© 2015, American Marketing Association  Journal of Marketing Research  PrePrint, Unedited  All rights reserved. Cannot be reprinted without the express  permission of the American Marketing Association. 

Using EEG to predict consumers’ future choices Ariel Telpaz1, Ryan Webb2, and Dino J Levy3, 4

Ariel Telpaz, PhD: Affiliation: 1 Faculty of Industrial engineering and management, Technion - Israel Institute of Technology Address: Technion City Haifa, 32000, Israel Tel: 972-546969568 Email: [email protected]

Ryan Webb, PhD: Affiliation: Rotman School of Management, University of Toronto Address: 105 St. George St, Toronto, Ontario, Canada, M5S3E6 Tel: 4169784418 Email: [email protected]

Dino J Levy, PhD: Affiliation: 3 Marketing Department, Recanati Business School and 4Sagol School of Neuroscience, Tel-Aviv University Address: 55 Haim Levanon st. Tel Aviv University, Ramat Aviv 69978, Israel Tel: 972-3--6409565 Email: [email protected]



2

Using EEG to predict consumers’ future choices It is now established that neural imaging technology can predict preferences over consumer products. However the applicability of this method to consumer marketing research remains in question, partly because of the expense required. In this article, we demonstrate that neural measurements made with a relatively low-cost and widely available measurement method — Electroencephalogram (EEG) — can predict future choices over consumer products. In our experiment, subjects viewed individual consumer products in isolation, without making any actual choices, while we measured their neural activity with EEG. After these measurements were taken, subjects then made choices between pairs of the same products. We find that neural activity measured from a mid-frontal electrode displays an increase in the N200 component and a weaker theta band power that correlates with a more preferred good. Using state-of-the-art techniques for relating neural measurements to choice prediction, we demonstrate that these measures predict subsequent choices. Moreover, the accuracy of prediction depends on both the ordinal and cardinal distance of the EEG data: the larger the difference in EEG activity between two goods, the better the predictive accuracy. Keywords: EEG, Choice prediction, Consumer Neuroscience, Theta power, N200



3

INTRODUCTION

Over the past 15 years, our understanding of the neuroscience underlying decision-making has advanced rapidly (see Glimcher 2011; Glimcher and Fehr 2013 for reviews), raising hopes that measurements of neural activity — and a deeper understanding of neural mechanisms — can be applied to marketing research. Two promising avenues for such a contribution have been previously identified (Ariely and Berns 2010). First, there is the possibility that insight from neuroscience might improve the marketing message for existing products. Second, there is the possibility that neuroscience can provide insight into how products are valued before they even exist in the marketplace, improving product design.

Both of these avenues rely on the proposal that neuroscience will reveal information about consumer preference that is unobtainable through conventional methods. There is certainly room for improvement. Previous studies have demonstrated that different preference elicitation methods can result in different subject responses (Buchanan and Henderson 1992; Day 1975; Griffin and Hauser 1993; McDaniel et al. 1985). The use of questionnaires for evaluating consumers’ preferences, attitudes, and purchase intent can result in a biased or inaccurate result (Fisher 1993; Neeley and Cronley 2004). A verbal statement of preferences can also generate conscious or unconscious biases. In some cases, consumers decline to state their actual preferences (for reasons such as discretion or shame), and in other cases consumers cannot verbalize a justification for their preferences (Johansson et al. 2006; Nisbett and Wilson 1977).



4 It can also be difficult (or sometime impossible) to directly elicit a consumer’s preferences

through choices. This may arise due to high product cost, ethical considerations, or the fact that the product does not yet exist. This forces the marketer to examine hypothetical choices with hypothetical rewards, yielding a potential bias in which responses are overstated compared to incentive-compatible choices (Blumenschein et al. 2008; Cummings et al. 1995; Johannesson et al. 1998; List and Gallet 2001; Murphy et al. 2005) or plans (Ariely and Wertenbroch 2002; O'Donoghue and Rabin 2008; Tanner and Carlson 2009). These results are bolstered by neuroscientific evidence suggesting variations in value computations between real and hypothetical choice situations (Kang and Camerer 2013; Kang et al. 2011).

