Vol 441|11 May 2006|doi:10.

1038/nature04676

LETTERS
Neurons in the orbitofrontal cortex encode
economic value
Camillo Padoa-Schioppa1 & John A. Assad1

Economic choice is the behaviour observed when individuals analyse each cell in relation to the choice pattern recorded in the same
select one among many available options. There is no intrinsically session.
‘correct’ answer: economic choice depends on subjective prefer- Our recordings focused on area 13 in the OFC. Figure 2 illustrates
ences. This behaviour is traditionally the object of economic the activity of one representative neuron. The cell’s activity does not
analysis1 and is also of primary interest in psychology2. However, depend on whether juice A is offered on the left or on the right
the underlying mental processes and neuronal mechanisms are (Fig. 2b). It also does not depend on whether the monkey chooses
not well understood. Theories of human and animal choice1–3 have the juice on the left or the juice on the right (that is, makes an eye
a cornerstone in the concept of ‘value’. Consider, for example, a movement to the left or to the right; Fig. 2c). However, the cell’s
monkey offered one raisin versus one piece of apple: behavioural activity varies with the offer type. This is consistent across
evidence suggests that the animal chooses by assigning values to the neuronal population. We recorded the activity of 931 cells
the two options4. But where and how values are represented in the and we analysed their neuronal responses in seven time windows
brain is unclear. Here we show that, during economic choice, (see Methods). We tested the activity of each cell in each time
neurons in the orbitofrontal cortex5–18 (OFC) encode the value of window with a three-way analysis of variance (ANOVA, with factors:
offered and chosen goods. Notably, OFC neurons encode value [position of juice A] £ [movement direction] £ [offer type],
independently of visuospatial factors and motor responses. If a P , 0.001). Rarely do responses depend on either the spatial con-
monkey chooses between A and B, neurons in the OFC encode the figuration of the offers or the motor output (,5% neurons). In
value of the two goods independently of whether A is presented on contrast, the activity of 505 (54%) neurons varies significantly
the right and B on the left, or vice versa. This trait distinguishes depending on the offer type in at least one time window. Pooling
the OFC from other brain areas in which value modulates activity time windows, a total of 1,379 responses are significantly modulated
related to sensory or motor processes19–25. Our results have broad by the offer type (Supplementary Fig. S2).
implications for possible psychological models, suggesting that The cell shown in Fig. 2d has a U-shaped response similar to that
economic choice is essentially choice between goods rather than hypothesized for a neuron encoding the chosen value. For this
choice between actions. In this framework, neurons in the OFC session V(A) ¼ 1.9V(B). Accordingly, the activity of the cell is
seem to be a good candidate network for value assignment
underlying economic choice.
In our experiments, monkeys choose between two types of juice
(A and B; where A is preferred) offered in different amounts. For
example, in the session shown in Fig. 1, the monkey chooses between
water (juice A) and unsweetened Kool-Aid (juice B). Offer types
include 1B:2A, 1B:1A, 2B:1A, 3B:1A, 4B:1A, 6B:1A, and 10B:1A, and
the ‘forced choices’ 0B:1A and 3B:0A. Behaviourally, we observe a
trade-off between juice type and juice quantity. The monkey chooses
A when 1B, 2B, or 3B are available as alternatives, it is roughly
indifferent between the two juices when offered 4B:1A, and it chooses
B when 6B or 10B are available.
We interpret this pattern of choice in terms of the ‘relative value’ of
the two juices4: in this case, the value of 1A is roughly equal to the
value of 4B. Fitting a sigmoid curve provides the better estimate
V(1A) ¼ V(4.1B), where V(x) indicates the value of x. Assuming a
linear value function, we obtain V(A) ¼ 4.1V(B). This equation puts
different quantities of juices A and B on the same value scale. On this
basis, we can compute for each trial the value of the juice chosen by
the monkey. Expressing values in units of V(B), the variable ‘chosen
value’ is about 4 when the monkey chooses 1A or 4B. When the
monkey chooses 2A, the chosen value is about 8. When the monkey Figure 1 | Experimental design. a, Trial structure (see Methods). b, Example
chooses 6B, 10B or 3B, the chosen value is respectively equal to 6, 10 of behavioural choice pattern. The plot shows the percentage of trials in
or 3. Hence, we can make specific hypotheses regarding the neuronal which the monkeys chose juice B (y axis) for various offer types (x axis). A
representation of juice values. In different sessions and with different sigmoid fit provides the measure of the relative value n* ¼ 4.1. The dotted
juices, we record different behavioural choice patterns. We then red circle indicates the saccade target chosen by the monkey.

