Content uploaded by Enrique Strelow
Author content
All content in this area was uploaded by Enrique Strelow on Apr 15, 2020
Content may be subject to copyright.
Multiple “buy buttons”in the brain: Forecasting chocolate sales at
point-of-sale based on functional brain activation using fMRI
Simone Kühn
a,
⁎, Enrique Strelow
b
, Jürgen Gallinat
a
a
UniversityClinic Hamburg-Eppendorf, Clinic and Policlinic for Psychiatry and Psychotherapy, Martinistraße 52, 20246 Hamburg, Germany
b
Justus-Liebig University of Gießen, Department of Business Administration and Economics, Licher Straße 74, 35394 Gießen, Germany
abstractarticle info
Article history:
Received 7 December 2015
Revised 5 May 2016
Accepted 6 May 2016
Available online xxxx
We set out to forecast consumer behaviour in a supermarket based on functional magnetic resonance imaging
(fMRI). Data was collected while participants viewed six chocolate bar communications and product pictures
before and after each communication. Then self-reports liking judgement were collected. fMRIdata was extract-
ed from aprioriselected brain regions: nucleusaccumbens, medial orbitofrontalcortex, amygdala, hippocampus,
inferior frontal gyrus, dorsomedial prefrontal cortexassumed to contributepositively and dorsolateral prefrontal
cortex and insula were hypothesized to contribute negatively to sales. The resulting values were rank ordered.
After our fMRI-based forecast an instore test was conducted in a supermarket on n = 63.617 shoppers. Changes
in sales werebest forecasted by fMRI signal during communication viewing, second bestby a comparison of brain
signal during product viewing before and after communication and least by explicit liking judgements.
The results demonstrate the feasibility of applying neuroimaging methods ina relatively small sample to correct-
ly forecast sales changes at point-of-sale.
© 2016 Published by Elsevier Inc.
Keywords:
neuroeconomics
neuromarketing
forecasting sales
reward processing
fMRI
The notion of using neuroscientific methods in marketing research
has become more and more established, as popular science book titles
such as “Neuromarketing: Understanding the Buy Buttons in Your
Customer's Brain”,“Brainfluence: 100 Ways to Persuade and Con-
vince Consumers with Neuromarketing”,“Unconscious Branding:
How Neuroscience Can Empower and Inspire Marketing”suggest.
However, in daily practice of marketers classical explicit (meaning
consciously accessible to the participant) market research tech-
niques, such as focus groups and surveys, are commonly used to pro-
vide answers to questions such as which of three different
advertisements is going to be successful on the market (Ariely and
Berns, 2010). The use of implicit (meaning not necessarily con-
sciously accessible to the participant) neuroscientific methods in
contrast, such as electroencephalography (EEG) focussing on event
related potentials (ERPs) or functional magnetic resonance imaging
(fMRI) measuring the so-called blood oxygen level dependent
(BOLD) effect in the brain seems still quite rare. Instead the common
practice of current neuromarketing approaches relies on deriving
abstract knowledge and principles from neuroscientific studies to
guide consulting. This may originate from the fact that most of the
published studies on neuromarketing are rather academic and do
not lend themselves to immediate application to hands-on market-
ing problems. However, these previous reports may provide a valid
basis for the selection of brain regions that could potentially be
relevant in forecasting of shopper behaviour in response to adver-
tisement –and more precisely perhaps: forecasting actual consumer
decisions at the point-of-sale.
In a seminal experimental study Knutson and colleagues asked par-
ticipants to observe pictures of products in an MRI scanner and provide
manual responses to indicate a purchase decision, at the end of the ex-
perimenttwo trials were selected to count and subjects actually bought
the items (Knutson et al., 2007). The fMRI data indicated that activity in
nucleus accumbens (NAcc) was associated with a preference for the
product, whereas high prices elicited activation in the insula (Ins) and
reduced activity in the medial orbitofrontal cortex (mOFC). Although
the purchase decision was situated in a laboratory context, and was
based on the presentation of different pictures of products only,the de-
cision whether to buy for the suggested price was accompanied by
higher activity in NAccand mOFC and the decision not to buy by Ins ac-
tivity. This implies that when trying to predict purchases Ins activity
should be considered as a prohibiting factor. Most recently, a study
succeeded in forecasting aggregated market-level elasticities of televi-
sion ads from NAcc (ventral striatal) brain activity measured by
means of fMRI (Venkatraman et al., 2015). Likewise real-life success of
people in the request of microloans on the Internet have been forecast-
ed by neural activity in NAcc in response to photographs of the re-
questers (Genevsky and Knutson, 2015). Similarly, we have recently
demonstrated that NAcc activity is also observed when participants
see their favourite brand label and when anticipating the receipt of a
particular Coke drink (Kuhn and Gallinat, 2013). Likewise, willingness
to pay has been shown to be associated with activity in mOFC in an
NeuroImage xxx (2016) xxx–xxx
⁎Corresponding author.