Since the marketing message in many campaigns is presented with the hope that it will affect consumers’ preferences, attitudes, and/or actual purchases sometime in the future, all the factors above confound the task of evaluating consumer preferences and limit the ability to predict choice at the time of the purchasing decision. Therefore, finding a cost-effective tool that can predict consumers’ future behavior in response to marketing messages and forecast future preferences over novel goods will be beneficial in consumer marketing applications.

Substantial recent progress directly addresses these two avenues for neuroscientific methods in marketing research. Evidence from functional magnetic resonance imaging (fMRI) suggests that the same brain areas that represent values in a choice situation – primarily the medial prefrontal cortex (mPFC) and striatum (for three recent meta-studies, see Bartra et al. 2013; Clithero et al. 2009; Levy and Glimcher 2012) - also represent values when subjects are evaluating individual goods in the absence of choice behavior (Falk et al. 2012; Lebreton et al.



5

2009; Levy et al. 2011; Smith et al. 2014; Tusche et al. 2010).1 The magnitude of these signals correlate with the trial-by-trial likelihood that a consumer will choose a particular good, and can be used to predict subsequent choices using a fully cardinal choice model referred to as the Neural Random Utility Model (Webb et al. 2013). This model extends the choice prediction results of the familiar Random Utility framework (Becker et al. 1964; McFadden 1973) to neural measurements, with the important distinction that there are no unobservable latent variables. In doing so, it characterizes neural sources of the stochasticity observed in choice behavior (Huettel and Payne 2009; Yoon et al. 2009) and improves upon choice prediction results.

These results are in line with many studies demonstrating that activity in the mPFC and striatum correlate with various value-related attributes, and correlate with known methods for estimating the values subjects place on choice objects - ranging from consumable goods, to money lotteries, charitable donations, durable goods, social preferences, and political preferences (for reviews see: Bartra et al. 2013; Grabenhorst and Rolls 2011; Kable and Glimcher 2009; Levy and Glimcher 2012; Padoa-Schioppa 2011; Platt and Huettel 2008; Rushworth 2008). Importantly, these same areas are also active for the valuation of novel goods that the consumer has never before experienced (Barron et al. 2013).

However the applicability of these findings to consumer marketing research remains in question, with the current cost of obtaining and operating an fMRI scanner preventing their broad application. Most prominently, an fMRI scanner has a very large fixed cost component. It

1

We note that each of these “non-choice” studies also find activity in other areas, varying from the dorsomedial prefrontal cortex (dmPFC), the insula, the anterior and posterior cingulate cortex (ACC, PCC), hippocampus, and parietal cortex. However the mPFC and Striatum are the only regions common across these studies, and the only regions identified in the meta-studies referenced above (which include the “non-choice” studies).



6

is expensive to purchase (~$1M-$2M), expensive to keep operational ($100K-$150K for insurance, maintenance, and support staff), expensive to locate (requiring a customized room/building), and immobile. Compared to the fixed-cost component, the marginal cost of running an fMRI experiment is relatively low, but still on the order of $500 per experiment. These relatively high costs severely limit the use of fMRI in both academic and commercial applications.

There are also technical limitations to fMRI, primarily a relatively low temporal resolution on the order of 2 seconds (Huettel et al. 2004). This resolution makes it difficult to examine the rapid dynamics of neural signals that are relevant for the neural mechanisms underlying value representation. A faster sampling rate might convey predictive information for consumers’ valuation and choice, information that is blurred by fMRI. For instance, consumers can make decisions for consumable goods in as little as a third of a second (Milosavljevic et al. 2011). It may well be the case that a particular, rapid, component of the neural signal has more indicative and predictive power for consumers’ preferences than the more global signal of fMRI.

To address these concerns, in our study we use an alternative neuroscientific tool called the electroencephalogram (EEG). From a fixed cost standpoint, EEG is orders of magnitude cheaper than fMRI (roughly $50K compared to $1-$2M per unit), requires little support and maintenance, and is widely available in neuroscience laboratories. The marginal cost of running an EEG experiment is only a few dollars, more than an order of magnitude cheaper than an fMRI experiment. From a technical standpoint, EEG also has a very high sampling rate (on the order of 1-2ms, Luck 2005) which enables identification of very fast changes in the neural signal over



7

short time scales (on the order of 50ms, Luck 2005) that may carry strong predictive information about consumer preferences and choice behavior.

