How the Opinion of Others Affects Our Valuation of Objects

Summary The opinions of others can easily affect how much we value things. We investigated what happens in our brain when we agree with others about the value of an object and whether or not there is evidence, at the neural level, for social conformity through which we change object valuation. Using functional magnetic resonance imaging we independently modeled (1) learning reviewer opinions about a piece of music, (2) reward value while receiving a token for that music, and (3) their interaction in 28 healthy adults. We show that agreement with two “expert” reviewers on music choice produces activity in a region of ventral striatum that also responds when receiving a valued object. It is known that the magnitude of activity in the ventral striatum reflects the value of reward-predicting stimuli [1–8]. We show that social influence on the value of an object is associated with the magnitude of the ventral striatum response to receiving it. This finding provides clear evidence that social influence mediates very basic value signals in known reinforcement learning circuitry [9–12]. Influence at such a low level could contribute to rapid learning and the swift spread of values throughout a population.

). Activations are whole-brain cluster corrected Z statistic maps (Z > 2.3, p < 0.05), which were overlaid onto the standard MNI brain. Due to the activation"s location along a sulcus/fissure it is challenging to interpret. Peak activation can be assigned to either side of the sulcus/fissure, depending on which atlas is consulted (Table S2). The significant activation (observing the entire blob) follows the parieto-occipital fissure and anterior calcarine sulcus and extends into both posterior cingulate cortex and visual areas [1]. MNI coordinates (mm) of view in Panel A: x = -8; y = -60. Within the region of the anterior activation, prior metaanalyses of fMRI studies have revealed consistent activation during tasks of autobiographical memory, imagining oneself in the future and theory of mind [2], supporting a theory that this area mediates mental projection of the self to other times, places and perspectives [3]. Self-referencing social stimuli such as being liked by others [4] and agreement/conformity with normative opinion [5] also activate this region. From this perspective, activation might have resulted from self-reference while evaluating agreement with others. On the other hand, the posterior areas of activation could be assigned to occipital regions, including intracalcarine cortex, lingual gyrus, and precuneus. Moreover, there was additional activation in what is clearly occipital cortex (peak (mm) -18 -86 -14) ( Table S1). The Oxford-Harvard Cortical Structural Atlas and Jülich Histological Atlas [6,7] place this most posterior activation in occipital fusiform cortex and V4. Activity within the visual system has been previously shown to be modulated by value [e.g. 8,9,10] and therefore may have been affected reward from agreement with reviewers. Stimulus "value" is but one of several known factors that bias activation levels within the regions of visual cortex that represent incoming sensory information from the visual field (see [8] for discussion on this topic). Overall, this increase in visual activation may represent mechanisms that bias spatially selective areas of visual cortex in favor of more valuable stimuli. Alternatively, this activation may relate to other attention-related processes. As Panel B shows, when reviewers preferred the alternative song, their pictures were moved to the opposite side of the display. In this case, subjects may have looked at both sides of the display (and both song tokens) more often than when reviewers preferred the same song token as the subject. Multiple visual stimuli has been previously shown to result in net suppression of BOLD activity in visual areas, across hemispheres and particularly in V4, as neighboring stimuli compete for neural representation [11][12][13]. This could therefore result in relatively more activity with reviewer agreement but would not, to our knowledge, affect the interpretation of other findings in the study. Table S1. All fMRI activations: cluster size (voxels), Z score at peak voxel, peak MNI coordinates (mm), and associated anatomy from [1,7]