E-mail address: skuehn@uke.de (S. Kühn).
YNIMG-13178; No. of pages: 7; 4C: 3, 4
http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
1053-8119/© 2016 Published by Elsevier Inc.
Contents lists available at ScienceDirect
NeuroImage
journal homepage: www.elsevier.com/locate/ynimg
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
experiment where subjects had to place bids on the right to eat different
foods (Plassmann et al., 2007, 2010). Interestingly, in these willingness
to pay studies the dorsolateral prefrontal cortex (DLPFC) seemed to play
a similar role as the mOFC when participant were explicitly asked to es-
timate how much money they would be willing to spend on food. How-
ever, this may be due to a working memory process elicited by the task
that requires recall of previous bids to put the current bid into perspec-
tive. In particular since a study investigatingthe effects of electoralcam-
paigns on changes in attitude towards the political candidate in
questionrevealed that higher activation of DLPFC leadmore negative at-
titude changes (Kato et al., 2009). In line with the interpretation that
mOFC signals pleasantness (Kuhn and Gallinat, 2012), mOFC activation
during viewing of the electoral campaigns predicted positive attitude
changes towards the candidate. Asimilar pattern of results has beende-
scribed in a study where participants had to make binary choices be-
tween either coffee or beer brands. When one of the shown products
was the market leader this was accompanied by a decrease in DLPFC
and in increase of mOFC activation (Deppe et al., 2005).
Another brain structure that has also previously been implicated in
reward processing, next to its more widely known involvement of pro-
cessing emotions, is the amygdala (Amyg) (Hampton et al., 2007;
Jenison et al., 2011). Anatomically, the Amyg is connected with the
mOFC, which puts it into an ideal position to influence the computation
of values (Price, 2003). While trying to predict the outcome of simulated
political voting judgements, candidates that were voted for more fre-
quently elicited stronger activation in bilateral amygdala when seen
by the voters (Rule et al., 2010). In a very similar study seeing the can-
didates that lost real votes elicited stronger activity in the Ins (Spezio
et al., 2008).
In addition to the above-mentioned brain regions that have quite
frequently been reported to be involved in consumer decisions, we
thought about the potential role of the hippocampus (HC), the inferior
frontal gyrus (IFG) and dorsomedial prefrontal cortex (dmPFC) that
may signal a higher degree of personal involvement during the percep-
tion of communications. The HC is known to be activated when memo-
ries are newly formed (Paller and Wagner, 2002). In a study that
investigated the superior persuasive effects of expert statements the
resulting enhanced memory effects were also ascribed to a higher acti-
vation of the HC (Klucharev et al., 2008). Recently it has been shown
that higher engagement of fronto-temporal regions, including the HC
lead to better remembrance of health messages presented an advertise-
ment context (Seelig et al., 2014).
The IFG has been implicated in a multitude of cognitive functions, in-
cluding speech production and response inhibition, however, the IFG
has likewise been identified as an integral part of the brains action un-
derstanding and imitation network (Kuhn et al., 2013). In monkeys
so-called mirror neurons have been found using single-cell recordings
in area F5 of which the human equivalent is assumed to be situated in
the human pars opercularis of the IFG (Molenberghs et al., 2012).
These mirror neurons are active when an animal acts but also when
an animal observes the very same action performed by someone else
(Molenberghs et al., 2009). It has been speculated that these neurons
are important in the process of understanding actions of other people
and by helping to simulate actions thereby contributing to theory of
mind abilities, namely the capacity to attribute mental states to oneself
and to others (Gallese et al., 2004). Similarly, the dmPFC has been impli-
cated in theory of mind processes and mentalizing (Schurz et al., 2014).
It has been suggested that the role of the dmPFC in metalizing is in
thinkingabout theself and in simulatingmental states for similar others
(Mahy et al., 2014) and may therewith reflect a higher degree of
engagement when activated during perception of advertisement
messages.
The aim of the present study was, to forecast changes of sales of a
highly popular chocolate bar in response to communications placed di-
rectly at point-of-sale using fMRI methods. In order to do so we exposed
a small sample of individuals to communications of that chocolate bar
and measured BOLD signal in the above-mentioned eight ROIs (NAcc,
mOFC, DLPFC, Ins, Amyg, HC, IFG, dmPFC). We used this fMRI signal to
rank order thetested communications according to their potential to in-
crease sales at the population level in a supermarket. In order to com-
pare predictive performance of BOLD signal with subjective report, we
additionally asked participants to rank the ads from least favourite to
most favourite.
