Neuron
Article
Neural Predictors of Purchases
Brian Knutson,1,
* Scott Rick,2
G. Elliott Wimmer,1
Drazen Prelec,3
and George Loewenstein2
1
Psychology and Neuroscience, Stanford University, Building 420, Jordan Hall, Stanford, CA 94305, USA
2
Social and Decision Sciences, Carnegie Mellon University, 208 Porter Hall, Pittsburgh, PA 15213, USA
3
MIT Sloan School of Management, Massachusetts Institute of Technology, E56-320, Cambridge, MA 02139, USA
*Correspondence: knutson@psych.stanford.edu
DOI 10.1016/j.neuron.2006.11.010
SUMMARY
Microeconomic theory maintains that purchases
are driven by a combination of consumer
preference and price. Using eventrelated
fMRI, we investigated how people weigh
these factors to make purchasing decisions.
Consistent with neuroimaging evidence suggesting
that distinct circuits anticipate gain
and loss, product preference activated the nucleus
accumbens (NAcc), while excessive prices
activated the insula and deactivated the
mesial prefrontal cortex (MPFC) prior to the purchase
decision. Activity from each of these
regions independently predicted immediately
subsequent purchases above and beyond selfreport
variables. These ﬁndings suggest that
activation of distinct neural circuits related to
anticipatory affect precedes and supports consumers’
purchasing decisions.
INTRODUCTION
The decision of whether to purchase a product is the fundamental
unit of economic analysis. From the bazaar to
the Internet, people typically consider characteristics of
available products, determine their cost, and then decide
whether or not to purchase. The success of economic theory
rests on its ability to characterize this repeated and
elementary decision process. Neuroeconomic methods
offer the hope of separating and characterizing distinct
components of the purchase decision process in individual
consumers.
In addition to being attracted to preferred products,
consumers avoid prices that seem excessive. Many incentive
schemes for promoting purchasing appear designed
to diminish the salience of payments (e.g., credit
cards) or to create the illusion that products have no
cost (e.g., frequent ﬂyer mileage) (Prelec and Simester,
2001). To explain these phenomena, recent behavioral
economic theories have postulated a hedonic competition
between the immediate pleasure of acquisition and an
equally immediate pain of paying (Prelec and Loewenstein,
1998). The notion that people consider prices as a
potential loss can be contrasted with a different economic
account in which people represent prices as potential
gains of alternative products that could be purchased for
the same amount of money (Deaton and Muellbauer,
1980).
The idea that purchase decisions involve a tradeoff between
the potential pleasure of acquisition and the pain of
paying is consistent, however, with recent neuroscientiﬁc
evidence that distinct neural circuits related to anticipatory
affect provide critical input into subsequent decisions
(Bechara et al., 1996; Kuhnen and Knutson, 2005). Mounting
neuroimaging evidence suggests that activity in different
neural circuits correlates with positive and negative
anticipatory affect. In the absence of choice, anticipation
of ﬁnancial gains activates the nucleus accumbens
(NAcc) and correlates with self-reported positive arousal,
whereas gain outcomes activate the mesial prefrontal cortex
(MPFC) (Knutson et al., 2001b). These ﬁndings have
been interpreted to indicate that NAcc activation correlates
with gain prediction, while MPFC activation correlates
with gain prediction errors (Knutson et al., 2003). Further
ﬁndings suggest that anticipation of physical pain
activates the insula, among other areas, and that insula
activation also correlates with self-reported negative
arousal (Buchel and Dolan, 2000; Paulus et al., 2003).
Thus, insula activation has been hypothesized to play
a critical role in loss prediction (Paulus and Stein, 2006).
Emerging evidence also suggests that activation in these
circuits may inﬂuence subsequent choice. For instance,
during an investing task involving choice of risky (e.g.,
stocks) or riskless (e.g., a bond) alternatives, NAcc activation
preceded switching to risk-seeking strategies (in
which anticipated gain should outweigh anticipated
loss), while insula activation preceded switching to riskaverse
strategies (in which anticipated loss should outweigh
anticipated gain) (Kuhnen and Knutson, 2005).
Compared with choices involving purely ﬁnancial risks,
the purchase of products represents a less constrained
decision scenario, because products can potentially
vary along inﬁnite dimensions. However, decisions to purchase
may recruit common anticipatory affective mechanisms.
For instance, a growing number of fMRI studies
have explored neural correlates of product preference.
Speciﬁcally, men who view pictures of sports cars versus
less desirable types of cars show increased mesolimbic
activation (midbrain, NAcc, MPFC) (Erk et al., 2002).
Both men and women who view pictures of preferred versus
nonpreferred drinks show increased MPFC activation
Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc. 147
(Paulus and Frank, 2003). Both men and women who taste
preferred versus nonpreferred drinks also show greater
MPFC activation (McClure et al., 2004). Finally, men who
view pictures of preferred versus nonpreferred brands of
beer show increased MPFC activation, and women who
view pictures of preferred versus nonpreferred brands of
coffee also show increased MPFC activation (Deppe
et al., 2005). Together, these ﬁndings implicate mesolimbic
dopamine projection areas in the representation of
anticipated gain, but do not clarify different roles of these
distinct projection areas (Knutson et al., 2005). Additionally,
subjects did not actually purchase products in any
of these studies. Preference may lead to purchasing, but
only when the price is right.
The goal of this study was to determine whether distinct
neural circuits respond to product preference versus excessive
prices, and to explore whether anticipatory activation
extracted from these regions could independently
predict subsequent decisions to purchase. Subjects
were scanned while engaging in a novel SHOP (i.e.,
‘‘Save Holdings Or Purchase’’) task, which consisted of
a series of trials, identical in temporal structure, in which
subjects could purchase products (example trial, Figure 1).