In this article, we rigorously examine if EEG measurements of neural activity — recorded while subjects view individual consumer goods on a computer screen without making any choices — can be used to predict both rank-ordered preference ratings and actual choices in a subsequent behavioral choice task. We demonstrate that this is indeed the case. We show that specific spatial and temporal components of the EEG signal correlate with subjects’ future rankordered preferences and can be used to predict subsequent choices. To our knowledge, this is the first EEG study to demonstrate a basic principle: we can use measured neural activations to predict choices without the need to ask consumers anything.

LITERATURE REVIEW

Link Between EEG Recordings and Valuation and Choice There have been several studies linking EEG activity with some aspect of consumer preferences. One of the first studies that used EEG data in consumer research, conducted by Ambler et al. 2004), demonstrated a link between EEG activity in the parietal cortex and the familiarity rating of a good. Evidence for hemispheric asymmetry in the EEG signal and preferences has also been uncovered. Subjects with greater resting activity in left-frontal electrodes (as reflected by lower power within the alpha EEG band, 8–13 Hz) selected more pleasant stimuli in a subsequent behavioral task compared to subjects with greater resting activity in right-frontal electrodes (Sutton and Davidson 2000). A related study also examined the relationship between



8

hemispheric asymmetry in the EEG signal and an aspect of preferences (namely risk aversion), but in the absence of choice (Gianotti et al. 2009). In the initial phase of this experiment, subjects sat quietly in a room while their baseline or “tonic” neural activity was measured. After the measurements, subjects then engaged in a behavioral task to elicit their preferences for risk. Gianotti and colleagues found that higher tonic activity in the right prefrontal cortex, measured before the behavioral task, correlated with a higher level of risk aversion (an avoidance-related behavior) as measured in the behavioral task. Importantly, this study demonstrated that EEG activity, measured in the absence of choice behavior, can be used to predict a preference trait.

Several studies have also demonstrated that EEG activity, measured concurrently with choice, is related to choice behavior. For instance, both gamma band (20–45 Hz) and alpha band (8-13 Hz) oscillations were correlated with subjects’ choices of consumer goods in specific time epochs and brain locations (Braeutigam et al. 2004). In a more recent study, Ravaja et al. 2013) demonstrated that relatively greater left frontal activation (in the alpha band), measured just a few seconds before choice, predicted the affirmative decision to purchase a given consumer good. Higher perceived need for a product and higher perceived product quality (as measured by a questionnaire answered at the end of the decision phase) were also associated with greater relative left frontal activation (also in the alpha band). However, note that in both of these studies subjects made actual choices during the EEG recording. Therefore, it is still an open question whether we can use EEG data during passive viewing of goods in order to predict choices over some substantial time horizon.



9 An important step forward on this question is a recent study by Vecchiato et al. (2011). The

authors recorded EEG activity while subjects viewed video commercials and subsequently related these measurements to the responses from a questionnaire regarding the pleasantness of the same commercials (conducted two hours following the EEG session). This study demonstrated that theta and alpha band activities were related to the subsequent pleasantness ratings, with activity in the left frontal cortex related to “pleasant” commercials and activity in the right frontal cortex associated with “unpleasant” commercials. Although this study demonstrated a link between EEG recordings and a subsequent behavioral response, the use of pleasantness ratings might not be correlated with the actual valuation and subsequent choice of a good, as noted above. Additionally, there was no means to assess predictive power and/or the precision of predictions. Therefore, in the current study we sought to overcome these limitations and demonstrate the applicability of using EEG for predicting consumer choices.

Technical Aspects of EEG Measurement Neurons in the brain communicate via electrical impulses. EEG measures the oscillations of the resulting electrical potentials (voltages) with electrodes located on the human scalp. Each electrode reflects the summation of the synchronous activity of thousands or millions of neurons that have similar spatial orientation. Because voltage fields fall off with the square of distance, activity from deep brain areas is more difficult to detect than activity near the skull. Hence, most of the measured EEG signal originates from cortical rather then subcortical areas. It is well established that behaviors and mental processes are the result of complex interactions between multiple brain areas in various spatial and temporal scales. Only part of this dynamic activity can be measured at the macroscopic level by scalp EEG (Luck 2005; Nunez and Srinivasan 2006).