1
Department of Neurobiology, Harvard Medical School, 220 Longwood Avenue, Boston, Massachusetts 02115, USA.

223
© 2006 Nature Publishing Group
LETTERS
low when the monkey chooses 1A or 2B (in units of V(B), chosen
value < 2), it is higher when the monkey chooses 2A or 4B (chosen
value < 4), and it is highest when the monkey chooses 3A or 6B
(chosen value < 6). A linear regression of this response on the
variable chosen value provides R 2 ¼ 0.86. Similar U-shaped
responses are frequent in the OFC, and Fig. 3a–c illustrates three
more examples. We also find other types of responses. For example,
neuronal responses often reflect the value of one of the two juices
alone. Figure 3d, e shows two cells in which activity covaries with
the value of A offered and the value of B offered, respectively. We
label these responses as related to the variable ‘offer value’. Other
frequently observed responses vary in a binary fashion depending
on the type of juice chosen by the monkey, independently of the
amount (Fig. 3f). We interpret these responses as related to the
variable juice ‘taste’.
Figure 2 | Activity of one neuron. a, Rasters. Each line represents one trial
and each small dot represents one spike. Trials, arranged by offer type, are
aligned at the ‘offer’ (left) and at the ‘juice’ (right). The blue highlight marks
the post-juice time window. ‘Sacc’ indicates the time of the saccade.
b, Activity profiles shown separately for trials in which juice A is offered on
the left (red) or on the right (green). The cell activity does not depend on the
spatial configuration of the visual stimulus. c, Activity profiles shown
separately for trials in which the monkey chooses the juice offered on the left
(red) or on the right (green). The cell activity does not depend on the
direction of the eye movement. d, The top panel shows the choice pattern
recorded in this session (n* ¼ 1.9). The bottom panel shows the activity of
the cell (^s.e.m.) recorded in the pre-offer (light grey, control) and post-
juice (black) time windows. Note that the response does not reflect simple
physical properties of the visual stimulus, such as the number of squares
displayed on the monitor. For example, offer types 1B:3A and 3B:1A, which
are visually identical except for the colour of the squares, elicit very different
activation.
224
© 2006 Nature Publishing Group
NATURE|Vol 441|11 May 2006
Although many neurons seem to encode chosen value, offer value
or juice taste, the relation between their activity and these three
variables could be subordinate to a correlation with other
behavioural variables. For example, neurons in the OFC might
encode the number of squares on the monitor (or variables pro-
portional to number, such as juice quantity, or absolute luminance of
the visual stimulus). To cast a wide net, we examine the linear
dependence of neuronal data on 19 possible variables (Supplemen-
tary Fig. S1). For example, we analyse the variables ‘chosen number’,
‘total number’ and ‘total value’. We include in this analysis 1,379
responses significantly modulated by the offer type, and we regress
each response separately on each variable. Collectively, the 19 vari-
ables explain 1,227 (89%) neuronal responses. However, the 19
variables are often highly correlated.
To identify a few variables that best describe the neuronal popu-
lation, we adapt procedures for variable selection commonly used in
multilinear regression in the presence of multi-collinearity. Both
the stepwise and the best-subset methods identify the variables offer
value, chosen value and taste, which explain well the large majority of
responses (1,085/1,379 ¼ 79% responses, with mean R 2 ¼ 0.63). A
post-hoc analysis indicates that the explanatory power of these three
variables is significantly higher than that of challenging alternatives.
Furthermore, data from the two monkeys analysed separately pro-
vide statistically indistinguishable results. Finally, a bilinear
regression analysis indicates that in 890/1,085 (82%) cases, adding
a second variable or a quadratic value term does not improve the
regression significantly (Supplementary Results S5 to S10 and Sup-
plementary Figs S4 to S11). We conclude that, as a population, OFC
responses indeed encode the variables offer value, chosen value and
taste.
We next turn to a specific analysis of U-shaped responses (Figs 2d,
3a–c). In our experiments, relative values were generally stable within
any recording session. However, the relative value of any given pair of
juices could vary from day to day. For example, the relative value of
apple juice versus peppermint tea varied between 1.5 and 3. This
variability provides a further opportunity to test the neuronal
encoding of value; specifically, U-shaped responses should reflect
this variability. For this analysis, we test the entire neuronal popu-
lation with the regression function a0 þ aA ðmA Þ þ aB ðmB Þ; where mA
and m B represent the amounts of juices A and B chosen by the
monkey. We define a response to be U-shaped if both a A and a B differ
from zero (P , 0.01). If U-shaped responses indeed encode the value
of the chosen juice, the slope ratio k* ¼ a A/a B should be, for each
U-shaped response, equal to the relative value (n*) measured in that
session. Most importantly, the slope ratio k* obtained for different
responses should covary with n*. To test this prediction, we compute
the regression k* ¼ b 0 þ b 1n* separately for every juice pair. Aver-
aging across juice pairs, we obtain b 0 (^s.e.m.) ¼ 20.13 (^0.15)
and b 1 ¼ 1.05 (^0.15), consistent with the identity k* ¼ n*. This
result demonstrates that U-shaped responses do not reflect the
quantity of any particular juice ingredient (for example, sugar).
Rather, they encode the value monkeys assign to the juice
they choose to consume (Supplementary Results S10 to S12 and
Supplementary Figs S12 to S14).
The timing of neuronal activation is as follows: the average
neuronal activity peaks shortly after the offer, declines during the
delay, is low before and during the eye movement, and has two
secondary peaks at juice delivery (Supplementary Fig. S3). To
appreciate how the variables offer value, chosen value and taste are
represented in the OFC over time, we analyse the activity of each cell
in 50-ms, non-overlapping time bins. Figure 4 shows the number of
cells encoding each of the three variables at different times. Remark-
ably, the time profile of different variables seems to reflect the mental
processes the monkey presumably undertakes during a trial. Shortly
after the offer, when the monkey assigns values to the two juices,
neurons encoding the offer value (that is, the value of one juice or the
other) are most prevalent. Also during the delay, many neurons
NATURE|Vol 441|11 May 2006
Figure 3 | Activity of six neurons. For each cell, the top panel shows the
choice pattern, with the relative value indicated on the top left. The bottom
panel shows the activity of the cell. a–c, Responses encoding the chosen
value. The response in c is negatively correlated with the chosen value (high
activity for low value). d–e, Responses encoding the value of juice A offered
(d) and the value of juice B offered (e). We refer to these responses as related
to the offer value. f, Response encoding the juice taste. Here we separate trials
encode the chosen value (that is, the value of the juice the monkey will
eventually consume), even though the choice is still covert (because
the ‘go’ signal has not been given yet). Finally, after the monkey has
indicated its choice, before and after juice delivery, many neurons
encode the taste of the chosen juice.
Conceptually, responses encoding the chosen value are particularly
interesting because, in addition to being independent of the visuo-
motor contingencies of the task, they are also independent of the
specifics of the good, namely juice type and juice amount. These
responses encode economic value in a non-specific way. Further
research is necessary to establish whether this result generalizes to
other kinds of goods, such as non-comestible goods26. The interpre-
tation of offer value responses is made more cautiously because,
assuming linear value functions, the value of a given amount of juice
is proportional to the juice quantity.
‘Value’ is known to modulate the activity of neurons in several
sensory and motor areas19–25. For example, neurons in the lateral
intraparietal area activate when monkeys plan a saccade towards a
particular location of the visual field; their response is enhanced
Figure 4 | Time course. We assign each neuron to one of the three variables
only if the regression slope is significantly different from zero (P , 0.01),
and we include all 931 neurons in the analysis. The dotted line indicates
chance level (9.31).
© 2006 Nature Publishing Group
LETTERS
in which the monkey chose juice A (diamonds) or juice B (circles). The
response reflects the chosen juice type independently of the amount.
Responses were recorded in the post-offer (a, d, e, blue), pre-juice (b, cyan),
and post-juice (c, f, black) time windows. For each cell, the curves in light
grey show the activity in the pre-offer time window. Error bars represent
s.e.m.
when the eye movement is associated with higher value20. On this
basis, it has been proposed that parietal neurons encoding the value
of all possible courses of action form a common path for decision-
making, and that their activity is actually the subject of economic
theory27. According to this ‘action-based’ model, economic choice is
fundamentally choice between actions. That neurons in the OFC
encode the economic value of offered and chosen goods per se, as
opposed to reflecting value as a modulation of visuomotor processes,
suggests an alternative ‘good-based’ model, according to which
economic choice is fundamentally choice between goods. In this
view, choice is made between goods, and a suitable motor action is
subsequently planned and executed.
Several arguments seem to favour the good-based model. From a
computational perspective, a modular design separating the mental
operations of ‘choosing’ and ‘moving’ is more parsimonious28,29. In
addition, values processed in the OFC are logically sufficient for
good-based choice. The action-based model would thus imply that,
during economic choice, the nervous system operates in a compu-
tationally inefficient way, while undertaking all the processes needed
to choose efficiently. Finally, a vast literature links choice in various
domains to the OFC. For example, human patients and monkeys
with OFC lesions can present eating disorders and hyperorality6–8,
abnormal risk-seeking and gambling behaviour9,10, and impulsivity,
altered personality and abnormal social behaviour6,11. In contrast,
parietal lesions typically result in visuospatial deficits such as
hemi-neglect or Balint’s syndrome30. In conclusion, together with
other lines of evidence, the present results support a good-based
psychological model of economic choice behaviour.
METHODS
Each trial begins with the monkey fixating the centre of a computer monitor
(Fig. 1a). After 1.5 s, two sets of squares appear on opposite sides of the fixation
point (‘offer’). The colour of the squares indicates the juice type and the number
of squares indicates the juice amount. For example, a monkey offered three blue
squares versus one yellow square chooses between three drops of peppermint tea
and one drop of grape juice. After a randomly variable delay (1–2 s), two saccade
targets appear near the offers (‘go’). The monkey indicates its choice with an eye
movement, and must maintain fixation on the target for an additional 0.75 s
before juice delivery (‘juice’). The trial is aborted if the monkey breaks fixation
before the go signal. The amounts of the two offered juices (0–10 drops) vary
pseudo-randomly. For a given offer type, left/right positions are counter-
balanced (that is, the monkey may be offered 1A on the left and 3B on the
225
LETTERS
right, or vice versa). A variety of different pairs of juices are used in different
sessions (Supplementary Methods).
We analyse cell activity in the following time windows: 0.5 s pre-offer (a
control time window); 0.5 s post-offer; late delay (0.5–1.0 s after the offer); 0.5 s
pre-go; reaction time (from go to saccade); 0.5 s pre-juice; and 0.5 s post-juice.
For the statistical analysis, we separate for each offer type trials in which the
monkey chooses juices A and B (Supplementary Methods).
Received 28 November 2005; accepted 24 February 2006.
Published online 23 April 2006.
1. Samuelson, P. A. Foundations of Economic Analysis (Harvard Univ. Press,
Cambridge, 1947).
2. Kahneman, D. & Tversky, A. (eds) Choices, Values and Frames (Russell Sage
Foundation–-Cambridge Univ. Press, Cambridge/New York, 2000).
3. Kagel, J. H., Battalio, R. C. & Green, L. Economic Choice Theory: An Experimental
Analysis of Animal Behavior (Cambridge Univ. Press, Cambridge/New York,
1995).
4. Padoa-Schioppa, C., Jandolo, L. & Visalberghi, E. Multi-stage mental process
for economic choice in capuchins. Cognition 99, B1–-B13 (2006).
5. Ongur, D. & Price, J. L. The organization of networks within the orbital and
medial prefrontal cortex of rats, monkeys and humans. Cereb. Cortex 10,
206–-219 (2000).
6. Pasquier, F. & Petit, H. Frontotemporal dementia: its rediscovery. Eur. Neurol.
38, 1–-6 (1997).
7. Hodges, J. R. Frontotemporal dementia (Pick’s disease): clinical features and
assessment. Neurology 56, S6–-S10 (2001).
8. Butter, C. M., McDonald, J. A. & Snyder, D. R. Orality, preference behavior, and
reinforcement value of nonfood object in monkeys with orbital frontal lesions.
Science 164, 1306–-1307 (1969).
9. Rahman, S., Sahakian, B. J., Hodges, J. R., Rogers, R. D. & Robbins, T. W.
Specific cognitive deficits in mild frontal variant frontotemporal dementia. Brain
122, 1469–-1493 (1999).
10. Bechara, A., Tranel, D., Damasio, H. & Damasio, A. R. Failure to respond
autonomically to anticipated future outcomes following damage to prefrontal
cortex. Cereb. Cortex 6, 215–-225 (1996).
11. Raleigh, M. J. & Steklis, H. D. Effect of orbitofrontal and temporal neocortical
lesions of the affiliative behavior of vervet monkeys (Cercopithecus aethiops
sabaeus). Exp. Neurol. 73, 378–-389 (1981).
12. Izquierdo, A., Suda, R. K. & Murray, E. A. Bilateral orbital prefrontal cortex
lesions in rhesus monkeys disrupt choices guided by both reward value and
reward contingency. J. Neurosci. 24, 7540–-7548 (2004).
13. O’Doherty, J., Kringelbach, M. L., Rolls, E. T., Hornak, J. & Andrews, C. Abstract
reward and punishment representations in the human orbitofrontal cortex.
Nature Neurosci. 4, 95–-102 (2001).
14. Knutson, B., Taylor, J., Kaufman, M., Peterson, R. & Glover, G. Distributed
neural representation of expected value. J. Neurosci. 25, 4806–-4812 (2005).
15. Rolls, E. T., Sienkiewicz, Z. J. & Yaxley, S. Hunger modulates the responses to
gustatory stimuli of single neurons in the caudolateral orbitofrontal cortex of
the macaque monkey. Eur. J. Neurosci. 1, 53–-60 (1989).
16. Roesch, M. R. & Olson, C. R. Neuronal activity related to reward value and
motivation in primate frontal cortex. Science 304, 307–-310 (2004).
226
© 2006 Nature Publishing Group
NATURE|Vol 441|11 May 2006
17. Wallis, J. D. & Miller, E. K. Neuronal activity in primate dorsolateral and orbital
prefrontal cortex during performance of a reward preference task. Eur.
J. Neurosci. 18, 2069–-2081 (2003).
18. Tremblay, L. & Schultz, W. Relative reward preference in primate orbitofrontal
cortex. Nature 398, 704–-708 (1999).
19. Kawagoe, R., Takikawa, Y. & Hikosaka, O. Expectation of reward
modulates cognitive signals in the basal ganglia. Nature Neurosci. 1, 411–-416
(1998).
20. Platt, M. L. & Glimcher, P. W. Neural correlates of decision variables in parietal
cortex. Nature 400, 233–-238 (1999).
21. Leon, M. I. & Shadlen, M. N. Effect of expected reward magnitude on the
response of neurons in the dorsolateral prefrontal cortex of the macaque.
Neuron 24, 415–-425 (1999).
22. Shidara, M. & Richmond, B. J. Anterior cingulate: single neuronal signals related
to degree of reward expectancy. Science 296, 1709–-1711 (2002).
23. Ikeda, T. & Hikosaka, O. Reward-dependent gain and bias of visual responses
in primate superior colliculus. Neuron 39, 693–-700 (2003).
24. Roesch, M. R. & Olson, C. R. Impact of expected reward on neuronal activity in
prefrontal cortex, frontal and supplementary eye fields and premotor cortex.
J. Neurophysiol. 90, 1766–-1789 (2003).
25. McCoy, A. N., Crowley, J. C., Haghighian, G., Dean, H. L. & Platt, M. L. Saccade
reward signals in posterior cingulate cortex. Neuron 40, 1031–-1040 (2003).
26. Deaner, R. O., Khera, A. V. & Platt, M. L. Monkeys pay per view: adaptive
valuation of social images by rhesus macaques. Curr. Biol. 15, 243–-248 (2005).
