Activity in the brain’s valuation and mentalizing networks is associated with propagation of online recommendations

Word of mouth recommendations influence a wide range of choices and behaviors. What takes place in the mind of recommendation receivers that determines whether they will be successfully influenced? Prior work suggests that brain systems implicated in assessing the value of stimuli (i.e., subjective valuation) and understanding others’ mental states (i.e., mentalizing) play key roles. The current study used neuroimaging and natural language classifiers to extend these findings in a naturalistic context and tested the extent to which the two systems work together or independently in responding to social influence. First, we show that in response to text-based social media recommendations, activity in both the brain’s valuation system and mentalizing system was associated with greater likelihood of opinion change. Second, participants were more likely to update their opinions in response to negative, compared to positive, recommendations, with activity in the mentalizing system scaling with the negativity of the recommendations. Third, decreased functional connectivity between valuation and mentalizing systems was associated with opinion change. Results highlight the role of brain regions involved in mentalizing and positive valuation in recommendation propagation, and further show that mentalizing may be particularly key in processing negative recommendations, whereas the valuation system is relevant in evaluating both positive and negative recommendations.


Supplementary information Participant Exclusions
Forty participants (28 females) between the ages of 18 and 24 (M = 20.9, SD = 2.1) were recruited for a single three-hour study appointment, incorporating a one-hour fMRI scan.
Participants met standard fMRI eligibility criteria, including being right-handed, not currently taking any psychoactive medications, no history of psychiatric or neurological disorders, not currently pregnant, no metal in their body contraindicated for MRI, and not suffering from claustrophobia. All runs from one participant and run 3 from three participants were excluded due to data corruption. Further, all runs from one participant and one run from one participant were excluded due to excessive movement. This resulted in thirty-eight participants included for analysis, with partial data from four participants.
Additional information about data acquisition Due to scheduling issues at our scanner center, 33 participants were scanned on a TIM Trio scanner, and the remaining 7 on a Prisma scanner. Models controlling for scanner type showed no significant or meaningful differences from those reported. The fMRI scan for the task of interest was obtained as part of a larger study that included another task. In the current paper, we focus on a single task (the App Rating Task) that spanned 3 runs. Participants completed all 3 runs of the App Rating Task first, and then completed the second task (more information on the second task obtained as part of the same protocol can be found on: https://github.com/cnlab/article_sharing_task).
Due to hypotheses not of interest in the current investigation, each participant read 40 written recommendations from one peer reviewer who had high ego-betweenness centrality and 40 written recommendations from one peer reviewer who had low ego-betweenness centrality.
Main results using human-coded sentiment scores To validate our machine-learning sentiment classifier, we ran additional models that parallel the main analyses using the human-coded sentiment scores. We found that our results that use human-coded sentiment scores largely support the findings using the machine-learning sentiment classifier that we report in the main manuscript.
Behavioral data analysis. We defined recommendation rating change in an analogous manner to how we defined it in the main manuscript. Accordingly, we defined recommendation rating change as being positive (+1) if the participant changed their initial ratings in the direction of the sentiment of the peer recommendation, negative (-1) if the participant changed their initial ratings away from the sentiment of the peer recommendation, and zero (0) if participants did not change their ratings. For this purpose, peer recommendations were classified into binary categories as either "positive" or "negative" by using the sentiment scores produced by the human coders, which ranged from 0-100 (0 being the most negative and 100 being the most positive). Thus, if the human-coded sentiment scores indicated that the recommendation was more likely to be positive than negative (>50), then it was categorized as positive (and vice versa). Thus, if participants changed their initial recommendation of a "5" to a final recommendation rating of a "3" after reading a peer recommendation that was classified as "positive", then the recommendation rating change was calculated as "+1". Paralleling the method that we used in the main manuscript, to determine the relationship between peer recommendation sentiment scores and participants' recommendation rating change, we ran a multi-level linear regression predicting the participants' recommendation rating change from the sentiment scores of the peer recommendations: recommendation rating changeij = B0 + B1sentimentij + μ0i + ν0j + ϵij, where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between human-coded sentiment score and recommendation rating change; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants. Participants and mobile apps were treated as random effects with intercepts allowed to vary randomly, accounting for nonindependence in the data due to repeated measures from each participant.