Methods
Participants
Eighteen healthy female subjects with a mean age of 39.9 years
(SD = 10.72, range: 23–56 years) participated on the basis of informed
consent. We recruited women exclusively since they are known to be
the typical buyers of chocolate in supermarkets. All womendid indicate
that they regularly buy the product we intended to investigate (3.3
times per week on average, SD = 5.3, range: 1–12). The study was
conducted according to the Declaration of Helsinki, with approval of
the local ethics committee. All subjects had normal or corrected-
to-normal vision and normal hearing. No subject had a history of
neurological, major medical, or psychiatric disorder. All participants
were right-handed as assessed by the Edinburgh handedness question-
naire (Oldfield, 1971).
fMRI Procedure
During the fMRI experiment, participants were presented with
different pictures including a product picture as well as six different
communications that were generated by the company (the toothbrush
picture was intended as a control communication). They were
instructed to watch the pictures carefully. Each trial started with a
fixation cross presented for a jittered duration of 6–8 s (steps of
500 ms), then a product picture was presented for 2 s (see Fig. 1A for
a schematic overview) and after another fixation cross presented for
the jittered duration of 6–8 s (steps of 500 ms), a communication was
shown for 3 s (Fig. 1B). Finally, after a third fixation cross presentation
(jittered duration of 6–8 s in steps of 500 ms) the product picture was
shown again for 2 s. Taken together each trial had an average length
of 28 s. The experiment consisted of 2 runs, each containing 36 trials,
with an overall duration of approximately 17 min per run. Therewith
each of the six communications was shown 12 times across the experi-
ment. The sequence of stimuli was randomized and we controlled for
transition probabilities of events.
After the fMRI session participants were interviewed and asked to
order the communications according to their liking, while seeing all
six in front of them. Before and after scanning participants were asked
how hungry they felt and how strong their craving for sweets was at
that particular moment in time. Participants responded using a scale
ranging from 1 “notatall”to 8 “very much”.
Scanning Procedure
Images were collected on a 1.5 T Avanto MRI scanner system
(Siemens Medical Systems, Erlangen, Germany) using a 20-channel
head coil. First, high-resolution anatomical images were acquired using
a three-dimensional T1-weighted magnetization prepared gradient-
echo sequence (MPRAGE) based on the ADNI protocol (www.adni-info.
org); repetition time = 2.560 s, echo time = 5.05 ms, flip angle =7°;
256 × 256 × 192 matrix, resolution 1x1x1 mm
3
voxel size. Whole
brain functional images were collected on the same scanner using a
T2*-weighted EPI sequence sensitive to BOLD contrast using sparse sam-
pling (TR = 2000 ms, TE = 35 ms, image matrix = 64 × 64, FOV =
192 mm, flip angle =80°, slice thickness = 3.5 mm, 29 near-axial slices,
aligned with the AC/PC line using parallel imaging implemented in
GRAPPA).
2S. Kühn et al. / NeuroImage xxx (2016) xxx–xxx
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
fMRI Data Pre-processing and Main Analysis
The fMRI data were analysed using SPM8 software (Wellcome
Department of Cognitive Neurology, London, UK). The first 4 volumes
of all EPI series were excluded from the analysis to allow the
magnetisation to reach a dynamic equilibrium. Data processing started
with slice time correction and realignment of the EPI datasets. A mean
image for all EPI volumes was created, to which individual volumes
were spatially realigned by means of rigid body transformations. The
structural image was co-registered with the mean image of the EPI se-
ries. Then the structural image was normalised to the Montreal Neuro-
logical Institute (MNI) template for the random effects analysis. The
normalisation parameters were then applied tothe EPI images to ensure
an anatomically informed normalisation. A commonly applied filter of
8 mm FWHM (full-width at half maximum) was used. Low-frequency
drifts in the time domain were removed by modelling the time series
for each voxel by a set of discrete cosine functions to which a cut-off
of 128 s was applied. The statistical analyses were performed using
the general linear model (GLM). We modelled each picture (product
and presentation of thecommunication) by means of a separate regres-
sor. These vectors were convolved with a canonical haemodynamic re-
sponse function (HRF)and its temporal derivativesto form regressors in
a design matrix. Furthermore, six movement regressors were entered
into the GLM. The parameters of the resulting general linear model
were estimated. We built contrasts between the six different
communication regressors and the implicit baseline as well as between
product presentation after the six different communications compared
to productpresentations before the communications. We extracted per-
cent signal change values in anatomically pre-defined ROIs from these
contrasts: bilateral insula (AAL atlas, (Tzourio-Mazoyer et al., 2002)),
bilateral amygdala (AAL atlas), bilateral medial orbitofrontal cortex
(AAL atlas: frontal middle orbital and gyrus rectus), bilateral hippocam-
pus (AAL atlas), bilateral inferior frontal gyrus (AAL atlas), bilateral
dorsomedial prefrontal cortex (AAL atlas: middle superior frontal),
bilateral dorsolateral prefrontal cortex (Brodmann area 9 and 46) and
bilateral accumbens (Harvard-Oxford subcortical atlas as used in FSL).