27. Glimcher, P. W., Dorris, M. C. & Bayer, H. M. Physiological utility theory and
the neuroeconomics of choice. Games Econ. Behav. 52, 213–-256 (2005).
28. Simon, H. A. The architecture of complexity. Proc. Am. Phil. Soc. 106, 467–-482
(1962).
29. Pinker, S. How the Mind Works (Norton, New York, 1997).
30. Zigmond, M. J., Bloom, F. E., Landis, S. C., Roberts, J. L. & Squire, L. R.
Fundamental Neuroscience Ch. 53 (Academic, San Diego, 1999).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank K. Irwin, T. LaFratta, D. Averbuch, J. LeBlanc,
S. Peled and J. Harper for technical assistance and animal care; E. Brown for a
discussion on the statistical analysis; and T. Herrington, D. Freedman and E. Bizzi
for comments on the manuscript. This work was supported by postdoctoral
fellowships from the Lefler Foundation and from the Harvard Mind/Brain/
Behavior Initiative (to C.P.-S.) and by a grant from the National Institute of
Neurological Disorders and Stroke (to J.A.A.).
Author Contributions C.P.-S. performed all aspects of the study, including the
design of the experiment, collecting and analysing the data, and writing the
manuscript. J.A.A. assisted in experimental design, data analysis and manuscript
preparation.
Author Information Reprints and permissions information is available at
npg.nature.com/reprintsandpermissions. The authors declare no competing
financial interests. Correspondence and requests for materials should be
addressed to C.P.-S. (camillo@alum.mit.edu).
Neurons in Orbitofrontal Cortex Encode Economic
Value
By Camillo Padoa-Schioppa and John A. Assad
Supplementary Information
Supplementary Methods
Subjects, setup, and experimental design
One male (“V”, 9.5 Kg) and one female (“L”, 6.3 Kg)
rhesus monkey (Macaca mulatta) participated in the
experiments. Neither animal had been used for
previous studies. During the experiments, monkeys sat
in a monkey chair in a darkened room. Their head was
restrained, and their eye position was monitored
continuously using a scleral eye-coil system1 (River-
bend Instruments). A computer monitor was placed 57
cm in front of the monkeys. The behavioral task was
controlled by custom-written software through a
computer interface (ITC-18, Instrutech Corporation).
Each trial began with the appearance of a fixation
point (0.2° of visual angle) placed in the center of the
monitor. The monkey directed gaze to the fixation
point. After 1.5 s, two sets of squares appeared on the
left and on the right sides of the fixation point (offer).
The color of the squares indicated the juice type, and
the number of squares indicated the juice amount.
After the offer, the monkey maintained fixation for a
1-2 s randomly variable delay, until a go signal. The
go was signaled by the appearance of two saccade
targets (0.2° of visual angle) placed 7° to the left and
to the right of the center fixation point. After the go,
the monkey had 1.5 s to indicate its choice by making
an eye movement towards one of the saccade targets.
The monkey then maintained fixation onto the chosen
target for an additional 0.75 s before juice delivery.
The trial was aborted if the monkey broke center
fixation at any time before the go signal, or after target
acquisition but before juice delivery. Trials were
separated by a 1.5 s inter-trial interval. Center fixation
was imposed within 1°.
Sets of 0-10 squares indicating the offers were located
around the (initially not visible) saccade targets, within
4.2° of visual angle. The side of each individual square
subtended 1.05°. The spatial configuration of a given
set of squares (for example, 3 squares on the right)
remained the same throughout the experiments. The
same color was associated with any given juice type
throughout the experiments.
We used the following juices, listed in roughly
decreasing order of preference (associated colors in
parenthesis): high-sugar lemon Kool-Aid (dark
yellow), grape juice (bright green), fruit punch
(magenta), apple juice (diluted to 1/2 with water, dark
green), cranberry juice (diluted to 1/3 or 1/4 with
water, pink), water (white), milk (red), peppermint tea
(bright blue), tea (light brown), low-sugar agua frescas
Kool-Aid (light red), low-sugar tamarind Kool-Aid
(dark brown), slightly salted water (0.60 g/l or 0.65 g/l
concentration, light gray).
Recording sessions lasted typically 200-500 trials. The
number of trials per offer type presented in each trial
block was determined at the beginning of each session,
and we generally tried to have more trials for offer
types near the animal’s (foreseen) indifference point.
Typically, this resulted in 20-65 trials for each offer
type in each recording session. The left/right
configuration of each offer type was pseudo-
randomized and counterbalanced within each trial
block. Each day, monkeys completed 2-8 sessions
(typically 5-6), with different pairs of juices used in
different sessions.
For the current experiments, we wanted to have a
stable relative value within any recording session.
Indeed, the relative value was generally stable (though
3% of sessions were excluded from the analysis for
unstable behavior). However, juice values could vary
considerably from session to session within each day.
In particular, early in the day monkeys were thirstier,
and relative values tended to differ less from unity.
Later in the day, monkeys generally became more
selective and the relative value of less sweetened
juices decreased. As a result, some variability was also
present in the relative value measured for any given
pair of juices in different days. For any given juice
pair, the range of variability was typically twofold or
less. The order of preference of different juices (i.e.,
their ranking) remained fairly stable across sessions
and across days for both monkeys (though it was not
exactly the same for the two monkeys).
Training lasted 6-8 months. During the training,
monkeys initially indicated their choice by moving a
bidirectional lever by hand. For most of the training,
we used only water offered in variables amounts (e.g.,
1 drop versus 2 drops). Monkeys spent most of the
training learning non-specific aspects of the task, for
example that objects on the monitor carried
information about the to-be-received liquid, that a
hand movement towards the right was associated to
the visual stimulus placed on the right, and, later in
training, that fixation should be maintained during the
S.1
delay. Monkeys were familiarized with each juice
separately for 1-3 single-juice sessions. The most
specific aspect of the task, namely the trade-off
between juice type and juice quantity, did not require
training, as monkeys showed it spontaneously and
immediately. This was consistent with our previous
observations in capuchins using solid foods in full
sight2. For the current experiments, we wanted to have
relative values away from unity. At the same time, we
wanted relative values to lie within the tested range of
offer types, so that, if the less preferred juice was
offered in sufficiently large amount, the monkey
would choose it. Prior to recordings, we tested
numerous juice types to find suitable pairs.
Juices were delivered through a three-line juice tube.
Each juice line was controlled by a separate solenoid
valve. We routinely calibrated the three juice lines
individually so that the valve-opening times
corresponded to the desired multiple of juice quantum.
During recordings, we used quanta of 80 µl and 65 µl
for monkeys V and L, respectively.
Surgery and recordings
Under general anesthesia and aseptic surgery, we
implanted a head-restraining device and a recording
chamber on the skull of the monkeys, and a scleral eye
coil1. In both monkeys we used a large oval custom-
made chamber (main axes 50x30 mm). The chambers
were centered on stereotaxic coordinates (A30, L0),
with the longer axis parallel to a coronal plane. Thus,
chambers covered the frontal lobes bilaterally.
Following surgery, monkeys were given antibiotics
(cefazolin, 20 mg/kg) and analgesics (buprenorphine,
0.005 mg/kg; flunixin, 1 mg/kg) for 3 days.
Throughout the experiments, we strictly followed the
NIH Guide for the Care and Use of Laboratory
Animals and the guidelines of the Harvard Medical
School Standing Committee on Animals.
For recordings, we used tungsten electrodes (125-µm
diameter, 5±1 MΩ initial impedance, Frederick Haer
& Co.). Electrodes were advanced with custom-built
motorized micro-drives. The micro-drives, consisting
of a titanium housing containing a 0-80 trapped screw,
were anchored to a Teflon grid placed on the chamber.
A brushless miniature precision gear motor (5.8-mm
outer diameter, Micro Precision Systems) drove the
trapped screw and was controlled remotely. The
system allowed up to 0.5 µm of depth resolution.
Typically four (and up to six) electrodes were used
each day. We usually advanced electrodes by pairs
(one motor for two electrodes), with the two electrodes
placed 1 mm apart. Electrical signals from each
channel were amplified and band-pass filtered
(custom-designed miniature amplifiers, low frequency
cutoff 400 Hz, high frequency cutoff 6 kHz) and
recorded at 20 kHz by a dedicated processor (Power
1401, Cambridge Electronic Design). Action potentials
were detected online by threshold crossing and
waveforms were saved to disk for subsequent analysis.
Spike sorting was performed semi-manually using
commercially available software (Spike 2, version 5,
Cambridge Electronic Design). We routinely used
multiple algorithms, including template matching,
clustering on waveform measurements, and principal
component analysis.
Structural MRI scans (1-mm sections) were obtained
for both monkeys before and after placing the
recording chambers. Scans were performed in a 3.0 T
magnet (General Electric) and three-dimensional
reconstruction was performed off-line (Slicer-3D,
www.slicer.org).
Recordings areas
Recordings were performed on the left hemisphere of
monkey V and on the right hemisphere of monkey L.
For their sulcal pattern, both hemispheres were
classified as of type II of Chiavaras and Petrides3.
The position of the chamber allowed us to reach OFC
through nearly straight dorso-ventral penetrations.
Because the chambers were large, we were able to
record from an extended region of OFC. Based on
exploratory recordings, we identified in both monkeys
a region of interest where neuronal responses were
modulated by the offer type in our task. In monkey V,
the region of interest was centered in stereotaxic
coordinates (A32.5, L–9.0) and extended for 6 mm
rostro-caudally and 5 mm medio-laterally. In monkey
L, the region of interest was centered in stereotaxic
coordinates (A33.5, L8.5) and extended for 6 mm
rostro-caudally and 2 mm medio-laterally. Based on
the MRI scans and on the sequence of gray and white
matter encountered during electrode penetrations, we
located the regions of interest in the lateral bank of the
medial orbital sulcus and in the medial part of the
posterior orbital gyrus. Comparing our reconstruction
to the architectonic subdivision of Carmichael and
Price4,5, we tentatively identified our regions of
interest with area 13m.
All the neurons described here were recorded from the
regions of interest. Apart from this criterion and apart
from imposing good and stable electrical isolation,
neurons were not otherwise selected prior to data
S.2
collection or during the analysis. We did not attempt to
identify the cortical layer of recordings. However,
because we reached OFC dorso-ventrally from the
white matter, we presume that the lower layers are
largely represented in our dataset.
Analysis of choice pattern and 3-way ANOVA
Unless otherwise specified, we always assume linear
value functions. Indicating with V(x) the value of x,
the linear assumption means that a multiple q of a unit
volume of juice X has a value V(qX) = q V(X).
Data are analyzed in Matlab (PC version 6.5, Math-
Works). In each session, we fit the choice pattern with
a sigmoid function, which for our data generally
provides an excellent fit (R2>0.95). The flex of the
sigmoid determines the relative value n* such that
V(A) = n* V(B). From this equation, we can compute
the value of any amount of juice A and B, up to a
scaling factor. For example, expressing values in units
of V(B), the value of b drops of juice B is b, and the
value of a drops of juice A is n*a. Because the
analysis of neuronal responses in relation to the juice
values is based on linear regressions (see below), the
results do not depend on the particular units used to
express value.
For quantitative analyses, neuronal firing is averaged
over the following time windows: pre-offer (0.5 s be-
fore the offer, a control time window); post-offer (0.5 s
after the offer); late delay (0.5-1.0 s after the offer);
pre-go (0.5 s before the go); reaction time (RT; from
the go to the saccade); pre-juice (0.5 s before the
juice); and post-juice (0.5 s after the juice).
We first analyze single-trial data from each neuron
with a 3-way ANOVA (factors: [position of juice A] x
[movement direction] x [offer type]), separately in
each time window. We impose a significance level of
p<0.001. We choose a relatively conservative
threshold because of the large number of responses
analyzed (931 neurons x 7 time windows). Because
factors [position of juice A] and [movement direction]
are rarely significant, we collapse data across these
two factors in all subsequent analyses. We also use the
results of the ANOVA as a screening criterion for sub-
sequent analyses, which (unless otherwise indicated)
we restrict to neurons and time windows for which the
factor [offer type] yields a significant effect.
We divide trials into “trial types” based on the offer
type and the choice. For example, a monkey facing the
offer type 3B:1A can choose either 1A or 3B,
corresponding to the two trial types (3B:1A, 1A) and
(3B:1A, 3B). In principle, there are twice as many trial
types as offer types in any given session. However,
many trial types are “empty,” because for most offer
types the choice of the monkey is univocal (for
example, in a given session, a monkey offered 3B:1A
might always choose 3B). In subsequent analyses, we
include only trial types with two or more trials.
Neuronal activity is averaged across trials separately
for each trial type. Hereafter, the term “response”
refers to the average activity of one neuron in one time
window, as a function of the different trial types. Note
that, for sake of clarity, in figures 2d and 3a-e of the
main text we grouped trials by offer type, not by trial
type.
Neuro-econometric analysis: defined variables and
correlation matrix
What variables are encoded in OFC? As illustrated in
the main text (figure 3), many responses seem quali-
tatively related to the variables chosen value, value A
offered, value B offered, and taste. For a quantitative
analysis, however, we want to consider alternative hy-
potheses, namely other variables with which neuronal
responses could a priori correlate. For example, we
want to test the hypothesis that some responses might
encode the other value (i.e., the value of the non-
chosen juice), or the total value, or the value A chosen;
etc. In addition, we want to test variables that might
capture the decision process, such as the value
difference (chosen–other) value and the value ratio
(other/chosen) value. Most importantly, we want to
distinguish between variables related to value and
variables related to physical properties of the
juice/stimulus, such as quantity, volume, number, etc.
These physical properties are all proportional to each
other, and we collectively refer to them with the
variable number.(a) To put value and number on equal
a
In our experiments, value is subjective but operationally-defined.
The neuronal correlates of value and number cannot be dissociated
easily using a single good (e.g., a single juice) because in that case
the two variables are inextricably inter-related (assuming a linear
value function, value and number are in fact proportional to each
other). Introducing uncertainty does not help, because the
confusion remains between expected value and expected number.
However, value and number can be dissociated using two different
juices, insofar as the relative value of the two juices (i.e., the ratio
V(A)/V(B)) differs from unity. For example, in a session in which
1A=3B, a smaller number (say 2A) may have higher value than a
larger number (say 4B). Of course, a correlation between value and
number always remains (after all, a large quantity of any given
juice is always more valuable than a small quantity of that same
juice). For this reason dissociating definitively the neuronal
correlates of value and number ultimately requires the “neuro-
econometric” analysis described below.
S.3
footing in the analysis, we want to test in the number
