The persuasion network is modulated by drug-use risk and

The persuasion network is modulated by drug-use risk

and predicts anti-drug message effectivenessRichard Huskey1 J Michael Mangus2 Benjamin O Turner3

and Rene Weber2

1School of Communication Cognitive Communication Science Lab The Ohio State University OH 43210 USA2Department of Communication Media Neuroscience Lab University of California Santa Barbara CA 93106USA and 3Wee Kim Wee School of Communication and Information Nanyang Technological University637718 Singapore

Correspondence should be addressed to Rene Weber Department of Communication University of California Santa Barbara 4405 Social Sciences amp MediaStudies Santa Barbara CA 93106-4020 USA E-mail renewcommucsbedu

Abstract

While a persuasion network has been proposed little is known about how network connections between brain regions contributeto attitude change Two possible mechanisms have been advanced One hypothesis predicts that attitude change results fromincreased connectivity between structures implicated in affective and executive processing in response to increases in argumentstrength A second functional perspective suggests that highly arousing messages reduce connectivity between structuresimplicated in the encoding of sensory information which disrupts message processing and thereby inhibits attitude changeHowever persuasion is a multi-determined construct that results from both message features and audience characteristicsTherefore persuasive messages should lead to specific functional connectivity patterns among a priori defined structures withinthe persuasion network The present study exposed 28 subjects to anti-drug public service announcements where arousal argu-ment strength and subject drug-use risk were systematically varied Psychophysiological interaction analyses provide supportfor the affective-executive hypothesis but not for the encoding-disruption hypothesis Secondary analyses show that video-levelconnectivity patterns among structures within the persuasion network predict audience responses in independent samples(one college-aged one nationally representative) We propose that persuasion neuroscience research is best advanced by consid-ering network-level effects while accounting for interactions between message features and target audience characteristics

Key words persuasion functional connectivity elaboration likelihood model fMRI public service announcements

Introduction

By most accounts the modern scientific investigation of per-suasion dates back to the Second World War when scholarssuch as Lasswell Hovland and Lazarsfeld sought to understandhow propaganda influenced attitudes and motivated behaviorchange (Rogers 1994) Since eight decades of research haveyielded a nuanced understanding of the mechanisms thatunderlie the persuasion process Indeed a large body of workhas illuminated how structural message characteristics and

argumentation contribute to message persuasiveness amongdifferent audiences (eg Petty and Cacioppo 1986)

However only recently have investigators sought to under-stand the neural basis of persuasive message processing Oneof the earliest studies used functional magnetic resonanceimaging (fMRI) to evaluate the ways in which American andKorean audiences evaluated text- and video-based persuasivemessages (Falk et al 2009) The results demonstrated that struc-tures commonly implicated in social affective and executive

Received 6 July 2017 Revised 21 August 2017 Accepted 16 October 2017

VC The Author (2017) Published by Oxford University PressThis is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (httpcreativecommonsorglicensesby-nc40) which permits non-commercial re-use distribution and reproduction in any medium provided the original work is properly citedFor commercial re-use please contact journalspermissionsoupcom

Social Cognitive and Affective Neuroscience 2017 1902ndash1915

doi 101093scannsx126Advance Access Publication Date 11 November 2017Original article

Downloaded from httpsacademicoupcomscanarticle-abstract121219024617753by Ohio State University-Moritz Law Library useron 02 December 2017

processing are recruited while evaluating persuasive messagesToday the literature has developed such that a structural con-sensus is emerging A synthesis of 20 persuasion-neurosciencestudies implicated structures involved in reward processingself-reflection social pain executive processing and saliencedetection as key components of the so-called persuasion network(Kaye et al 2016)

Despite reference to a persuasion network the majority ofwork in this domain has in fact ignored network structure andhas instead used the traditional mass-univariate approach thataddresses questions of how much activity is elicited within dif-ferent brain regions (but not how those regions interact withone another Friston 1994 2011)