For the extraction we used the Matlab scripts from MarsBaR (http://
marsbar.sourceforge.net/). Then we computed an average forecast
value by summarizing the BOLD signal in the following way:
fMRI sales forecast value ¼NAcc2þmOFC2þAmyg þHC þIFG
þdmPFC−DLPFC−Ins
WeighingNAcc and mOFC more strongly by multiplying it times two
to account for the stronger and more consistent literature-based
evidence for the involvement of these two brain regions in purchase
decisions (Knutson et al., 2007; Plassmann et al., 2007, 2008, 2010;
Schaefer et al., 2011; Schaefer and Rotte, 2007). We did this separately
Fig. 1. (A) Schematic drawing of the experimental paradigm, (B) Depiction of the six different communicationsused in the present study(top to bottom: woman, couple, hands without
text, hands with text, group, toothbrush), (C) Photograph of the quarter palette placement at point-of-sale in the supermarket.
3S. Kühn et al. / NeuroImage xxx (2016) xxx–xxx
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
for the BOLD signal during communications and the BOLD change from
product presentation before and after the communication.
Validation of fMRI findings in a supermarket
After the fMRI data was acquired and the data analysis and there-
with the forecast of sales was finished, the pretested communications
were tested at a point-of-sale of the product in a German supermarket.
The communications were placed behind a quarter palette of the prod-
uct in the direct neighbourhood and in addition to the regular product
placement in the shelf (Fig. 1C). Six weeks were selected during the
year that were not influenced by festive periods, holiday seasons or
the fact that the product was on special offer and assessed sales of the
product from the quarter palette on each day allowing a direct compar-
ison of the sales between the different communications tested. The
communications were tested in the following order: group, couple,
hands, hands with text, woman and toothbrush each for one week.
The presentation of theproduct on a quarter palette was nothing entire-
ly new for thecustomers since this is a common additional placement in
supermarkets in Germany. Therefore we think we can exclude that the
first time exposure lead to a significant increase in sales. Due to the fact
that we selected comparable weeks some placements were presented
weeks apart, which may further reduce order effects. Since the choco-
late bar is so popular in Germany we do not think that customers are
likely to have tried the product the first time within our acquisition pe-
riod in the supermarket. However future studies may consider to re-
peatedly and randomly assess sales in response to communications in
order to exclude effects of the order of presentation.
Post-hoc comparison of fMRI ROIs and their association with sales
After the sales data was acquired we computed post-hoc Spearman
rho correlation coefficients to explore the association between the sep-
arate ROIsand different models of ROI combination and sales. We would
like to point out that correlationsderived from a sample of n = 6 may be
highly error-prone, however they may offer a way to compare the value
of different ROIs against one another in forecasting sales.
Results
Interview results
When participants were explicitly asked for their opinion on which
communication they liked best five participants chose thehands without
text (Fig. 1B middle left), four chose the toothbrush (Fig. 1B lower right),
three chose the couple (Fig. 1B upper right) or the group (Fig. 1Blower
left), two chose the woman (Fig. 1B upper left) and one chose the
hands with text (Fig. 1Bmiddleright)(Fig. 2A).
fMRI results based on our a priori hypothesis
We extracted BOLD signal from ROIs in NAcc, mOFC, Amyg, DLPFC,
Ins, HC, IFG and dmPFC. Based on our a-priori hypotheses and derived
from previous literature of reward processing we added the signal of
all brain regions except for DLPFC and Ins, which we subtracted, and
weighted mOFC and NAcc by inserting it twice in the formula. The
resulting activation scores were not related to pre- or post-test hunger
or sweet craving ratings (pN0.12; hunger before scan: 3.25 (SD =
1.96), hunger after scan: 4.69 (2.12), sweet craving before scan: 4.83
(1.98), sweet craving after scan: 5.06 (2.31)). Based on this we comput-
ed the average fMRI-derived sales forecast across participants and rank
ordered the resulting values.
First we did this for the BOLD signal measured during the presenta-
tion of the communication where the group ranked first, the woman
ranked second, the toothbrush ranked third, the couple ranked fourth,
the hands without text ranked fifth, and the hands with text ranked last
(Fig. 2B).
Second we repeated this analysis but this time extracting BOLD
signal measured during the presentation of the product. In order to
evaluate changes caused by showing the communication in between
two product presentations we subtracted BOLD signal during prod-
uct viewing after from the same signal before viewing the communi-
cation (fMRI sales forecast value during product viewing after
communication minus fMRI sales forecast value during product
viewing before communication). These average fMRI-derived sales
forecast change scores were again rank ordered and resulted in the
Fig. 2. Rankingof the six communications based on (A) the explicit judgementof the participants,(B) BOLD signal extracted from eight regions of interest andcomputed by means of our a
priori proposed fMRI-derivedsales prediction value (=NAcc*2 + mOFC*2+ Amyg + HC + IFG+ dmPFC -DLPFC-Ins),(C) BOLD signal changefrom seeing the product after compared to
before thecommunication based on the proposed sales prediction value,(D) behaviouraldata from a field study measuring actual product salewhen the product was offered on a quarter
palette with the corresponding communication in the back.
