures
the frequency of risk taking in the low-volatility condition.
When the frequency of risk taking (in both conditions) is low,
most of the data points fall below the 45-degree line. Conversely,
when the frequency of risk taking is high, most of the data points
are above the 45-degree line. This pattern is consistent with the
model: the likelihood functions in the high-volatility condition
generally lead to less discriminability between payoffs, compared
to the low-volatility condition. As such, the model tends to predict
that the probability of risk taking is less extreme—that is, closer
to 50%—in the high-volatility condition.
Our design also enables us to examine how coding varies
within subjects over time when faced with a change in the en-
vironment. Although our model does not make predictions about
(A)
1
0.8

0.6
0.4
0.2
0
-3 -2 -1 0 1 2 3 4
(B)
1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
FIGURE V
Frequency of Risk Taking across Volatility Conditions
Panel A: the graph plots, for each volatility condition, the empirical frequency of
risk taking against the difference in expected values between the risky lottery and
the certain option, namely, pX − C. The frequency of risk taking is computed as
the proportion of trials on which subjects choose the risky lottery. Data are pooled
across subjects over all common trials in the first condition, and thus represent
between-subjects comparisons. For each volatility condition, we bin the running
variable, pX − C, to its nearest integer value, and plot the mean for each bin. The
length of the vertical bar inside each data point denotes two standard errors of the
mean. Standard errors are clustered by subject. Panel B: each point represents one
of the 30 common trials in the first condition. The x-axis measures the frequency of
risk taking in the high-volatility condition, while the y-axis measures the frequency
of risk taking in the low-volatility condition. Inside each data point, the length of
the vertical bar denotes two standard errors of the mean frequency of risk taking in
the low-volatility condition. The length of the horizontal bar denotes two standard
errors of the mean frequency of risk taking in the high-volatility condition.
adaptation, we can test whether behavior changes in the direc-

tion predicted by efficient coding as subjects experience a shift
in the environment. Recall that halfway through the risky choice
task, on trial 301, the payoff distribution switches. Thus, by rees-

timating the regression in Table I, column (2) using data from
both conditions for each subject, we can measure the effect within
subjects. Column (3) shows that the coefficients on the interac-
tion terms have the predicted sign, though the effects are weaker
compared to those from the between-subjects tests. One reason for
these weaker effects is that at the beginning of the second condi-
tion, subjects may still be adapted to the first condition. To allow
for longer adaptation in the second condition, we exclude trials
from the first half of the second condition, namely, trials 301–450.
Column (4) shows that the magnitudes of the interaction effects
do indeed get larger.
Columns (5) and (6) further disaggregate the data based on
whether the subject experiences the low- or high-volatility condi-
tion first. Column (5) shows that for those subjects who begin with
the low-volatility condition, the coefficients on X × high and C ×
high are both significant at the 1% level. This result is important
because it rules out an alternative theory whereby subjects en-
code payoffs with noise, and through repeated experience with a
payoff, the amount of noise decreases over time. Such a “learning
from experience” theory can explain our between-subjects results,
because subjects in the low-volatility condition experience low-
volatility payoffs more frequently than do subjects in the high-
volatility condition. However, this alternative theory cannot ex-
plain the within-subject result in column (5): under this theory,
behavior in the second (high-volatility) condition should be less
noisy, as subjects experience the same set of 30 common trials
for a second time. However, we find that the effect has the op-
posite sign, and is therefore consistent with noisy and efficient
coding.
Column (6) provides results for those subjects who experience
the high-volatility condition first. The coefficients on X × high
and C × high have the predicted sign, but the effects become
weaker and are no longer statistically significant. We conjecture
that it is easier for subjects to detect a change in the environment
when moving from a low-volatility to high-volatility condition,
because “outlier payoffs” that are never experienced in the first
condition begin to appear in the second condition. In contrast,
when moving from a high-volatility to low-volatility condition,
the information that signals a change in environment is less

salient.
2. Response Times. Not only are subjects more sensitive to

payoffs in the low-volatility condition, they also implement deci-
sions more quickly in this condition. Among common trials, sub-
jects take an average of 2.02 seconds in the high-volatility condi-
tion versus 1.79 seconds in the low-volatility condition (p-value =
.001 for a within-subject test). This difference in response times
rules out the hypothesis that subjects perceive payoffs more pre-
cisely in the low-volatility condition because they spend more time
on each decision.
Response times can also be used to understand how subjects
adapt to a new payoff distribution. Specifically, we examine the
time series evolution of response times over the course of the
entire 600 trials during the risky choice task. Figure VI shows
the response time data, disaggregated by which condition a sub-
ject experienced first; here the figure includes data from noncom-
mon trials, in addition to common trials. The upper panel plots,
for each of the 600 trials, the response time averaged across sub-
jects who first go through the low-volatility condition, followed by
the high-volatility condition. We see a spike in response time at
trial 301, which is the beginning of the high-volatility condition.
This spike in response time may be due to the fact that subjects
begin to experience novel and hence salient payoffs that they had
not seen in the first 300 trials. As a result, these payoffs signal
a change in environment and presumably restart the adaptation
process.
In contrast, the lower panel plots the response time averaged
across subjects who first go through the high-volatility condition,
followed by the low-volatility condition. In this cut of the data,
we find no corresponding spike in response time at the begin-
ning of the second condition. Here, subjects do not observe salient
“outlier” payoffs: every payoff in the second condition is in the
support of the payoff distribution from the recently experienced
first condition. We speculate that this extra difficulty in adapting
to the low-volatility distribution may be partly responsible for the
stronger within-subject results in column (5), compared to those
in column (6) of Table I.
Although our theory does not make predictions about re-
sponse times, the data strongly suggest that response times are
systematically related to the DM’s prior distribution. A natural
FIGURE VI Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

Evolution of Response Times over the 600 Trials from the Risky Choice Task
(Experiment 1)
Each dot represents a trial-specific response time averaged across subjects. The
upper panel presents data from subjects who experienced the low-volatility con-
dition first. The lower panel presents data from subjects who experienced the
high-volatility condition first. The vertical line denotes the onset of the change in
environment. Data include both common and noncommon trials.
way to incorporate response times into our framework would be

to allow the DM to draw a sequence of noisy signals, {Rx,i }i=1
S
for a given payoff X. The time that the DM takes to execute a

decision would then reflect the number of signals drawn, which
is a common interpretation of sequential sampling models from
mathematical psychology (Ratcliff 1978; Bogacz et al. 2006; Kra-
jbich, Armel, and Rangel 2010) and more recently from economics
(Woodford 2014; Fudenberg, Strack, and Strzalecki 2018; Hébert

and Woodford 2019).

III.C. Design of the Perceptual Choice Task
Recall that all the implications of our model are driven by the
noisy encoding of X and C. In particular, we make two simplifying
assumptions: (i) there is no noise in encoding the probability p or
the $0 payoff, and (ii) there is no noise in computing the product of
p and E[ X̃|Rx ]. In reality, there is likely to be noise in both of these
processes, which could be responsible for some of the experimental
results discussed already.
To provide a more targeted test of the key efficient coding
mechanism, each subject participates in a second “perceptual
choice task.” In this task, subjects still need to perceive X but
do not need to perceive the probability p or integrate probabilities
with perceived payoffs. Given that the noisy encoding of payoffs
is sufficient to generate our main theoretical predictions in Sec-
tion II, we expect to find evidence that the perception of X still
depends on the recent stimulus distribution.
Our perceptual choice task is informed by work from the lit-
erature on perception of symbolic numbers (Moyer and Landauer
1967). We build on the design of Dehaene, Dupoux, and Mehler
(1990), who present subjects with an Arabic number between 31
and 99 on each trial of their experiment. The subjects’ task is to
classify whether the Arabic numeral presented on the screen is
larger or smaller than the reference level of 65. Dehaene, Dupoux,
and Mehler (1990) find that as the stimulus gets closer to the refer-
ence level, accuracy decreases and response times increase. These
results are consistent with the noisy encoding of Arabic numer-
als, which lies at the foundation of the model of risky choice we
present in Section II.
One notable feature of the Dehaene, Dupoux, and Mehler
(1990) experiment is that the stimulus distribution is held con-
stant throughout the experiment. Here, we exogenously vary the
stimulus distribution across two conditions: a high-volatility con-
dition and a low-volatility condition. In the high-volatility condi-
tion, subjects are presented with an Arabic numeral, which we
denote by X, that is drawn uniformly from integers in the set [31,
99]\{65}. In the low-volatility condition, X is drawn uniformly
from integers in the set [56, 74]\{65}. In each condition, subjects
are asked to classify whether X is above or below the reference