However cognitive neuroscience has increasingly come torecognize that a complete understanding of how psychologicalprocessess (including persuasion) are instantiated in the brainrequires studying the brain at the network level (Bassett andGazzaniga 2011 Bullmore and Sporns 2012 Petersen andSporns 2015 Weber et al 2015 Bassett and Sporns 2017Cooper et al 2017) Accordingly prominent calls have forcefullyadvocated that our understanding of the neural basis of psycho-logical processes can only advance by moving beyondlsquoblobologyrsquo (eg Poldrack and Farah 2015) and adopting a net-worked perspective Applying this perspective to the persuasionnetwork requires analytical procedures that illuminate thetask-modulated relationship between one network structureand another (Friston 2011) To date a small handful of persua-sion neuroscience studies have addressed this issue using psy-chophysiological interaction analyses (PPI Friston et al 1997)Among these two (not necessarily incompatible) functionalviews of the network connectivity patterns that underpin per-suasion are beginning to emerge Each is briefly summarizedbelow

Affective-executive hypothesis

Ramsay et al (2013) proposed a mechanism where evaluationsof persuasive messages are modulated by affective and execu-tive processing In their study among both non-drug and drugusers strong argument strength (AS) persuasive messagesyielded increased functional connectivity between left inferiorfrontal gyrus (lIFG a putatively lsquoexecutiversquo region) and theamygdala (AMY a putatively lsquoaffectiversquo region) and insula com-pared to weak AS messages A related study among smokersobserved a negative relationship between structures commonlyimplicated in affective and executive processing in response toaversive anti-smoking (and therefore self-relevant) imagescompared to aversive but not self-relevant images (Dinh-Williams et al 2014) Specifically a seed region of interest (ROI)in the lIFG showed negative connectivity with the anterior cin-gulate cortex (ACC) medial prefrontal cortex (MPFC) precuneusand insula This connectivity pattern suggests that issueinvolvement modulates the way in which audiences process apersuasive message such that smokers engage in defensiveprocessing in an attempt to down regulate negative feelingsassociated with a threatening message Unfortunately thisstudy did not include a non-smoking comparison group so itis difficult to fully integrate this finding with the study byRamsay et al Nevertheless these two results account for keyvariables known to modulate persuasion specifically AS andaudience issue-involvement (Petty and Cacioppo 1986) andshow that each variable contributes to different neural process-ing patterns

Similarly a study by Cooper et al (2017) examined connectiv-ity patterns in sedentary individuals in response to (individuallyrelevant) messages promoting physical activity and found thatconnectivity strength between a subregion within the MPFCand ventral striatum (VS) and AMY during message exposurepredicted decreases in sedentary behavior Taken togetherthese three studies provide converging evidence that affective-executive processing underlies attitude and behavior change

Encoding-disruption hypothesis

The previous studies evaluated variables such as AS and issueinvolvement At the same time message sensation value (MSVa measure of message arousingness or intensity) has also beenhypothesized as a mechanism that can disrupt cognitive proc-essing thereby limiting the extent to which a persuasive mes-sage is encoded and ultimately recalled (eg Lang 2006)Experimental fMRI work has shown that low-compared to high-MSV persuasive ads are better remembered elicit increasedactivity in temporal and prefrontal cortex and correspond tolower levels of occipital cortex activation (Seelig et al 2014)

The same study also tested the hypothesis that increases inMSV disrupt message encoding and therefore should result indecreased functional connectivity between structures impli-cated in sensory processing and message encoding The resultswere ambiguous in that a negative relationship was observedbetween inferior lateral occipital cortex (iLOC) and the middletemporal (MTG) and inferior frontal gyrus (IFG) regardless ofvariation in MSV although this negative relationship wasweaker when MSV was low One possible explanation for thisfinding is that audience characteristics modulate audience reac-tions to MSV For instance Wang et al (2014) show thatincreases in MSV lead to approach tendency among high-riskaudiences whereas low-risk audiences show avoidanceresponses

The present study

As these studies indicate and a large body of theoretical andempirical evidence demonstrates issue involvement modulatespersuasive message processing (Petty and Cacioppo 1986) Theliterature shows that being at-risk for drug use constitutes aform of issue involvement where those that are at higher-riskfor using a drug hold stronger counter-attitudinal views whenpresented with an anti-drug message compared to those thatare at lower-risk (Cappella et al 2003) This increases the proba-bility that at-risk and therefore issue-involved audiencesengage in defensive message processing (ie counterarguing)when presented with a counter attitudinal message This defen-sive processing makes it difficult to use self-report measures toidentify which messages are likely to result in attitude or behav-ior change (Weber et al 2013 2014) Identifying what messagefeatures contribute to attitude change among this audience iscritically important

