Cognitive sciences and technologies: from neuron to cognition

Размещено на http://www.allbest.ru/

The Government of the Russian federation

Federal State Autonomous Educational Institution of Higher Professional Education

National Research University - Higher School of Economics

Institute of Cognitive Neurosciences

Master's program

Final qualifying work - master thesis

M400 Study of Marketing Associations

Cognitive sciences and technologies: from neuron to cognition

Student group № MKN181:

Kuznetsova Elizaveta

Scientific advisor:

Shestakova Anna

Director of Institute of Cognitive Neuroscience, Ph.D.

Scientific co-advisor:

Alexei Gorin

Junior Researcher

Moscow 2020

Annotation

semantic magnetoencephalography linguistic association

The study focuses on magnetoencephalographic (MEG) research of human perception of marketing associations in the framework of a fundamental linguistic concept of semantic associations, manifested in a well-known N400 correlate of event related potential (ERP) N400. The N400 is a component of an EEG signal, evoked by an anomalous or semantically incongruent word in a sentence or string of words. Using N400, recent neurobranding and neuropricing studies showed promising results of monitoring brand perception and willingness to pay, correspondingly. We tested the hypothesis that marketing and semantic associations reflected in the magnetoencephalographic equivalent of N400, M400, share similar brain mechanisms, that can be probed using the source analysis of spaciotemporal dynamics of cortical responses. Our hypothesis was falsified as MEG visualization indicated activation of different neural networks: we found difference of neural responses to incongruencies of information between marketing and semantic contexts. Our data bring new evidence for the existence of a distinct mechanism of processing marketing associations as compared with linguistic information.



Table of contents

Introduction

1. Literature review

1.1 Neural approaches to study marketing associations

1.2 N400. The correlate of semantic incongruence

1.3 Neuropricing

1.4 M400 or Magnetoencephalography for semantic abnormalities

2. Methods

2.1 Experimental procedure and study design

2.2 Stimuli

2.3 Participants

2.4 MEG data processing

2.5 Preprocessing

2.6 Sensor space

2.7 Head modelling

2.8 Source space and analysis

3. Results

3.1 Sensor space

3.2 Source space

4. Discussion

Conclusion

References



Introduction

With the development of Internet and communication technologies top companies started to put more effort into building powerful brand perceptions. They realized that the way consumers perceive brands is a key determinant of long-term business-consumer relationships (Fournier, 1998). Since then the concept of brand associations has been widely explored. Marketing experts David Aaker and Kevin Keller are considered to be the first who proposed the classical theory of brand association (Aaker, 1991; Keller, 1993). K. Keller refers to brand image as “perceptions about a brand as reflected by the brand associations held in consumer memory”. His theory includes perceptions of brand quality and attitudes toward the brand (Keller, 1993; Bridges and Keller, 1998). Similarly, Aaker proposes that brand associations are anything linked in memory to a brand (Aaker 1991, 1996a). Different approaches to study brand associations have been proposed later on and now we can confidently say that in the third decade of XXI century research interest in branding have reached its apogee. However, contemporary methods of marketing research (surveys, focus groups, in-depth interviews, etc.) may be more effective if cognitive research technologies are used in the planning and analysis of the experiment. It is objectivity and the ability to test implicit effects of brand perception that are the advantage of neurovisualization (or neuroimaging) when studying consumer behaviour, i.e. almost every individual in modern society.

So far, only a few EEG neuromarketing studies employing ERP approach (Herbes et al., 2015; Gorin et al., submitted) have become available for publics. MEG studies using the M400 component to study semantic associations (Halgren et al., 2002; Hulten et al., 2019) are extremely rare too. To our best knowledge, not a single MEG study of marketing associations can be found in the literature. To be mentioned, MEG has superior resolution in space as compared to EEG. In temporal domain, EEG and MEG both excel fMRI. Thus, our study continues the line of neuromarketing studies but uses a more advanced methods (MEG) that permits to go beyond the electrophysiological responses to the localization of the sources of an M400 response, which is a MEG equivalent of the famous N400 correlated of semantic associations.

In sum, our interdisciplinary research in the field of neuromarketing may contribute to the development of fundamental ideas about the lexical and semantic processing of information by the brain using a novel MEG approach.



1. Literature review

1.1 Neural approaches to study marketing associations

Classic methods for brand associations rapidly develop together with new technologies. Nowadays social media has provided an opportunity to create a dialog between the customer and a manufacturer via advertisement. Customers can influence the brand image with the help of online communication providing the manufacturer with information about their preferences and suggestions about the product. These interactions are useful for Companies as previous studies showed that customer engagement generates additional value for the brand itself (Schau, Muсiz, & Arnould, 2009). The simplest way to engage customers into communication is to create particular brand associations via marketing campaigns. It may imply anything that target audience connects to the brand e.g. images, emotions, values, etc. The main aim of manufacturer is to develop sustainable associations between the brand and some particular information about it.