FIGURE VII
Design of the Perceptual Choice Task in Experiment 1
The task consists of two blocks of trials: one block contains trials from the high-
volatility condition, and the other block contains trials from the low-volatility
condition. The order of the blocks is randomized across subjects. Each block begins
with 60 “adaptation trials,” followed by 340 “test trials.” On each trial, subjects
are incentivized to classify whether the number shown on the screen is greater or
less than the reference level of 65.
level of 65. Each subject completes both conditions, and we ran-

domize the order of conditions across subjects. Figure VII gives a
schematic of the design.
In all other respects, the perceptual choice task design follows
closely the design of the risky choice task. In each condition, there
is an initial set of 60 trials that are intended to allow subjects to
adapt to a given distribution. As outlined in our preregistration
document, we only analyze behavior after the adaptation phase
in the subsequent 340 test trials. To generate a clean comparison
across conditions, our main analysis focuses on those trials in the
test phase for which the stimulus numbers fall in the range of com-
mon support across the two conditions [56, 74].12 As in the risky
12. For the high-volatility condition, we designate 90 out of the 340 test trials
as common trials. These 90 common trials are created by sampling each element
in the low-volatility condition five times. The remaining 250 trials in the high-
volatility condition are drawn with 50% probability from a uniform distribution
over [31, 55] and with 50% probability from a uniform distribution over [75, 99].
For the low-volatility condition, we designate the entire 340 test trials as common
choice task, the restriction to common trials is crucial because it

allows us to cleanly identify the effect of the prior distribution on
behavior.
We pay subjects based on both the accuracy and speed of

their classifications. Specifically, subjects earn a payoff of $(15 ×
accuracy − 10 × avgseconds), where accuracy is the percentage of
correctly classified trials, and avgseconds is the average response
time (in seconds) across all trials in the perceptual choice task. We
incentivize fast responses in this task (but not in the risky choice
task) to avoid a “ceiling effect” in the choice data where subjects
would approach 100% accuracy. Although the ceiling effect is not
problematic on its own, it would cause difficulty in detecting any
differences in the choice data across experimental conditions. The
average earning for the perceptual choice task was $8.70.
III.D. Results from the Perceptual Choice Task

We begin by reporting results for between-subjects tests using
all common trials from the first condition. Subjects correctly clas-
sify the number on 93.5% of trials (std. err.: 0.7%) with an average
response time of 0.573 seconds (std. err.: 0.012 seconds). Two out of
the 150 subjects exhibit an average response time of only 0.05 and
0.10 seconds, which indicates that they used a guessing strategy.
Therefore, we exclude them from all subsequent analyses (their
average accuracy rates were 51.8% and 55.0%, respectively).13
Figure VIII, Panel A plots for each value of X the proportion of
trials that subjects classified X as greater than the reference level
of 65. Consistent with previous research on numerical cognition,
we see that subjects exhibit errors in classification; moreover, the
errors increase as X approaches 65. Although it is unsurprising
that subjects exhibit errors, the fact that the frequency of errors
correlates with |X − 65| provides evidence that coding is noisy and
that number comparison is more difficult when the numbers are
closer together.
The novel aspect of our design that enables us to test for
efficient coding is the manipulation across the two volatility
conditions. Among trials for which X ∈ [56, 74], we find that sub-
jects exhibit significantly greater accuracy in the low-volatility
trials as they all fall in the range [56, 74]. This procedure ensures that X is drawn
according to its population distribution in both conditions.
13. Including these two subjects in our subsequent analyses does not affect
any of the main results.
(A)
1
0.8

0.6
0.4
0.2
0
30 40 50 60 70 80 90 100
(B)
0.7
0.6
0.5
30 40 50 60 70 80 90 100
FIGURE VIII
Classification Accuracy and Response Times for the Perceptual Choice Task
Panel A: the x-axis denotes the integer X that is presented on each trial, and the
y-axis denotes the proportion of trials on which subjects classified X as greater than
65. Panel B: the y-axis denotes the average response time for subjects to execute
a decision, for trials on which subjects responded correctly. Data are pooled across
subjects over all test trials in the first condition, and thus represent between-
subjects comparisons. The length of the vertical bar inside each data point denotes
two standard errors of the mean. Standard errors are clustered by subject.
condition, compared with the high-volatility condition (95.0%

versus 92.3%, with p-value <.001 under a mixed-effects linear re-
gression). Not only are subjects more accurate in the low-volatility
condition, they respond significantly faster (0.576 seconds versus
0.611 seconds, with p-value = .03 under a mixed-effects linear
regression). This result is analogous to our finding from the
risky choice task, in which subjects are more sensitive to payoffs

and respond faster in the low-volatility condition. Figure VIII,
Panel B shows that the average response time for trials on which
subjects responded correctly increases as X approaches 65, and

that across the distribution of X, the average response time is
shorter in the low-volatility condition.
Next we test for a difference in slope between the two choice
curves presented in Figure VIII, Panel A. Efficient coding predicts
a steeper slope for the choice curve from the low-volatility condi-
tion. Before proceeding to the test, we note that in the perceptual
choice task, conditions (9) no longer hold because the DM needs
to only encode one payoff, and thus her prior is only over one di-
mension. We therefore base our analyses for the perceptual choice
task on the assumption that the DM maximizes mutual informa-
tion.14 To formally test for a difference in slope, Table II presents
results from a set of mixed-effects logistic regressions. The depen-
dent variable takes the value of 1 if the subject classifies X as
above 65, and 0 otherwise.
In column (1), the coefficient on X − 65 is significantly posi-
tive, indicating that subjects’ propensity to classify X as greater
than 65 is increasing in X. More important, the coefficient on the
interaction term, (X − 65) × high, is significantly negative, indi-
cating that choices are noisier on trials in the high-volatility con-
dition. Columns (2) and (3) examine only numbers inside the 60s
decade and only numbers outside the 60s decade, respectively; in
both cases, the coefficient on the interaction term remains signifi-
cant. Last, columns (4)–(6) provide within-subject tests by pooling
data across both conditions for each subject. For each of these three
specifications, the coefficient on the interaction term remains sig-
nificant at the 1% level.
III.E. Model Estimation

In this section, we structurally estimate the model to test for
a connection between the two tasks. For each subject, we first
estimate the model using data from the risky choice task, and
14. In the perceptual choice task, we incentivize fast responses; therefore, the
coding rule from equation (5) does not necessarily maximize expected financial
gain. At the same time, the coding rule from equation (10) continues to maximize
mutual information. For this reason, we opt for the assumption that the DM
maximizes mutual information in this task.
194
TABLE II
FREQUENCY OF CLASSIFYING X AS GREATER THAN 65 IN THE PERCEPTUAL CHOICE TASK
Between-subjects tests Within-subject tests
56 X 74 60 X 69 56 X 59 56 X 74 60 X 69 56 X 59
Dependent variable: or or
“Classify X as greater 70 X 74 70 X 74
than 65” (1) (2) (3) (4) (5) (6)
X − 65 0.855∗∗∗ 1.096∗∗∗ 0.578∗∗∗ 0.792∗∗∗ 1.051∗∗∗ 0.558∗∗∗