domain every variable tested in the value domain. In
summary, we want to analyze the 19 variables defined
in figure s1.
These 19 variables are often highly correlated with
each other. To estimate the typical correlation between
any two variables X and Y, we proceed as follows. For
each session, we compute the correlation coefficient:
r r r r
ρ ( X ,Y ) = x ⋅ y x2 ⋅ y2
session
r r
where x and y are vectors of values taken by
variables X and Y for different trial types. So defined,
the correlation coefficient varies between –1 and +1.
Most informative in this context is the absolute value,
which we compute and average across sessions:
ρ ( X ,Y ) = ρ ( X ,Y )
session sessions
Repeating for all pairs of variables, we obtain a
symmetric matrix ρ of elements ρ ( X ,Y ) that vary
between 0 and 1. Hereafter, we refer to ρ as the
“correlation matrix.”
The 19 variables defined in figure s1 cast a wide net.
But ultimately we would like to test whether few
variables can capture the neuronal activity of OFC,
and identify those that best do so. To achieve this goal,
we proceed in two steps. First, we assume that
individual OFC responses encode each only one
variable and that value functions are linear. We
perform independent linear regressions of each
response on each variable, and we apply methods of
variable selection. We thus identify value A offered,
value B offered, chosen value, and A|B chosen as the
variables that best account for the dataset. These
variables explain well the large majority of OFC
responses. Second, having identified for each response
the variable encoded “at the first order,” we relax the
two initial assumptions. We test with standard multi-
regression methods whether adding a second variable
or a quadratic value term improves the regression, and
we observe that for the large majority of responses this
is not the case. In other words, both the “one response-
one variable” assumption and the assumption of linear
value functions are, by and large, warranted. We
conclude that indeed OFC responses encode the
variables identified by the variable-selection
procedures.
The next sections and the Supplementary Results
detail the methods and results of these analyses.
Neuro-econometric analysis: variable selection
The variable-selection analysis is based on the
assumption that individual responses encode each (at
most) one variable. For this analysis, we regress each
response separately on each of the 19 variables. Each
regression provides a slope and the corresponding R2.
We consider a variable capable of “explaining” a
given response if the regression slope is significantly
different from zero (p<0.05), and we conventionally
set R2 = 0 for variables that do not explain the respon-
se. In general, any given response can be explained by
multiple variables. Largely, this is because variables
are often highly correlated with each other (for
example, this is the case for variables chosen value,
total value and chosen number), a situation referred to
as multi-collinearity6,7.
In the presence of multi-collinearity, it is often
possible to identify a small subset of variables that
account for much of the data. However, the problem of
identifying the appropriate subset is in principle not
trivial. For our analysis we adapt two methods of
variable selection routinely used in the case of multi-
linear regressions, namely the “stepwise selection”
method and the “best-subset” method6,7. Notably, our
situation differs from that typically found in multi-
linear regression, where any single response depends
on multiple variables. Each of our responses is
sampled in few data points (typically 8-10 trial types),
many fewer than the number of variables we want to
test (19 variables). However, we can capitalize on the
large number of responses available (a total of 1379
responses pass the ANOVA screening criterion) and
identify the most relevant variables through a
population analysis (see below).
In addition to the 19 original variables, we define two
“collapsed” variables value A|B offered and value A|B
chosen, as follows. The variable value A|B offered is
taken to explain a given response if at least one of the
two variables value A offered and value B offered
explains the response. The R2 of the collapsed variable
is equal to the higher R2 obtained from of the two
original variables (and equal to zero if none of the
original variables explains the response). Similarly, the
variable value A|B chosen is defined by collapsing
variables value A chosen and value B chosen.
Neuro-econometric analysis: second order encoding
To explore the possibility that individual OFC
responses might encode a mixture of variables, we
proceed formally as follows8.
S.4
Consider a response encoding at the first order the
variable X with R 2 = R X2 (i.e., a response explained
by X better than by any other selected variable,
with RX2 > 0 ). To establish whether adding a second
variable Y to the regression provides a significantly
better account, we compute
FY | X = (n − 3) ∗ ( R XY
2
− R X2 ) (1 − R XY
2
)
In the equation, R X2 is obtained from the linear regres-
sion on X only; R XY 2
is obtained from the bi-linear
regression on X and Y; and n is the number of trial
types (data points in the regression). We compute
FY | X for each of the variables Y we wish to test as
potentially encoded at the second order, and we take
{ }
the maximum F = max FY | X . The degrees of
freedom of F are 1 for the numerator and n − 3 for the
denominator. We then set a threshold F* corresponding
to a desired p* = 0.01 (we choose this threshold be-
cause residuals are not pre-screened and because we
test each response with multiple potential second order
variables). If F passes the criterion, we identify the
second order variable encoded by the response. If F
does not pass the criterion, we conclude that the
response does not encode any second order variable (at
least among those tested).
For second order encoding, we test some of the
variables tested for first order encoding. In addition,
we test value-encoding responses with quadratic value
terms. Hence, with this analysis we scrutinize the
validity of the two assumptions underlying the variable
selection procedures, namely that individual OFC
responses encode each only one variable and that
value functions are linear.
Analysis of time course
To analyze the time course by which neurons in OFC
encode variables value A offered, value B offered,
chosen value, and A|B chosen, we define 50-ms non-
overlapping time bins, aligning trials separately at the
time of the offer and at the time of juice delivery. For
each 50-ms time bin, we subject each cell in the
population of Ntot = 931 neurons to independent linear
regressions on the four selected variables. We assign a
neuron to a variable if the regression slope differs
significantly from zero in that time bin. If more than
one variable has a non-zero slope (a rare case), we
assign the neuron to the variable with highest R2. For
this analysis, we do not screen neuronal data prior to
regression, because small time bins result in spike
counts mostly equal to 0 or 1. For this reason, we
choose a slightly more conservative threshold (p<0.01)
to identify regression slopes as significantly non-zero.
In each time bin, the number of neurons expected to be
assigned by chance to each class is Nchance = p* Ntot =
9.31. For each time bin, we then compute:
Nvalue A|B offered = Nvalue A offered + N value B offered – Nchance
Supplementary Results
Neuronal database, linear regressions, and
correlation matrix
We recorded the activity of 931 neurons (375 from
monkey V; 556 from monkey L). Figure s2 reports the
number of cells for which each main factor in the
ANOVA has a significant effect. In particular, the
factor [offer type] yields a significant effect in at least
one time window for 505 (54%) neurons in total (211
= 56% from monkey V; 294 = 53% from monkey L).
We refer to these as “task-related” cells. Figure s3
shows the average neuronal activity separately for
task-related cells and for “other” (i.e., not task-related)
cells. It can be observed that the average activity
profile of task-related cells peaks early after the offer,
slowly decays during the delay, and has two secondary
peaks before and after juice delivery. In contrast, the
average activity profile of other cells is lower and
essentially flat throughout the trial.
A total of 1379 responses are significantly modulated
by [offer value] in the ANOVA (656 from monkey V;
723 from monkey L). We further analyze these respon-
ses in relation to the variables defined in figure s1. In
total, 1227/1379 (89%) responses are explained by at
least one of the 19 variables (567/656 = 86% from
monkey V; 660/723 = 91% from monkey L). Methods
for variable selection are applied to this dataset.
Figure s4 illustrates the results of individual linear
regressions obtained for one particular response (i.e.,
the activity of one neuron in one time window as a
function of the different trial types). The two top left
panels show the behavioral choice pattern and the
neuronal response plotted with respect to offer type. In
the other panels, the same neuronal response is plotted
against each of the 19 variables. A blue regression line
indicates that the slope is significantly non-zero
(p<0.05); if so, the respective R2 is indicated in the top
left corner of the panel. In the top left panels, the
activity of the cell as a function of the offer type has a
U-shaped profile, qualitatively similar to that
S.5
hypothesized for a response encoding the chosen
value. In fact, the variable chosen value explains the
response (R2 = 0.90). However, the response is also
explained by several other variables, for example total
value (R2 = 0.72), chosen number (R2= 0.50), etc.
The fact that multiple variables can explain the same
response is not surprising, since the variables are in
general not independent. This point is illustrated in
figure s5, which shows the correlation matrix. In the
figure, different shades of gray represent different
values of correlation ( ρ ) with black corresponding to
ρ = 1 and white corresponding to ρ = 0 . A small
white circle or cross indicates ρ > 0.8 . For example,
the variables chosen value and total value are highly
correlated with each other and with other variables
(e.g., total number, max number, and chosen
number)—a case of multi-collinearity.
Ideally, to provide a concise description of the dataset,
we would like to achieve two goals: first, to assign
each response to only one variable; second, to identify
a small subset of variables that can explain as many
responses as possible. Clearly, achieving the second
goal helps to achieve the first. But how can we select
the “right” subset of variables? In the following
sections, we first make a qualitative case for variable
selection and then we illustrate the results obtained
with statistically principled procedures.
Variable selection: Qualitative analysis
Consider the response analyzed in figure s4. Our initial
hypothesis that U-shaped responses might reflect the
chosen value is bolstered by the fact that the variable
chosen value provides the best fit (i.e., the highest R2).
However, at this level of analysis, it would be
unreasonable to rule out other variables (for example
total value) that provide a slightly lower R2. A
potentially powerful approach is to consider the entire
population of responses. For example, we might
discover that whenever (or most often when) both
variables chosen value and total value provide a non-
zero slope, the variable chosen value provides a
slightly higher R2. If that were the case, we could
conclude that neuronal responses genuinely encode the
chosen value, and that the variable total value has no
additional explanatory power. This argument is
obviously complicated by the fact that, as we observed
qualitatively (figure 3, main text), not all responses in
our dataset encode the same variable. In addition,
variables have more than just pairwise correlation and
cannot be simply considered two at the time. So we
must analyze the entire population of responses and all
the variables at once.
Figure s6 illustrates two complementary ways to
derive a population analysis from the individual linear
regressions. In figure s6a, we compute for each time
window and for each variable the number of responses
explained. For example (top left), in the post-offer
time window, the variable total value explains 130
responses. Because, as we noted, more than one
variable may explain any given response, any given
response may appear in multiple bins in this plot.
Some trends emerge clearly. First, as indicated by the
ANOVA, more responses are modulated by the trial
type in early time windows (post-offer and late delay)
and late time windows (pre-juice and post-juice), as
compared to the peri-movement time windows (pre-go
and RT). Second, focusing on the post-offer time
window we observe that a group of variables explain a
large number of responses. For example, the variables
total value, chosen value, total number, max number,
chosen number, (max–min) number, and value B
offered all explain more than 100 responses.
Inspection of figure s5 reveals that these variables are
highly inter-correlated. So the picture in figure s6a
clearly contains a high degree of redundancy, which
we may hope to resolve with further analysis. Third, in
addition to these variables, which are prevalent in the
post-offer, pre-juice and post-juice time windows,
there is another group of variables, including (chosen–
other) number, A|B chosen, and value B chosen, which
are common in the pre-juice and post-juice time
windows, but not in earlier time windows. Again,
inspection of figure s5 reveals that these variables are
highly inter-correlated, a redundancy that we may
hope to resolve with further analysis.
A complementary way to derive a population analysis
from the individual regressions is to consider for each
response only the variable providing the best fit (i.e.,
the highest R2). By doing so, we essentially force each
response in only one bin. Figure s6b shows the result
of this analysis. Note that much of the redundancy of
figure s6a seems naturally resolved. For example,
many more responses are best fit by the variable
chosen value than by any of the variables total value,
total number, etc. that are highly correlated with
chosen value. Likewise, many more responses are best
fit by the variable A|B chosen than by either (chosen–
other) number, or value B chosen. Thus, although the
population picture obviously remains complex, figure
s6b suggests that fewer variables than the 19 initially
considered may be sufficient to account for most
responses.
S.6
In some sense, the pictures presented in figures s5a
and s5b represent two opposite and extreme ways to
analyze the population of responses. On the one hand,
in figure s6a we forgo the possibility of ranking the
quality of the linear fits (i.e., the R2) and, by doing so,
we give up the possibility of a concise description. In
fact, it could be argued that we spuriously added
redundancy by including “unreasonable” variables in
the analysis. And, indeed, one could easily come up
with more variables that would further complicate the
picture, to little or no advantage. On the other hand, in
figure s6b we ignore that a given variable may
sometimes provide a satisfactory explanation for a
given response, even if another variable provides a
better fit. That is clearly excessive: because neuronal
data are noisy, responses that genuinely encode a
given variable (say chosen value) might at times be
best fit by another, highly correlated variable (say total
value). In other words, it is likely that some responses
that would genuinely belong to one bin in figure s6b
“spilled over” to another bin because of noise. So here
too it could be argued that we spuriously added
complexity to the picture by including “unreasonable”
variables in the analysis. And again, one could
certainly further complicate the picture by adding
more variables, to little or no advantage.
In conclusion, an ideal account of the dataset would on
the one hand make use of the information of figure
s6b, namely that some variables (say chosen value) do
consistently better than others (say total value) in
fitting the population of neuronal responses. On the
other hand, an ideal account would also use the
information embedded in figure s6a, namely that one
variable might provide a satisfactory (albeit
suboptimal) explanation for many responses. But how
can we identify a small subset of variables to explain a
large number of responses? This problem closely
resembles that of variable selection encountered in
multi-linear regressions in the presence of multi-
collinearity. Methods commonly used for multi-linear
regressions6,7 can be adapted to our case.