Understanding the brain mechanisms of processing of brand associations gives an opportunity to go deeper in customer's mind and understand the effectiveness of marketing campaigns. A great scope of work was made to measure the effectiveness of marketing campaigns as they seem important for the brand equity. But the data gathered from traditional methods of research (questionnaires, focus groups or in-depth interviews) may be unreliable. That is why researchers started to employ neurobiological markers such as hemispheric asymmetry or event-related potentials (ERPs) in search for objective correlates of the effectiveness of marketing campaigns (Baker, 2003, Shapiro, 1999, Zaltman, 2000).

ERPs can be extracted from electroencephalogram (EEG) by averaging of stimulus-locked events at an individual or a group level (Handy, 2005). Some ERPs such as N400, P3a or even N270 appeared to be very promising in neuromarketing research (Shang et al., 2018; Wang et al., 2012). For example, Ma and colleagues used N270 (Ma et al., 2007) and P300 (Ma et al., 2008) components to reveal that incongruent brand associations with brand name evoke lower amplitude of P300 and higher for N270. In experiments Ma used beverage categories and found out that people tend to accept familiar brand extensions of the products from the similar categories rather that from different ones. Participants were asked to respond on the keyboard whether the product could be an extension of the particular brand they have seen before. This process needs conscious evaluation of the product so N270 and P300 were taken as ERP components (Ma et al., 2007). Later Ma et al. (2014) also ran N400 study for brand associations and achieve the results which go in accordance with our knowledge of N400 as a marker of semantic processing.

Wang et al. (2012) recorded the language specific N400 brain signal in response to the product's attributes which were atypical to the category of the brand in the context of brand extension (when the established brand name is used to market a new product). In their study, participants were presented with a pair of sequential stimuli consisting of a soft drink brand name and two product types: a beverage - the typical product extension for that particular brand - and clothing - an atypical product for the same brand. The atypical brand extension triggered the frontal, fronto-central, and central N400. In accordance with the conceptual idea that brands are represented as mental categories (Aaker and Keller, 1990), Wang and colleagues reasoned that the N400 elicited in their study reflected an integration and conceptual process related with the mental category.

Further, Gorin et al. studied whether the N400 observed in the aforementioned product-brand associations can be explained by the same neurobiological mechanisms. In their within-group EEG study, Gorin et al. compared semantic N400 in response to incongruency of sentence endings with the neural response to the degree of incongruency of associations to brands. To test whether in the same subjects incongruent brand associations would elicit the N400 brain response resembling semantic N400, they constructed two-word phrases in which brand names were proceeded by congruent or incongruent adjectives that played the role of brand associations. Sentences lacking any marketing context served as control stimuli and evoked classical centro-parietally distributed N400 response. The sensitivity of the brain activity to the incongruent brand associations was manifested in the fronto-centrally distributed N400 response within the same time window as in the semantic response. In addition, correlation analysis of the N400 amplitudes showed association between the semantic- and brand-related brain responses. In sum, EEG results of this study not only support the N400 hypothesis of mental categorization for brand representations but suggest similarity of neural mechanisms of the semantic and brand-related N400. In this research we also focused on N400 and are going to describe this potential in more detail.

1.2 N400. The correlate of semantic incongruence

The N400 is a component of an EEG signal, evoked by an anomalous or semantically incongruent word in a sentence or string of words. This electrophysiological response was first observed by Kutas and Hillyard (1980). The authors confirmed that the observed negativity was evoked not only by the unexpected action deviation of the potential but also by a correlate connected to the semantic processing of the stimuli. A variety of publications consider the N400 in the paradigm of lexical or sentential meanings (Brown & Hagoort 1993; Chwilla et al., 1995; Halgren et al., 2002; Kiefer, 2002; Kutas & Federmeier, 2011; Lau et al., 2008; Van Berkum, 1999).

Moreover, King and Kutas (1998b) found that the duration of the left anterior negativity, occurring within 200-400 ms after the stimulus onset, correlates with the frequency of the word in this language. The brain, therefore, more quickly recognizes words that are better known. At the sentence level, the amplitude of N400 declines, depending on how well the word is put into context. When the semantics of a word is completely out of context, for example, “He drinks coffee with sugar and socks”, the amplitude of N400 is the largest (Kutas and Hillyard, 1984). Additional evidence to support the hypothesis that N400 characterizes the process of semantic integration can be found in the literature on aphasia. Swaab, Brown and Hagurt (1997) showed that in patients with moderate or severe aphasia N400 begins with a delay and its amplitude is significantly weakened; however, patients with equivalent lesions of the right hemisphere do not have the same patterns of N400. Importantly, N400 studies have been also expanded into non-linguistic contexts, including picture sequences, photos of objects in a visual scene, and short videos of everyday events (reviewed in Sitnikova et al. 2008; Kutas & Federmeier, 2001).

N400 reflects the search for information in memory, so it can be concluded that information about a word exists in a certain mental space, the parameters of which are determined by our experience. Thus, if we have such experience and the word or connection between two or more categories is familiar amplitude of N400 would not be high. Getting closer to the main topic of the research, it is worth to mention that big manufacturing Companies use very similar mechanisms. They aim to create this experience of the Brands they have among the customers to increase loyalty and familiarization with the product.