10

In the current study, we examined the EEG response to commercial goods using two common methods, both of which have been shown to represent mental processes that emerge in reaction to various stimuli. The first method is the event related potential (ERP), which measures the changes in the voltage level in response to a stimulus presented as a function of time. Because the temporal resolution of these measurements is on the order of tens of milliseconds, ERPs can accurately measure when rapid processing activities take place in the human brain, and can provide information about a broad range of cognitive and affective processes (Luck 2005; Nunez and Srinivasan 2006). With regard to decision processes, such as categorizing and evaluating a stimulus, two well-known ERP components have been identified: the P300 wave component (i.e. a positive deflection in the scalp potential starting 300ms after stimulus presentation, see reviews by Polich 2007; Soltani and Knight 2000), and the N200 wave component (i.e. a negative deflection in the scalp potential starting 200ms after stimulus presentation; see review by Folstein and Van Petten 2008).

The second method we employ is termed event related spectral perturbations (ERSP). Similar to the ERP technique, ERSP measures the response to a stimulus over time, but it divides the EEG signal into different frequency bands. The ERSP method then examines if, and to what extent, there is a change in the power of a given frequency band across time. Importantly, the measured change in power provides both a temporal and a spatial code, which adds additional valuable information to the ERP data. The frequency spectrum is usually subdivided into frequency bands: delta (1-4 Hz), theta (5-8 Hz), alpha (8-12 Hz), beta (14-30 Hz) and gamma (40 Hz).



11

There is ample data linking changes in these frequency bands to various cognitive processes such as changes in mental state (Moretti et al. 2004), changes in attention allocated to a task (Klimesch 1999), memory processes (Klimesch 1996), motivation and emotional processes (Knyazev 2007), different sleep stages (Keenan 1999) and consciousness levels (John 2002), among others. For example, the alpha band has been associated with attention focusing (Prime et al. 2003), the theta band with inhibition of elicited responses (Kirmizi-Alsan et al. 2006; Yamanaka and Yamamoto 2010), and the beta band with alertness (Pfurtscheller and Lopes da Silva 1999). However, it is important to emphasize that each frequency band can be associated with many cognitive processes and one can not conclude a particular mental process is active simply by examining changes in a specific frequency band (Poldrack 2006).

It is also important to remember that because the spatial resolution of the EEG signal is very poor, any conclusions regarding the exact localization of the signal should be taken very cautiously and should not be used as conclusive evidence that an identified brain area is related to a measured behavior.

METHODS

Our study follows the three-stage experimental procedure laid out in Levy et al. (2011). In the first stage, subjects received a general description of the study procedure and familiarized themselves with ten consumer goods. In the second stage, neural activity was measured with EEG while subjects viewed pictures of the goods they encountered in stage one. The aim of this



12

stage was to acquire independent measurements of neural activity for each good in isolation. In the third stage, subjects were presented with pairs of the consumer goods, made binary choices between all the goods they saw during the EEG stage, and were also asked to rank order the goods according to their preferences. We now describe these stages in detail.

Stage 1: Familiarization with the goods The experimenter briefly described each good and subjects were invited to examine them (the products were in their original packages). Subjects were not informed of the actual prices of the goods. After all goods were presented, we informed subjects that at the end of the experiment they would get the product they wanted most. The consumer goods used in the study were randomly chosen from the on-line website of one of Israel's largest retail stores (‘HomeCenter’). The goods were: 1) white digital stereo headphones, 2) a white plastic kettle, 3) a pink bulb desk lamp, 4) a red optical wireless mouse, 5) a red and black 16GB USB flash drive, 6) a magnetic message board, 7) a rainbow colored hammock, 8) a white & blue steam iron, 9) a pink yoga mat, and 10) a yellow fry pan. Importantly, the average price of a good was 80 NIS, ranging from 70 to 90 NIS. This limits the possibility that the value differences we observe are due to differences in purchasing price. A full description of the goods, including images, is given in the Web Appendix.