Subjects saw a labeled product (4 s), saw the product’s
price (4 s), and then chose either to purchase the product
or not (by selecting either ‘‘yes’’ or ‘‘no’’ presented randomly
on the right or left side of the screen; 4 s), before ﬁxating
on a crosshair (2 s) prior to the onset of the next trial.
Timing for each trial period was intentionally limited to
minimize distractions and maximize affective engagement
in the task (Slovic et al., 2002). We predicted that during
product consideration, preference would activate neural
circuits associated with anticipated gain (Knutson et al.,
2001a). We also predicted that during price presentation,
excessive prices would activate circuits associated with
anticipated loss (Paulus et al., 2003), as well as deactivate
brain regions previously associated with balancing potential
gains against losses (de Quervain et al., 2004). Finally,
we predicted that activation extracted from these regions
prior to the purchase decision would predict, above and
beyond self-report variables, whether individuals would
subsequently choose to purchase a product. Thus, this
work represents an initial attempt to distinguish neural correlates
of consumer reactions to preference versus price
information and use brain activation to predict purchasing.
RESULTS
Behavior
Subjects purchased 23.58 ± 13.31 (mean ± SEM) out of 80
products total (i.e., approximately 30% of the products
they saw). The percentage of products purchased did
not signiﬁcantly differ for men versus women. Therefore,
these groups were combined in subsequent analyses.
Subjects who purchased several products from one set
were more likely to purchase several from the other set
[r = 0.62, t(25) = 3.91, p < 0.001]. Thus, subsequent prediction
analyses controlled for individual ﬁxed effects. Purchasing
was also consistent over repeated presentations
of the same product [r = 0.83, t(25) = 47.73, p < 0.001].
Figure 1. SHOP Task Trial Structure and Regressors
For task structure, subjects saw a labeled product (product period; 4 s), saw the product’s price (price period; 4 s), and then chose either to purchase
the product or not (by selecting either ‘‘yes’’ or ‘‘no’’ presented randomly on the right or left side of the screen; choice period; 4 s), before ﬁxating on
a crosshair (2 s) prior to the onset of the next trial. In regression models, preference was correlated with brain activation during the product and price
periods, price differential was correlated with brain activation during the price period, and purchasing was correlated with brain activation during the
choice period.
148 Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc.
Neuron
Neural Predictors of Purchases
Speciﬁcally, 87% of products purchased during the ﬁrst
presentation were also purchased during the second presentation,
while 95% of products not purchased during
the ﬁrst presentation were also not purchased during the
second presentation. Reaction time did not differ between
products that were purchased [median ln(rt) = 6.70] versus
not purchased [median ln(rt) = 6.69; t(25) = 0.54, n.s.]. For
purchased products, (log) reaction time correlated negatively
with preference [r = À0.21, t(25) = 3.47, p < 0.01],
while for unpurchased products, (log) reaction time correlated
positively with preference [r = 0.11, t(25) = 3.04, p <
0.01]. Thus, subjects deliberated longer prior to purchasing
a product for which they had a relatively weak preference,
as well as prior to not purchasing a product for
which they had a relatively strong preference, suggesting
that reaction time indexed response conﬂict. To control for
these associations, reaction time was included in regression
models during the choice period as a covariate (Knutson
et al., 2005).
Localization
Preference was correlated with activation in the NAcc (TC:
±12, 13, À2; Figure 2), as well as other regions (Table 1),
during product and price periods. Price differential (i.e.,
the difference between what the subject was willing to
pay and the displayed price of the product) was correlated
with activation in MPFC (À4, 59, À3; Figure 2), as well as
other regions (Table 2). Purchasing was correlated with
deactivation of the bilateral insula (TC: ±32, 10, 9; Figure 2)
during the choice period, as well as activation of other
regions (Table 3). In summary, as predicted, NAcc activation
was positively correlated with preference during the
Figure 2. Correlated Activation, Volumes of Interest, and Corresponding Activation Time Courses
(Top row, left to right) Conjoined correlations of NAcc activation with preference during product and price periods; MPFC activation with price differential
during the price period; insula activation with decision to purchase during the choice period (n = 26). (Middle row, left to right) Volumes of
interest superimposed on structural images of the bilateral NAcc, bilateral MPFC, and right insula. (Bottom row, left to right) Bilateral NAcc activation
time courses for trials in which products were subsequently purchased versus not; bilateral MPFC activation time courses; and right insula activation
time courses (white, predicted divergence; ***, product period; $$$, price period; ???, choice period; all lagged/shifted right by 4 s; n = 26, *p < 0.05;
error bars = SEM).
Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc. 149
Neuron
Neural Predictors of Purchases
product and price periods (i.e., when the product was displayed);
MPFC activation was positively correlated with
price differential during the price period (i.e., when the
price was displayed); and insula activation was negatively
correlated with purchasing during the choice period (i.e.,
when subjects chose to purchase or not). Data were extracted
from the predicted regions and submitted to veriﬁcation,
prediction, and validation analyses. Data were
also extracted from other regions identiﬁed in the localization
analysis and submitted to prediction analyses in order
to determine whether they would add to the predictive
power of the three hypothesized regions.
Veriﬁcation
Comparison of volume-of-interest (VOI) activation time
courses veriﬁed that bilateral NAcc activation distinguished
between purchase versus nonpurchase trials during
the product period as predicted (p < 0.001), as well as
during subsequent price and choice periods (p < 0.001;
p < 0.05). Bilateral MPFC activation distinguished between
purchase versus nonpurchase trials during the
price period as predicted (p < 0.001), as well as during
the second half of the previous product period (p < 0.01)
and the subsequent choice period (p < 0.001). Right, but
not left, insula activation also distinguished between purchase
versus nonpurchase trials during price period as
predicted (p < 0.01), as well as during the subsequent
choice period (p < 0.01; Figure 2). Therefore, based on
predicted and subsequently veriﬁed initial points of significant
divergence in activation, we included bilateral average
NAcc activation during the product period, bilateral
average MPFC during the price period, and right insula activation
during the price period in subsequent prediction
analyses.