The choice of the neural correlate of semantic memory - the N400 - seems to be promising as branding focuses on the dynamics and plasticity of brand associations, building new associations or influencing old ones and thus changing loyalty, recognition, memory, and values for brands. While neuroscience of brand associations is in its beginning, the N400 research in language is ample (for a review, see Lau et al., 2008). In the current study, we strove to go further and check whether the more general mechanism of semantic memory, namely the congruency effect probed by the N400, can be sensitive to the congruency of price associations.

1.3 Neuropricing

Price of the product significantly influences its competitiveness. One can increase competitiveness by reducing the price, but there is another more promising method which helps to make the product desirable even with a high price - marketing campaigns. These marketing strategies in turn depend on the understanding how much the consumer is willing to pay for particular benefits of the product (WTP). Behavioral economics provided us with a bunch of methods to WTP estimation. Most of them were described and nicely presented in the paper of Breidert and colleagues (2006) (Fig.1).

Figure 1. Classification framework for methods to measure willingness-to-pay (Breidert et al., 2006)

Methods can be firstly divided into revealed and stated preferences. Revealed preferences imply price-response data gathered from either market itself or conducted experiments. Experiments can be further divided into laboratory and field experiments. Often, the Becker-DeGroot-Marschak auction is used to measure WTP. Taking closer look at stated preferences one can notice that they imply survey-based techniques, whether direct (when the respondent is explicitly asked how much would they be willing to pay for particular product) or indirect (when ranking procedure is applied). In neuromarketing, surveys as well as questionnaires and interviews have been widely used as a common tool to measure WTP. However, neuroscience research now gains its popularity as neuroimaging studies have already proved its effectiveness as an innovative field of marketing research (Mu?ller, 2010; Pradeep, 2010; Shiv and Yoon, 2012) and can fulfill the revealed preferences methodology by broadening the variety of experiments that might be conducted. Plassmann et al. (2012: 31) claimed that analyzing brain patterns promises “better predictions of decision-making behavior across domains” and that decoding these patterns “will be a turning point for consumer neuroscience research.” And in fact, the predictability of consumer`s decisions was more accurate using neuroscientific methods than self-reporting instruments (Knutson et al., 2007; Berns and Moore, 2012). Moreover, the sample size needed to conduct neuromarketing research is less than for traditional methods (Berns and Moore, 2012).

In paper of Knutson and colleagues (2007), authors tried to predict the impact of price on the decision to purchase high demand products such as popular DVD sets and household items. They used functional magnetic resonance imaging (fMRI) approach and compared the results of neuronal correlates with self-reported answers. In another study applying neuroscience-based method to measure purchasing decision for green energy authors developed a tool to measure WTP using neural signals from an electroencephalogram (EEG) (Mu?ller, 2012; Thadeusz, 2013). The idea is that time-locked EEG signals are an ideal tool to measure match-mismatch signals from the brain. For example, if you listen to monotonous audio stimuli and then deviant stimulus appears, average EEG signals for deviant signal will significantly differ from EEG signal for standard one. The same results were obtained on visual and language stimulus as well. Moreover, most time-locked signals are recorded within the first 500 ms after a stimulus onset. This prevents information from being distorted by tactical considerations or social influences which usually appear later on. So, in the recent study by Herbes et al. (2015) authors implied the method of neuropricing to describe and mitigate common WTP methodology biases such as anchoring effect, systematic bias, scope sensitivity and the most important strategic behavior. A participant of the study was exposed to a picture explaining the energy product, a price, the adjective word `cheap' or `expensive' and gave a manual response [Yes or No] whenever he or she agreed or disagreed with the overall statement presented in pictures, for example `[Energy product] - [Price] - [Expensive]'. The main result was that in general consumers were ready to pay 15% more for green electricity products than non-green. The study also showed that neuropricing methodology is indeed able to avoid common WTP biases and can be used not only for testing the effects of product characteristics on willingness to pay, but also for evaluating WTP effects of specific messages in marketing communications. Thus, this approach may lead to a better understanding of consumer behavior and hence facilitate more successful marketing campaigns of goods.

Recently, our group has found that N400 can be sensitive to the strength price associations (Gorin et al., in prep.). They hypothesized that N400 can be elicited by the mere incongruence of the product and price information. Gorin et al. tested whether the excessive or low prices would evoke larger N400 responses than the market prices. In order to choose the right price, they dwell upon the concept of price fairness (Conzelmann, 2014; Thadeusz, 2013), which is associated with willingness to pay (WTP) and widely used in population behavior. In order to access whether the excessive or small price will evoke the N400 response, they used the price fairness task proposed by Herbes et al. (2015). Each trial of this task had identical temporal structure: 20 subjects first saw the product image (two differently priced mobile phones, presented during 1s, which was followed by the 0.8 s exposition of randomly chosen price (small, right or big). After price presentation the condition-defining words - `cheap' or `expensive' - appeared and participants were asked to answer `yes' by the button press if they agree with the correspondence between condition-defining word and the given price for particular phone and `no' if they do not agree. The researchers analyzed the brain activity that was locked to the beginning of the final condition-defining words. The same experiment was replicated one more time using another set of the stimuli (Apple iPhone XR and Xiaomi Mi A2). EEG data analysis was carried out offline using the Brainstorm software package interface. In both experiments, the excessive and small prices evoked N400-like responses in the time window of 300-400ms from the presentation of the condition-defining word (Fig 2a, 2b). The N400 activity had central and centro-parietal distribution. For each of the four products, the polarity of the N400 response to the incongruence of the price interval and the condition-word changed to the opposite from the point of the fair (right) price which in its turn well corresponds to the averaged market price.

