LETTERS
Neural correlates, computation and behavioural
impact of decision confidence
Adam Kepecs
1
, Naoshige Uchida
1,2
, Hatim A. Zariwala
1,3
& Zachary F. Mainen
1,4
Humans and other animals must often make decisions on the basis
of imperfect evidence
1,2
. Statisticians use measures such as P values
to assign degrees of confidence to propositions, but little is known
about how the brain computes confidence estimates about deci-
sions. We explored this issue using behavioural analysis and
neural recordings in rats in combination with computational
modelling. Subjects were trained to perform an odour categoriza-
tion task that allowed decision confidence to be manipulated by
varying the distance of the test stimulus to the category boundary.
To understand how confidence could be computed along with the
choice itself, using standard models of decision-making
3–6
, we
defined a simple measure that quantified the quality of the evid-
ence contributing to a particular decision. Here we show that the
firing rates of many single neurons in the orbitofrontal cortex
match closely to the predictions of confidence models and cannot
be readily explained by alternative mechanisms, such as learning
stimulus–outcome associations
7–10
. Moreover, when tested using a
delayed reward version of the task, we found that rats’ willingness
to wait for rewards increased with confidence, as predicted by the
theoretical model. These results indicate that confidence esti-
mates, previously suggested to require ‘metacognition’
11,12
and
conscious awareness
13,14
, are available even in the rodent brain,
can be computed with relatively simple operations, and can drive
adaptive behaviour. We suggest that confidence estimation may be
a fundamental and ubiquitous component of decision-making.
Rats were trained on a two choice odour mixture categorization
task (Fig. 1a). On each trial, a binary mixture of two pure odorants
(A, caproic acid; B, 1-hexanol) was delivered at one of several con-
centration ratios (Fig. 1b), which were randomly interleaved from
trial-to-trial
15
. Choices were rewarded at the left choice port for mix-
tures A/B . 50/50 and at the right choice port for A/B , 50/50
(Fig. 1b). By varying the distance of the stimulus to the category
boundary (50/50) we could vary the difficulty of the decision
(Fig. 1c, d). Although the reward contingencies were deterministic,
subjects experienced varying degrees of decision uncertainty due to
imperfect perception of stimuli and/or knowledge of the category
boundary.
To explore the neural correlates of decision confidence, we
recorded single neuron activity in the orbitofrontal cortex (OFC;
Supplementary Fig. 1), a brain region implicated in decision-making
under uncertainty
16–20
. We reasoned that neural activity related to the
subject’s confidence in the outcome of a choice should occur while
the subject is anticipating the trial outcome, and therefore focused
our analysis on this delay period (Fig. 2a). The firing rates of many
OFC neurons were modulated by stimulus difficulty during the anti-
cipation period. Figure 2b, c shows the activity of a neuron that fired
more intensely following more difficult decisions. By replotting the
same data as a function of the choice accuracy associated with each
stimulus type, it can be seen that this neuron fired more vigorously
when the likelihood of an upcoming reward was lower (Fig. 2d). A
large fraction of OFC neurons, like this example, fired more intensely
for stimuli closer to the category boundary (120/563 at P , 0.05,
Wilcoxon signed-rank test). A smaller fraction (66/563) showed
the opposite tuning, firing at a higher intensity for easy stimuli, those
far from the category boundary (Fig. 2e, f).
The observed modulation of firing rate by stimulus difficulty is
consistent with previous findings that the response of many OFC
neurons correlates with the expected values associated with reward
predictive cues
7–10
. Surprisingly, however, when we compared correct
and incorrect choices for the same stimulus (for example, the 68/32
mixture), we found that many neurons showed different firing rates
even before the outcome was delivered. Figure 3a, b shows an
example of a neuron that tended to fire more when the rat had
1
Cold Spring Harbor Laboratory, 1 Bungtown Road, Cold Spring Harbor, New York 11724, USA.
2
Department of Molecular and Cellular Biology and Center for Brain Science, Harvard
University, Cambridge, Massachusetts 02138, USA.
3
Allen Institute for Brain Science, Seattle, Washington 98103, USA.
4
Champalimaud Neuroscience Programme, Instituto
Gulbenkian de Cie ˆnc ¸ia, 2780-901 Oeiras, Portugal.
a
b
50
100
60
80
100
0
c
d
B
A
0 32 44 56 68 100
0
100
32
68
44
56
56
44
68
32
100
0
Choice A Odour Choice B
Left choice (%) Accuracy (%)
Odour mixture (% A)
Figure 1 | Odour mixture categorization task. a, Schematic of the
behavioural paradigm. To initiate a trial, the rat enters the central odour port
and after a pseudorandom delay of 0.2–0.5 s a mixture of odours is delivered.
Rats respond by moving to the left or right choice port, where a drop of water
is delivered after a 0.3–2 s waiting period for correct choices. b, Stimulus
design. c, Performance of one rat discriminating between mixtures of caproic
acid (A) and 1-hexanol (B) in a single session. Error bars (s.e.m.) are hidden
by markers. Colours are used to represent odour mixtures, with different
blue and green blends representing different odour mixture ratios. d, Choice
accuracy as a function of odour mixture. Data across three rats are plotted as
mean 6 s.e.m.
Vol 455 | 11 September 2008 | doi:10.1038/nature07200
227
©2008 Macmillan Publishers Limited. All rights reserved