Pre-scanning
One week prior to scanning, subjects submitted a list of twenty songs that could be purchased from an online music store. Each was a song that the subject desired but did not yet own. On arrival to the centre, subjects had their photo taken and rated each of their 20 songs for desirability on a scale from 1 (I do not want this song) to 10 (I really want this song). Subjects also looked at pictures of two music "experts" and read descriptions of the two them, as follows: Dave is a respected musician and sometimes London DJ. He has been listening to and playing music as long as he can remember. He owns a massive collection of music from over 50 countries, but he also listens to the top 40 at work. He is an avid drummer and plays guitar. When DJing, he creatively mixes samples from anything from hip hop to the Beatles and describes his music taste as "eclectic but with a good ear for quality sounds." Michelle is a music writer. Michelle is always listening to music. She reviews albums for UK and USA music magazines, interviews up-and-coming artists and often has access to music well before the general public. She describes her music taste as very open, and listens to a wide variety. She likes new and independent artists, but admits that she also listens pop music while out and about in town and with friends.
Subjects were asked to rate each reviewer from 1 (not at all) to 7 (very much) for how much the person could be trusted to pick music that the subject would like. No comparisons were made between experts during the study. Descriptions were created to communicate a degree of expertise across a broad range of popular music tastes.
Subjects were informed that the two experts had listened to the 20 songs and provided reviews for each. Reviews were preferences between each of the 20 subject-provided songs and an alternative song, provided by the experimenter. Each subject-provided song was reviewed six times (relative to six different alternative songs). Subjects received instructions for the task and answered a series of questions to confirm that their task was understood. Each subject confirmed that they believed the reviews were real.

Task and Timing
The task was programmed and run using Presentation v.12 (Neurobehavioural Systems). Visual displays were back-projected to a display in the scanner. Subjects viewed the displays via a mirror placed above their eyes. Responses (from the right hand) were collected using by two fibre-optic button boxes.
Each trial (see Figure 1) began with a choice for the subject. We presented subjects with two songs at the top of the screen. One was a song that the subject provided.
The other was an alternative, provided by the experimenter. The alternative was a Canadian or Scandanavian pop song, which was real but unknown to the subject (confirmed after the scan session). Song choices were randomly assigned to the left and right side of the display. Pictures of the experts were arranged vertically down the centre of the display. A picture of the subject appeared at bottom of the screen, beneath the expert pictures. The words "I prefer" were placed under each photo.
The subject"s task was to move their own picture beneath the song they desired the most. Subjects pressed the left button to move their picture left, or the right button to move it right. A scrambled picture of the subject was placed under the song they did not choose. Subject-provided songs appeared equally-often on left and right sides of the display. Subjects were told that the song that they chose had a slightly (less than 5%) higher chance of being chosen for a token at the end of the trial to provide motivation to pick their real preference. Each song actually had a 50% chance of being chosen. Subjects knew that the songs with the most tokens at the end of the task were to be purchased for them and placed on a CD. There was a time limit of 2 seconds to make a choice. If no choice was made, a large "X" appeared on the screen for the remainder of the trial. This phase is termed the "review outcome." Next, the songs alternately changed color between green and white (every 50ms, for 1s). Finally, a song was chosen for a token and appeared at the bottom of the screen. This phase was the "object outcome." Review outcomes were completely independent from object outcomes. During instruction, subjects confirmed that expert choices did not predict which song token would be received. The subject received a token for each of their submitted songs as often as they received an alternative. The order of trials was optimized to provide maximum efficiency for detection of Blood Oxygenation Level Dependent (BOLD) activity related, independently, to different review and object outcomes. For these purposes, it was not possible to use real expert reviews, and confederate reviews were used in their place. Each participant confirmed that they believed the reviews were real. As a result, trials could be placed close together in time with a brief minimum of 3 seconds between each modeled event (see section on fMRI analysis) reducing subject time in the scanner but still controlling for nonlinearities of the BOLD signal [20].
Decisions appeared at time 0 of each trial. Review outcomes appeared at 3 seconds, and songs began to flash at 4 seconds. Object outcomes were presented at 5 seconds and remained on display for 2 seconds. A fixation cross was displayed for 2 seconds between each trial.
There is no non-social equivalent, to our knowledge, to a human opinion. Even a "computer" from which one could make accurate inferences of subjective human value only acts as an indirect inference of the human opinions used to program it, and thus computer reviews would still remain "social." For this reason, we saw little merit in providing an artificial "non-social" control in this study.

Post-scanning
After completing the task, subjects rated each of their 20 songs for desirability for a second time. Subjects were also asked if they had learned more about the reviewers or more about the songs. The 10 songs for which the subject had the most tokens (from the object outcome of the task) were purchased for the subject.

Conditions
Only trials in which subjects chose the same song as they had provided a week prior were included in the analysis to prevent analysis of subject errors. Key independent variables were: 1. Review outcome i. RS (experts chose the subject-preferred song) ii. RA (experts chose the alternative) iii. RSPLIT (split; one expert chose the subject-preferred song; the other chose the alternative).