(0.047) (0.061) (0.019) (0.028) (0.038) (0.012)
(X − 65) × high − 0.209∗∗∗ − 0.279∗∗∗ − 0.081∗∗∗ − 0.147∗∗∗ − 0.236∗∗∗ − 0.077∗∗∗
(0.051) (0.069) (0.025) (0.014) (0.023) (0.011)
high 0.288∗∗∗ 0.405∗∗∗ − 0.041 0.259∗∗∗ 0.300∗∗∗ 0.146∗∗
(0.052) (0.070) (0.101) (0.037) (0.047) (0.064)
Constant 0.065** 0.071** 0.269∗∗∗ 0.036∗ 0.077∗∗∗ 0.119∗∗
(0.030) (0.036) (0.067) (0.020) (0.023) (0.048)
Observations 31,230 15,522 15,708 63,210 31,580 31,630
Notes. The table reports results from mixed-effects logistic regressions in which the dependent variable takes the value of 1 if the subject classifies the integer X as larger than
THE QUARTERLY JOURNAL OF ECONOMICS
65, and 0 otherwise. The integer X is drawn uniformly from the set [31, 99]\{65} in the high-volatility condition, while it is drawn uniformly from the set [56, 74]\{65} in the
low-volatility condition. The dummy variable, high, takes the value of 1 if the trial belongs to the high-volatility condition, and 0 if it belongs to the low-volatility condition. There
are random effects on the independent variable X − 65 and the intercept. Standard errors of the fixed-effect estimates are clustered at the subject level and reported in parentheses.
∗ , ∗∗ , and ∗∗∗ indicate significance at the 10%, 5%, and 1% level, respectively.

then we estimate the model using data from the perceptual choice
task.
1. Estimation of the Risky Choice Task. Recall that the one

free parameter of the model, n, denotes the number of binary
readings that are used to generate the noisy signal Rx (and Rc ).
That is, n represents the amount of the subject’s limited percep-
tual resources: the greater n is, the more precise is the represen-
tation of each payoff. We note that our model only accounts for
noise in the perception of X and C. Additional sources of noise
in the decision process that are outside the model—for exam-
ple, noise in perceiving p—may also be captured by the struc-
tural estimate of n, because it is the only free parameter of the
model.
To estimate n, we use maximum likelihood. Specifically, for
each subject, we maximize the following log likelihood func-
tion over n, using choice data from the test phase of the first
condition:
300
LL(n|y) = yt · log(Prob(yt |n)) + (1 − yt ) · log(1 − Prob(yt |n)),
t=31
(20)
where y ≡ {yt }300 t=31 and yt denotes the subject’s choice on trial t;
yt = 1 if the subject chooses the risky lottery, and yt = 0 if the
subject chooses the certain option. In addition, Prob(yt |n) denotes
the model-predicted probability of choosing the risky lottery given
n, Xt , and Ct ; it is computed using equation (17). We maximize the
log likelihood function in equation (20) by searching over integer
values of n in [5, 40]. We find that the average estimate of n,
across subjects, is 8.9 with a standard deviation of 9.7, indicating
substantial heterogeneity.
Our baseline model of efficient coding assumes linear utility;
as such, the DM’s optimal decision rule depends on E[ X̃|Rx ] and
E[C̃|Rc ]. However, the model can easily be integrated with stan-
dard nonlinear utility functions. To generalize the baseline model,
we assume that the DM maximizes expected utility, with a util-
ity function U(·) = (·)α . Under this assumption, the DM chooses
the risky lottery if and only if p · E[( X̃)α |Rx ] > E[(C̃)α |Rc ].15 The
optimal coding rules presented in equations (5) and (6) are then
15. As in equation (17), we also assume that the DM randomly chooses between
the risky lottery and the certain option when p · E[( X̃)α |Rx ] = E[(C̃)α |Rc ].
replaced by16
⎡ ⎛ X
⎞ ⎤2
2 α−1
⎢ ⎜π f (x) (x) dx ⎟⎥
3 3

⎢ ⎜ X ⎟⎥
(21) θ (X) = ⎢sin ⎜ Xl ⎟⎥
⎣ ⎝2 u
2 α−1 ⎠⎦
f (x) 3 (x) 3 dx
Xl
and
⎡ ⎛ C
⎞ ⎤2
2 α−1
⎢ ⎜π f (c) (c) dc ⎟⎥
3 3
⎢ ⎜ Cl ⎟⎥
(22) θ (C) = ⎢sin ⎜ C ⎟⎥ .
⎣ ⎝2 u
2 α−1 ⎠⎦
f (c) 3 (c) 3 dc
Cl
Finally, the probability of risk taking from equation (17) is re-

placed by
Prob(risk taking|X, C)

n
n

= 1 p·E[( X̃)α |Rx ]>E[(C̃)α |Rc ] · f (Rx |X) · f (Rc |C)
Rx =0 Rc =0
n
n
1
(23) + 1 p·E[( X̃)α |Rx ]=E[(C̃)α |Rc ] · f (Rx |X) · f (Rc |C) .
2
Rx =0 Rc =0
We estimate this generalized model using the same maxi-

mum likelihood procedure as before, but for each subject we now
estimate two parameters, n and α. We find that the best-fitting pa-
rameter pair, averaged across subjects, is (n, α) = (10.1, 0.93). This
result indicates that the average subject exhibits a modest degree
of intrinsic risk aversion with the small stakes in our experiment.
We also quantitatively assess the validity of our linear-utility as-
sumption by running an Akaike information criterion (AIC) test
at the subject level. For each subject, we compare the AIC across
the baseline model in which we constrain α = 1 and the gen-
eralized model. We find that 54% of subjects are best fit using
the restricted model; therefore, our baseline assumption of linear
utility is not overwhelmingly restrictive. At the same time, the
16. The generalization of the coding rules to allow for nonlinear utility follows
the analysis of this issue in Payzan-LeNestour and Woodford (forthcoming).
generalized model with intrinsic risk aversion provides a better

fit to a substantial number of subjects.
2. Estimation of the Perceptual Choice Task. We now estimate

the model for each subject using data from the perceptual choice
task. Following the procedure from the previous section, for each
subject we maximize the following log likelihood function over n,
using choice data from the test phase of the first condition:
400
LL(n|z) = zt · log(Prob(zt |n)) + (1 − zt ) · log(1 − Prob(zt |n)),
t=61
(24)
where z ≡ {zt }400

t=61 and zt denotes the subject’s choice on trial t; zt =
1 if the subject classifies the stimulus Xt on trial t as greater than
65, and zt = 0 if the subject classifies Xt as less than 65. The term
Prob(zt |n) denotes the model-predicted probability that the subject
classifies Xt as greater than 65. We maximize the log likelihood
function in equation (24) by searching over integer values of n in
[5, 40]. The best-fitting value of n, averaged across all subjects, is
15.8, with a standard deviation of 13.1.
Given that each subject completes both the risky choice task
and the perceptual choice task, our design enables us to compare
the latent structural parameter n across the two tasks. Two ob-
servations are worth noting. First, the average value of n is lower
in the risky choice task than in the perceptual choice task. This
difference is likely driven by the fact that the risky choice task
is more complex, and hence additional sources of noise enter the
decision process (e.g., encoding the probability p and integrating p
with E[ X̃|Rx ] or E[( X̃)α |Rx ]); as noted above, the structural estima-
tion may then account for these outside sources of noise through
a lower value of n.
Second, we test for a correlation between the estimated n
from each task across subjects. This test is important because it
allows us to assess whether errors in numerical discrimination
from the perceptual choice task can explain variation in the risky
choice task. We find a modest but significant rank correlation
of 0.30 between the estimated n from each task when allowing
α 1 (p-value < .001). The correlation remains significant at
0.26 when using the estimated n from the restricted model where
α = 1 (p-value = .001). The results are also robust to using
Pearson correlations for both models (p-value = .002 for the
unrestricted model, and p-value = .001 for the restricted model).

These positive correlations demonstrate that, across subjects,
variation in perception partly explains variation in risk-taking
behavior.