Unfortunately our understanding in this area is incompleteOf the four fMRI PPI studies to-date two only evaluate at-riskparticipants (Dinh-Williams et al 2014 Cooper et al 2017) Theremaining two studies (Ramsay et al 2013 Seelig et al 2014)include both not-at-risk and at-risk participants in their sam-ples however they do not consider the different ways each riskgroup processes persuasive messages Furthermore each ofthese studies has focused narrowly on just one aspect of per-suasive messages (eg MSV AS)

R Huskey et al | 1903

Importantly theoretical and empirical evidence demon-strates that each of these variables interacts with issue involve-ment which leads to different message evaluations (Weberet al 2013) These different evaluations correspond to differen-ces in neural processing (Weber et al 2014) with real-worldimplications Specifically neural activity in the middle frontalgyrus (MFG) and superior temporal gyrus (STG) among issue-involved audiences is predictive of message evaluations amongindependent samples

Moreover not one of these four studies has been subjectto empirical replication In this study we attempt to replicateand extend the encoding-disruption and affective-executivehypotheses by examining the degree to which they are influ-enced by audience issue involvement Should the affective-executive hypothesis be true we expect that a seed ROI in theMPFC will show network connections with AMY and VS (Cooperet al 2017) while a seed ROI in the IFG should show connectivitywith AMY ACC MPFC precuneus and insula (Ramsay et al2013 Dinh-Williams et al 2014) particularly when argumentstrength or self-relevance is high By comparison the networkmodel proposed by the encoding-disruption hypothesis indi-cates that a seed ROI in the iLOC should show stronger connec-tivity with the MTG and IFG when MSV is low

Accordingly our study interrogates network models pro-posed in the literature while extending these models to includeseed ROIs that have shown to differ by issue involvement (ieMFG STG Weber et al 2014)

This extension is motivated by the fact that we currentlylack a firm understanding of how message and audience char-acteristics modulate functional connectivity within the persua-sion network and how this network connectivity contributesto message evaluation let alone attitude or behavior changeFurther and as others have pointed out elsewhere much of thepersuasion neuroscience literature is data-driven in nature(Vezich et al 2016) which adds additional complications whentrying to map neural processing to specific cognitive functionsTo address these gaps we draw on the Elaboration LikelihoodModel (ELM) which argues that message features (ie MSV AS)interact with issue-involvement to modulate persuasive mes-sage processing (Petty and Cacioppo 1986) To test these rela-tionships subjects at high- and low-risk for drug use wereexposed to anti-drug public service announcements (PSAs) thatvaried systematically along MSV and AS dimensions We showthat (i) neural processing among high- and low-risk audiencescorresponds to distinct network connectivity patterns and (ii)that these patterns provide support for the affective-executivehypothesis but not also for the encoding-disruption hypothesisFurthermore we modify the brain-as-predictor approach(Berkman and Falk 2013) to identify how brain-network pat-terns for individual PSAs predict of message perceptionsResults from this network-as-predictor analysis demonstratethat variation in connectivity at the individual video level is pre-dictive of perceived anti-drug messagesrsquo effectiveness in twoindependent samples (one college age one nationally represen-tative) further suggesting that there is utility in examining theneural underpinnings of persuasion at the network level

Methods

Previous reporting and general method

The subjects materials measures experimental proceduresfunctional magnetic resonance imaging (fMRI) scanner parame-ters and pre-processing pipeline used in this study are reported

by Weber et al (2014) Briefly the imaging data reported hereinwere pre-processed using FEAT (fMRI Expert Analysis Tool v60)from the Oxford Center for Functional MRI of the Brain (FMRIB)Software Library (FSL v50) using a standard pipeline (Weberet al 2015) The study adopted a 2x2x2 mixed factorial designwhich manipulated AS (highlow) and MSV (highlow) aswithin-subjects factors and subject drug-use risk (highlow) asbetween-subjects factors A total of 28 subjects split evenlyamong highlow drug-use risk (assessed using the risk for mari-juana use scale Cappella et al 2003) were exposed to 32 30-sec-ond anti-drug PSAs (drawn from the Annenberg School forCommunication at the University of Pennsylvania antidrug PSAarchive) and eight control conditions four of which featuredPSAs where the video and audio had been scrambled (therebypreserving luminosity auditory levels and low-level visual fea-tures such as cuts but removing semantic meaning) and fourthat showed a black screen with no visual or auditory featuresPSAs and control conditions were split evenly between two runsin a counterbalanced order