Variable selection: Stepwise method
One simple way to handle our problem is to select
variables one at the time. First, we select the variable
that provides the highest number of best fits (the
darkest bin in figure s6b). In our case, this variable is
A|B chosen. We explore the entire dataset and we
remove all the responses that can be explained by this
variable. Then we iterate the procedure by selecting a
second variable, then a third variable, etc. We continue
the procedure as long as any newly selected variable
explains at least a certain percentage (for example 5%)
of otherwise unexplained responses. At the end of the
procedure, we classify responses based on the R2 (i.e.,
we assign responses explained by more than one
selected variable to the variable with highest R2). This
method is called “stepwise selection.”6,7 (b)
We apply the stepwise selection method to our dataset.
In figure s7, panels on the left represent the population
of unexplained responses at different iterations of the
procedure. Tables of best fit are shown, so that the top
panel is the same as shown in figure s6b. For each
iteration, a red asterisk indicates the selected variable;
blue dots indicate variables excluded by the 5%
selection criterion. The first four iterations select
variables A|B chosen, chosen value, value A offered,
and value B offered, and no other variable reaches the
5% selection criterion in subsequent iterations.
In the analysis shown in figure s7, we keep separate
the 19 variables defined in figure s1. However, it could
be argued that in some cases different variables really
represent the same class of response. For example, the
distinction between value A offered and value B
offered is somewhat arbitrary, because one particular
juice may be labeled “A” in a given session and “B” in
another session. Furthermore, if a selection procedure
identified only one of the two variables, the result
would be difficult to interpret. Thus it is probably
more correct to combine the variables value A offered
and value B offered using the “collapsed” variable
value A|B offered (see Supplementary Methods).
Likewise, we can use the other collapsed variables
value A|B chosen. This leaves us with 17 variables.
We re-analyze our dataset using collapsed variables.
The stepwise procedure selects the three variables
value A|B offered, chosen value, and A|B chosen,
confirming our previous result.
In summary, the stepwise method applied to our
dataset selects three variables: chosen value, value A|B
offered, and A|B chosen. These three variables explain
b
The stepwise selection method is used with the following caveat.
The “marginal explanatory power” of one particular selected
variable X is the number of responses that are explained by X and
that are not explained by any other selected variable. The 5%
selection criterion sets a threshold on the marginal explanatory
power of newly selected variables. However, the marginal
explanatory power of a selected variable X generally drops over
iterations, because variables selected after X may explain some of
the responses previously explained only by X. Eventually, it is
possible that a selected variable X only explains a small percentage
(less than 5%) of otherwise unexplained responses. For a correct
procedure, we check at each iteration that all of the selected
variables actually meet the 5% selection criterion, and we exclude
previously selected variables that fail to meet the criterion.
S.7
1085 responses, corresponding to 79% of all responses
significantly modulated by offer type (1379), and to
88% of responses explained by all 17 variables (1227).
In other words, in limiting ourselves to three variables,
we “lost” only 12% of responses. Having thus
identified the variables encoded in OFC, we classify
responses based on the R2 (i.e., we assign responses
explained by multiple selected variables to the variable
with highest R2). Figure s8 (top) summarizes the final
result of this classification. Note that, to provide a
more intuitive interpretation of the variables, in the
main text we renamed variable value A|B offered as
offer value and variable A|B chosen as taste. The
histograms in figure s8 (bottom) illustrate the
distribution of R2 obtained as a result of the final
classification. In general, selected variables capture the
variability of individual responses remarkably well
(mean R2 = 0.63).
In the following sections, we show that alternative
procedures for variable selection yield essentially the
same results.
Variable selection: Best-subset method
One potential limitation of the stepwise method for
variable selection is that the results obtained may be
“path-dependent,” and the method is not guaranteed to
select the best possible subset of variables. This goal
can be achieved using the “best-subset” method6,7.
The idea of the best-subset method is to compute for
each possible subset of d variables the corresponding
number of responses explained; to identify the best
subset of d variables as the subset that explains the
maximum number of responses; and to repeat the
procedure for d = 1, 2, 3, … If n(d) is the number of
responses explained as a function of d, the number d*
of variables necessary to characterize the population
can be determined either by an “elbow” in the function
n(d), or with a threshold criterion (e.g., imposing that
at least 85% of responses be explained).
Figure s9 shows the results of analyzing our dataset
with the best-subset method. In the top panel, the x-
axis represents the number d of variables included in
each subset, and the y-axis represents the percentage
n(d) of responses explained by the best subset. The
bottom table indicates the variables included in the
best subset for various d. Several points should be
noted. First, although n(d) does not present a clear
“elbow,” selecting only three variables seems
reasonable, as additional variables add limited
explanatory power. Second, the three variables
identified by the stepwise selection method, namely
chosen value, value A|B offered, and A|B chosen, are
indeed the best possible subset of three variables.
Third, these three variables are also included in the
best subset of four variables and in the best subset of
five variables. This is not necessarily expected,
because in general the best subset of d+1 variables
might not include all or any of the variables included
the best subset of d variables. The fact that the three
variables that make up the best subset for d = 3 are
also included in the best subset for d = 4 and for d = 5
is a sign of robustness of the result.
In conclusion, the variable-selection analysis using the
best-subset method confirms the results obtained with
the stepwise selection method. In both cases, the three
selected variables are chosen value, value A|B offered,
and A|B chosen.(c, d)
Variable selection: Over-fitting and post hoc analysis
Although it guarantees optimality, the best-subset
procedure does not provide a measure of reliability of
the result. For example, in our case, the method indi-
cates that, among the subsets of three variables, chosen
value, value A|B offered, and A|B chosen explain the
highest number of responses, but we do not know how
well these three variables do compared to the possible
alternatives. In other words, we might have a problem
of over-fitting. In multi-linear regressions, the standard
way to handle this problem is to repeat measures and
to analyze multiple datasets6,7 (see below). But in
addition to repeating measures, our case also lends
itself to an informative post hoc analysis.
In the post hoc analysis, we want to test selected
variables against highly correlated but discarded
variables. For example, the best-subset method selects
the variable chosen value. We want to establish
whether selecting this variable is significantly better
than selecting the variable total value, which is highly
c
The stepwise selection method and the best-subset method differ
in how different time windows are analyzed. Using the stepwise
method, responses from different time windows are analyzed
separately but in parallel. Variables are selected for the number of
best fits by time window, but variables can then explain responses
in any time window. In contrast, using the best-subset method,
responses from different time windows are pooled together.
Keeping time windows separate may or may not be viewed as a
desirable feature. In any case, it is reassuring that the two methods
provide converging results.
d
One important advantage of the best-subset method is that results
do not depend on “irrelevant” variables. That is to say, including in
the analysis variables that turn out to have little explanatory power
has no effect on the outcome. This ensures that we don’t “add
noise” to the procedure by testing “unreasonable” variables.
S.8
correlated with chosen value and is therefore a
presumably “challenging” alternative. That chosen
value, and not total value, is part of the best subset
ultimately means that chosen value has a higher
marginal explanatory power: disregarding responses
explained by either value A|B offered or A|B chosen,
the number of responses explained by chosen value
and not by total value is greater than the number of
responses explained by total value and not by chosen
value. To establish whether this inequality is true in a
statistical sense, we can formally proceed as follows.
Consider the sub-population of responses explained by
chosen value (variable X) and/or by total value
(variable Y), and not explained by any other variable
in the best subset (i.e., value A|B offered and A|B
chosen). Of this sub-population, a number nX of
responses are explained by X and not by Y; a number
nY of responses are explained by Y and not by X; and
the other responses are explained by both X and Y.
The best-subset method indicates that nX>nY, but it is
possible that nX is actually quite close to nY. The
problem of determining whether the inequality nX>nY
is statistically significant is equivalent to asking how
unlikely it is to draw nX heads out of nX+nY tossings
of a fair coin, and can be addressed with a simple
binomial test. For this particular case, we have nX =
58 and nY = 21, from which we infer that the marginal
explanatory power of chosen value is significantly
higher than that of total value (p<10-5).
We repeat this procedure for all the pairs of variables
(X,Y) for which X is a variable in the best subset and
Y is a variable highly correlated to X (i.e.,
ρ ( X ,Y ) > 0.8 ; see figure s5). Additionally, we test
chosen value against chosen number. The results are
presented in figure s10. In these pairwise comparisons,
all the variables included in the best subset do
significantly better than the challenging alternatives
(maximal p<10-5). The only exception is that the
variable A|B chosen fails this test against the variable
value A|B chosen (p = 0.12). As discussed below, this
degree of ambiguity remains largely unresolved in our
analysis. For all other aspects, the post hoc analysis
confirms our previous conclusions.
Variable selection: Other procedures and conclusions
We employed a number of alternative procedures to
select variables. Essentially, they all confirm the
results described in previous sections. For example, in
the spirit of repeating measures, we find that data from
the two monkeys analyzed separately yield statistically
indistinguishable results. Another variant consists in
weighting the contribution of each response to the
explanatory power of any variable with the
corresponding R2. The variable selection analyses
using this alternative metrics yields results identical to
the ones obtained with the binary “explained/not-
explained” metrics. Finally, correcting regressions for
unequal variance also provides identical results.(e)
Using collapsed variables, all these procedures
consistently indicate that the best subset includes value
A|B offered and chosen value, and either A|B chosen or
value A|B chosen, with these two latter variables
having statistically indistinguishable marginal
explanatory power. Using non-collapsed variables, all
the procedures provide statistically consistent results.
However, using non-collapsed variables, we find that
the marginal explanatory power of A|B chosen is
significantly higher than that of either value A chosen
or value B chosen.
Although the issue between A|B chosen and value A|B
chosen remains largely unresolved, two arguments
seem to favor the former variable. First, in the analysis
of non-collapsed variables, A|B chosen does signi-
ficantly better than either value A chosen or value B
chosen taken alone Second (and partly related), A|B
chosen, with only one intrinsic degree of freedom (i.e.,
the relative value of the two juices), is more
parsimonious than value A|B chosen, which has two
intrinsic degrees of freedom (i.e., the relative value,
and which one of the two juices is coded for). In
summary, although these arguments are largely
heuristic, it seems preferable to summarize our data in
terms of the variable A|B chosen.(f)
e
The least-squares method, used here for all regressions, is based
on two assumptions: gaussianity (i.e., data must come from
gaussian distributions) and homoscedasticity (i.e., the gaussians
must all have the same variance). Cortical single-trial spike counts
are known to violate both these assumptions [ref 9]. Responses
analyzed here (i.e., averages across trials) approximately satisfy
gaussianity (for the central limit theorem). However, responses
generally do not satisfy homoscedasticity. In cases of unequal
variance (heteroscedasticity), the least squares method can be
corrected by weighting the residual associated to each data point
mi with a term proportional to 1/σi, where σi is the standard
deviation of the distribution from which mi is drawn [ref 8]. In our
case, mi is the response measured for a particular trial type and we
can estimate σi with the standard error. Notably, 1/σi is
proportional to the square root of the number of trials. This
suggests that, in our case, it might be preferable not to correct for
unequal variance, because trial types close to the indifference
point, which are in many respects the most informative, are also
those for which we have fewer trials (because monkeys “divide”
their choices between the two juice types). For this reason, we
generally use uncorrected linear regressions in all our analyses.
f
Consistent with the results of a recent study of gustatory
responses in this brain area (Pritchard et al., 2005), in our final
classification, we observe A|B chosen (i.e., taste) responses in
S.9
In conclusion, having tested a large number of hypo-
theses with a variety of variable selection procedures,
we identified value A|B offered, chosen value, and A|B
chosen as the three variables encoded by OFC respon-
ses. These three variables describe our dataset well,
and significantly better than challenging alternatives.
Second order encoding
The variable selection analysis presented in previous
sections rests on two key assumptions: the assumption
that OFC responses encode each only one variable and
the assumption that value functions are linear. We now
put these two working hypotheses under scrutiny.
For second order encoding, we test variables selected
at the first order (that is, we test whether responses
encode mixtures of selected variables). We also test
the two variables value A chosen and value B chosen
that were excluded last in the variable selection
analysis, as well as the variable chosen number.
Finally, for responses encoding at the first order value
A offered, value B offered, and chosen value, we test
the corresponding quadratic term (value A offered)2,
(value B offered)2, and (chosen value)2. We extend this
analysis to the population of 1085 responses classified
at the first order (figure s8) and we pool together
responses from different time windows. The results are
presented in figure s11.
In the two tables, rows indicate first order variables
and columns indicate second order variables. The
rightmost column indicates responses that do not
encode any second-order variable. The top table
(figure s11a) reports the number of responses. The
bottom table (figure s11b) reports the same data as
percentages (normalized by the number of responses
explained by each first-order variable). In general, we
observe that the vast majority of responses do not
encode second order variables, independently of the
variable encoded at the first order. Over the entire
population, 837/1085 (77%) responses do not encode
second order variables. If we exclude from this
analysis the quadratic value terms, we find that
890/1085 (82%) responses do not encode second order
variables. We interpret this result as a substantial
justification for the one response-one variable