Table 1. Conjoint Activation Foci for Preference
Preference Set 1 Z Set 2 Z R A S
R anterior
cingulate / med FG
3.63 4.73 8 33 24
L anterior cingulate 3.85 4.82 À10 37 À6
R DLPFC 3.39 4.60 37 33 11
L DLPFC 3.24 4.09 À45 37 4
L medial
frontal gyrus
4.30 4.94 À4 31 30
L superior
frontal gyrus
3.35 3.71 À21 25 51
R anterior insula 3.22 5.63 29 21 À1
L anterior insula 4.05 4.27 À27 21 À3
R NAcc 3.47 4.30 10 12 À1
L NAcc 4.97 4.35 À10 11 0
R caudate 3.29 3.96 7 5 5
L caudate 3.78 3.85 À12 11 10
L globus pallidus 3.24 4.77 À8 2 À1
L posterior cingulate 3.48 4.79 À3 À34 31
Whole brain, p < 0.001 uncorrected for each product set, conjoint
for both product sets. Cluster = 3 voxels. Predicted activations
in italics, Product + Price Periods.
Table 2. Conjoint Activation Foci for Price Differential
Price differential Set 1 Z Set 2 Z R A S
R frontopolar cortex 3.43 4.22 30 60 À4
L frontopolar cortex 3.24 4.02 À13 68 À1
L MPFC 3.61 3.35 À4 59 À3
R MPFC 3.85 3.12 4 46 À6
R anterior cingulate 4.03 3.41 3 34 1
L parahippocampal gyrus 3.50 3.46 À9 À47 5
Whole brain, p < 0.001 uncorrected for each product set, conjoint
for both product sets. Cluster = 3 voxels. Predicted activations
in italics, Price Period.
Table 3. Conjoint Activation Foci for Purchase
Choice Set 1 Z Set 2 Z R A S
L VMPFC 3.97 4.71 À5 62 À4
4.39 4.10 À4 38 À11
L middle
frontal gyrus
4.69 5.54 À19 34 42
L inferior
frontal gyrus
4.08 4.66 À44 15 24
R OFC 3.92 4.04 30 29 À9
L OFC 3.58 3.58 À34 33 À10
R insula À3.48 À4.91 32 9 9
L insula À3.13 À4.67 À31 9 11
À3.38 À4.92 À38 À2 15
À3.34 À4.11 À38 À16 20
L postcentral gyrus À3.44 À4.69 À49 À19 21
À3.30 À3.40 À25 À35 58
posterior cingulate 4.36 5.57 0 À30 34
R inferior
parietal lobule
À3.75 À4.80 55 À35 27
L inferior
parietal lobule
À4.36 À4.07 À49 À31 28
L paracentral lobule À3.71 À4.13 8 À41 53
L posterior cingulate 3.93 4.73 À8 À52 20
L superior
parietal lobule
3.94 3.79 À25 À69 46
Whole brain, p < 0.001 uncorrected for each product set, conjoint
for both product sets. Cluster = 3 voxels. Predicted activations
in italics, Choice Period.
150 Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc.
Neuron
Neural Predictors of Purchases
Prediction
We ran logistic regressions to determine whether brain activation
could predict purchases, and whether it could do
so above and beyond self-report variables. We hypothesized
that NAcc activation during the product period, as
well as MPFC activation and right insula deactivation during
the price period, would predict subsequent decisions
to purchase. Regressions pooled all subjects and choices,
but included subject ﬁxed effects to control for individual
differences in percentage purchased (although models
without ﬁxed effects yielded similarly signiﬁcant results).
The ﬁrst model regressed the purchase decision on preference
and price differential, two self-report measures
provided by subjects post-scan. Both self-report variables
positively predicted purchasing. The second model regressed
the purchase decision on brain activation extracted
from the three VOIs at predicted time points prior
to the purchase decision (i.e., NAcc activation during the
product period, and MPFC and insula activation during
the price period). Brain activation from all three regions
signiﬁcantly predicted purchasing. Speciﬁcally, bilateral
NAcc activation during the product period and bilateral
MPFC activation during the price period signiﬁcantly predicted
subsequent decisions to purchase, while right insula
activation during the price period signiﬁcantly predicted
subsequent decisions not to purchase (all p <
0.001). A third model combined both self-report and brain
activation variables. Brain activation variables still signiﬁcantly
predicted decisions to purchase, although coefﬁcient
magnitudes decreased by about half. These ﬁndings
implied that activation in these three brain regions correlated
with the two self-report variables (Supplement 10).
Nevertheless, the third model also demonstrated that,
above and beyond self-reported preference and price differential,
NAcc, MPFC, and right insula activation independently
and signiﬁcantly predicted subsequent decisions
to purchase a product (as indicated by a higher
R2
) and that including both self-report and brain parameters
signiﬁcantly improved the ﬁt of the model to the
data (as indicated by a lower Akaike Information Criterion
value [AIC]) (Table 4). In a further analysis, brain activation
extracted from other regions identiﬁed in localization analyses
did not signiﬁcantly add to the ability of the hypothesized
brain activation variables to predict purchasing
(Supplement 4). Speciﬁcally, other regions correlated
with preference did not enhance the ability of NAcc activation
during the product period to predict purchasing, and
other regions correlated with price differential did not enhance
the ability of MPFC or right insula activation during
the price period to predict purchasing.
Validation analyses established the robust generalizability
of brain activation to predict purchasing in other
scenarios. Speciﬁcally, relative to chance prediction
(50%), leave-one-out cross-validation indicated that the
brain activation variables accurately predicted purchasing
at 60% (±1%; p < 10À10
).