IV. EXPERIMENT 2: SHAPE MANIPULATION
In the experiments reported in the previous section, we fo-
cused on manipulating the range—and hence the volatility—of
the payoff distribution while holding the mean constant. In this
section, we investigate whether manipulating the shape of the
payoff distribution, while holding the range constant, affects risk
taking in the manner predicted by efficient coding. Such a ma-
nipulation allows us to test for a bias in perception, which can
systematically change the average level of risk taking.
IV.A. Design of Experiment 2

As in the risky choice task from Experiment 1, here we design
an experiment in which subjects are presented with choice sets
of the form {(X, 0.5; 0, 0.5); (C, 1)}. There are two experimental
conditions: in one condition, X is drawn from a weakly decreasing
distribution over the range [2, 8]; in the other condition, X is drawn
from a weakly increasing distribution over the same range. The
specific distributions that we use are shown in Figure II, Panel
B.17
Our design ensures that large values of X—for example, those
between $7 and $8—are frequent in the increasing condition but
are rare in the decreasing condition. For both experimental con-
ditions, we set the distribution of C to be that of pX. This is an
important feature of the design because it implies that for both
conditions, the expected value of the risky lottery and the cer-
tain option are the same. Thus, while the expected value of the
risky lottery is higher in the increasing condition compared to
the decreasing condition, the same is true for the certain option.
Therefore, any observed difference in risk taking across condi-
tions provides evidence for an endogenous shift in the likelihood
function—rather than just a shift in the prior.18 In other words,
17. The increasing and the decreasing distributions take the form of
7 1
equations (18) and (19). The parameter values are: Xl = 2, Xu = 8, h = 25 , l = 125 ,
i = 4.5, and Xd = 5.5.
Xm m
18. Online Appendix A provides a brief proof of this statement.
our design helps us target a test of efficient coding, rather than

just noisy coding.
Figure II, Panel B shows that under efficient coding, the de-
creasing distribution of X generates perception that is insensitive

and biased downward for large values of X. On the other hand,
the increasing distribution generates perception that is approxi-
mately unbiased for large values of X. This systematic difference
in perception of large values of X across conditions will affect the
DM’s appetite for risk.
To test the predicted difference in risk taking, we again cre-
ate a set of common trials that subjects face in both conditions.
Specifically, we create 8 common trials, where we fix C at $2.70
and vary X from $7.13 to $7.99 in approximately $0.12 increments;
see Online Appendix Table D.3 for exact values.19
There are 300 trials per experimental condition, and subjects
are randomized into whether they first experience the increas-
ing condition or the decreasing condition. In each condition, sub-
jects first make choices on 60 “adaptation trials.” Subjects then go
through eight consecutive test blocks, where each test block con-
tains 30 trials. At the end of each test block, we insert a common
trial. For example, the first common trial from the first condi-
tion occurs on trial 90, and the last common trial occurs on trial
300. The order of the common trials is randomized at the subject-
condition level. On noncommon trials, we draw payoffs according
to the increasing or decreasing payoff distribution, depending on
the condition to which the trial belongs.20
Our main testable prediction is that a subject’s average de-
mand for the risky lottery on common trials is greater in the
increasing condition, compared to the decreasing condition. Re-
call that on common trials, X takes values between $7 and $8. For
19. Our common trials focus on large values of X because these values lead
to a difference in risk taking across the two experimental conditions that remains
substantial even when subjects exhibit a strong degree of intrinsic risk aversion.
By contrast, small values of X lead to a difference in risk taking across the two
conditions that diminishes when subjects’ degree of risk aversion is sufficiently
high. Moreover, we fix the value of C at $2.70 so that it has a high density in
both conditions. As a result, the difference in the perception of C across the two
conditions is minimal and only has a small impact on the difference in risk taking
across conditions.
20. The design in this task is similar to that of Payzan-LeNestour and Wood-
ford (forthcoming) who insert a “test trial” every 40 trials, although their design is
implemented in the perceptual choice where a subject is incentivized to discrimi-
nate between shades of gray.
the increasing condition, these values of X are frequent outcomes

because they come from the high density part of the increasing
prior distribution. As such, the perceptual bias in X is minimal:
v(X) ≈ X. However, in the decreasing condition, these large values

of X are rare outcomes. They come from the long right tail of the
decreasing prior distribution and hence lead to a perceptual bias
that is substantially negative: v(X) < X. Efficient coding there-
fore predicts that a subject will perceive the risky lottery on each
common trial to be less attractive in the decreasing condition,
compared to when it is presented in the increasing condition.
We preregister the experiment and recruit 200 subjects from
Prolific, an online data collection platform; see Online Appendix B
for the preregistration document. As outlined in our preregistra-
tion document, all of our statistical tests in Experiment 2 are con-
ducted within subjects.21 Given that Experiment 2 is conducted
online, we impose a 10-second time limit on each trial to promote
engagement. If a subject does not respond within the 10-second
time limit, the computer randomly chooses one of the two options.
Each subject completes the task and is paid according to one ran-
domly selected trial, in addition to a $6.50 participation fee. The
average earning for this task, including the participation fee, was
$9.27. The experimental instructions for Experiment 2 are pro-
vided in Online Appendix B.
IV.B. Results from Experiment 2

We begin our analysis by applying the following four exclu-
sion criteria outlined in the preregistration document. First, we
exclude subjects who failed to correctly answer at least one of the
two comprehension quiz questions. Second, we exclude trials for
which subjects failed to respond within the 10-second time limit.
Third, we exclude subjects who chose the risky option on less than
2.5% or more than 97.5% of noncommon trials.22 Finally, we ex-
clude trials for which subjects exhibited excessively fast response
21. We choose a within-subject design because of the added statistical power

that it provides, which is useful for two reasons. First, in pilot testing on Pro-
lific, we observed substantial variation across subjects in the frequency of choos-
ing the risky lottery. Second, in this experiment we restrict our analysis to only
8 common trials per condition; in comparison, the risky choice task from Experi-
ment 1 contains 30 common trials per condition.
22. This exclusion criterion is motivated by the observation that in pilot testing
on Prolific, a small fraction of subjects exhibited almost zero variation in behavior
across the entire 600 trials, indicating a strategy that does not depend on the
TABLE III
FREQUENCY OF CHOOSING THE RISKY LOTTERY IN EXPERIMENT 2 (SHAPE
MANIPULATION)

All common All common First half of Second half of
Dependent variable: trials trials each condition each condition
“Choose risky lottery” (1) (2) (3) (4)
increasing prior 0.075∗∗∗ 0.075∗∗∗ 0.088∗∗∗ 0.070∗∗∗

(0.023) (0.023) (0.029) (0.024)
X 0.045∗
(0.024)
Constant 0.690∗∗∗ 0.351∗ 0.684∗∗∗ 0.695∗∗∗
(0.029) (0.188) (0.032) (0.030)
Observations 2,278 2,278 862 1,416
Notes. The table reports results from mixed-effects linear regressions in which the dependent variable
takes the value of 1 if the subject chooses the risky lottery, and 0 otherwise. The dummy variable, increasing
prior, takes the value of 1 if the trial belongs to the increasing-prior condition, and 0 if it belongs to the
decreasing-prior condition. Only data from common trials are included. The variable C is constant among
common trials, and therefore is not included in the regressions as a control variable. There are random effects
on the independent variable X and the intercept. Standard errors of the fixed-effect estimates are clustered
at the subject level and reported in parentheses. ∗ , ∗∗ , and ∗∗∗ indicate significance at the 10%, 5%, and 1%
level, respectively.
times, defined as less than 0.5 seconds. After applying these four
exclusion criteria, 151 subjects remain with a total of 85,703 trials,
of which 2,278 are common trials. All regression results presented
below are robust to using the full sample without applying the ex-
clusion criteria.
Our main hypothesis involves testing whether, for a fixed
value of C, a subject’s appetite for risk is higher when large val-
ues of X are more frequent. Table III presents results from mixed-
effects linear regressions in which the dependent variable takes
the value of 1 if the subject chooses the risky lottery and 0 oth-
erwise. All regressions in Table III include only common trials.
Column (1) shows that the frequency of choosing the risky lot-
tery is 7.5% higher in the context of the increasing distribution,
compared with the decreasing distribution (p-value = .001). As
in Experiment 1, this result is based on a comparison where the
choice sets are fixed and only the context varies across conditions.
We emphasize that the significant difference in risk taking cannot
be driven by the mere fact that X has a higher mean in the in-
creasing condition, because this effect is exactly offset by a higher
values of X and C. So as not to select on the dependent variable, this criterion is

based on choices from noncommon trials, which never enter our main analyses.
mean of C in the increasing condition. Thus, we interpret the em-

pirical finding in column (1) as a consequence of an endogenous
shift in the likelihood function, rather than an exogenous shift in
the prior mean of X.