assumption. Similarly, for most value-encoding
responses, adding a quadratic term does not improve
the regression significantly. Repeating the analysis
with only quadratic terms as potential second order
168/931 (18%) of recorded neurons, most prevalently at the time
of juice delivery (i.e., in pre-juice and post-juice time windows).
variables, we find that in 757/817 (93%) cases qua-
dratic terms fail to provide a statistically appreciable
gain. Again, we interpret this result as a justification
for the assumption of linear value functions.(g)
In summary, the analysis of second order encoding
provides post facto support for the two main
assumptions underlying the variable selection analysis,
namely the one response-one variable assumption and
the assumption of linear value functions. This result
concludes our “neuro-econometric” analysis.
Relationship between slope ratio and relative value
During the experiments, we used a large number of
juice pairs. In each session we measured the relative
value n* from the behavioral pattern of choice,
through a sigmoid fit. Naturally, we measured
different relative values for different juice pairs. For
example, monkeys generally had a mild preference for
grape juice over fruit punch (typically, n*<2), but a
strong preference for grape juice over peppermint tea
(typically, n*≥3). In addition, the relative value of any
given juice pair could vary from session to session and
from day to day. For example, the relative value of ½
apple juice over peppermint tea varied over many days
in the range n*∈(1.5,3). If U-shaped responses indeed
encode the chosen value, we should expect them to
reflect this variability. The following analysis confirms
this prediction.
To avoid any bias, we want to identify U-shaped
responses independently of the specific variable that
they might encode. We proceed using a bi-linear
regression:
fr = a 0 + a A ( # A) + a B ( # B) (1)
where fr is the firing rate of the neuron, and (#A) and
(#B) are, respectively, the number of drops of juice A
and juice B chosen by the monkey. (Note that in any
given trial either (#A) = 0 or (#B) = 0.) We then define
a response to be “U-shaped” if the regression slopes aA
and aB are both significantly different from zero
(p<0.01). (For this analysis, we do not pre-screen
responses with the ANOVA.)
g
Of course, the presence and nature of quadratic value terms is a
consequential issue in economic theory, where convexity of the
value function plays a fundamental role. The present results
indicate that the assumption of linearity is adequate for our dataset
and at the relatively coarse level of our analysis. However, the
present results do not exclude the possibility that a consistent
departure from linearity may emerge upon more refined
examination, or in experiments that span a wider range of values.
S.10
As illustrated in figure s12 for one particular response,
the regression slopes aA and aB are in general different
from each other. Since both aA ≠ 0 and aB ≠ 0, Eq.1
can be re-written as follows:
fr = a 0 + m (k * ( # A) + ( # B) ) (2)
with m = aB and k* = aA/aB. The hypothesis that U-
shaped responses encode the chosen value leads to a
simple prediction regarding the slope ratio k*, which
should be (statistically) equal to the relative value of
the two juices. Critically, k* should co-vary with n*.
At least for the response in figure s12, the relationship
k* ≈ n* holds true. Indeed, from the bi-linear
regression we obtain k* = 3.0 (±1.4) (±s.d.), while
from the behavioral choice pattern we obtain n* = 3.2.
Applying the bi-linear regression criterion to all 931
neurons, we identify 254 U-shaped responses. Figure
s13 illustrates the relationship between k* and n*
observed for this population. The scatter plot in figure
s13a includes data from all sessions, pooling together
all pairs of juices. In the plot, the x-axis represents the
behaviorally measured relative value n*, the y-axis
represents the neuronal slope ratio k*, and both axes
are plotted in log scale. Each dot represents one
response. Defining “A” as the preferred juice is
equivalent to imposing n* > 1. In principle, the slope
ratio k* could assume any possible value. However,
nearly always k* > 1. Moreover, we observe a signifi-
cant correlation between k* and n* (p<10-12).
Most importantly, the relationship k* ≈ n* can be
observed when the analysis is restricted to responses
recorded with individual pairs of juices, as shown in
figure s13b for one particular pair of juices (A = ½
apple, B = peppermint tea). During the experiments,
We used a total of 25 different juice pairs. The number
of U-shaped responses recorded with any given juice
pair varied between 1 and 40. Restricting our analysis
to the 7 pairs of juices for which we have >10 respon-
ses, we test whether the relationship k*≈ n* holds true
using the regression function:
k * = b 0 + b1 n * (3)
For the pair of juices shown in figure s13b, the
regression indicates b0 = 0.08 and b1 = 1.18, with 95%
confidence intervals b0∈(-0.21, 0.37) and b1∈(0.77,
1.59). Congruent results are obtained for all 7 pairs of
juices. Figure s13c illustrates in particular for different
juice pairs (y-axis), the value of the regression slope b1
(x-axis), together with the 95% confidence intervals.
Notably, the values of b1 are distributed around the red
dashed lined corresponding to b1 = 1. Averaging across
the 7 juice pairs, we obtain <b1> = 1.05 ± 0.15 (mean
± s.e.m.). With respect to the intercept, averaging
across juice pairs we obtain <b0> = -0.13 ± 0.15. These
results are consistent with the predicted identity k* =
n*.
In summary, the neuronal “U” shapes recorded in OFC
closely match the behavioral choice pattern on a juice-
by-juice and session-by-session basis, a phenomenon
naturally captured by the concept of economic value.
Ingredient-based hypothesis
One concern is whether U-shaped responses, which
vary with the quantity of both juices A and B chosen
by the monkey, may simply encode the quantity of one
particular ingredient, or combination of ingredients,
present in both juices. By “ingredient,” we mean any
compound contained in the juice, for example a
compound that would elicit a taste response (e.g.,
water, sugar, citric acid, etc.)10-13. The strongest
argument against the ingredient-based hypothesis
follows from the relationship k* ≈ n* found in the
previous section. To appreciate this point, let us refer
to the cartoon shown in figure s14.
To summarize the results of the previous section, we
showed that for any two juices A and B, in the scatter
plot of k* versus n*, data lie on a diagonal line (figure
s14, gray line) and cannot be described by a horizontal
line. We can now examine different variants of the
ingredient-based hypothesis.
The first variant is the hypothesis that U-shaped
responses all encode the quantity of one particular
ingredient. If this is true, given the ingredient and two
juices A and B, the two regression slopes aA and aB
should be proportional to the concentrations ρA and ρB
at which the ingredient is present in the juices.
Therefore, the slope ratio k* should be equal to the
concentration ratio ρA/ρB, independently of the relative
value n* recorded in the session. Consider for example
the ingredient water. Because for any amount of juice
the quantity of water is equal to the juice volume, if U-
shaped responses all encode the quantity of water,
neuronal data should lie on the horizontal line k* ≈ 1
(figure s14, blue line), contrary to what we observe. In
other words, the sole fact that neuronal data
overwhelmingly lie in the quadrant {k*>1, n*>1} rules
out the possibility that U-shaped responses all encode
the quantity of water consumed by the monkey. (This
is an alternative way to rule out the variable chosen
number.) Consider now another ingredient, for
example sugar, which we may assume to be present in
given juices A and B at a certain concentration ratio
S.11
ρA/ρB. If U-shaped responses all encode the quantity of
sugar chosen by the monkey, we should observe
neuronal data lying on the horizontal line k* ≈ ρA/ρB,
independently of the relative value n* recorded in the
session. However, our observations contradict this
prediction. More generally, no matter what ingredient
one may want to test, the same argument can always
be made. If OFC responses all encode that particular
ingredient, given two juices A and B, neuronal data
should always lie on a horizontal line corresponding to
the concentration ratio, contrary to our observations. In
conclusion, the simple hypothesis that U-shaped
responses all encode one single ingredient can be
rejected.
A second variant of the ingredient-based hypothesis is
that U-shaped responses all encode a particular linear
combination of multiple ingredients, such as water,
sugar, etc. However, any linear combination of
horizontal lines in figure s14 is itself a horizontal line,
such as the brown dashed line in the figure. Therefore,
if U-shaped responses all encode a linear combination
of ingredients, neuronal data should lie on some
horizontal line, contrary to our observations. Hence,
this variant of the ingredient-based hypothesis can be
also rejected.
A third variant is the hypothesis that different U-
shaped responses encode different ingredients. For
example, the linear relationship between k* and n*
could be perhaps explained if U-shaped responses with
large k* encode, say, sugar (red line), while U-shaped
responses with small k* encode another ingredient, say
ingredient X (green line). However, this hypothesis
can only be true assuming that we happened to record
from sugar-encoding neurons on days in which n* was
large, and that we happened to record from X-
encoding neurons on days in which n* was small—a
seemingly impossible coincidence. Thus this variant
can also be rejected.
The fourth and last variant of the ingredient-based
hypothesis is that different U-shaped responses encode
different linear combinations of ingredients. This
variant combines the second and third variants, and
can be rejected for the same reasons.
Studies of gustatory responses in various areas of the
orbitofrontal cortex, particularly those of Rolls and
colleagues, found neurons encoding the taste of one
particular juice, whose responses decreased following
selective satiation of that juice12,14. We emphasize that
the relationship between k* and n* observed in our
data cannot be accounted for by the ingredient-based
hypothesis taken in combination with the phenomenon
described by Rolls. To appreciate this point, consider
again the response shown in figure s12. According to
the ingredient-based hypothesis, the response is U-
shaped because the activity of this neuron encodes the
quantity of some ingredient X present in both juices A
and B. (Here we refer to the first variant of the
hypothesis, but the argument is valid more generally.)
In the context of the ingredient-based hypothesis, the
phenomenon described by Rolls can be described in
terms of reduced responsiveness, or neuronal de-
sensitization. For example, under de-sensitized
conditions, the response of the neuron in figure s12 to
any given quantity of ingredient X may be reduced by
half. Imagine now to record from this same neuron
under conditions of reduced responsiveness. Since the
activity of the neuron encodes the quantity of X
consumed by the monkey, and since X is present in
both juices, reduced responsiveness will affect both
trials in which the monkey chooses juice A and trials
in which the monkey chooses juice B. For example, if
the neuronal responsiveness is halved, we will observe
half-sized responses when the monkey chooses A, and
half-sized responses when the monkey chooses B. As a
consequence, in figure s12, both slopes aA and aB will
be halved. But if both slopes are halved, their ratio
does not change. More generally, if both slopes are
scaled by the same factor, their ratio does not change.
As quantified in Eq.2, this means that any change in
responsiveness to the encoded ingredient affects the
scaling parameter m, which multiplies both (#A) and
(#B), but does not affect the ratio parameter k*. In
summary, even assuming changes in responsiveness of
the kind described by Rolls, the ingredient-based
hypothesis predicts that k* should be constant and
independent of n*, contrary to our observations. Thus
the ingredient-based hypothesis cannot be “salvaged”
by appealing to changes in neuronal responsiveness a
la Rolls.
To conclude, our analysis demonstrates that U-shaped
responses do not encode the quantity of any particular
ingredient or combination of ingredients, but rather the
value the monkey assigns to the juice it chooses to
consume.
Conclusions
We showed that OFC responses do not depend on the
visuomotor contingencies of the task. Assuming that
OFC responses encode each only one variable and that
value functions are linear, we showed that OFC
responses are best described as encoding value A|B
offered, chosen value, and A|B chosen. In the main
text, we refer to these variables respectively as offer
S.12
value, chosen value, and taste. The explanatory power
of these three variables is high (mean R2 = 0.63) and
significantly higher than that of challenging
alternatives. Conversely, we showed that in the large
majority of cases, it is sufficient to assume that OFC
responses encode each only one variable, if that
variable is one of the three characteristic of this area.
Indeed, taking into consideration the possibility that
OFC responses encode a mixture of variables
generally does not provide a significantly better
account. We also showed that for the large majority of
OFC responses recorded in our experiment it is
adequate to assume a linear value function. In a
separate analysis of U-shaped neuronal responses we
showed a statistical identity between the slope ratio
and the behaviorally measured relative value. For any
given pair of juices, the two quantities co-vary on a
session-by-session basis, which rules out any
ingredient-based hypothesis. Finally, we showed that
the timing by which OFC neurons encode offer value,
chosen value and taste corresponds well to the mental
operations that monkeys presumably undertake during
economic choice. Specifically, neurons encoding the
offer value—an operation necessary to make a
choice—are the most prevalent shortly after the offer
is presented to the monkey. Neurons encoding the
chosen value are frequently observed during the delay,
when the monkey has presumably internally made a
choice, but before the choice is actually revealed.
Finally, neurons encoding the taste of the chosen juice
are most prevalent immediately before and after juice
delivery. Taken together, these results suggest that
neurons in OFC provide a substrate for value
assignment during economic choice.
Supplementary References
1. Judge, S. J., Richmond, B. J. & Chu, F. C.
Implantation of magnetic search coils for
measurement of eye position: an improved method.
Vision Res 20, 535-8 (1980).
2. Padoa-Schioppa, C., Jandolo, L. & Visalberghi, E.
Multi-stage mental process for economic choice in
capuchins. Cognition 99, B1-B13 (2006).
3. Chiavaras, M. M. & Petrides, M. Orbitofrontal
sulci of the human and macaque monkey brain. J
Comp Neurol 422, 35-54 (2000).
4. Carmichael, S. T. & Price, J. L. Architectonic
subdivision of the orbital and medial prefrontal
cortex in the macaque monkey. J Comp Neurol
346, 366-402 (1994).
5. Carmichael, S. T. & Price, J. L. Connectional
networks within the orbital and medial prefrontal
cortex of macaque monkeys. J Comp Neurol 371,
179-207 (1996).
6. Dunn, O. J. & Clark, V. Applied statistics: analysis
of variance and regression (Wiley, New York,
1987).
7. Glantz, S. A. & Slinker, B. K. Primer of applied
regression & analysis of variance (McGraw-Hill,
Medical Pub. Division, New York, 2001).
8. Neter, J., Wasserman, W. & Kutner, M. H. Applied
linear statistical models: regression, analysis of
variance, and experimental designs (Irwin,
Homewood, IL, 1990).
9. Rieke, F., Warland, D., de Ruyter van Steveninck,
R. & Bialek, W. Spikes: exploring the neural code
(MIT Press, Cambridge, MA, 1997).
10.Scott, T. R. & Plata-Salaman, C. R. Taste in the
monkey cortex. Physiol Behav 67, 489-511 (1999).
11.Brand, J. G. in Tasting and smelling (eds.
Beauchamp, G. K. & Bartoshuk, L.) 1-24
(Academic Press, San Diego, CA, 1997).
12.Pritchard, T. C. et al. Gustatory neural responses in
the medial orbitofrontal cortex of the old world
monkey. J Neurosci 25, 6047-56 (2005).
13.Zigmond, M. J., Bloom, F. E., Landis, S. C.,
Roberts, J. L. & Squire, L. R. Fundamental
neuroscience (Academic Press, San Diego, CA,
1999).
14.Rolls, E. T., Sienkiewicz, Z. J. & Yaxley, S.
Hunger modulates the responses to gustatory
stimuli of single neurons in the caudolateral
orbitofrontal cortex of the macaque monkey. Eur J
Neurosci 1, 53-60 (1989).
S.13
 Figure s1