To sum up, Gorin et al., found that incongruency of the price and the product evoked the N400-like response. This study showed that the N400 component of visual ERP can serve as an index price fairness.

Figure 2. Average ERPs for 2 price spans used in the experiment. The potentials were averaged across all subjects and 6 electrodes included to the cluster 1 (iPhone + Nokia)



Figure 3. Average ERPs for 2 price spans used in the experiment. The potentials were averaged across all subjects and 6 electrodes included to the cluster 1 (iPhone + Xiaomi)

The main aim of this research is to localize the origins of neuromarketing (namely, product-price) associations in the human brain and compare it with neural correlate of semantic associations using powerful MEG approach. Here we chose to use the association of price (cheap or expensive) with the product. As a common external clue, price has a more complex influence on purchase intention (Roy et al., 2014). On the one hand, price is an important signal of money loss. On the other hand, price has always been considered as a symbol of quality (Daliri et al., 2014). When see the price consumer should always make a decision whether the brand provides good value for the money and whether there are reasons to buy this brand over competitors.

1.4 M400 or Magnetoencephalography for semantic abnormalities

MEG is a relatively new neuroimaging technique discovered in 1980s. In comparison with EEG MEG measures not the electrical current by itself but magnetic fields produced by this current in the neurons in the brain. It is believed that David Cohen was the first person who registered magnetic fields in a living organism. The magnetic field produced by each individual neuron is extremely small, even by neuronal populations it is not bigger than 10^-15 T. But arrays of SQUIDs (superconducting quantum unit interference devices) placed in liquid helium with the temperature of 4 Kelvin allow to detect very weak magnetic fields a few centimeters away from the surface of the MEG helmet. MEG is non-invasive and non-contact method, where magnetometers and gradiometers are used as sensors. Magnetometers are more sensitive to the deeper layers of the brain, but at the same time they pick up extraneous noise very well. Gradiometers better filter this noise, but in turn, they are slightly less sensitive to the magnetic field of the brain. The main generator of the MEG signals as well as for EEG are postsynaptic currents in pyramidal neurons. But MEG is more sensitive to tangential currents, in the walls of cortical fissure, while EEG to perpendicular ones. This fact makes these two methods complement each other - the usage of both techniques in research allows to obtain a complete picture of spatiotemporal patterns of brain activity under particular conditions.

However, there are characteristics that make MEG better than EEG - its spatial resolution and scalp penetrance. Magnetoencephalography allows to precisely (with an accuracy of 1 mm) detect the location of magnetic field sources in the brain, moreover, to do it over time with an accuracy of a millisecond. Secondly, poorly-conducting human skull distorts EEG signal significantly while it is completely permeable for magnetic field. It is worth to mention also the convenience of MEG for the patient - there is no need to wear a hat (EEG) - the person just need to place the head inside the helmet and stay still, the equipment is not noisy (MRI) and there is no need to inject radioactive substances (PET). All these advantages make MEG an extremely expensive technique both because of its complicated structure and the need for permanent maintenance.

Analysis of MEG data also challenges the researcher who runs magnetoencephalographic experiments. The main aim is usually to determine the location of electric activity within the brain from induced magnetic fields outside the head - so called inverse problem. But this problem does not have a unique solution - any surface recording can be explained by an infinite number of different configurations of internal sources. However, in neurocognitive experiments, it is possible to achieve a uniqueness of a solution by introducing restrictions based on cognitive, perceptual or other properties of the experimental problem, as well as on our knowledge of the functional anatomy of the brain.

A widely used source-modeling technique implies calculating a set of equivalent current dipoles (ECDs), which assumes the underlying neuronal sources to be focal. This dipole fitting procedure is non-linear and over-determined, since the number of unknown dipole parameters is smaller than the number of MEG measurements (Huang et al., 2006). Another popular model - distributed source model (DSM) divides the source space into a grid containing a large number of dipoles. The inverse problem is to obtain the dipole moments for the grid nodes. The inverse solution of DSM is highly underdetermined because the number of unknown dipole moments is much greater than the number of MEG sensors. To reduce the ambiguity of the solution, additional constraints are needed (Hдmдlдinen & Ilmoniemi, 1994). Although achieving precise source еstimаtes is difficult, MEG allows robust discrimination between activity generated in the left versus the right hemisphere and anterior versus posterior cortical areas over time, and can thus provide supporting evidence for the time course of activity in areas identified with fMRI.