The significant difference in risk taking also holds when we
add a linear control for X across the eight common trials, as shown
in column (2). Note that we do not include any controls for C be-
cause it is fixed at $2.70 across all common trials. Columns (3)
and (4) demonstrate that the main result in column (1) continues
to hold when we condition only on trials in the first half or the
second half of each condition. Overall, the systematic increase in
risk taking as we shift from the decreasing to increasing distribu-
tion demonstrates that diminishing sensitivity can arise, in part,
as a consequence of an optimal allocation of perceptual resources.
Our results from this experiment complement recent empir-
ical evidence from a perceptual decision-making task on “outlier
blindness” (Payzan-LeNestour and Woodford forthcoming). Those
authors provide data that subjects are less accurate in classifying
shades of gray when the stimuli under consideration are outliers
according to the subjects’ prior. Analogously, we show that when
a subject faces an outlier payoff—a large payoff from the decreas-
ing distribution—the perceptual system is “blinded” and cannot
accurately perceive such a high value. Crucially, efficient coding
provides a directional prediction about the misperception of out-
liers; our data are consistent with this prediction of a negative
perceptual bias for large payoffs in the decreasing condition.
V. DISCUSSION
V.A. Adaptation Dynamics
The theoretical framework in Section II is a static model of
risky choice, and hence does not tackle the important question re-
garding how the DM learns the prior distribution. Most empirical
tests of efficient coding in sensory perception assume full adap-
tation to the prior distribution (Laughlin 1981; Wei and Stocker
2015), and this assumption has also been recently invoked in pa-
pers on efficient coding in value-based decisions (Rustichini et al.
2017; Polanı́a, Woodford, and Ruff 2019). Following this litera-
ture, we have assumed that subjects in our experiments are fully
adapted to the population distribution after completing an ini-
tial set of preregistered “adaptation trials.” Yet we emphasize this
assumption is not trivial, particularly because the DM’s learn-

ing problem is more complex than in standard settings, where
Bayesian inference would typically generate convergence. The ad-
ditional layer of complexity is due to the DM’s inability to observe

the sequence of objective payoffs, and as a result, the DM must
learn from the history of perceived payoffs (Robson and White-
head 2018; Aridor, Grechi, and Woodford 2020; Młynarski and
Hermundstad 2021).
Moreover, even if we assume that subjects can fully adapt to
the environment through experience, this leaves open an impor-
tant question regarding how much experience is needed to learn
the prior distribution. For example, previous experimental work
from HWP shows that convergence takes place after about 200
trials. To explore these dynamics in our data, we analyze the pat-
terns of risky choice over the course of Experiment 1. Specifically,
we compare behavior across the first and second half of the first
experimental condition; that is, we compare behavior on trials
31–165 with behavior on trials 166–300. For each of these two
subsamples, we estimate the mixed-effects linear regression spec-
ified in Table I. If the treatment effect becomes stronger in the
second half of the first condition, this would provide evidence that
adaptation is not complete by trial 165.
Online Appendix Table D.4 shows that the estimated coeffi-
cients on X × high and C × high are significant at the 5% level in
both subsamples, indicating that the treatment effect is present
in both halves of the first condition. Nonetheless, we do not de-
tect any significant difference in the strength of the treatment
effect between the two subsamples. Thus, we cannot rule out the
possibility that full adaptation has taken place by trial 165.
An alternative way to assess the speed of adaptation is by
looking at the time series of response times from Figure VI. One
can see signatures of learning dynamics as response times fall
sharply at the beginning of the experiment; importantly, we find
that the decrease in response times continues after the first 30
adaptation trials. In fact, response times asymptote around trial
200, consistent with the experimental evidence from HWP. Al-
though the decrease in response times may reflect learning about
the task in general—rather than about the prior exclusively—the
spike in average response time in the upper panel of Figure VI
(at trial 301) strongly suggests that response times capture infor-
mation about the adaptation process. We also emphasize that not
all past observations are likely to receive the same weight when
subjects form a prior. For example, if subjects are more likely to

recall past payoffs that were experienced more recently or payoffs
that are more similar to those in the current choice set, then they
may overweight these recent or similar stimuli when forming a

prior. As such, research from the memory literature is likely to
have important implications for efficient coding of economic stim-
uli (Kahana 2012; Bordalo, Gennaioli, and Shleifer 2020; Wachter
and Kahana 2021).
V.B. Multidimensional Efficient Coding

The efficient coding model we present in Section II assumes
that the capacity constraint parameter n is fixed and does not
vary with the prior distribution. It is conceivable, however, that
the DM optimally chooses to allocate a larger capacity n toward
the dimensions of X and C when the payoff volatility along these
dimensions is higher. Intuitively, when the payoff volatility along
the two dimensions of X and C becomes higher, these dimensions
could receive more resources from a “third” dimension, which we
can think of as representing other task demands besides perceiv-
ing X and C. Indeed, this is an implication from the multidi-
mensional efficient coding models of Woodford (2012a, 2012b) and
Dewan (2020)—although the performance objectives in those mod-
els slightly differ from what we assume in Section II.
Interestingly, our experimental data are consistent with this
implication: when separately estimating the capacity constraint
parameter n across the low- and high-volatility conditions using
the risky choice data from Experiment 1, we find n = 7.05 for
the low-volatility condition and n = 11.66 for the high-volatility
condition (p-value < .001 for a within-subject test). Moreover, the
rank correlation between the values of n estimated from the two
volatility conditions is 0.53 (p-value < .001), indicating that n is
a persistent trait at the subject level.
Although we do find that subjects allocate a larger capac-
ity toward perceiving X and C in the high-volatility condition,
this does not imply that choice sensitivity is higher in the high-
volatility condition, compared with the low-volatility condition.
Indeed, Figure IX shows that our main theoretical result, as pre-
sented earlier in Figure III, continues to hold when we set n = 7
for the low-volatility condition and n = 12 for the high-volatility
condition (these values of n are chosen to match the results from
the structural estimation).

FIGURE IX
Model-Predicted Probability of Choosing the Risky Lottery when the Capacity
Constraint Parameter n Changes across Volatility Environments
The graph plots, for each of the two volatility environments (low volatility:
Xl = 16, Xu = 24, Cl = 8, and Cu = 12; high volatility: Xl = 8, Xu = 32, Cl =
4, and Cu = 16), the probability of risk taking, defined in equation (17). The prior
distributions for X and C are uniform. The probability p that the risky lottery
pays X dollars is set to 0.5. The capacity constraint parameter n is set to 7 for the
low-volatility environment and 12 for the high-volatility environment. For each
volatility environment, we draw X uniformly from [16, 24] and C uniformly from
[8, 12]. We then compute, for a given X and C, the probability of risk taking. Finally,
we aggregate these probabilities for each level of pX − C.
How can the larger capacity parameter n that we estimate in

the high-volatility condition generate less choice sensitivity dis-
played in Figure IX? As we have done throughout the article,
here we analyze only those payoffs that are drawn from the com-
mon support of the high- and low-volatility distributions. This
restriction is crucial for identifying the effect of context, which
we implement in our experiment through the use of common tri-
als. But importantly, the optimal coding capacity allocated to the
common trials in the high-volatility condition is only a small frac-
tion of the overall capacity; a large remaining fraction of the cod-
ing capacity is consumed by perceiving payoffs in the tails of the
high-volatility distribution. Therefore, although the DM chooses
a larger overall capacity in the high-volatility condition, this ef-
fect is more than offset by the fact that we restrict our analysis
to only those payoffs in the common support of the two volatility
distributions.
V.C. Comparison with Alternative Theories