Table 4. Logistic Regression Models Predicting Decisions to Purchase or Not
Self report Brain activation Combined
Constant À16.14***a
À1.74 À16.05***
À6.12 (0.379)b
À0.30 (0.170) À6.12 (0.381)
Preference 18.95*** 18.79***
1.21 (0.064) 1.20 (0.064)
Price differential 12.35*** 11.97***
0.14 (0.011) 0.13 (0.011)
NAcc (bilateral) 5.59*** 2.79**
0.85 (0.153) 0.60 (0.217)
MPFC (bilateral) 7.27*** 3.42***
0.78 (0.107) 0.51 (0.150)
Insula (right) À5.31*** À2.53**
À0.85 (0.160) À0.61 (0.239)
Number of observations 3,909 3,909 3,909
Pseudo-R2
0.528 0.105 0.533
AICc
2279.6 4271.8 2259.7
Subjects with signiﬁcant ﬁxed effectsd
11 16 13
a
Signiﬁcance: ***p < 0.001; **p < 0.01; *p < 0.05.
b
Regression includes subject ﬁxed effects, Z-scores above coefﬁcients with standard errors in parentheses.
c
AIC, Akaike Information Criterion (lower score indicates better ﬁt to the data).
d
Out of 26; p < 0.01.
Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc. 151
Neuron
Neural Predictors of Purchases
DISCUSSION
The goal of this study was to characterize neural predictors
of purchasing. We hypothesized and found that activation
in regions associated with anticipating gain (the
NAcc) correlated with product preference, while activation
in regions associated with anticipating loss (the insula)
correlated with excessive prices. Further, activation in
a region implicated in integrating gains and losses (the
MPFC) correlated with reduced prices. Analyses of time
course data extracted from each of these regions indicated
that, while NAcc activation initially predicted subsequent
purchasing decisions during product presentation,
insula and MPFC activation initially predicted subsequent
purchasing decisions during price presentation. Even after
controlling for retrospective self-reported preference
and purchasing price, activation in these brain regions independently
predicted decisions to purchase. Validation
analyses indicated that the ability of brain activation to
predict purchasing would generalize to other purchasing
scenarios. The results did not vary signiﬁcantly as a function
of subjects’ sex, and were replicated across two different
sets of products. Together, these ﬁndings suggest
that activation of distinct brain regions related to anticipation
of gain and loss precedes and can be used to predict
purchasing decisions.
While several other brain regions have been implicated
in decision-making in both comparative and human research,
these regions did not play central roles in the present
study, perhaps due to speciﬁc aspects of the SHOP
task (Supplement 4). For instance, the parietal cortex
has been implicated in decision-making (Huettel et al.,
2005; Platt and Glimcher, 1999). However, while many decision-making
tasks that elicit parietal activation involve
a spatial component, the SHOP task minimizes spatial demands
by presenting items and prices sequentially in the
center of the screen. Thus, parietal activation may play
a role in mapping evaluative information to spatial action
plans, but it was not observed in relation to the variables
of interest in this task. Anterior cingulate activation has
also been implicated in decision-making (Volz et al.,
2005), but may be more related to conﬂict between potentially
competing courses of action (Botvinick et al., 1999).
For instance, in a ﬁnancial decision-making task, while anterior
cingulate activation was related to response conﬂict,
it did not predict subsequent risk-seeking versus riskaverse
choices (Kuhnen and Knutson, 2005). Similarly, in
the SHOP task, anterior cingulate activation was greatest
in situations involving high response conﬂict (e.g., long reaction
times to choose, items with high preference but
also high price, etc.), but did not signiﬁcantly add to other
brain regions’ ability to predict purchasing decisions.
Thus, while anterior cingulate activation may facilitate
conﬂict resolution, activation in this region does not necessarily
predict how conﬂict will be resolved. Amygdalar
and orbitofrontal cortex activation have also been implicated
in decision-making (Bechara et al., 2000), but did
not signiﬁcantly predict purchasing in this task. While
activation of these regions has been most robustly elicited
in situations involving learning, the SHOP task is designed
to minimize learning demands. Unlike studies that focus
on how preferences are established, this study instead
focused on how people decide to purchase based on
already-established preferences.
The study’s design and additional analyses ruled out
a number of alternative accounts of brain activation prior
to the purchase decision. First, anticipatory activation
could not be attributed to increased motor preparation
prior to purchasing, since subjects had to press a button
to indicate either the choice to purchase or not to purchase
a product, and did not know which button would
indicate purchasing versus not purchasing until the
choice prompt appeared (thus sequestering motor preparation
and execution to the choice period). Further, reaction
times did not differ between decisions to purchase
or not to purchase. Self-reported familiarity with products
(collected in a subset of 20 subjects) also could not account
for the correlation between NAcc activation and
preference, since additional localization analyses indicated
that familiarity was not correlated with NAcc activation
during the product and price periods. Also, prediction
analyses including familiarity as an independent
variable did not reduce the ability of NAcc activation to
predict purchasing and only slightly increased overall
model ﬁt (Supplement 5). Additionally, the few trials involving
previously-owned items were omitted from analyses.
Similarly, price alone could not account for the correlation
between MPFC activation and price differential,
since additional localization analyses indicated that price
was not signiﬁcantly correlated with MPFC activation
during the price period, and prediction analyses including
price did not reduce the ability of MPFC activation to predict
purchasing, although inclusion of price did increase
overall model ﬁt (Supplement 6). These ﬁndings are consistent
with the idea that people do not react as much to
absolute price as to the price relative to what they think is
acceptable for a given product (Thaler, 1985) (making it
difﬁcult to determine whether prices are high or low without
knowing their associated product). Finally, anticipatory
activation could not be attributed to a global increase
in neural recruitment (as might be expected in
the case of general arousal), because while NAcc and
MPFC showed increased activation prior to purchasing,
insula instead showed decreased activation prior to
purchasing.