Studies of semantic priming and semantic anomaly that use ECD source analysis uniformly report a response that localizes to the left mid-posterior middle temporal gyrus (MTG), superior temporal sulcus (STS) and superior temporal gyrus (STG), that has an onset of ~250 ms in auditory presentation (Uusvuori et al., 2007) and ~300-350 ms in visual presentation (Service et al., 2007; Halgren et al., 2002), and that has a peak latency of 410-450 ms. Other MEG studies have estimated the source of the semantic anomaly effect using a distributed-source model based on the cortical surface (Dale et al., 2002). Early effects (at 250-500 ms) were observed in the left planum temporal and left MTG and inferior temporal cortex (IT) (Halgren et al., 2002). Additional areas -- the left anterior temporal and inferior frontal cortex and the right orbital and anterior temporal cortex -- were implicated in the later part of the anomaly response, which may reflect either the latter part of the N400 effect or the post-N400 positivity. However, other studies using distributed-source models have found somewhat different results (Maess et al., 2006), so clarification is needed. Nevertheless, all methods of MEG source analysis have converged on the finding that the left mid-posterior temporal cortex is one source of the N400 effect. Furthermore, a study using simultaneous recordings of ERP and the event-related optical signal (EROS) corroborated the MEG source estimates: left-hemisphere responses to semantically anomalous sentence endings were observed in the mid-posterior STS and MTG at 200-400 ms, with contributions from anterior temporal and inferior frontal areas only after 500 ms (Tse et al., 2006).

To sum up, a great amount of studies was done to investigate the neuronal mechanism of brand processing via EEG. But electrophysiological methods are much scarcer than magnetoencephalographical (Hulten et al., 2019). EEG is driven by volume currents and therefore is influenced to inhomogeneity of the sculp and scull. MEG is more suited for source analysis because it measures magnetic field generated by the primary currents with the help of the SQUID sensors without a contact with the scull and is not distorted by inhomogenties of conducting media and it allows more precise cortical source reconstruction (Hamalainen and Ilmoniemi, 1994).

To our knowledge there were quite a few studies in which evoked-fields potentials (EFPs) were used to investigate semantic incongruency via magnetoencephalography (Brennan and Pylkkanen, 2012; Uusvuori et al., 2007) and significantly less MEG studies investigating the localization of M400. To our best knowledge, no MEG studies can be found for Brand associations or neuropricing. Our study uses an advanced MEG method that permits to go beyond the electrophysiological responses to localize the origins of brand associations in the human brain and compare it with neural correlate of semantic associations.



2. Methods

2.1 Experimental procedure and study design

The experiment consisted of two experimental blocks during which participants were exposed to price association statements (price or marketing condition) and short sentences (control or `semantic association' condition). In each condition, the congruent word or price combinations were randomly interspersed with incongruent ones.

The price condition (Fig. 5) included 240 trials with the following structure: first, the product image was presented (iPhone XS or Nokia) on the screen for 1.3 seconds, then price taken randomly from one of the possible price ranges appeared for the same period of time. After the price presentation the participant was exposed to the context-defining words-questions - `Cheap' or `Expensive'. The trial ended with the subject's button press `yes' when they agreed and `no' when disagreed with the correctness of the proceeding expression. For example, the combination of Apple iPhone XS (right price equals 50 000-70 000/ rubles 700-1,000 USD) + 7 000 Rubles (100 USD) + the `expensive' adjective would be denoted as the incongruent condition, whereas the combination of Apple iPhone XS + 7 000 Rubles (100 USD) + `cheap' would be denoted as congruent condition. The distance between participant's head and the display was 1.5 m. Fixation cross duration is 0,3 seconds, price and photo duration - 1.3 seconds and no limit of time to answer the question. After each sentence the jitter interval of 0.1 - 0.3 seconds was presented. All words were written in black letters on white background. In test condition the analyzed brain activity was locked to the beginning of the final condition-defining word.



Figure 4. Experimental procedure of the neuropricing block

The control condition (Fig. 4) included 160 simple semantic statements, both congruent and incongruent, for example, `Grandmother knits scarves for children' or 'Grandmother knits apples for children'. Each word remained on the screen for 300 ms. Each sentence was presented word by word and followed by the jitter interval of 0,1-0,3 seconds before the next one. In one third of trials participant was asked to answer the question «Was there any sense in the previous sentence» to maintain participant's concentration on the stimulus.

Figure 5. An example of the stimulus material in the semantic association block

2.2 Stimuli

In our study, the high-resolution pictures of Apple iPhone XS were used as a marketing stimulus, which were followed by the price from 1 out of 5 ranges created (Fig. 6). There were 40 sentences in total in each price range which expectedly should give rise to a greater N400 amplitude (Range 1: 500 - 7000 rub. and Range 5: 200000 - 300000 rub.) and just a few where the price is expectedly normal for this model of iPhone (40000 - 100000 rub.). Nokia phone was taken as a filler stimulus that is why the amount of trials in all 5 Nokia price ranges was reduced to minimum in order to not overload the participant with a big number of trials. The following procedure was applied to construct the pairs of brand and its price association.

Figure 6. Price ranges for marketing stimuli used in the experiment (iPhone and Nokia)

Semantic stimuli were made in collaboration with HSE Center for Language and Brain. The experiment included 40 experimental units (160 sentences in the whole experiment). Each unit consisted of 4 sentences with 2 different contexts. In one sentence within the same context the target word was congruent and made sense, while in the other sentence the target word was taken from another context from the same unit and, consequently, was incongruent. Thus, in each block we had 2 different contexts and 2 different target words (for example, see Fig. 7). To reduce the confounding component of different words parameters (such as frequency, length, sound complexity) all target words had a stress on the second vowel (out of two) and were matched for frequency. Moreover, all sentences consisted of 4 words (not including propositions) and the target word was always the 3^rd out of 4 words in the sentence.