In this section, we discuss alternative models of behavior and
how their predictions relate to our experimental results. As noted

in Section I, the main prediction we test in Experiment 1—that
choice sensitivity to payoff values increases when the dispersion of
potential values decreases—is also shared by models of “normal-
ization” (Rangel and Clithero 2012; Louie, Glimcher, and Webb
2015).23 A particularly relevant class of normalization models are
those in which value is normalized based on the range of potential
stimuli (Soltani, De Martino, and Camerer 2012; Rustichini et al.
2017). Under this class of models, the subjective value of a payoff
depends only on the payoff itself and the range of potential
payoffs. The results from Experiment 1 strongly support the pre-
dictions of range normalization models. In particular, our results
highlight the interpretation that normalization can implement
normative principles of efficient coding.24 It is worth noting that
Experiment 2 offers a test between our efficient coding model and
range normalization models. In Experiment 2, we hold the range
of payoffs constant across experimental conditions. Therefore,
our finding of greater risk taking in the increasing condition
cannot be explained by range normalization models, suggesting
that other statistics of the distribution besides the range do affect
coding. See Online Appendix C for more discussion.
In the decision-by-sampling (DbS) model by Stewart, Chater,
and Brown (2006), the DM’s subjective value of a stimulus is given
by its rank within a distribution of values recalled from memory.
23. Several experiments have found evidence consistent with normalization

of value signals in the brain (e.g., Tobler, Fiorillo, and Schultz 2005; Padoa-
Schioppa 2009). For behavioral evidence consistent with normalization,
see Soltani, De Martino, and Camerer (2012), Khaw, Glimcher, and Louie (2017),
and Zimmermann, Glimcher, and Louie (2018). Recent behavioral economic the-
ories also invoke normalization to explain several prominent patterns of context-
dependent choice (Glimcher and Tymula 2019; Landry and Webb 2021).
24. Not all models of normalization are grounded in principles of optimization;
some are developed to describe the decision process and its outcome. For example,
in the Soltani, De Martino, and Camerer (2012) model, range normalization is
assumed, and its implications are shown to provide a good description of decoy
effects in risky choice (though normalization takes place over values on a single
experimental trial, rather than over the history of trials experienced). In more
recent models, such as Rustichini et al. (2017), normalization is the outcome of
an optimization procedure. Relatedly, the idea that the DM puts less weight on
attributes that exhibit greater variability is a key assumption in the relative
thinking model by Bushong, Rabin, and Schwartzstein (2021).
To the extent that the distribution of recalled values is related

to the prior distribution that we focus on in this article, DbS and
efficient coding models make qualitatively similar predictions. In
fact, Bhui and Gershman (2018) show that efficient coding can

serve as a normative foundation for DbS. We interpret the data
from both of our risky choice experiments as novel evidence con-
sistent with the core mechanism in DbS.25
Kőszegi and Rabin (2007) (KR) offer a model of risky choice
where the reference point is given by rational expectations about
outcomes from a reference lottery. At a basic level, KR and effi-
cient coding share the feature that expectations shape the DM’s
perception of a lottery payoff. Yet an important distinction exists
between the two models. In efficient coding, the driving force of
the model is the DM’s expectation of a payoff value, after it has
been presented in the choice set. In KR, however, the driving force
is the DM’s expectation over which payoff value she will receive
as a future, unrealized outcome from the lottery. Therefore, con-
ditional on a choice set, KR does not predict a change in behavior
as the prior distribution varies. We further elaborate this point in
Online Appendix C.
Salience theory is an alternative model of risky choice that is
also grounded in principles of perception and delivers context-
dependent behavior (Bordalo, Gennaioli, and Shleifer 2012).
Salience theory and models of efficient coding are fundamen-
tally linked. Under salience theory, attention is drawn to payoffs
that are very different from a reference payoff. Consistent with
this assumption, the multidimensional efficient coding models of
Woodford (2012a, 2012b) and Dewan (2020) imply that when
the payoff volatility along a dimension increases, more coding
25. Our evidence also speaks to a recent controversy in interpreting tests of

the DbS model. Stewart, Reimers, and Harris (2015) and Walasek and Stewart
(2015) claim to find supporting evidence for DbS by manipulating the distribution
of payoffs across choice sets, similar to the manipulation in our design. How-
ever, a reanalysis of their experimental evidence finds that neither paper can be
interpreted as supporting DbS (Alempaki et al. 2019; André and de Langhe forth-
coming). The issue arises from the fact that behavior was analyzed on different
choice sets across experimental conditions. In contrast, our design has the im-
portant advantage of presenting a collection of choice sets that are common to
both conditions, and the common choice sets reflect the statistical properties of the
payoff distribution assumed in the theory. Thus, our results should help restore
faith in the empirical validity of DbS. Moreover, our design provides a template
for future experimental tests of the DbS theory.
resources will flow toward that dimension, and therefore extreme

payoffs will receive more weight in the decision process.
The two models, however, differ with respect to their primi-
tive assumptions. Bordalo, Gennaioli, and Shleifer (2012) appeal

to Weber’s law of diminishing sensitivity, in part, as a justification
for their definition of salience; importantly, in their model, Weber’s
law is an exogenous assumption. In contrast, Weber’s law arises
endogenously under efficient coding for prior distributions that
are decreasing. It is also notable that for prior distributions that
are increasing in payoff values, efficient coding predicts an “anti-
Weber’s” law. This difference leads the two models to generate dif-
ferent predictions in many environments. For example, salience
theory does not predict that risk taking will increase when the
DM is adapted to an increasing payoff distribution, as we find in
Experiment 2. Nor does it deliver stochastic choice, where the de-
gree of stochasticity changes systematically with the prior—as we
observe in Experiment 1. At the same time, there are extant em-
pirical patterns in the literature that salience theory can explain
and the baseline efficient coding model from Section II cannot,
such as the dependence of risk taking on the correlation between
mutually exclusive lotteries.
V.D. Instability of Preference Parameter Estimates

Our main experimental results are related to but fundamen-
tally distinct from much work in experimental economics that
documents how risk taking depends systematically on the real-
ized lottery outcomes from previous choices (Thaler and Johnson
1990; Weber and Camerer 1998; Imas 2016). Importantly, we find
that even when realized lottery outcomes are not presented to
subjects, the distribution of previous choice sets still causes sys-
tematic variation in behavior. As a result, efficient coding may
provide a distinct source of variation of behavior in typical lab ex-
periments in which preference parameters are elicited by present-
ing subjects with a sequence of choice sets (Broomell and Bhatia
2014).
The causal effect of past choice sets on risk taking is par-
ticularly relevant for newer methods of preference elicitation in
which the ordering of choice sets is tailored in real time to a sub-
ject’s history of choices. For example, Toubia et al. (2013) and
Chapman et al. (2019) present subjects with choice sets that max-
imize the information gain for estimating preference parameters
from prospect theory. Efficient coding suggests that the optimal

choice set to present to a subject should condition not only on
the history of the subject’s choices but also on the history of the
presented choice sets. Conditioning on this extra aspect of the sub-

ject’s past experience should therefore aid in further optimizing
experimental designs.
VI. CONCLUSION
We have experimentally tested the hypothesis that efficient
coding, a core principle from neuroscience, is a driving force in
decision making under risk. Our results provide strong evidence
that the DM’s willingness to take risk depends systematically on
the payoff distribution to which she has recently adapted. Our ex-
perimental data are consistent with the noisy perception of lottery
payoffs; moreover, we find that the noise distribution varies as an
optimal response to a change in the environment. In Experiment
1, we show that risky choice becomes noisier as the volatility of
the payoff distribution increases. In Experiment 2, we find that
the level of risk taking changes with the shape of the payoff dis-
tribution, which highlights the role that efficient coding plays in
generating perceptual biases. Together, our data indicate that risk
taking is systematically unstable across environments, in a man-
ner that closely mimics the instability of sensory perception.
Our results raise a number of important directions for future
work. There is a strong need to understand how the DM adapts
to a given environment based on the history of perceived pay-
offs. This mechanism of course depends on the DM’s prior beliefs
about payoffs—which we manipulate in our experiments—but it
also depends on higher-order priors about the rate at which the
environment changes. For example, if the DM expects the environ-
mental distribution to change rapidly, then adaptation will also
likely take place at a fast pace (Behrens et al. 2007; Nassar et al.
2012). Theory is already being developed along this direction, but
future experimental evidence of the adaptation process will be
critical in guiding further development of such theory (Robson and
Whitehead 2018; Aridor, Grechi, and Woodford 2020; Młynarski
and Hermundstad 2021).
Another important direction for future research is to test the
implications of efficient coding outside the laboratory. A challenge
here is to measure the prior distribution to which the DM has
adapted. A more refined theory of adaptation will be integral for
guiding empirical work in the field, as it will shed light on the

relevant time scale for forming prior beliefs, and hence percep-
tions. Institutional factors will also probably shape the relevant
time scale for adaptation. For example, in financial markets, a

stock’s price distribution over the past 52 weeks is typically salient
to investors and therefore may be a good candidate for investors’
prior distribution. We ourselves expect that future progress on
the topic of efficient coding will benefit from the close interplay
between theory, experimental tests, and empirical validations in
the field.
UNIVERSITY OF SOUTHERN CALIFORNIA, UNITED STATES