To maximize subject engagement and minimize distraction,
SHOP task trials progressed at a fairly rapid pace
(i.e., 4 s per trial phase). While prior studies indicate that
hemodynamic responses in the VOIs (i.e., NAcc, MPFC,
insula) typically peak from 4–6 s after stimulus onset (Knutson
et al., 2003), the hemodynamic response can rise as
early as 2 s after stimulus onset, raising potential concerns
about the separability of signals during different trial
phases. These concerns might speciﬁcally affect inferences
about localization (e.g., MPFC activation depends
on the revelation of price, but not product, information)
152 Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc.
Neuron
Neural Predictors of Purchases
and prediction (e.g., MPFC activation begins to predict
purchasing after price, but not choice, information is revealed).
However, reanalysis indicated that models in
which the price differential regressor was lagged forward
(by 4 s into the product period) no longer exhibited correlated
activation in the MPFC or right insula (Supplement 7).
This ﬁnding suggests that the correlation of MPFC activation
with price differential depended upon the revelation of
price information. Additionally, brain activation extracted
from the MPFC and right insula during product presentation
(i.e., prior to price presentation by 4 s) did not signiﬁcantly
predict purchasing. Two additional experiments
were conducted to verify the dependence of MPFC and
right insula results on the appearance of price information
(Supplements 8 and 9). In a ﬁrst additional experiment (n =
8 males), lagging the appearance of price information (by 4
s) also lagged the correlation of MPFC and right insula activation
with price differential. Similarly, lagged MPFC activation
also best predicted subsequent purchasing (Supplement
8). In a second additional experiment (n = 8
males), lagging the onset of the choice period (by 4 s)
did not alter MPFC or right insula activation’s correlation
with price differential, and activation in these regions during
the price period continued to predict purchasing (Supplement
9). Together, these ﬁndings suggest that whereas
NAcc activation reﬂected subjects’ reaction to products,
MPFC and insula activation reﬂected subjects’ reaction
to price information. In all experiments, NAcc activation
begins to predict purchasing during the product period,
while MPFC and insula activation begin to predict purchasing
during the price period.
Activation in the hypothesized regions (i.e., NAcc,
MPFC, and insula) conformed to most, but not all, predictions.
Activation in all three regions correlated more robustly
with subjective variables (i.e., product preference,
price differential) than with objective variables (i.e., product
identity, price) (Supplement 11). NAcc activation correlated
strongly with product preference, discriminating
between eventually purchased and not purchased products
as soon as the product was displayed, while MPFC
activation correlated strongly with price differential, and
did not discriminate between eventually purchased and
not purchased products until the price was displayed.
These ﬁndings are consistent with distinct gain prediction
accounts of NAcc function and gain prediction error accounts
of MPFC function (Knutson et al., 2003). While right
insula showed deactivation during the price period, activation
in this region did not signiﬁcantly correlate with
price differential, although it did correlate nonsigniﬁcantly
in the predicted direction (conjoined Z = À1.36), and it also
discriminated between eventually purchased and unpurchased
products. These ﬁndings are not inconsistent
with a loss prediction account of insula function (Paulus
and Stein, 2006), since validation analyses further indicated
that insula deactivation predicted purchasing, but
they may suggest an inﬂuence of other factors besides excessive
price on insula activation (e.g., responses to nonpreference
or a more prolonged response). Thus, the
speciﬁcity of the insula response to excessive prices remains
to be clariﬁed by future research.
The present ﬁndings have several implications. With respect
to neuroscience, by implicating common circuits
(i.e., NAcc, MPFC, and insula) in decisions to purchase diverse
products, the ﬁndings are consistent with a ‘‘common
currency’’ account of purchasing (Knutson et al.,
2005; Montague and Berns, 2002; Shizgal, 1997). However,
they additionally suggest that decisions to purchase
may involve distinct dimensions related to anticipated
gain and loss rather than just a single dimension related
to anticipated gain. These ﬁndings not only add to prior
studies of product preference, but also link to studies of
social decisions that implicate NAcc activation in the intention
to cooperate (King-Casas et al., 2005; Rilling
et al., 2002) and insula activation in the intention to defect
(Sanfey et al., 2003). Thus, these ﬁndings further illustrate
the power of the neuroeconomic approach to elucidate
distinct neuropsychological components that may exert
consistent collective inﬂuences on subsequent purchasing
decisions. They also suggest that even commonplace
purchasing decisions can be deconstructed with methods
adopted from psychology, economics, and neuroscience.
With respect to economic theory, the ﬁndings support
the historical notion that individuals have immediate affective
reactions to potential gain and loss, which serve as inputs
into decisions about whether or not to purchase
a product (Kuhnen andKnutson, 2005).This ﬁnding hasimplications
for understanding behavioral anomalies, such
as consumers’ growing tendency to overspend and undersave
when purchasing with credit cards rather than cash.
Speciﬁcally, the abstract nature of credit coupled with deferred
payment may ‘‘anaesthetize’’ consumers against
the pain of paying (Prelec and Loewenstein, 1998). Neuroeconomic
ﬁndings thus might eventually suggest methods
of restructuring institutional incentives to facilitate
increased saving.
Finally, the results illustrate an innovative technical application
of fMRI in which brain activation is used to predict
purchasing decisions on-line. Whether added information
from fMRI data is more cost-effective than
simple self-report remains to be established, and may depend
upon future methodological and technical advances.
fMRI prediction methods may eventually prove most useful
in situations when people’s behavior and self-reported
preferences diverge.
In summary, this study provides initial evidence that
speciﬁc patterns of brain activation predict purchasing.
Prior to the purchase decision, preference elicits NAcc activation,
while excessive prices can elicit insula activation
and MPFC deactivation. Anticipatory neural activation in
these regions predicts subsequent purchasing decisions.