4S. Kühn et al. / NeuroImage xxx (2016) xxx–xxx
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
following order: the couple ranked first, followed by the group,the
woman,hands with text,toothbrush and then the hands without text
(Fig. 2C).
Actual sales in supermarket
In a field study we measured actual sales of the product at point-of-
sale on a quarter palette, each communication over a period ofone week
devoid of holidays, festivities, or special offers of the product. The best
selling communication was the group with 59 products from 10.318
customers in that particular week(each 175th shopper bought), follow-
ed by the woman with 57 products when 10.442 visited the market
(each 183rd shopper bought). Third ranked was the couple with 52
sold products from 10.666 customers (each 205th shopper bought),
then the toothbrush with 53 of 10.908 shoppers (each 206th shopper
bought), the hands with text with 51 or 10.764 customers (each 211st
shopper bought) and finally the hands without text with 45 products
sold when 10.519 visited the market (each 234th shopper bought)
(Fig. 2D).
Post-hoc comparison of fMRI ROIs their potential to forecast sales
In order to explore the potential value of the separate ROIs in fore-
casting sales we ran post-hoc correlations between the ROIs as well as
different alternative combinations of the ROIs and sales. The values
from our a priori hypothesized weighted formula resulted in a high cor-
relation coefficient (r
s
= 0.94, p = 0.005, see Fig. 3). When considering
each ROI separately the following associations were observed: NAcc:
r
s
= 0.43, p = 0.40, mOFC: r
s
= 0.94, p = 0.005, DLPFC: r
s
=0.54,
p=0.27,Ins:r
s
= 0.60, p = 0.21, Amyg: r
s
= 0.14, p = 0.79, HC:
r
s
= 0.66, p = 0.16, IFG: r
s
= 0.37, p = 0.47, dmPFC: r
s
= 0.09, p =
0.87. The results show that the mOFC is by far the most predictive ROI
when considered in isolation. When the formula contains all a priori se-
lected ROIs but without the weighting, the correlation coefficient
dropped considerably (r
s
= 0.37, p = 0.47). When the weighting was
only applied to mOFC, not to NAcc, the association likewise dropped
(r
s
= 0.89, p = 0.019), indicating that the NAcc does play a crucial
role, at least when considering more than one ROI at the same time.
Interestingly, when running a linear regression model, which unfor-
tunately could not be fully estimated due to the small sample size, the
enter, forward and backward method include mOFC (standardized
beta coefficient 0.81), Ins (−0.54), HC (0.52), dmPFC (0.27), NAcc
(0.21) but exclude the ROIs DLPFC, Amyg, and IFG. In order to explore
the role of the Ins in more depth, since its contribution seemed to be
positive in the individual bivariate correlation analyses but negative in
the regression analyses and we likewise assumed its influence on sales
to be negative by subtracting its signal from the other brain regions ac-
tivity we computed our formula once without subtractingIns (r
s
= 0.89,
p = 0.019) and once with adding instead of subtracting Ins (r
s
=0.77,
p=0.072).
Discussion
Within the scope of the present study we were able to forecast
changes in sales of a popular chocolate bar by testing six different com-
munications placed in a supermarket by means of fMRI. We chose a par-
ticularly well-known chocolate bar brand in order to ensure that the
customers in the supermarket were well acquainted with it and have
probably all tasted and most likely bought the product before. We ex-
posed a small sample of individuals to the communications and to a
product picture presented before and after the communication while
situated in an MRI scanner and quantified BOLD signal in eight apriori
defined ROIs (NAcc, mOFC, DLPFC, Ins, Amyg, HC, IFG, dmPFC). After
scanning we asked participants explicitly for a ranking according to
their liking of the different communications. Then we tested each of
the communications at point-of-sale in a supermarket (one week
each) and counted sales of the chocolate bar on a population level. Actu-
al sales (Fig. 2D) were best forecasted by a ranking of BOLD signal com-
puted across the apriorisetofROIsduringtheperceptionofthe
communication (Fig. 2B). The first and the second rank were forecasted
correctly then 3rd and 4th rank were exchanged as well as 5th and 6th.
The change in BOLD signal during product viewing after compared to
before the communication (Fig. 2C) was shown forecasted the actual
behaviour (Fig. 2D) second best and predictability from participants
self-report turned out to perform the worst (Fig. 2A).