CALIFORNIA INSTITUTE OF TECHNOLOGY, UNITED STATES
SUPPLEMENTARY MATERIAL
An Online Appendix for this article can be found at The
Quarterly Journal of Economics online.
DATA AVAILABILITY
Data and code replicating the tables and figures in this article
can be found in Frydman and Jin (2021) in the Harvard Dataverse,
https://doi.org/10.7910/DVN/PYQXAD.
REFERENCES
Alempaki, Despoina, Emina Canic, Timothy L. Mullett, William J. Skylark, Chris
Starmer, Neil Stewart, and Fabio Tufano, “Reexamining How Utility and
Weighting Functions Get Their Shapes: A Quasi-Adversarial Collaboration
Providing a New Interpretation,” Management Science, 65 (2019), 4841–4862.
André, Quentin, and Bart de Langhe, “No Evidence for Loss Aversion Disappear-
ance and Reversal in Walasek and Stewart (2015),” Journal of Experimental
Psychology: General (forthcoming), https://doi.org/10.1037/xge0001052.
Aridor, Guy, Francesco Grechi, and Michael Woodford, “Adaptive Efficient Coding:
A Variational Auto-encoder Approach,” Columbia University Working Paper,
2020.
Barberis, Nicholas, “Psychology-Based Models of Asset Prices and Trading Vol-
ume,” in Handbook of Behavioral Economics, Vol. 1, Douglas Bernheim, Ste-
fano DellaVigna and David Laibson, eds. (Amsterdam: North Holland, 2018),
Ch. 2, 79–175.
Barlow, Horace, “Possible Principles Underlying the Transformations of Sensory
Messages,” in Sensory Communication, Walter A. Rosenblith, ed. (Cambridge,
MA: MIT Press, 1961), 217–234.
Behrens, Timothy, Mark Woolrich, Mark Walton, and Matthew Rushworth,
“Learning the Value of Information in an Uncertain World,” Nature Neuro-
science, 10 (2007), 1214–1221.
Bhui, Rahul, and Samuel Gershman, “Decision by Sampling Implements Efficient

Coding of Psychoeconomic Functions,” Psychological Review, 125 (2018), 985–
1001.
Bogacz, Rafal, Eric Brown, Jeff Moehlis, Philip Holmes, and Jonathan D. Cohen,
“The Physics of Optimal Decision Making: A Formal Analysis of Models of

Performance in Two-Alternative Forced-Choice Tasks,” Psychological Review,
113 (2006), 700–765.
Bordalo, Pedro, Nicola Gennaioli, and Andrei Shleifer, “Salience Theory of Choice
under Risk,” Quarterly Journal of Economics, 127 (2012), 1243–1285.
———, “Memory, Attention, and Choice,” Quarterly Journal of Economics, 135
(2020), 1399–1442.
Broomell, Stephen B., and Sudeep Bhatia, “Parameter Recovery for Decision Mod-
eling Using Choice Data,” Decision, 1 (2014), 252–274.
Bushong, Benjamin, Matthew Rabin, and Joshua Schwartzstein, “A Model of Rel-
ative Thinking,” Review of Economic Studies, 88 (2021), 162–191.
Carandini, Matteo, and David J. Heeger, “Normalization as a Canonical Neural
Computation,” Nature Reviews Neuroscience, 13 (2012), 51–62.
Chapman, Jonathan, Erik Snowberg, Stephanie Wang, and Colin Camerer, “Loss
Attitudes in the U.S. Population: Evidence from Dynamically Optimized Se-
quential Experimentation (DOSE),” New York University Abu Dhabi Working
Paper, 2019.
Dehaene, Stanislas, The Number Sense (Oxford: Oxford University Press, 2011).
Dehaene, Stanislas, Emmanuel Dupoux, and Jacques Mehler, “Is Numerical Com-
parison Digital? Analogical and Symbolic Effects in Two-Digit Number Com-
parison,” Journal of Experimental Psychology: Human Perception and Perfor-
mance, 16 (1990), 626–641.
Denrell, Jerker C., “Reference-Dependent Risk Sensitivity as Rational Inference,”
Psychological Review, 122 (2015), 461–484.
Dewan, Ambuj, “Costly Multidimensional Information,” New York University
Working Paper, 2020.
Enke, Benjamin, and Thomas Graeber, “Cognitive Uncertainty,” Harvard Univer-
sity Working Paper, 2021.
Friedman, Daniel, “The S-Shaped Value Function as a Constrained Optimum,”
American Economic Review, 79 (1989), 1243–1248.
Frydman, Cary, and Lawrence J. Jin, “Replication Data for: ‘Efficient Code
and Risky Choice’,” (2021), Harvard Dataverse, https://doi.org/10.7910/
DVN/PYQXAD.
Fudenberg, Drew, Philipp Strack, and Tomasz Strzalecki, “Speed, Accuracy, and
the Optimal Timing of Choices,” American Economic Review, 108 (2018), 3651–
3684.
Gabaix, Xavier, and David Laibson, “Myopia and Discounting.” NBER Working
Paper no. 23254, 2017.
Girshick, Ahna R., Michael S. Landy, and Eero P. Simoncelli, “Cardinal Rules: Vi-
sual Orientation Perception Reflects Knowledge of Environmental Statistics,”
Nature Neuroscience, 14 (2011), 926–932.
Glimcher, Paul, and Agnieszka Tymula, “Expected Subjective Value Theory
(ESVT): A Representation of Decision under Risk and Certainty,” SSRN Work-
ing Paper, 2019.
Hébert, Benjamin, and Michael Woodford, “Rational Inattention When Decisions
Take Time,” Stanford University Working Paper, 2019.
Heng, Joseph, Michael Woodford, and Rafael Polanı́a, “Efficient Sampling and
Noisy Decisions,” eLife, 9 (2020), e54962.
Imas, Alex, “The Realization Effect: Risk-Taking after Realized versus Paper
Losses,” American Economic Review, 106 (2016), 2086–2109.
Kahana, Michael J., Foundations of Human Memory (Oxford: Oxford University
Press, 2012).
Kahneman, Daniel, and Amos Tversky, “Prospect Theory: An Analysis of Decision
under Risk,” Econometrica, 47 (1979), 263–291.
Khaw, Mel Win, Paul W. Glimcher, and Kenway Louie, “Normalized Value Coding
Explains Dynamic Adaptation in the Human Valuation Process,” Proceedings
of the National Academy of Sciences, 114 (2017), 12696–12701.
Khaw, Mel Win, Ziang Li, and Michael Woodford, “Cognitive Imprecision and
Small-Stakes Risk Aversion,” Review of Economic Studies, 88 (2021), 1979–