2.
Object outcome: i. S (subject gains a token for their preferred song) ii. A (subject gains a token for the alternative song).
These variables formed a 2*3 design matrix (Figure 1) Figure S1). Binf was used as a between-subject regressor for subsequent fMRI analysis. A linear regression was also performed to test the effect of unanimous review frequency (which varied from 4 to 7 per song) on change in a song value. Gender effects on Binf were tested with an independent samples t-test.

Functional Magnetic Resonance Imaging (fMRI)
Standard fMRI acquisition, and preprocessing were used in this study.

Acquisition
Scanning took place at the Wellcome Trust Centre for Neuroimaging in London, UK.
Subjects were scanned at 3 Teslas with a Siemens MAGNETOM Trio scanner (Siemens Medical Solutions, Erlangen, Germany) fitted with a 12-channel head coil.
Field maps were acquired with a standard double echo gradient echo field map sequence (TE, 10.0 and 12.46 ms), using 64 slices covering the whole head (voxel size, 3*3*2 mm with 1mm gap between slices). Functional data was collected as T2-weighted echo planar images (EPI) in descending slice acquisition order. Each volume (voxel size: 3*3*3mm; TE, 30 ms; TR, 3360ms) contained 48 slices, covering the whole brain. BOLD sensitivity losses in the orbitofrontal cortex due to susceptibility artifacts were minimized by applying a z-shim gradient moment of -1.4 mT/m*ms, a slice tilt of -30°, and a positive PE gradient polarity [21]. 176-slice whole-brain anatomical scans (matrix, 256*256; 1mm slice thickness; TE, 2.48ms; TR=7.92ms; flip angle, 16 o ; TI=910ms) were acquired using a modified driven equilibrium Fourier transform (MDEFT) sequence with optimized parameters as described previously [22] for co-registration with the EPI data. Images were reconstructed by performing a standard 3D Fourier Transform, followed by modulus calculation. No data filtering was applied in k-space or in the image domain.

Preprocessing
Image unwarping, and motion correction was performed using statistical parametric mapping (SPM 5; Wellcome Trust Centre for Neuroimaging; www.fil.ion.ucl.ac.uk/spm) on Matlab (version 7.1, MathWorks). EPI images were generated off-line from the complex k-space raw data using a generalized reconstruction method based on the measured EPI k-space trajectory to minimize ghosting. They were then corrected for geometric distortions caused by susceptibilityinduced field inhomogeneities. A combined approach was used which corrects for both static distortions and changes in these distortions due to head motion [23,24].
The static distortions were calculated for each subject from a field map that was processed using the FieldMap toolbox as implemented in SPM5. Using these parameters, the EPI images were then realigned and unwarped with a procedure that allows the measured static distortions to be included in the estimation of distortion changes associated with head motion. The remaining preprocessing and was carried out with the FMRIB"s Software Library (FSL) version 5.63 [25]. Brain matter was segmented from non-brain using a mesh deformation approach [26]. High pass temporal filtering was applied using a Gaussian-weighted running lines filter, with a cut-off of 50s [27]. Each volume was smoothed with a Gaussian filter (full-width halfmaximum of 5mm). Independent Component Analysis was used to visually identify and remove artifacts in the data using Multivariate Exploratory Linear Optimized Decomposition into Independent Components (MELODIC) software [28].

Single Subject General Linear Models and fMRI Analysis
Modeling and statistical analysis of fMRI data was carried out with the FEAT (FMRI Expert Analysis Tool, www.fmrib.ox.ac.uk/fsl) version 5.63 [25]. A standard general linear model (GLM) was used for individual subject analyses. The GLM was fit in prewhitened data space (to account for autocorrelation in the FMRI residuals [29]).
Regressors corresponding to each condition of the 2*3 design matrix (Figure 1) (plus their temporal derivatives) were included in the model as stick functions placed midway through the "object outcome" display period. Decisions, trials in which subjects took longer than 2s to respond, and trials in which the subject chose the alternative song (i.e. not their pre-submitted song) were included in the model as independent regressors but not used in further analysis. Regressors were convolved with the FSL default haemodynamic response function (HRF, gamma function, delay = 6s, standard deviation = 3 s). High-pass temporal filtering (50s) was also applied to the regressors. GLM results were estimated [30] and transformed, after spatial normalization, into standard (MNI152) space [27].