The unique feature of the present study is, that this is the first study
forecasting actual sales in a supermarket on a population level from
neuroimaging data of a relatively small and feasible sample of partici-
pants. Furthermore the ROIs were selected aprioribased on the
neuromarketing literature that is currently available. In these respects
the present study goes beyond most previously published studies in
neuromarketing that successfully tackle brain regions that are related
to certain cognitive and affective processes during product viewing or
simulated shopping situations (Knutson et al., 2007; Plassmann et al.,
2008; Schaefer et al., 2006), but do not actually forecast consumer be-
haviour extending out of the sample to a supermarket environment.
Moreover it is rarely the case that predictive studies actually formulate
a forecast first and test this a priori forecast independently of the out-
come. Even advanced machine learning algorithms –that have by
now also been used in a neuromarketing context (Calvert and
Brammer, 2012; Smith et al., 2014; Tusche et al., 2010)–need to take
the actual outcome, e.g. the simulated decision to buy a product, into
account before being able to establish an algorithm that can forecast be-
haviour. An independent test of the resulting algorithm on a completely
new set of data includingnew participants is however almost never un-
dertaken. A frequently observed phenomenon in machine learning
studies is, that the respective algorithm forecasts behaviour very well
under the exact conditions under which the training data set had been
acquired, but does not generalize to novel experimental setups
(Pereira et al., 2009). Our a priori literature-based ROI selection may
be at an advantage here because it is less selective and less targeted to
our particular data set and therewith may provide a higher potential
to generalize across different domains and therewith perform better
in a context of different types of communications and products. Future
studies should be undertaken to demonstrate generalizability across
Fig. 3. Scatterplot depicting the post-hoc association between our a priori proposed fMRI-
derived sales pr ediction valu e (=NA cc*2 + mOFC*2 + Amyg + HC + IFG + dmPFC
-DLPFC-Ins), and actual product sale expressed in percent of the customers that bought
the product on the communication.
5S. Kühn et al. / NeuroImage xxx (2016) xxx–xxx
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
different consumer goods and test the value of the proposed formula in
other contexts.
To our knowledge only two studies have used a similar design to the
one used in the present study. One predicted behaviour in the health
sector (Falk et al., 2012). Smokers saw anti-smoking television cam-
paigns in an fMRI scanner and from this data the study predicted the
call volume of a line taking phone calls from smokers needing help
with quitting, measured after each of the presented campaign was pub-
licly launched. The prediction was based on a single a priori defined ROI
consisting of a small subregion of the medial prefrontal cortex in
Brodmann's area 10, which would correspond to a subsegment of our
mOFC ROI. This ROI was the result from a previous study of the authors
in which they investigated persuasion of messages on the need to use
sunscreens to avoid cancer (Falk et al., 2010). The reported results of
the prediction of smokers' help-seeking behaviour point into a similar
direction as our present findings on shopper behaviour, namely that
self-report judgements are not predictive of actual behaviour on the
population level, whereas brain activity during viewing of the com-
munication is. The second study compared six methods used in
neuromarketing research namely self-reports, implicit measures,
eye tracking, biometrics, electroencephalography, and fMRI and
found that NAcc activity in fMRI best predicted advertising elasticity
(Venkatraman et al., 2015). The present study is very similar to the
second study, however in contrast we prospectively forecasted
real-life sales in a supermarket with communications that have
never been previously shown.
In post-hoc analyses we focussed on the predictive value of thesingle
ROIs that we summarized in our formula. Albeit our mOFC was consid-
erably larger than the ROI used by Falk et al. (2010) it likewise predicted
the consumer behaviour fairly well, exactly to the same degree as the
result of our more complex ROI formula. Also NAcc seemed to play an
important role in our proposed formula since removing it, decreased
the formulas predictive value, even though the selective predictive
value of NAcc was fairly low. We hypothesize that it is preferable to
take a combination of larger brain regions into account, that are
known to play an important role in approach behaviour in order to
avoid overgeneralization of single results. However, future research is
needed to confirm this hypothesis. Interestingly, none of the other
ROIs that we included in ourformula were predictive of later supermar-
ket sales when looked at individually. However, combination the
information of the individual ROIs was highly predictive of sales. Re-
markably, although Ins on its own was positively (but insignificantly)
associated with sales when looking at the bivariate correlation of both
variables, adding instead of subtracting Ins in our formula worsened
the predictive value to insignificance. Since excluding Ins from the for-
mula altogether also lead to a drop in predictive value we conclude
that Ins indeed contributed negatively to the prediction of product
sales as we hypothesized before data acquisition in the supermarket.
At first sight it might seem odd that the forecast is actually better
outside thetested sample (in the supermarket) than within the sample
when taking the self-report data of the very same subject into account.
However, many previous studies on consumer behaviour have indicat-
ed that consumer preferences are frequently constructed in the mo-
ment people are being asked and are often based on consumers
actions, which means that the causal path is not that preferences deter-
mine behaviour, but instead behaviour determines preferences (Ariely
and Norton, 2008). When consumers were for example asked to choose
jam or tea after tasting two differentkinds they were later on oftentimes
not able to tell when the product was switched unbeknownst of them.