2013.
Kőszegi, Botond, and Matthew Rabin, “Reference-Dependent Risk Attitudes,”
American Economic Review, 97 (2007), 1047–1073.
Krajbich, Ian, Carrie Armel, and Antonio Rangel, “Visual Fixations and the Com-
putation and Comparison of Value in Simple Choice,” Nature Neuroscience, 13
(2010), 1292–1298.
Landry, Peter, and Ryan Webb, “Pairwise Normalization: A Neuroeconomic Theory
of Multi-Attribute Choice,” Journal of Economic Theory, 193 (2021), 105221.
Laughlin, Simon, “A Simple Coding Procedure Enhances a Neuron’s Information
Capacity,” Zeitschrift fur Naturforschung, 36 (1981), 910–912.
Louie, Kenway, Paul W. Glimcher, and Ryan Webb, “Adaptive Neural Coding: From
Biological to Behavioral Decision-Making,” Current Opinion in Behavioral
Sciences, 5 (2015), 91–99.
Ma, Wei Ji, and Michael Woodford, “Multiple Conceptions of Resource Rationality,”
Behavioral and Brain Sciences, 43 (2020), e15.
Moyer, Robert S., and Thomas K. Landauer, “Time Required for Judgments of
Numerical Inequality,” Nature, 215 (1967), 1519–1520.
Młynarski, Wiktor F., and Ann M. Hermundstad, “Efficient and Adaptive Sensory
Codes,” Nature Neuroscience, 24 (2021), 998–1009.
Nassar, Matthew R., Katherine M. Rumsey, Robert C. Wilson, Kinjan Parikh,
Benjamin Heasly, and Joshua I. Gold, “Rational Regulation of Learning Dy-
namics by Pupil-Linked Arousal Systems,” Nature Neuroscience, 15 (2012),
1040–1046.
Natenzon, Paulo, “Random Choice and Learning,” Journal of Political Economy,
127 (2019), 419–457.
Netzer, Nick, “Evolution of Time Preferences and Attitudes toward Risk,” Ameri-
can Economic Review, 99 (2009), 937–955.
Padoa-Schioppa, Camillo, “Range-Adapting Representation of Economic Value in
the Orbitofrontal Cortex,” Journal of Neuroscience, 29 (2009), 14004–14014.
Payzan-LeNestour, Elise, and Michael Woodford, “Outlier Blindness: A Neuro-
biological Foundation for Neglect of Financial Risk,” Journal of Financial
Economics (forthcoming), https://doi.org/10.1016/j.jfineco.2021.06.019.
Polanı́a, Rafael, Michael Woodford, and Christian Ruff, “Efficient Coding of Sub-
jective Value,” Nature Neuroscience, 22 (2019), 134–142.
Rangel, Antonio, and John A. Clithero, “Value Normalization in Decision Making:
Theory and Evidence,” Current Opinion in Neurobiology, 22 (2012), 970–981.
Ratcliff, Roger, “A Theory of Memory Retrieval,” Psychological Review, 85 (1978),
59–108.
Rayo, Luis, and Gary S. Becker, “Evolutionary Efficiency and Happiness,” Journal
of Political Economy, 115 (2007), 302–337.
Robson, Arthur J., “The Biological Basis of Economic Behavior,” Journal of Eco-
nomic Literature, 39 (2001), 11–33.
Robson, Arthur J., and Lorne A. Whitehead, “Adaptive Hedonic Utility,” Simon
Fraser University Working Paper, 2018.
Rustichini, Aldo, Katherine E. Conen, Xinying Cai, and Camillo Padoa-Schioppa,
“Optimal Coding and Neuronal Adaptation in Economic Decisions,” Nature
Communications, 8 (2017), 1208.
Soltani, Alireza, Benedetto De Martino, and Colin Camerer, “A Range-
Normalization Model of Context-Dependent Choice: A New Model and Evi-
dence,” PLoS Computational Biology, 8 (2012), e1002607.
Steiner, Jakub, and Colin Stewart, “Perceiving Prospects Properly,” American Eco-
nomic Review, 106 (2016), 1601–1631.
Stewart, Neil, Nick Chater, and Gordon Brown, “Decision by Sampling,” Cognitive
Psychology, 53 (2006), 1–26.
Stewart, Neil, Stian Reimers, and Adam J. L. Harris, “On the Origin of Utility,
Weighting, and Discounting Functions: How They Get Their Shapes and How
to Change Their Shapes,” Management Science, 61 (2015), 687–705.
Stigler, George J., “The Economics of Information,” Journal of Political Economy,
69 (1961), 213–225.

Thaler, Richard, and Eric Johnson, “Gambling with the House Money and Trying
to Break Even: The Effects of Prior Outcomes on Risky Choice,” Management
Science, 36 (1990), 643–660.
Tobler, Philippe N., Christopher D. Fiorillo, and Wolfram Schultz, “Adapting Cod-
ing of Reward Value by Dopamine Neurons,” Science, 307 (2005), 1642–1645.
Toubia, Olivier, Eric Johnson, Theodoros Evgeniou, and Philippe Delquié, “Dy-
namic Experiments for Estimating Preferences: An Adaptive Method of Elic-
iting Time and Risk Parameters,” Management Science, 59 (2013), 613–640.
Wachter, Jessica A., and Michael J. Kahana, “A Retrieved-Context Theory of Fi-
nancial Decisions,” University of Pennsylvania Working Paper, 2021.
Walasek, Lukasz, and Neil Stewart, “How to Make Loss Aversion Disappear and
Reverse: Tests of the Decision by Sampling Origin of Loss Aversion,” Journal
of Experimental Psychology: General, 144 (2015), 7–11.
Weber, Martin, and Colin Camerer, “The Disposition Effect in Securities Trading:
An Experimental Analysis,” Journal of Economic Behavior and Organization,
33 (1998), 167–184.
Wei, Xue-Xin, and Alan A. Stocker, “A Bayesian Observer Model Constrained by
Efficient Coding Can Explain ‘Anti-Bayesian’ Percepts,” Nature Neuroscience,
18 (2015), 1509–1517.
———, “Lawful Relation between Perceptual Bias and Discriminability,” Proceed-
ings of the National Academy of Sciences, 114 (2017), 10244–10249.
Woodford, Michael, “Inattentive Valuation and Reference-Dependent Choice,”
Columbia University Working Paper, 2012a.
———, “Prospect Theory as Efficient Perceptual Distortion,” American Economic
Review Papers and Proceedings, 102 (2012b), 41–46.
———, “Stochastic Choice: An Optimizing Neuroeconomic Model,” American Eco-
nomic Review, 104 (2014), 495–500.
———, “Modeling Imprecision in Perception, Valuation and Choice,” Annual Re-
view of Economics, 12 (2020), 579–601.
Zimmermann, Jan, Paul W. Glimcher, and Kenway Louie, “Multiple Timescales of
Normalized Value Coding Underlie Adaptive Choice Behavior,” Nature Com-
munications, 9 (2018), 3206.

Efficient Coding and Risky Choice

Uploaded by

Copyright:

Available Formats

Efficient Coding and Risky Choice

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Efficient Coding and Risky Choice

Uploaded by

Copyright:

Available Formats

EFFICIENT CODING AND RISKY CHOICE∗

CARY FRYDMAN AND LAWRENCE J. JIN

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

∗ We are grateful to Andrei Shleifer (the editor), four anonymous referees,

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

1. For experimental evidence consistent with efficient coding in sensory per-

KLW model, which assumes that the DM observes noisy signals

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

payoffs is smaller.2 This prediction is inconsistent with most

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

2. This prediction is derived under the assumption that coding resources do

volatility condition, and therefore our results cannot be driven

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

sets across the two conditions, we find evidence consistent with

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

4. For alternative approaches to endogenizing the value function in prospect

The rest of the article is organized as follows. In Section II, we

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

II. THE MODEL

II.A. Choice Environment

(1) p · U (X) + (1 − p) · U (0) > U (C).

Conditional on X, C, and p, the DM’s choice is nonstochastic.

by the literature in sensory perception, where a common finding

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

6. In the choice environment we study here, our interpretation is that the

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

Expected financial gain

We focus on the performance objective in equation (2) because

X can then be defined by

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

Importantly, we see that the coding rules depend explicitly on

(7) max I(X, Rx ),

where the mutual information I(X, Rx ) is defined as the difference

Interestingly, there exists a class of priors for which the three

(i) pX and C are i.i.d.

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

where Xl , Xu , Cl , and Cu are the bounds of the uniform distribu-

The expressions in equation (12) show that the likelihood func-

Xl , Xu , Cl , and Cu . This dependence of the likelihood function on

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

II.C. Value Function and Implications for Choice

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

That is, v(X) represents the subjective valuation of X averaged

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

where Xl = 2, Xu = 8, Xm d = 5.5, h = 7 , and l = 1 . The lower graph plots the

To formalize this prediction, we compute the probability of

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

II.D. Increasing and Decreasing Payoff Distributions

We specify the decreasing distribution by

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

III. EXPERIMENT 1: VOLATILITY MANIPULATION

III.A. Design of the Risky Choice Task

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

order of the two conditions is randomized across subjects. Each

Downloaded from https://academic.oup.com/qje/article/137/1/161/6354795 by KIT Library user on 07 February 2022

is more natural than telling subjects the payoff distributions they