Name Used
in Main Text Collapsed Variable Name Definition

total value chosen value + other value

chosen value ................................................................... chosen value
other value
(chosen–other) value
(other/chosen) value
total number
max number
chosen number
min number
other number
(max–min) number
(chosen–other) number
(min/max) number
(other/chosen) number
............ value A offered
offer value .............. value A|B offered .... { ............ value B offered
taste .................................................................................... A|B chosen
............ value A chosen
value A|B chosen .... { ............ value B chosen
Value of the chosen juice
Value of the non-chosen juice
chosen value – other value
(other value) / (chosen value)
max number + min number
Maximal offered number
Chosen number
Minimal offered number
Non-chosen number
max number – min number
chosen number – other number
(min number) / (max number)
(other number) / (chosen number)
Value of juice A offered
Value of juice B offered
Binary: 1 if A chosen, 0 if B chosen
Value of juice A chosen
Value of juice B chosen

Figure s1. Tested variables. We tested variables related to value, variables related to number, and variables related to
one of the two juices. Note that having assumed a linear value function (see Supplementary Methods), the value of one
of the two juices offered or chosen is proportional to the number of drops of juice. The two collapsed variables are
indicated in the bottom left of the table. In the main text, we re-labeled the variable value A|B offered = offer value, and
the variable A|B chosen = taste.
 Figure s2

position move offer

juice A direction type
pre-offer 0 0 1
post-offer 18 14 284
late delay 4 6 207
pre-go 4 8 133
RT 5 5 93
pre-juice 3 14 329
post-juice 5 11 332
total 39 58 1379
at least 1 33 46 505

Figure s2. 3-way ANOVA. We recorded the activity of 931 cells. We analyze each cell in each time window with a 3-
way ANOVA (factors [position of juice A] x [movement direction] x [offer type]). The three columns in the table indicate
the number of cells for which each of the main factors has a significant effect (p<0.001). The bottom row indicates the
number of cells whose activity pass the ANOVA test in at least one time window. The factor [offer type] is significant for
a total of 1379 responses.
 Figure s3

offer juice

10
average firing rate (sp/s)

505 task-related
426 other cells
0
-500 0 500 1000 -1000 -500 0 500 1000 1500
time (ms) time (ms)

Figure s3. Average activity profile. Neuronal activity was analyzed in 10ms-large, non-overlapping time bins. For each
cell, we included in this analysis all trials and all trial types. We then averaged the resulting activity profile across cells.
This analysis was done separately for task-related cells (i.e., cells that passed the ANOVA criterion in at least one time
window; blue color) and for other cells (i.e., cells that did not pass the ANOVA criterion in any time window; red color).
100% total value chosen value other value (chosen–other) value
1A = 3.2B 40 40 40 40
75% R 2 = 0.72 R 2 = 0.90 R2 = 0.56
30 30 30 30
50%
20 20 20 20
25%

0% 10 10 10 10
0:1 1:3 1:2 1:1 2:1 3:1 4:1 6:1 3:0
10:1
0 0 0 0
0 5 10 15 0 5 10 0 2 4 6 -5 0 5 10
40
(other/chosen) value total number max number chosen number
40 40 40 40
30 R2 = 0.41 R2 = 0.41 R2 = 0.50
30 30 30 30
20
20 20 20 20

10 10 10 10 10

0 0 0 0 0
0:1 1:3 1:2 1:1 2:1 3:1 4:1 6:110:13:0 0 1 2 0 5 10 0 5 10 0 5 10

offers (#B:#A) min number other number (max–min) number (chosen–other) number
40 40 40 40
R 2 = 0.40 R2 = 0.33
30 30 30 30