Recommendation rating change and sentiment (Human-Coded)
Paralleling the main results, participants changed their ratings in alignment with the human-coded sentiment of the peer recommendations (B = 0.012, t (2240) = 18.51; p <.001), and the effects were greater for peer recommendations higher in negativity than positivity (B = -0.003, t (1972) = -7.018, p <.001).
Brain activity and sentiment. We ran analyses examining whether the neural activity in the mentalizing and value also correlated with human coded sentiment scores: mean brain activityij = B0 + B1sentimentij + μ0i + ν0j + ϵij, where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between human-coded sentiment score and brain activity; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target regions of interest, with separate models run for mentalizing and valuation systems.
Paralleling results in the main manuscript, we found that the relationship between mean activity in the mentalizing regions and human coded sentiment scores was marginally significant (B = -0.004, t (2013) = -1.650, p = 0.099), and that the relationship between mean activity in the valuation regions and human coded sentiment scores was not significant (B = 0.002, t (1491) = 0.825, p = 0.409).

Brain activity and recommendation rating change
We next ran analyses examining whether neural activity in the mentalizing and value systems correlated with trials where participants changed their initial ratings toward that of the peers, using the recommendation rating change variable calculated from human coded sentiment scores: where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between brain activity and recommendation rating change; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target regions of interest, with separate models run for mentalizing and valuation systems.
We found that mean activity in the mentalizing and value regions was associated with recommendation rating change, though the relationship with mentalizing was marginal We next ran analyses predicting recommendation rating change from the interaction of the human coded sentiment scores and mean brain activity: recommendation rating changeij = B0 + B1brain activity + B2sentiment + B3brain where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between brain activity and recommendation rating change, B2 is an unstandardized regression coefficient capturing the average slope of the relationship between human-coded sentiment scores and recommendation rating change, B3 is an unstandardized regression coefficient capturing the average slope of the interaction effect of brain activity and sentiment on recommendation rating change; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target regions of interest, with separate models run for mentalizing and valuation systems).
We found a directional trend of an interaction between the sentiment of the review and neural activity in the mentalizing system in predicting recommendation rating change (see Table   S1 below), but not an interaction between the sentiment of the review and neural activity in the valuation system in predicting recommendation rating change (see Table S2 below). To complement our main results that extracted percent signal change in all the voxels of our mentalizing and valuation ROIs (as defined from Neurosynth), we ran the same analyses using the subregions in each of the ROIs. We used regions.connected_regions from the nilearn package in Python 3 42 to extract 10 contiguous clusters from the mentalizing network and 2 contiguous clusters from the valuation network. We then repeated the main analyses as described in the Methods section of the main manuscript. We first examined whether neural activity in each of our subregions was associated with the sentiment of the peer recommendations (Tables   S3 and S4). where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between sentiment and brain activity; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target ROI. where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between sentiment and brain activity; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target ROI.
We next examined the relationship between subregions of the mentalizing and valuation ROIs and recommendation rating change (see Tables S5 and S6). is an unstandardized regression coefficient capturing the average slope of the relationship between brain activity and recommendation rating change; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target ROI). where B0 is the overall intercept, representing the grand mean across all observations, B1 is an unstandardized regression coefficient capturing the average slope of the relationship between brain activity and recommendation rating change; subscript i refers to participant, j refers to app, and μ0i and ν0j represent the random errors for the deviation of the mean intercept for each participant and app from the grand mean intercept, respectively, and εij is the random error for each app rating within participants; "brain activity" represents activity in the target ROI).