In these cases they were giving reasons for choosing a product that
they had actually never chosen (Hall et al., 2010). The present data
goes beyond this notion and suggests that fMRI is not necessarily a
good predictor for the verbalizable preference judgements but rather
for consumers' actual purchase actions in the supermarket.
To summarize the present findings,wewereabletoforecastchoco-
late bar sales in a supermarket based on BOLD fMRI signal during
viewing of communications that were later on placed behind a quarter
palette at point-of-sale. The ROIs consisting of NAcc, mOFC, Amyg, HC,
IFG, dmPFC that were assumed to contribute positively to later sales
and DLPFC and Ins that were predicted to contribute negatively to
sales, were selected aprioriand based on prior neuroimaging literature
therefore uninfluenced by the resulting sales changes acquired in a Ger-
man supermarket. Interestingly, the forecast, a rank order based on the
signal during communication viewing, was superior to the forecast
based on the changes during product viewing after as compared with
before the communication was shown. However both were more accu-
rate than the explicit self-report judgement of the participants. Post-hoc
correlation and regression analyses show that mOFC reaches a similar
predictive value as our formula, but we speculate that the combination
of multiple ROIs may lead to forecasts that generalize to new datasets.
Future research is needed to test this explicitly. The present results
demonstrate the feasibility to use neuroimaging methods in a relatively
small sample of participants to forecast the influence of communica-
tions on the actual consumer behaviour at the point-of-sale.
References
Ariely, D., Berns, G.S. , 2010. Neuromarketing: the hope and hype of neuroimaging in
business. Nat. Rev. Neurosci. 11, 284–292.
Ariely, D., Norton, M.I., 2008. How actions create–not just reveal –preferences. Trends
Cogn. Sci. 12, 13–16.
Calvert, G.A., Brammer, M.J., 2012. Predicting consumer behavior: using novel mind-
reading approaches. IEEE Pulse 3, 38–41.
Deppe, M., Schwindt, W., Kugel, H., Plassmann, H., Kenning,P., 2005. Nonlinear responses
within the medial prefrontal cortex reveal when specific implicit information influ-
ences economic decision making. J. Neuroimaging 15, 171–182.
Falk, E.B., Berkman, E.T., Mann, T., Harrison, B., Lieberman, M. D., 2010. Predi cting
persuasion-induced behavior change from the brain. J. Neurosci. 30, 8421–8424.
Falk, E.B., Berkman, E.T., Lieberman, M.D., 2012. From neural responses to population be-
havior: neural focus group predicts population-level media effects. Psychol. Sci. 23,
439–445.
Gallese, V., Keysers, C., Rizzolatti, G., 2004. A unifying view of the basis of social cognition.
Trends Cogn. Sci. 8, 396–403.
Genevsky, A., Knu tson, B., 2015. Neural Affective Mechanis ms Predict Market-Level
Microlending.Psychol.Sci.26,1411–1422.
Hall, L., Johansson, P., Tarning, B., Sikstrom, S., Deutgen, T., 2010. Magic at the market-
place: Choice blindness for the taste of jam and the smell of te a. Cognition 117,
54–61.
Hampton, A.N., Adolphs, R., Tyszka, M.J., O'Doherty, J.P., 2007. Contributions of the amyg-
dala to reward expectancy and choicesignals in human prefrontal cortex. Neuron55,
545–555.
Jenison, R.L., Rangel, A., Oya, H., Kawasaki, H., Howard,M.A., 2011. Value encoding in sin-
gle neurons in the human amygdala during decision making. J. Neurosci. 31,
331–338.
Kato, J., Ide, H., Kabashima, I., Kadota,H., Takano, K., Kansaku, K., 2009.Neural correlatesof
attitude change following posi tive and negativ e advertisemen ts. Front. Behav .
Neurosci. 3, 6.
Klucharev, V., Smidts, A., Fernandez, G., 2008. Brain mechanisms of persuasion: how 'ex-
pert power'modulates memoryand attitudes. Soc.Cogn. Affect. Neurosci. 3, 353–366.
Knutson, B., Rick, S., Wimmer, G.E., Prelec, D., Loewenstein, G., 2007. Neural predictors of
purchases. Neuron 53, 147–156.
Kuhn, S., Gallinat, J., 2012. The neural correlates of subjective pleasantness. NeuroImage
61, 289–294.
Kuhn, S., Gallinat, J., 2013. Does taste matter? How anticipation of cola brands influences
gustatory processing in the brain. PLoS One 8, e61569.
Kuhn, S., Brass, M., Gallinat, J., 2013. Imitation and speech: commonalities within Broca's
area. Brain Struct. Funct. 218, 1419–1427.
Mahy, C.E., Moses, L.J., Pfeifer, J.H., 2014. How and where: theory-of-mind in the brain.