20 20 20 20

10 10 10 10

0 0 0 0
0 0.5 1 0 2 4 6 0 5 10 -5 0 5 10

(min/max) number (other/chosen) number value A offered value B offered

40 40 40 40
R2 = 0.32
30 30 30 30

20 20 20 20

10 10 10 10

0 0 0 0
0 0.5 1 0 2 4 6 0 5 10 0 5 10

A|B chosen value A chosen value B chosen

40 40 40

30 30 30

20 20 20

10 10 10

0 0 0
0 1 0 5 10 0 5 10

Figure s4. Linear regressions. The two top left panels show the behavioral choice pattern and one response plotted in
the same coordinates. The response is the same shown in figure 3a (main text), except that here we grouped trials by
trial type (squares for “A” choices; circles for “B” choices). In the other panels, the neuronal response (y-axes) is plotted
against each of the 19 variables (x-axes), and each dot represents one trial type. Values are expressed in units of V(B).
Blue regression lines indicate regression slopes significantly different from zero. For variables that explain the
response, the R2 is indicated on the top left corner in the panel. The highest R2 (chosen value) is marked in red.
 Figure s5

e be r er
valu valu
e
b er num ber numb
r ) ) e r m r ) m ) ed ed en n
e he en er er umb ber mber in) nu othe x) nu osen offer offer osen chos chose
a l ue n valu value n–ot /chos umb umb n u m nu m n– a /c h A B Bc h A B
e n e m
al v hose ther chos other otal n ax n hose in n ther max– chos min/ other ue alue ue alue
tot c o ( ( t m c m o ( ( ( ( va l v A | va l v

total value

chosen value

other value

(chosen–other) value

(other/chosen) value

total number

max number

chosen number

min number

other number

(max–min) number

(chosen–other) number

(min/max) number

other/chosen number

value A offered

value B offered

A | B chosen

value A chosen

value B chosen

Scale:
1
0

Figure s5. Correlation matrix. Elements of the correlation matrix vary between 0 and 1 (see Supplementary Methods).
Here the correlation matrix ρ is rendered in gray scale so that ρ=1 is represented in black (diagonal elements) and ρ=0
is represented in white. Small circles and crosses indicate matrix elements for which ρ>0.8 (excluding the diagonal).
We use circles for correlations that include one of the selected variables and crosses otherwise.
 Figure s6

a. explained e be r er
valu valu
e
b er num ber numb
d
e )
he r e n
)
er r be r r er
m
nu he r ) num e n) ere ere
d
n ose
n sen
a l ue valu value n–ot /chos umb umbe num umbe numb min) n–ot ax) /chos A off B off chose A ch B cho
n e n e m
al v hose ther chos other otal n ax n hose in n ther max– chos min/ other alue alue | B alue alue
tot c o ( ( t m c m o ( ( ( ( v v A v v
post-offer 130 126 47 79 20 124 130 111 13 21 116 73 4 17 78 112 50 70 94

late delay 68 65 38 47 24 76 73 65 19 26 55 40 4 17 45 63 43 46 50

pre-go 57 52 26 24 9 54 55 48 10 17 45 25 3 7 30 47 24 27 39

RT 38 41 12 24 10 41 43 39 8 8 32 23 2 7 24 34 18 20 25
Scale
pre-juice 110 121 27 76 14 124 133 161 11 30 119 141 5 50 69 108 127 87 148
161 cells
post-juice 130 139 23 81 10 119 126 160 12 23 113 147 1 44 77 100 122 92 147
0 cells

b. best fit v
e
alu valu
e be r
ber ) num ber num d
be r
e er ) n ) r be r r u m er u m n ) re red sen sen
l ue valu alue –oth hose mbe mber num mber umbe in) n –oth ax) n hose offe offe hosen cho cho
a n v en /c u u n u n –m e n m /c A B c A B
al v hose ther chos other otal n ax n hose in n ther max chos min/ other alue alue | B alue alue
tot c o ( ( t m c m o ( ( ( ( v v A v v
post-offer 20 47 10 17 2 6 21 16 4 5 14 3 1 3 20 19 16 18 10

late delay 21 17 7 14 6 3 9 9 3 5 7 7 2 1 13 16 10 19 10

pre-go 15 15 4 4 2 4 6 7 1 5 6 1 1 0 9 11 8 10 5

RT 7 12 3 8 2 3 6 5 1 2 4 3 0 2 6 6 5 6 2
Scale
pre-juice 18 33 3 7 2 2 9 39 1 4 16 20 0 9 23 10 51 26 23
52 cells
post-juice 19 43 2 10 2 6 12 32 0 2 13 23 1 6 24 8 53 23 26
0 cells

Figure s6. Qualitative analysis. Top. Numbers in the top panel represent for each variable (x-axis) and for each time
window (y-axis) the number of responses for which the linear regression provided a non-zero slope (i.e., the number of
responses explained by the variable). The color table underneath reports the same numbers in gray scale. Bottom.
Numbers in the bottom panel represent for each variable and for each time window the number of responses for which
the corresponding variable provided the best fit (highest R2). Again, the color table represents the same numbers in
gray scale.
 Figure s7

ue ue ber ber
val ) val r mber ) nummber num d d n
e r)
e en er r e
b r er nu e nu en er ere n ose sen
r ) e
alue valuvalue n–oth/chos umb umbe numumbenumb min) n–oth ax) chos A off B off choseA ch B cho
n n m /
al v se er ose er al n x n se n n er x– ose n/ er ue ue B ue ue
tot cho oth (ch (oth tot ma cho mi oth (ma (ch (mi (oth val val A | val val

% responses explained
100

90

80

70

60

50

40

30

20

10

0
0 2 4
number of variables

Figure s7. Stepwise selection method. See Supplementary Results.

 Figure s8

position move offer offer chosen taste other unclas-

juice A direction type value value class sified
0 0 1 pre-offer 0 0 0 1 0
18 14 284 post-offer 113 87 25 27 32
4 6 207 late delay 72 46 26 33 30
4 8 133 pre-go 46 37 16 15 19
5 5 93 RT 29 30 11 13 10
3 14 329 pre-juice 95 79 93 29 33
5 11 332 post-juice 90 93 97 24 28
39 58 1379 total 445 372 268 142 152
33 46 505 at least 1

Scale

113 cells

value A offered 0 cells

50

40
# responses

30

20 chosen value A|B c hosen

50 50
10
40 40
0
0 0 .5 1 30 30

20 20
value B offered
50
10 10
40
# responses

0 0
0 0 .5 1 0 0 .5 1
30
2 2
20 R R
10

0
0 0 .5 1
2
R

Figure s8. Population summary. Top. The figure illustrates the final result of the classification. The three columns on
the left repeat teh results of the s-way ANOVA (figure s2). The factor [offer type] is significant for a total of 1379
responses. The five columns on the right summarize the results of the variable selection analysis. Responses that
cannot be explained by any of the 19 variables appear in the rightmost column (unclassified). Variables value A|B
offered, chosen value and A|B chosen collectively explain 1085 responses (79% of the total, 88% of responses
explained by the 19 variables). Responses explained by one of the 19 variables but not explained by any of the
selected variables appear on the second column on the right (other class). The remaining three columns indicate the
number of responses classified as value A|B offered, chosen value, and A|B chosen, respectively. The color table
represents the same number in gray scale. The prevalence of different response classes varies over time windows:
value A|B offered is most prevalent shortly after the offer, chosen value is prevalent throughout the trial, and A|B
chosen is most prevalent late in the trial, before and after juice delivery. This time course can be also observed at much
higher resolution (figure 4, main text). Bottom. How well are responses accounted for in the final classification? The
histograms show for the four response classes the number of responses (y-axis) with the corresponding R2 (x-axis).
The responses included in the four histograms are 159, 286, 372 and 268, respectively. The mean (avg) and median
(med) of the four distributions are (avg=0.61, med=0.60), (avg=0.63, med=0.61), (avg=0.64, med=0.64) and (avg=0.64,
med=0.63), respectively.
 Figure s9

% responses 100
explained
80

60

40

20

0
0 1 2 3 4 5 6

number of variables

d=1 d=2 d=3 d=4 d=5

value A|B offered total value chosen value chosen value chosen value
selected variables

value A|B chosen value A|B offered value A|B offered value A|B offered

A|B chosen A|B chosen A|B chosen

chosen number other value
(chosen–other)
number

Figure s9. Best-subset method. See Supplementary Results.

 Figure s10

variable X variable Y nX nY p
chosen value total value 58 21 <10-5
chosen value (chosen–other) value 91 34 <10-7
chosen value chosen number 53 20 <10-5
value A|B offered total number 99 14 <10-10
value A|B offered max number 93 19 <10-10
value A|B offered (max–min) number 102 15 <10-10
value A|B offered value A|B chosen 100 19 <10-10
A|B chosen value A|B chosen 26 19 0.12

Figure s10. Post hoc analysis. The best subset of three variables includes chosen value, value A|B offered, and A|B
chosen. We tested these three variables separately against all other variables highly correlated with them (ρ>0.8). The
relevant pairs of variables are marked by small white circles in the correlation matrix in figure s5. The collapsed variable
value A|B offered is tested here against variables highly correlated either with value A offered or with value B offered. In
addition to the comparisons dictated by the ρ>0.8 criterion, we tested the variable chosen value against the alternative
variable chosen number. In the table, the two left columns indicate the tested variable and the alternative variable, the
next two columns indicate the marginal explanatory power of the two variables, and the right column indicates the result
of a binomial test. In essence, all tests indicate that selected variables have significantly higher explanatory power than
the challenging alternatives, except that the variable A|B chosen does not reach significance level against the
alternative value A|B chosen (bottom row).
 Figure s11

a. # responses 2 2
n n r 2
ere
d d
ere alue n ose ose mbe fered fered lue
of f of f v e c h c h nu of o f va
os
lue A lue B osen B ch lue A lue B osen lue A lue B osen ne
va va ch A | va va ch va va ch no
value A offered - 2 1 2 7 1 0 19 - - 127

value B offered 7 - 12 5 4 12 24 - 19 - 203

chosen value 12 15 - 13 4 3 4 - - 16 305

A|B chosen 12 2 22 - 10 8 12 - - - 202

b. % responses 2 2
d d en en er d d 2
f fere f fere alue e n h os h os u mb f fere f fere alue
A o e B o en v chos e A c e B c en
n A o e B o en v
lue lu os |B lu lu os lue lu os ne
v a v a c h A v a v a c h v a va ch no
value A offered - 1 1 1 4 1 0 12 - - 81

value B offered 2 - 4 2 1 4 8 - 7 - 72

chosen value 3 4 - 4 1 1 1 - - 4 83

A|B chosen 5 1 8 - 4 3 5 - - - 77

Figure s11. Second order encoding. See Supplementary Results.

 Figure s12

A = grape, B = 1/3 cranberry

100%
n* = 3.2
75%

50%
40
25%

0% 35
0:1 1:3 1:2 1:1 2:1 3:1 4:1 6:1 3:0
10:1

30
40 aA aB
25
30

20
20

10 15 k* = aA/ aB = 3.0 (± 1.4)

0 10
0:1 1:3 1:2 1:1 2:1 3:1 4:1 6:110:13:0
3 2 1 2 3 4 6 10
offers (#B:#A) #A chosen #B chosen →

Figure s12. One example of U-shaped response (same shown in figures 3a and s4). In the left panels, the behavioral
choice pattern and the neuronal response are plotted in the usual “ordinal” x-axis coordinates. From the behavioral
choice pattern, we obtain the relative value n*=3.2. In the right panel, the neuronal response is plotted in “cardinal” x-
axis coordinates, separately for trials in which the monkey chose juice A and juice B. The two regression slopes are
plotted in red. From the bi-linear regression, we obtain a slope ratio k* = 3.0 (±1.4) (±95% confidence interval).
 Figure s13

a all juice pairs b 1/2 apple : peppermint

8

4 4

slope ratio (k*)

2
2
slope ratio (k*)

1
k* = 0.08 + 1.18 n*
1

1/2 1 2 4
relative value (n*)

1/4

1/8 c regression slope b1 for different juice pairs

1 2 4 8
relative value (n*) regression: k* = b + b n*
0 1
mean b = 1.05 (±0.15)
1

grape : peppermint

grape : 1/3 cranberry

1/2 apple : peppermint

lemon k.a. : 1/2 apple

grape : 1/2 apple

fruit punch : 1/2 apple

1/2 apple : .60 g/l salt

-1 0 1 2 3
b1

Figure s13. Relationship between slope ratio k* and relative value n*. See Supplementary Results.
 Figure s14

sugar

X
slope ratio (k*)

water
1

1
relative value (n*)

Figure s14. Predicted relationship between k* and n*. If U-shaped responses encode chosen value, the values of k*
should lie on a diagonal (gray line). If they encode the quantity of some juice ingredient (e.g., water, sugar, or some
other ingredient X or Y), values of k* should lie on some horizontal line corresponding to the concentration ratio (blue,
red, green, and yellow lines). Similarly, if U-shaped responses encode the linear combination of multiple ingredient,
values of k* should lie on some horizontal line (e.g., brown dotted line).