The ﬁndings are consistent with the hypothesis that the
brain frames preference as a potential beneﬁt and price
as a potential cost, and lend credence to the notion that
consumer purchasing reﬂects an anticipatory combination
of preference and price considerations. A physiological
account of these factors may facilitate neurally
Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc. 153
Neuron
Neural Predictors of Purchases
constrained theories of human decision-making (Glimcher
and Rustichini, 2004). Such theories may not only help scientists
to decompose the components that go into decisions,
but also help to build neuroeconomic models that
better predict choice and inform policy.
EXPERIMENTAL PROCEDURES
Subjects
Twenty-six healthy right-handed adults (12 females; age range 18–26)
participated in the study. Along with the typical magnetic resonance
exclusions (e.g., metal in the body), subjects were screened for psychotropic
drugs and ibuprofen, substance abuse in the past month,
and history of psychiatric disorders (DSM IV Axis I) prior to collecting
informed consent. Subjects were paid $20.00 per hour for participating
and also received $20.00 in cash to spend on products during each of
the two sessions. In addition to the 26 subjects who were included in
the analysis, 6 subjects were excluded because they purchased fewer
than four items per session (i.e., <10%) and therefore failed to provide
sufﬁcient data to model, and eight subjects were excluded due to
excessive head motion (i.e., more than 2 mm from one whole brain
acquisition to the next during either of the two scanning sessions).
Task
To ensure subjects’ engagement in the SHOP task, one trial was randomly
selected after each scanning session to count ‘‘for real.’’ If subjects
had chosen to purchase the product presented during the randomly
selected trial, they paid the price that they had seen in the
scanner from their $20.00 endowment and were shipped the product
within two weeks. If not, subjects kept their $20.00 endowment. Subjects
were actually shipped products following 15 (29%) of the total of
52 sessions. Products were preselected to have above-median attractiveness
as rated by an independent sample drawn from the same
population prior to the study (Supplement 1). While products ranged
in retail price from $8.00–$80.00, to encourage purchasing, the prices
subjects saw in the scanner were discounted by 75% of retail value.
Consistent with preliminary studies, this discount led subjects to purchase
30% of the products on average, generating sufﬁcient instances
of purchasing to adequately power statistical modeling with standard
regressors. Consistent with the ﬁndings of pilot experiments, the discount
yielded an average price differential that did not signiﬁcantly differ
from $0.00 (i.e., mean price differential = À$0.54, SD = ±$8.56), indicating
that on average, subjects perceived half of the products to be
cheap, and the other half to be expensive (Supplement 1).
Subjects were instructed in the task and tested for task comprehension
prior to entering the scanner. After leaving the scanner, subjects
rated products on several dimensions (i.e., desirability, percentage
of retail price that they would be willing to pay for the product, and
whether or not they already owned the product). Post-scan ratings
of desirability and willingness to pay were used to create individualized
regressors that modeled subjective reactions to product preference
and price, as described in the fMRI Acquisition and Analysis subsection.
After the ratings, subjects were informed which trial had been randomly
selected to count for real.
To ensure that results would generalize across different product
sets, subjects participated in two SHOP task scanning sessions,
each separated by less than two weeks. During each session, subjects
were presented with opportunities to buy 40 different products twice,
allowing subsequent veriﬁcation of choice consistency. Order was
pseudorandomized within a session and repeated, such that each
task session took approximately 9 min, 20 s. Thus, a total of 80 different
products were used in the experiment. Half of the subjects encountered
one group of 40 products in the ﬁrst session and the other 40 in
the second session, while the other half of the subjects encountered
the two sets of 40 products in reverse order.
fMRI Acquisition and Analysis
Images were acquired with a 1.5-T General Electric MRI scanner using
a standard birdcage quadrature head coil. Twenty-four four-millimeter-thick
slices (in-plane resolution 3.75 3 3.75 mm, no gap) extended
axially from the mid-pons to the top of the skull, providing whole-brain
coverage and adequate spatial resolution of subcortical regions of interest
(e.g., midbrain, NAcc, OFC). Whole-brain functional scans were
acquired with a T2*-sensitive spiral in-/out- pulse sequence (TR = 2 s,
TE = 40 ms, ﬂip = 90
) designed to minimize signal dropout at the base
of the brain (Glover and Law, 2001) (Supplement 3). High-resolution
structural scans were also acquired to facilitate localization and coregistration
of functional data using a T1-weighted spoiled grass
sequence (TR = 100 ms, TE = 7 ms, ﬂip = 90
).
Analyses were conducted using Analysis of Functional Neural Images
(AFNI) software (Cox, 1996). For preprocessing, voxel time series
were sinc interpolated to correct for nonsimultaneous slice acquisition
within each volume, concatenated across runs, corrected for motion,
high-pass ﬁltered (admitting frequencies with period <90 s), and normalized
to percent signal change with respect to the voxel mean for
the entire task. Visual inspection of motion correction estimates conﬁrmed
that no subject’s head moved more than 2.0 mm in any dimension
from one volume acquisition to the next. All regression models included
six regressors indexing residual motion and six regressors
modeling baseline, linear, and quadratic trends for each of the two
runs.
Next, localization, veriﬁcation, prediction, and validation analyses
were conducted. For localization analyses, regressors of interest included
individualized and orthogonalized models of (1) product preference
during the product and price display periods, (2) price differential
(or the difference between the price subjects were willing to pay for
a product and the displayed price) during the price display period,
and (3) decision to purchase the product or not during the choice
period. Because product and price presentation overlapped in time,
product preference and price differential were modeled simultaneously
during price presentation to ensure that each of these regressors
would account for unique variance in brain activation prior to the
purchase decision. An additional regressor of noninterest included reaction
time during the choice period. Preference was derived from
a post-scan rating of desire to own each product (on a 1–7 Likert scale
later converted to a À3 to +3 scale) and modeled during the product
and price periods. Price differential was calculated as the price that
subjects reported they would be willing to pay for a product postscan
(rated as a percentage of retail price) minus price displayed for
that product during the scan, and was modeled during the price period.