Subjects in the fMRI sample (henceforth referred to as MRIS)were drawn from the undergraduate research pool at Universityof California Santa Barbara M(sd)agefrac14 203(15) Subjects didnot show any contraindication to fMRI scanning All proceduresconformed to the Declaration of Helsinki and were approved bythe Universityrsquos Institutional Research Board Subjects watchedeach PSA while undergoing fMRI scanning Participants pro-vided a continuous response measure (CRM not reported in thismanuscript) of message convincingness (not at all convincingmdashvery convincing) during the scanning session After the fMRIscan was complete subjects were then exposed to the PSAs fora second time in a session that took place outside of the scan-ner Self-report measures were collected during this sessionSubjects rated each PSA on perceived argument strength (pASZhao et al 2011) and perceived message sensation value (pMSVKang et al 2006)

Self-reported ratings of message effectiveness were eval-uated using the perceived message effectiveness (PME) scalewhich has been shown to correlate highly with actual PSA effec-tiveness and behavior change (Dillard et al 2007ab Bigsbyet al 2013) These PME ratings were drawn from two independ-ent studies where subjects rated PME for each PSA The firstsample (henceforth referred to as IS1) was among 599 collegestudents M(sd)agefrac14 1955(212) 738 female (Weber et al2013) The second sample (IS2) was nationally representative ofadolescents in the United States (Kang et al 2006) This sample(nfrac14 601 Mage frac14 1530 range 12ndash18) was balanced across malesand females

PPI analysis

Whole-brain functional connectivity patterns for a set of struc-tures in the persuasion network were evaluated using a seriesof PPI analyses (Friston et al 1997) As discussed above the seedROIs used in this study were selected a priori based on theempirical literature Seed ROIs were constructed in FSL by draw-ing a binary 5 mm sphere (in MNI152 space) around coordinatespreviously reported in the literature and included the medialprefrontal cortex (MPFC 4 56 4 Falk et al 2016) middlefrontal and superior termporal gyri (MFG 38 30 34 STG 56 418 Weber et al 2014) IFG (46 28 12 Ramsay et al 2013) andinferior lateral occipital cortex (iLOC 32 70 16 Seelig et al2014) Time series data (PHYS) were extracted from filteredfunctional data for each subject for each run within each seedROI by taking the arithmetic mean of all voxels within the

1904 | Social Cognitive and Affective Neuroscience 2017 Vol 12 No 12

corresponding binary mask First-level PPI general linear models(GLMs) were then fit separately using each seed ROI Followingestablished guidelines (Cohen et al 2003) each first-level GLMincluded explanatory variables (EVs) that encoded experimentalconditions (ie MSV HighgtMSV Low AS HighgtAS Low) controlconditions (C1frac14 scrambled C2frac14black screen) a PHYS EV andtwo- and three-way condition by PHYS interaction terms (equa-tion 1)

PPI frac14 b1MSV thorn b2ASthorn b3MSVxAS thorn b4PHYSthorn b5MSVxPHYSthorn b6ASxPHYSthorn b7MSVxASxPHYS thorn b8C1thorn b9C2thorn b10C1xPHYSthorn b11C2xPHYSthorn b0 thorn (1)

Following a hybrid generalized PPI approach (gPPI McLarenet al 2012) contrasts were constructed between interaction EVcoefficients (eg ASPHYSgtC1PHYS) Importantly while thefull model was estimated we deliberately do not consider theMSVASPHYSgtC1PHYS contrast The reason is that thesign of a three-way interaction does not have the same straight-forward interpretation as a two-way interaction (where a posi-tive sign indicates an increase and a negative sign a decrease inconnectivity) This renders any contrast of a three-way interac-tion EV coefficient nearly uninterpretable against a two-wayinteraction EV coefficient Still the analysis and results areappropriate given that we are replicatingextending existinghypotheses (ie affective-executive encoding-disruption) andtheir analytical paradigms