Single Subject Contrasts
The following contrast images (and their inverse contrasts) were generated from the GLMs.

Group-level Analysis
Group-level analysis was carried out with FLAME 1+2 (FMRIB's Local Analysis of Mixed Effects [30]. All subjects were modeled as a single group. A GLM was fit to the effects of the contrasts described above. This was done in two separate group analyses: 1. Group mean 2. Group mean + a between subject regressor of Binf All Z statistic maps were cluster corrected (contiguous clusters defined by Z > 2.3) with a whole brain cluster significance level of p < 0.05 [31][32][33]. To note the subthreshold bilateral effect of "review outcome" in the ventral striatum ( Figure 2B), this contrast was also analyzed with contiguous clusters defined by voxels Z > 2.0, p < 0.05 (cluster corrected).

Further Investigation into the Nature of Ventral Striatum Responses to Agreement
The fact that some of the individuals had a negative Binf ( Figure S1) and changed their value of songs in the opposite direction to net reviewer opinion made the mean group activation in ventral striatum with respect reviewer agreement more challenging to interpret as a reward response. If the ventral striatum signaled a reward with agreement, one could propose that the effect would be stronger or perhaps only present in those subjects with positive Binf values (i.e. those subjects whose opinions of the songs conformed to the opinions of reviewers). This is based on the assumptions that (a) if subjects conform their opinions to those of reviewers, their motivation is derived from the presence of a reward from agreement and (b) those that do not conform do not experience a reward from an agreement. We tested if those with negative Binf values still produced a ventral striatum response with agreement with experts (whole brain, Z > 2.3, p < 0.05 cluster corrected). Seven participants reduced their subjective value of a song as the number of positive reviews of that song increased (negative Binf). Group analysis using just these subjects still produced a ventral striatum response to "review outcome" [RSS+RSA]- [RAS+RAA] (right peak: 6, 14, -6; left peak -6, 16, 2) (Table S1). This result is discussed in the main text.
Testing the Potential Impact of Unanimous Reviews on Object value Lateral orbitofrontal cortex / anterior insula cortex activity during unanimous reviewer opinion relative to split opinions might be interpreted as mediating the impact of these unanimous reviews on song values. To test this interpretation, we completed regression analyses to see if the subject"s change in song value (and, separately, the absolute value of that change in song value) varied as (a) a function of the number of unanimous reviews received for each song and (b) a function of the net reviewer opinion of the song weighted by the number of unanimous reviews that made up that net reviewer opinion. No significant effect of unanimous reviews were observed in either case (Fs(1,19) < 0.53, p"s > 0.47). We also tested if BOLD activation from unanimous review outcomes (relative to split review outcomes; single subject contrasts 4 and 5) varied between subjects with tendencies to be influenced by reviewer opinion (Binf). Again, no significant relationship was observed. These results are discussed in the main text.
Testing the Relationship between Disagreement Activity and Influence Activity We tested if activations during disagreement with the experts (relative to agreement) ( Figure 3) predicted changes of ventral striatum activity due to social influence on object value (Figure 4) in the same subjects. Each active cluster correlating with Binf during disagreement was converted to a mask. These masks (right insula / central opercular cortex, left insula, dorsal anterior cingulate cortex, left prefrontal cortex, right prefrontal cortex, temporoparietal junction) were used to calculate the percent signal change within each of these regions, for each participant, during disagreement with the experts (relative to agreement with experts). The same was done to calculate the percent signal change for each subject within the area defined by Figure 4"s cluster resulting from the interaction between review outcome and object outcome activity (influence). We then completed a regression analyses to see if the percent signal change within any cluster during disagreement with reviewers predicted the percent signal change in ventral striatum due to social influence on object value. No significant relationships were observed.

Gender Effects
No significant effect of gender was observed on susceptibility to influence (Binf) (t(26) = 0.647; p < 0.543). Likewise, no effect of gender was observed in reported BOLD activations when entered as a between-subject factor.