Thus, positive price differential indicated that subjects perceived
a product as cheap (relative to their willingness to pay for it), while negative
price differential indicated that subjects perceived a product as
expensive (i.e., similar to ‘‘net utility’’ or ‘‘consumer surplus’’ in economics)
(Prelec and Loewenstein, 1998). Together, preference and
price differential modeled each subject’s individual subjective reaction
to each product and its associated price. The purchasing decision was
represented by a simple contrast between purchased versus unpurchased
products, and modeled during the choice period. Trials involving
products that were already owned were excluded from analyses.
Regressors of interest modeling preference, price differential, and purchasing
were convolved by multiplying the original ‘‘box car’’ function
regressors with a g-variate function approximating a hemodynamic response
prior to inclusion in regression models (Cohen, 1997).
Maps of contrast coefﬁcients for regressors of interest were transformed
into Z-scores, coregistered with structural maps, spatially normalized
by manually warping to Talairach space, slightly spatially
smoothed to minimize effects of anatomical variability (FWHM = 4
mm), and collectively submitted to a one-sample t test against the
null hypothesis of no activation to test for a group difference while controlling
for random effects. Separate group maps were constructed for
each of the two product sets. Each of these maps were thresholded at
p < 0.001 (uncorrected) and conjoined using an ‘‘AND’’ operation
154 Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc.
Neuron
Neural Predictors of Purchases
(Knutson et al., 2001a). Conjointly activated regions passing this
threshold thus identiﬁed replicable activations over both product
sets at p < 0.001 (Nichols et al., 2006). Brain regions of interest were
then examined for evidence of conjointly signiﬁcant activation across
both product sets.
For veriﬁcation analyses, bilateral spherical VOIs (8 mm diameter)
were constructed at centers of mass within the predicted regions
(e.g., NAcc, MPFC, insula), ensuring that equal amounts of data
were extracted for each subject in each region. VOIs were superimposed
on each subject’s brain and individually adjusted to ensure
that they only included gray matter in the predicted regions for each
subject (Supplement 3). Spatially averaged activation time courses
were extracted from each VOI and then divided by the average activation
for each VOI over the course of the entire experiment to derive
measures of percent signal change. Percent signal change time
courses were then averaged for purchase versus nonpurchase trials
within subjects. Paired t tests then compared percent signal change
for purchase versus nonpurchase trials across subjects. Percent signal
change was compared at predicted time points (i.e., the onset of product
presentation, price presentation, and choice presentation) lagged
by 4 s (to account for the lag in hemodynamic response), with a significance
threshold that corrected for comparisons at each trial period
(i.e., p < 0.017). These VOI percent signal change data were also
used in subsequent prediction analyses.
Prediction analyses also utilized time courses of percent signal
change for each VOI to predict purchasing. Speciﬁcally, prediction
analyses utilized logistic regression to determine whether NAcc activation
during the product period as well as MPFC activation and insula
deactivation during the price period (all lagged by 4 s) would predict
the subsequent decisions to purchase a product during the choice period,
both before and after controlling for self-report variables (i.e.,
preference and price differential). Logistic regression models also included
ﬁxed effects and motion correction estimates (Kuhnen and
Knutson, 2005) (however, analyses remained signiﬁcant and qualitatively
similar when ﬁxed effects were not included in the models).
The AIC was used to compare model ﬁt (Akaike, 1974). Additional analyses
used identical models, but included data extracted from additional
VOIs (e.g., other cortical regions).
For validation analyses, brain activation data used in logistic regressions
for the ﬁrst presentation of each product were submitted to
leave-one-out cross-validation analysis to establish generalizability
of brain activation as predictors of purchasing to other datasets (Kamitani
and Tong, 2005). While brain activation data was collected before
choice, self-report variables were not, so they were not included in
these analyses. Since fewer items were purchased than not purchased
(see Results), data for nonpurchase trials were randomly removed until
the ratio of purchase to nonpurchase trials (and therefore, the base rate
prediction) was 50% for each subject (n = 1150 observations total).
This adjustment of the prior probability for classiﬁcation was veriﬁed
with 10-fold cross-validation.
Supplemental Data
The Supplemental Data for this article can be found online at http://
www.neuron.org/cgi/content/full/53/1/147/DC1/.
ACKNOWLEDGMENTS
We thank Cliff Baum and George Wang for assistance in recruitment of
subjects and data analysis; Logan Grosenick for assistance with data
analysis; as well as Jeffrey C. Cooper, Samuel McClure, Hilke Plassmann,
and three anonymous reviewers for advice on earlier versions
of the manuscript. This research was supported in part by a National
Institute of Aging seed grant AG024957-02 to B.K., a National Science
Foundation Graduate Research Fellowship to S.R., the John Simon
Guggenheim Memorial Fellowship to D.P., and a MacArthur Foundation
network grant to G.L.
Received: August 23, 2006
Revised: October 13, 2006
Accepted: November 9, 2006
Published: January 3, 2007
REFERENCES
Akaike, H. (1974). A new look at the statistical model identiﬁcation.
IEEE Transactions on Automatic Control 19, 716–723.
Bechara, A., Tranel, D., Damasio, H., and Damasio, A.R. (1996). Failure
to respond autonomically to anticipated future outcomes following
damage to prefrontal cortex. Cereb. Cortex 6, 215–225.
Bechara, A., Damasio, H., and Damasio, A.R. (2000). Emotion, decision
making and the orbitofrontal cortex. Cereb. Cortex 10, 295–307.
Botvinick, M., Nystrom, L.E., Fissell, K., Carter, C.S., and Cohen, J.D.
(1999). Conﬂict monitoring versus selection-for-action in anterior cingulate
cortex. Nature 402, 179–181.
Buchel, C., and Dolan, R.J. (2000). Classical fear conditioning in functional
neuroimaging. Curr. Opin. Neurobiol. 10, 219–223.