A second-level analysis then combined runs for each subjectin a fixed effects model Finally the interaction terms were ana-lyzed in a series of third-level mixed effects models (FLAME 1Beckmann and Smith 2004 Woolrich et al 2004) Subject drug-use risk (highlow) was specified as two groups in FSLrsquos FEATtool (which lets variance estimates differ between groups) andthen encoded using indicator variables Contrasts at the third-level assessed mean functional connectivity by risk group aswell as in the high-riskgt low-risk and low-riskgthigh-risk con-trasts A cluster-based procedure was applied to correct formultiple comparisons (Worsley 2001) with a cluster definingthreshold of Zfrac14 23 and a cluster extent threshold of P lt 05 Allstructural localizations were identified using FSLrsquos probabilisticatlases and cross referenced with the Neurosynth database(Yarkoni et al 2011)

Network-as-predictor analysis

The previous section described a classic PPI analysis wheretask-modulated neural activity in a seed ROI is fit to every othervoxel in the brain Our network-as-predictor analyses had dif-ferent aims and a different analytical approach Here wewanted to identify network connectivity patterns specific toeach PSA and the degree to which variation in these patternswas predictive of PME in two independent samples To dothis we modified the brain-as-predictor analytical paradigm(Berkman and Falk 2013) Whereas earlier brain-as-predictorstudies used signal change in seed ROIs for each video (eg Falket al 2012 Weber et al 2014) to predict message effectivenessin independent samples we used connectivity parameter esti-mates (PEs) between seed and target ROIs Our seed ROIs werethe same as those reported above in the PPI Analysis sectionwhile target ROIs for the encoding-disruption and affective-executive hypotheses were selected based on previous empiri-cal results and therefore represent connectivity patterns thathave been reported in the literature A full list of seeds and theircorresponding target ROIs can be found in Table 1

Said differently this seed-target ROI-based analysis exam-ined how each PSA video modulated functional connectivityamong a-priori defined structures within the persuasion net-work (for subjects in the MRIS sample) and the extent to whichthis variation was predictive of PME-ratings for each PSA in twolarge independent samples (IS1 IS2) To that end a series ofgPPI models were constructed for each of the seed ROIs definedabove The first-level models (one model for each subject foreach run for each seed ROI) included EVs that encoded each PSAvideo and control condition (C1frac14 scrambled C2frac14black screen)a PHYS EV and then interaction terms (Equation 2) Contrasts atthe first-level followed a gPPI logic and were defined asVideoNPHYSgtC1PHYS

PPI frac14 b1Video1thorn b2Video2thorn thorn b16Video16thorn b17C1thorn b18C2thorn b19PHYSthorn b20Video1xPHYSthorn b21Video2xPHYSthorn thorn b35Video16xPHYSthorn b36C1xPHYSthorn b37C2xPHYSthorn b0thorn

These contrasts of parameter estimates were then carried for-ward into two fixed-effects second-level analyses (one analysisfor each run) At the second-level indicator variables encodedsubject drug-use risk (highlow) Planned contrasts evaluatedfunctional connectivity for each PSA by risk group A 5 mmbinary spherical mask was drawn for each a-priori defined tar-get ROI Finally FSLrsquos Featquery tool was used to extract medianfunctional connectivity parameter estimates (PEs) between seedand target ROIs for each PSA for each of the second-level

Table 1 Seed and Target ROIs used in the network-as-predictoranalysis

Structure Laterality Coordinates

Middle frontal gyrus Right 38 30 34Superior parietal lobe Left 40 48 48Angular gyrus Right 46 46 38Inferior parietal lobe Left 44 52 52Superior lateral occipital cortex Left 30 66 42Superior lateral occipital cortex Right 34 62 50Supramarginal gyrus Left 56 28 40Superior parietal lobe Medial 0 74 52MFPC Left 4 56 4Dorsomedial prefrontal cortex Left 10 52 44Posterior superior temporal sulcus Left 56 30 8Posterior superior temporal sulcus Right 68 14 6Temporal pole Left 56 8 18Temporal Pole Right 60 4 16Ventrolateral prefrontal cortex left 52 20 2Inferior frontal gyrus seed Left 46 28 12Amygdala Left 26 2 18Lateral occipital cortex Right 48 68 16Temporal occipital cortex Left 42 54 22Inferior lateral occipital cortex Left 32 70 16Inferior frontal gyrus Left 46 24 0Middle temporal gyrus Left 54 26 4Middle temporal Gyrus Right 44 32 2Superior temporal gyrus Right 56 4 18Posterior cingulate cortex Right 10 56 32Precuneus Right 8 54 48Primary somatosensory cortex Right 38 34 52Superior parietal lobe Right 28 52 58