Dev. Cogn. Neurosci. 9, 68–81.
Molenberghs, P., Cunnington, R., Mattingley, J.B., 2009. Is the mirror neuron system in-
volved in imitation? A short review and meta-analysis. Neurosci. Biobehav. Rev. 33,
975–980.
Molenberghs, P., Cunnington, R., Mattingley, J.B., 2012. Brain regions with mirror properties:
a meta-analysis of 125 human fMRI studies. Neurosci. Biobehav. Rev. 36, 341–349.
Oldfield,R.C., 1971. The assessment and analysis of handedness: the Edinburgh inventory.
Neuropsychologia 9, 97–113.
Paller,K.A.,Wagner,A.D.,2002.Observingthe transformationof experience into memory.
Trends Cogn. Sci. 6, 93–102.
Pereira, F., Mitchell, T., Botvinick, M., 2009. Machine learning classifiers and fMRI: a tuto-
rial overview. NeuroImage 45, S199–S209.
Plassmann, H., O'Doherty, J., Rangel, A., 2007. Orbitofrontal cortex encodes willingness to
pay in everyday economic transactions. J. Neurosci. 27, 9984–9988.
Plassmann, H., O'Doherty, J., Shiv, B., Rangel, A., 2008. Marketing actions can modulate
neural representation s of experienced pleasant ness. Proc. Natl. Acad. Sci. U. S. A.
105, 1050–1054.
6S. Kühn et al. / NeuroImage xxx (2016) xxx–xxx
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
Plassmann, H., O'Doherty, J.P. , Rangel, A., 201 0. Appetitive a nd aversive goal values
are encoded in the medial orbitofrontal cortex at the time of decision making.
J. Neurosci. 30, 10799–10808.
Price, J.L., 2003. Comparative aspects of amygdala connectivity. Ann. N. Y. Acad. Sci. 985,
50–58.
Rule, N.O., Freeman, J.B., Moran, J.M., Gabrieli, J.D., Adams Jr., R.B., Ambady, N., 2010. Vot-
ing behavior is reflected in amygdala response across cultures. Soc. Cogn. Affect.
Neurosci. 5, 349–355.
Schaefer, M., Rotte, M., 2007. Favorite brands as cultural objects modulate reward circuit.
Neuroreport 18, 141–145.
Schaefer, M., Berens, H., Heinze, H.J., Rotte, M., 2006. Neural correlates of culturally famil-
iar brands of car manufacturers. NeuroImage 31, 861–865.
Schaefer,M., Knuth, M., Rumpel,F., 2011. Striatalresponse to favoritebrands as a function
of neuroticism and extraversion. Brain Res. 1425, 83–89.
Schurz, M., Radua, J., Aichhorn, M., Richlan, F., Perner, J., 2014. Fractionating theory of
mind: a meta-analysis of functional brain imaging studies. Neurosci. Biobehav. Rev.
42, 9–34.
Seelig, D.,Wang, A.L., Jagannathan, K., Loughead, J.W., Blady,S.J., Childress, A.R., Romer, D.,
Langleben, D.D., 2014. Low message sensation health promotion videos are better
remembered and ac tivate areas of the brain associated with memory encod ing.
PLoS One 9, e113256.
Smith, A., Bernheim, B.D., Camerer, C., Rangel, A., 2014. Neural Activity Reveals Prefer-
ences Without Choices. Am. Econ. J. Microecon. 6, 1–36.
Spezio, M.L., Rangel, A., Alvarez, R. M., O'Doherty, J.P., Mattes, K., Todo rov, A., Kim, H.,
Adolphs, R., 2008. A neural basis for the effect of candidate appearance on election
outcomes. Soc. Cogn. Affect. Neurosci. 3, 344–352.
Tusche, A., Bode, S., Haynes, J.D., 2010. Neural responses to unattended products predict
later consumer choices. J. Neurosci. 30, 8024–8031.
Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N.,
Mazoyer, B., Joliot, M., 2002. Automated anatomical labeling of activations in SPM
using a macroscopic anatomical parcellation of the MNI MRI single-subject brain.
NeuroImage 15, 273–289.
Venkatraman, V., Dimoka, A., Pavlou, P.A., Vo, K., Hampton, W., Bollinger, B., Hershfield,
H.E., Ishihara, M., Winer,R.S., 2015. PredictingAdvertising SuccessBeyond Traditional
Measures: New Insights from Neuro physiological Methods and Mark et Response
Modeling. J. Mark. Res. 52, 436–452.
7S. Kühn et al. / NeuroImage xxx (2016) xxx–xxx
Please citethis article as: Kühn, S., et al., Multiple “buy buttons”in the brain: Forecasting chocolate sales at point-of-sale based on functionalbrain
activation using f, NeuroImage (2016), http://dx.doi.org/10.1016/j.neuroimage.2016.05.021