Stage 2: EEG Measurement After examining all the goods, the second stage of the experiment began. Each subject was seated in a comfortable chair in a dimly lit soundproof room and an EEG electrode cap was placed on their head. Subjects were asked to minimize head and body movements as much



13

as possible at the time of the recording. On a standard computer screen, images of all ten goods encountered in stage 1 were sequentially presented. Only one good was presented in each trial, and subjects were simply instructed to think about how much a good is worth to them. Note that during EEG recordings, subjects did not make any actual choices nor did they execute any other motor response.

Figure 1A depicts the visual presentation of a consumer good. On each trial a fixation cross was presented at the center of the screen for a randomly varied interval of 800 – 1200ms, followed by the presentation of a good for 2sec. The fixation period of the next trial started immediately at the offset of the previous trial. To improve the signal-to-noise ratio of our measurement, each good was presented 50 times, in a random order, resulting in 500 total trials. These 500 trials were divided into 10 blocks, consisting of 50 trials in each block. Subjects were allowed to take short breaks between the blocks. At the end of each block a message appeared on the screen stating that the subject can continue the task whenever she is ready by pressing the mouse button. The total time of the EEG recording stage was 25 minutes. Please see the Web Appendix for a full description of the technical details regarding the EEG recording and preprocessing of the signal. - INSERT FIGURE 1 HERE Stage 3: Choice Stage After the EEG recording was finished, the EEG electrode cap was removed and subjects waited 10 minutes before starting the behavioral choice task. Figure 1B depicts the visual presentation for this task. On each trial, two goods appeared simultaneously on a computer screen and the subject stated which good they preferred (under no time limit). All possible pairwise



14

comparisons of the 10 goods were presented, totaling 45 pairs, and each pairwise choice was repeated six times, totaling 270 randomly-ordered choice trials. The repetition of six choices is important for the measurement error correction we will employ in our choice prediction analysis. The location of each good on the screen (left or right) was also randomly altered.

To further validate the results of the behavioral choice task, we conducted two additional measures. First, subjects were asked to answer a brief computerized questionnaire, in which they were requested to rate how much they liked each good on a 7-point scale ranging from ‘dislike a lot’ (=1) to ‘like a lot’ (=7) and also to rate how much do they want each good on a 7-point scale ranging from ‘don't want at all’ (=1) to want very much (=7). Second, subjects had to rank order the goods from 1 (most preferred) to 10 (least preferred). Finally, in order to control for any possible ownership effects, we asked subjects to state, for each good, if they own a similar product. The goods on all questionnaires were randomly ordered. At the end of the experiment the participants selected their favorite good.

To ensure that the possession of similar goods will not effect the correlation between each of the questionnaires and the scores of the behavioral choice task, we conducted a Pearson partial correlation between the questionnaires and the choice preferences while using the possession of the goods as a binary control variable. The analysis revealed that possession of similar goods did not have any significant effect on the magnitude or significance of the correlations.

EEG Measurements



15

To analyze the EEG data, we used two distinct, well-established, methods (Nunez and Srinivasan 2006). In the first approach, termed ERP, the general waveform for each electrode (of the 19 electrodes in our setup) is averaged across all repetitions of the same event locked to stimulus presentation. Therefore in our study, we averaged the waveforms of all 50 presentations separately for each good and for each subject. This allowed us to look at the averaged waveforms across the different goods for each subject and examine whether there are systematic differences that predict future choices. The second approach, termed ERSP, looks at specific frequencies embedded within the general EEG signal. The ERSP analysis reflects changes across time in the power of specific frequency domains as response to stimulus presentation. Therefore for each frequency, average event-locked deviations from baseline activity (mean power) can be tracked (Makeig et al. 2004). We then examined whether specific frequencies could be used to predict subjects’ future choices.

In order to avoid the issue of multiple comparisons and post-hoc hypotheses, we used both previous literature and a basic visual and statistical analysis conducted on our first five subjects to determine which electrode to focus on, which ERP and ERSP components to analyze, and the duration of the time window. We continued with our remaining subjects -- and with the choice prediction exercises that constitute the main hypothesis of the study -- only after we decided on these basic aspects of the EEG analysis.