Figure 7. An example of one experimental unit in semantic condition used in the experiment

2.3 Participants

25 right-handed native speakers (17 women) with normal or corrected to normal vision with no history of head injuries, language disorders or mental illness were included in the analysis. Also, participants should not have any metal implants in the body in order not to damage the MEG machine. All participants signed informed consent and were instructed before the experiment.

2.4 MEG data processing

Preprocessing

Analysis was performed using the MATLAB-based Brainstorm package (Tadel et al., 2011) Across all subjects we cut off all bad parts of the recordings which Maxfilter could not fix, then data was band-pass filtered (1-50Hz), and bipolar EOG channels were reconstructed for vertical (VEOG) and horizontal (HEOG) eye movements from monopolar EOG recordings. Independent Component Analysis (ICA) was implemented to suppress cardiac and eye-movement artifacts. We preferred ICA rather than another powerful method of artifact correction, Signal Space Projection (SSP), because it is believed that SSP reduces the signal amplitude, while the amplitude is preserved, when ICA is being used. Moreover, signal-to-noise ratio (SNR) is more or less the same when applying any of these methods (ICA or SSP) for EEG and magnetometer channels, but it is slightly better for gradiometers, when ICA-based artifact correction is applied (Haumann, 2016). We ran separately analysis for magnetometers and gradiometers. Otherwise, the resulting components will be a mixture of both sensor types, so when the component is rejected and projected back to the sensor space, gradiometer activity should "leak" into the magnetometer data and vice versa.

Later on, for the event-related magnetic fields (ERF) analysis, continuous MEG recordings were divided into epochs (from -200 to 800 ms) centered on the onset of the stimulus presentation (t = 0) and then down sampled to 500 Hz to optimize the signal processing time. A baseline correction based on the prestimulus interval (from -200 to 0 ms) was applied to each trial to remove the direct current offset. ERFs were calculated for each trial type separately. After all trials were averaged across subjects and conditions.

2.5 Sensor space

We split wide N400 window (300-500 ms) into early (300-400 ms) and late (400-500 ms) spans. To compare ERFs in this time window, we performed a series of cluster-corrected permutation tests across magnetometers and gradiometers separately, comparing congruency conditions in the two neuropricing (low and high prices) and semantic block. In particular, word “expensive” was incongruent with the context in the low-price block (500 - 7000 rub.) and congruent in the high one (210 000 - 300 000 rub.) for iPhone XS model. As example, “iPhone XS costs 600 rub. Expensive?” The expected answer should be “No”, as this amount of money for iPhone XS is far less than the window of its fair price.

We used the Brainstorm interface of the Fieldtrip toolbox and performed a paired spatiotemporal cluster-corrected permutation test (the cluster inclusion threshold was set to p < 0.01 with 1000 permutations over two N400-related time windows, early (300-400 ms) and late (400-500)) comparing congruent and incongruent conditions. The cluster p values were defined as the probability of observing the cluster with the higher mass, separately for the positive and negative clusters. The test was calculated separately for the magnetometers and the gradiometers.

2.6 Head modelling

We collected individual T1-weighted MRIs for all subjects using a 1.5 T Siemens scanner and constructed individual 3-D brain models with the FreeSurfer software (http://surfer.nmr.mgh.harvard.edu/) toolbox. These models were imported into the Brainstorm workspace. To compare the MR models and MEG data in space, we used the Polhemus Fastrak reference point digitization system just before starting to record MEG data. Forward solutions for the individual head models were calculated using the overlapping spheres approach.

2.7 Source space and analysis

We used the Brainstorm implementation of the sLORETA estimation algorithm with a constrained orientation to find the cortical current density distribution underlying the observed ERFs. We used sLORETA method, which is based on the source distribution estimated from MNE and a posteriori standardization by the variance of each estimated dipole source, because it is well appropriate to distributed source models where the dipole activity is likely to extend over some areas of the cortical surface (Hassan et al., 2014). The absolute values of activation were then computed for each vertex and projected from the individual models onto the default anatomy model (15002 vertices) using the iterative closest neighbor search algorithm, as implemented in the Brainstorm software.

We compared activity over a set of regions of interest (ROI) that included areas, involved in the semantic N400 generation (left inferior frontal gyrus (lIFG), temporal pole, superior and middle temporal gyri (Lau et al., 2008) and their counterparts) and in value estimation (ventromedial prefrontal cortex, VMPFC, divided in left and right orbitofrontal and cingular cortices); bilateral precuneus was used as a control ROI (Fig. 8).

Figure 8. Set of the ROIs used in the experiment. For semantic N400: left inferior frontal gyrus (red), superior temporal gyri (purple), middle temporal gyri (brown); for marketing N400: orbitofrontal cortex (light green), cingular cortex (dark green); for control: bilateral precuneus (blue)

We computed mean of absolute current value in the early and late time windows for each ROI, and compared levels of activation between different word conditions (cheap and expensive) in the range blocks ang congruency conditions in the semantic block using FDR corrected permutation test as implemented in Brainstorm.