Cohen, M.S. (1997). Parametric analysis of fMRI data using linear
systems methods. Neuroimage 6, 93–103.
Cox, R.W. (1996). AFNI: Software for analysis and visualization of
functional magnetic resonance images. Comput. Biomed. Res. 29,
162–173.
de Quervain, D.J., Fischbacher, U., Treyer, V., Schellhammer, M.,
Schnyder, U., Buck, A., and Fehr, E. (2004). The neural basis of altruistic
punishment. Science 305, 1254–1258.
Deaton, A., and Muellbauer, J. (1980). Economics and Consumer
Behavior (Cambridge: Cambridge University Press).
Deppe, M., Schwindt, W., Kugel, H., Plassman, H., and Kenning, P.
(2005). Nonlinear responses within the medial prefrontal cortex reveal
when speciﬁc implicit information inﬂuences economic decision
making. J. Neuroimaging 15, 171–182.
Erk, S., Spitzer, M., Wunderlich, A.P., Galley, L., and Walter, H. (2002).
Cultural objects modulate reward circuitry. Neuroreport 13, 2499–
2503.
Glimcher, P.W., and Rustichini, A. (2004). Neuroeconomics: The consilience
of brain and decision. Science 306, 447–452.
Glover, G.H., and Law, C.S. (2001). Spiral-in/out BOLD fMRI for increased
SNR and reduced susceptibility artifacts. Magn. Reson.
Med. 46, 515–522.
Huettel, S.A., Song, A.W., and McCarthy, G. (2005). Decisions under
uncertainty: probabilistic context inﬂuences activation of prefrontal
and parietal cortices. J. Neurosci. 25, 3304–3311.
Kamitani, Y., and Tong, F. (2005). Decoding the visual and subjective
contents of the human brain. Nat. Neurosci. 8, 679–685.
King-Casas, B., Tomlin, D., Anen, C., Camerer, C.F., Quartz, S.R., and
Montague, P.R. (2005). Getting to know you: reptutation and trust in
a two-person economic exchange. Science 308, 78–83.
Knutson, B., Adams, C.M., Fong, G.W., and Hommer, D. (2001a). Anticipation
of increasing monetary reward selectively recruits nucleus
accumbens. J. Neurosci. 21, RC159.
Knutson, B., Fong, G.W., Adams, C.M., Varner, J.L., and Hommer, D.
(2001b). Dissociation of reward anticipation and outcome with eventrelated
FMRI. Neuroreport 12, 3683–3687.
Knutson, B., Fong, G.W., Bennett, S.M., Adams, C.M., and Hommer,
D. (2003). A region of mesial prefrontal cortex tracks monetarily rewarding
outcomes: Characterization with rapid event-related FMRI.
Neuroimage 18, 263–272.
Knutson, B., Taylor, J., Kaufman, M., Peterson, R., and Glover, G.
(2005). Distributed neural representation of expected value. J. Neurosci.
25, 4806–4812.
Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc. 155
Neuron
Neural Predictors of Purchases
Kuhnen, C.M., and Knutson, B. (2005). The neural basis of ﬁnancial
risk-taking. Neuron 47, 763–770.
McClure, S.M., Li, J., Tomlin, D., Cypert, K.S., Montague, L.M., and
Montague, P.R. (2004). Neural correlates of behavioral preference for
culturally familiar drinks. Neuron 44, 379–387.
Montague, P.R., and Berns, G.S. (2002). Neural economics and the
biological substrates of valuation. Neuron 36, 265–284.
Nichols, T., Brett, M., Andersson, J., Wager, T., and Poline, J.B. (2006).
Valid conjunction inference with the minimum statistic. Neuroimage
25, 653–660.
Paulus, M.P., and Frank, L.R. (2003). Ventromedial prefrontal cortex
activation is critical for preference judgments. Neuroreport 14, 1311–
1315.
Paulus, M.P., and Stein, M.B. (2006). An insular view of anxiety. Biol.
Psychiatry 60, 383–387.
Paulus, M.P., Rogalsky, C., and Simmons, A. (2003). Increased activation
in the right insula during risk-taking decision making is related to
harm avoidance and neuroticism. Neuroimage 19, 1439–1448.
Platt, M.L., and Glimcher, P.W. (1999). Neural correlates of decision
variables in the parietal cortex. Nature 400, 233–238.
Prelec, D., and Loewenstein, G.F. (1998). The red and the black:
Mental accounting of savings and debt. Marketing Science 17, 4–28.
Prelec, D., and Simester, D. (2001). Don’t leave home without it. Mark.
Lett. 12, 5–12.
Rilling, J., Gutman, D., Zeh, T., Pagnoni, G., Berns, G., and Kilts, C.
(2002). A neural basis for social cooperation. Neuron 35, 395–405.
Sanfey, A.G., Rilling, J.K., Aronson, J.A., Nystrom, L.E., and Cohen,
J.D. (2003). The neural basis of economic decision-making in the
Ultimatum Game. Science 300, 1755–1758.
Shizgal, P. (1997). Neural basis of utility estimation. Curr. Opin. Neurobiol.
7, 198–208.
Slovic, P., Finucane, M., Peters, E., and MacGregor, D. (2002). The affect
heuristic. In Heuristics and Biases: The Psychology of Intuitive
Judgment, T. Gilovich, D. Grifﬁn, and D. Kahneman, eds. (New York:
Cambridge University Press), pp. 397–420.
Thaler, R. (1985). Mental accounting and consumer choice. Marketing
Science 4, 199–214.
Volz, K.G., Schubotz, R.I., and von Cramon, D.Y. (2005). Variants of
uncertainty in decision-making and their neural correlates. Brain
Res. Bull. 67, 403–412.
156 Neuron 53, 147–156, January 4, 2007 ª2007 Elsevier Inc.
Neuron
Neural Predictors of Purchases