For our first five subjects, we compared the average ERP signal in response to a median split of the top 5 most-preferred goods in the sample (across all subjects) and the bottom 5 leastpreferred goods. A full report of this analysis can be found in the results section and Table WA1



16

in the Web Appendix. We found that the mean amplitude of the N200 component could differentiate between high- and low-preferred goods, however a non-significant P300 component was present in the ERP signal. Based on previous studies (Folstein and Van Petten 2008) and the known dynamics of the N200 component (Folstein and Van Petten 2008; Naatanen and Picton 1986; Sutton et al. 1965) we thus focused the remainder of our analysis on a 100ms time window (200ms-300ms following stimulus presentation) that was centered on the N200 peak amplitude (typically observed near 250ms).

With regard to electrode choice, ample data from EEG studies (Holroyd and Coles 2002; Nieuwenhuis et al. 2004; San Martin et al. 2010; Yeung and Sanfey 2004) and fMRI studies (Bartra et al. 2013; Grabenhorst and Rolls 2011; Kable and Glimcher 2009; Levy and Glimcher 2012; Padoa-Schioppa 2011; Platt and Huettel 2008; Rushworth 2008) suggest value representation is located in frontal areas. In accordance with this preliminary hypothesis, we identified that the strongest difference in N200 amplitude (between the top 5 most-preferred goods and the bottom 5 least-preferred goods) was evident in the front of the of the scalp map, with the strongest effect in electrode Fz. Because we decided to focus our analysis on the N200 component, and because this component is mainly evident in frontal central electrodes (Folstein and Van Petten 2008; Luck 2005; Nunez and Srinivasan 2006), we focused the rest of our analysis on electrode Fz — a central electrode located near the front of the brain.

This same strategy was repeated for the ERSP analysis. The average ERSP signal in the frequency range 0-40Hz could differentiate between a median split of the top 5 most-preferred goods and the bottom 5 least-preferred goods (again, see the results section and Table WA1 in



17

the Web Appendix). This guided us to focus on the theta band activity (5-8 Hz) within a timewindow of 100-400ms following stimulus presentation. This observation is in line with previous studies demonstrating a link between theta band activity and valuation (Cohen et al. 2007; Gehring et al. 2012). Based on previous literature we also hypothesized that alpha waves might also be related to subjects’ subsequent choices. However, we found no evidence to support this hypothesis (see Table WA1 in the Web Appendix).

The accumulation of these results led us to focus our subsequent analyses for the entire subject sample on the N200 and ERSP theta component in electrode Fz (in the time windows noted above). Importantly, in our subsequent analysis and choice prediction exercise, we did not run any statistical analysis on any other time windows or any other electrodes. To examine the robustness of our findings, at the end of the study we repeated the initial median-split analysis on the remaining ten subjects that were initially held out, and repeated our entire analysis on a control electrode described shortly. These results matched our initial findings reported above (see Table WA1 in the Web Appendix).

Control Electrode After conducting all of our correlation and choice prediction exercises on the frontal electrode Fz, we aimed to examine whether the predictive information of the EEG signal originates in more frontal areas (as would be expected from previous findings), or whether the predictive signal could be detected in other electrodes. Therefore we engaged in a control exercise by repeating all of the analyses in a more posterior but still centrally located electrode - Pz.



18 RESULTS

Establishing Consistency in Choices and Liking Ratings The first step in relating neural measurements to choice data is determining the consistency of choices over the six repetitions of each choice pair, and gauging the degree of stochasticity in choice behavior. The existence of consistent preferences and/or a clear rank-ordering over the consumer goods will presumably ensure a suitable range of valuations that can be measured via EEG.

The proportion of choice pairs that resulted in an even split between the two goods (each good was preferred in half a subject’s choices) was markedly low (0.02, SE = 0.01), suggesting a clear rank order preference of goods for all subjects. Furthermore, the proportion of the six repeated pairs in which the subjects switched their preference at least once was 0.25 (SE = 0.02) (over all subjects). This proportion is relatively low when you consider that this would be the probability of observing at least one switch out of six trials from the binomial distribution with a success probability of 0.795 on each trial.

To gauge how much of the switching arose from possibly inconsistent preferences, we also examined the proportion of stochastic transitivity violations (Tversky 1969). For each triplet of consumer goods {A,B,C}, such violations occur when P(A|{A,B})>=0.5, P(B|{B,C}) >=0.5 , but P(A|{A,C})