3. Results

3.1 Sensor space

To visually inspect temporal evolutions of the obtained ERFs, we computed mean power of the magnetometers activation (Fig. 9).

Figure 9. Mean power of the magnetometers activation for the neuropricing and semantic conditions

In the magnetometers space, the analysis revealed significant differences between words in early window for both price ranges (low: p = 0.032, cluster mass = 449; high: p = 0.006, cluster mass = 1532) and in the late window for low price range(p = 0.016, cluster mass = 459); in the semantic block, significant differences were revealed for the late window only (p = 0.002, cluster mass = 1686). Cluster parameters were given for the biggest clusters observed. In the gradiometer space, differenced in ERFs were insignificant for all conditions. Also, to be mentioned, the topographies of difference of ERFs in the marketing condition had a spatially similar direction of the field, that was archived but subtracting congruent condition from incongruent one. (Fig. 10).

Figure 10. Topographic maps of the magnetic field distribution in two time windows. Asterisks mark significance level of p < 0.05 (single), and p < 0.01 (double). Insets depict spatial distribution of the significant clusters

3.2 Source space

In the early window only marginal differences were detected in the low price range whereas in the high price range bilateral OFC and ACC as well as rIFG were activated higher in the incongruent condition, p < 0.05 (Fig. 11; Table 1). In the neuropricing blocks, no significant differences were found in the late time window (Fig. 12; Table 1).

In the semantic block, no differences were observed for the early window, but in the late time span almost all of the structures associated with N400 generation (left IFG, STG and MTG) showed significantly higher activity for incongruent condition together with bilateral ACC and rIFG, p < 0.05 (Fig. 8a, 8b; Table 1).

Figure 11. Distribution of sources of brain activity, 300-400 ms, common head model

Figure 12. Distribution of sources of brain activity, 400-500 ms, common head model



Table 1. Difference between incongruent and congruent conditions, t-values. Significant scores are in red

ROI	Time window
	300-400 ms	400-500 ms
	Low Prices	High Prices	Semantic	Low Prices	High Prices	Semantic
L MTG	-0.7	1.6	0.4	-0.8	0.9	2.5
R MTG	2.1	0.4	0.2	0.9	0.0	-0.9
L precuneus	-0.4	0.5	0.2	1.2	0.8	-0.4
R precuneus	-0.3	1.3	0.6	0.0	-0.4	-0.2
L temporalpole	0.3	2.4	-2.7	0.3	1.5	0.9
R temporalpole	0.1	2.2	-0.9	1.5	-0.8	-1.3
L OFC	2.5	2.8	-1.3	1.5	0.2	2.1
R OFC	2.3	3.9	-0.3	2.3	0.7	1.7
L IFG	1.4	1.0	-0.7	0.8	0.3	3.0
R IFG	3.3	2.6	0.3	1.6	1.2	2.6
L ACC	1.9	2.7	-0.1	1.8	0.4	2.9
R ACC	2.1	3.1	0.5	0.8	0.8	2.8
L STG	-1.1	0.9	0.5	-0.2	1.7	3.7
R STG	0.8	0.7	1.1	1.3	1.0	2.0



4. Discussion

The purpose of this study was twofold: (1) replicate previous neuromarketing N400 studies in the neuropricing context and (2) find and compare cortical sources the brain activity in response to incongruent neuropricing and semantic information.

To do so, we chose a widely used semantic paradigm where sentences with incongruent and congruent endings were interspersed and the participants' brain responses to incongruent messages are analyzed with regard to the start of the incongruency. Such studies (also called the N400 studies because they evoke the brain component 400 ms after the incongruency point) have provided exhaustive and rich background to address the question of processing semantic associations. Indeed, semantically violated sentences manifest themselves into negative deflection peaking at ~400ms after the stimuli onset (Kutas and Hillyard, 1980). In our study we showed that incongruent sentences evoked magnetic activity starting from 400 ms with the largest peak at about 450 ms. Our MEG results in the semantic condition is therefore in line with the theory of processing semantic information.

In the source analysis in the semantic condition we obtained significant activity in the left inferior frontal (lIFG), superior temporal (STG) and middle temporal gyrus (MTG). According to the model of semantic processing in the context proposed by Lau, Phillips, and Poeppel (2008), lexical representations are stored and activated in the middle temporal gyrus (MTG), nearby superior temporal gyrus (STG) and inferior temporal cortex (ITC). Importantly, the aforementioned model proposes the cortical network for semantics which integrates the following cortical areas in the processing of word meaning: anterior temporal cortex (ATC) and angular gyrus (AG), where the information is integrated into contextual representations; anterior inferior frontal gyrus (aIFG) which is responsible for controlled retrieval of lexical representations based on top-down information, and posterior IFG (pIFG) which mediates the choice between highly activated candidate representations (Lau et al., 2008). Our source modelling results of the semantic information processing corroborate this model. Moreover, some researchers report the co-modulation effect between lIFG and left posterior temporal cortex (lPTC) around 400ms after the stimulus onset (Hulten et al., 2018) and those who study semantic abnormalities in particular mention the activity in left superior temporal and posterior frontal regions (Kielar et al., 2015; Hulten 2018, Lau et al.,2018). Thus, according to the earlier research papers all regions activated in semantic condition in our research play a significant role in semantic processing, what let us assume that we successfully replicated previous sematic N400 studies.

In the neuropricing condition, irrespective to the monetary context (low and high monetary ranges - 500-7000 rubles and 210000-300000 rubles, respectively), the analysis revealed significant differences between congruent and incongruent conditions which was manifested in the N400 response in the early (300-400 ms), but not late (400-500 ms) time windows. These results corroborate the previous EEG findings of our group: one on the brand associations (Gorin et al. submitted) and another on neuropricing (Gorin et al., in prep).

In the study of brand associations, Gorin et al., introduced two types of visual stimuli: brand statements and ordinal semantic statements (control probes). The results show that both semantically incongruent statements and incongruent brand-product associations evoked larger negative deflection in the time window of 300-500 ms after the incongruency onset. The topographical distribution, however, differed between the semantic and neuromarketing contexts: the sensitivity of the brain activity to the incongruent brand associations was manifested in the fronto-centrally distributed N400 response within the same time window as in the semantic response. In addition, correlation analysis of the N400 amplitudes showed association between the semantic- and brand-related brain responses. Based on their findings, Gorin et al., concluded that mental categorization for brand representations suggest similarity of neural mechanisms of the semantic and brand-related N400.

Further, in the neuropricing study, Gorin et al. (in prep) studied whether the N400 component of visual ERP can serve as price sensitivity meter. Using images of phones of two famous brands - Nokia and Apple - they found that dynamics of the N400 component was sensitive to the price and specific for the two products. In both experiments, the excessive and low prices evoked N400-like responses in the time window of 300-400ms from the presentation of the condition-defining word. The N400 activity had central and centro-parietal distribution.

In addition to the N400 studies, some papers report other marketing-specific indicators such as P200. For example, in their EEG study of brand extension evaluation Ma et al. (2014) discovered two-stage categorization process reflected by the P2 and N400 components. They suggested that P2 could reflect the early low-level and similarity-based processing in the first stage, whereas N400 could reflect the late analytic and category-based processing in the second stage. In their study researchers used 300-450 ms time window for N400 component, corroborating our findings. In the EEG study of neurobranding associations, Wang et al. (2012) reported similar to our study time window of the N400 equal to 350-450 ms. The previous EEG study on neuromarketing associations by Gorin et al. (in prep.) reported the fronto-central maximum of the N400 component. The permutation test revealed a statistically significant difference between ERPs to incongruent and congruent brain attributes between 330 ms and 460 ms at the F3, Fz, F4, C3, Cz, and C4 electrodes. The latencies of the N400 response were comparable across the two conditions.

Thus, the latency finding of the N400 in the neuropricing context corroborates previous neuromarketing studies. However, to our best knowledge, no groups but our (Gorin et al, submitted) attempted to compare the semantic and brand association processing manifested in N400 in the same participants.

In the study by Gorin at el. of brand-product associations, the latency of the N400 did not differ between the two experiments: semantic and neuromarketing. In our study we find latency difference between the neuropricing and semantic N400. In the early window we find significant N400 for the neuropricing stimuli only, whereas in the late window of 400-500 ms - for semantic stimuli. This latency finding contradicts our previous neuromarketing results and leaves an open question whether this difference marks different neural mechanisms for perception of pure semantic and price information.

In the source space, in the marketing condition the largest activity was found in the early 300-400 ms in bilateral OFC and ACC as well as rIFG. Left IFG seems to be the only region that is implicated in the neuroanatomical model for semantic processing (Lau et al., 2008). We hypotheses that the IFG functional role in the neuropricing condition could be associated with selection between the two possible affirmative ending: expensive or cheap. Further, we shall briefly discuss the functional role of the OFC, ACC, and IFG in the processing of price-product associations observed in our study in the framework of decision-making theory and compare our results with the previous fMRI studies on WTP.

Although the role of medial orbitofrontal area or vmPFC is still debated in marketing research, some papers showed that vmPFC plays a key function in brand preferences tasks and decision-making (Paulus and Frank, 2003; Koenigs and Tranel, 2008). Paulus and Frank (2003) showed that vmPFC is activated when brand preference judgment is required using photographs of soft drinks. In fact, they found important activations in this brain region when participants selected preferred soft drinks in contrast with a visual discrimination task of the same stimuli (liquids contained in bottles or glasses). Koenigs and Tranel (2008) compared performance of healthy controls and patients with damaged vmPFC in their preferences mediated by information about the brand. Healthy people changed their preferences when brand-cued stimulus was presented while patients with vmPFC impairment were persistent with their initial choice ignoring the information about the brand. The results show that vmPFC is used not only in brand preferences tasks but also is necessary in the integration of information in the decision-making process. Our paradigm of neuropricing research also implies the integration of information about a price and a product. The participant should not only perceive the phone and its price, but correlate the phone with its brand. It might explain significant activity in two particular parts of vmPFC - orbitofrontal cortex and anterior cingulate cortex - obtained by source analysis in our research.