Notes Seed ROIs are shown in bold and target ROIs are listed A 5 mm binary

spherical mask was drawn around each seed and target ROI Coordinates are

shown in MNI152 space

contrasts These parameter estimates represent the strength offunctional connectivity between key structures in the persua-sion network for each risk group as modulated by each PSA

We then computed a zero order correlation matrix to exam-ine the relationship between PEs and PME ratings in our twoindependent samples for each risk group We considered PEsthat were significantly correlated with out-of-sample PME rat-ings as promising candidates in our network-as-predictorregression models The actual prediction regression models fol-lowed the analytical logic established in (Falk et al 2012 Weberet al 2014) with one notable change instead of transforming allPEs and PME evaluations into ranks for each PSA and risk groupwe used the original unranked continuous data in our predic-tion models (making sure that no outliers were present)Unranked data did not lead to any substantial differences (com-pared to ranked data) and in a few cases our modified proce-dure was slightly more conservative (that is it led to lowerprediction accuracies) As in (Falk et al 2012 Weber et al 2014)our network-as-predictor analysis employed a step-wise regres-sion procedure In a first step we estimated baseline regressionmodels which included the self-reported PSA evaluations in ourMRIS sample that are known to be predictive from traditionalpersuasion theory (pMSV pAS and the pMSVpAS interac-tion) All self-report predictors were group mean centered toreduce bias in predictor coefficients due to the inherent multi-collinearity between main- and interaction-effect predictorsWe then included the brain network PEs as predictors in a sec-ond step This procedure allowed us to identify the networkconnectivity PEs that increased PME prediction accuracy aboveand beyond self-reported PSA evaluations

Results

The primary aims of this study were to evaluate (i) differencesin functional connectivity patterns between risk groups whenresponding to anti-drug PSAs and (ii) whether these patternsprovide support the affective-executive and encoding-disruption hypotheses and the degree to which network con-nectivity patterns unique to each PSA are predictive of PME rat-ings in two independent samples We report on the whole-brainPPI results as they relate to the affective-executive andencoding-disruption hypotheses before turning to the network-as-predictor results

PPI results

In this study we constructed contrasts that encoded between-group comparisons as well as mean within-group connectivitypatterns With the exception of just one case between-groupcomparisons did not survive statistical thresholding Thereforeprimarily qualitatively different connectivity patterns emergedby risk-group Said differently unless otherwise specified dif-ferences reported are within risk-group and do not representthe more stringent between risk-groups contrast However it isworth noting that our analytic decision to contrast connectivitypatterns in each risk group by an active control and then con-trast the resulting group-level PEs against each other is conser-vative and makes it more difficult to detect group-leveldifferences A less-conservative approach would have beento contrast group-level connectivity patterns without alsoaccounting for an active baseline

However this would make it more difficult to distinguishbetween functional connections simply associated with watch-ing a PSA (and are therefore driven by low-level audiovisual

features) and those that account for the semantic content of aPSA For clarity we split the reporting of our PPI results accord-ing to the different hypotheses they address

Affective-executive hypothesis This hypothesis argued that differ-ences in AS should modulate functional connectivity betweenstructures implicated in affective and executive processing andthat these connectivity patterns should differ by issue-involvement Accordingly and consistent with existing litera-ture the results presented in this subsection correspond to theASPHYSgtC1 PHYS contrast When seeding from the MPFC(Table 2 Figure 1) among high-drug risk subjects the MPFCshowed functional connections with the dorsoanterior insulacentral operculum temporal pole putamen (dorsal striatum)and orbitofrontal cortex (OFC) For subjects in the low-drug riskgroup the MPFC showed connections with the inferior parietallobe (IPL) primary somatosensory cortex supramarginal gyrussuperior frontal gyrus and pre-motor cortex

For low-risk subjects the MFG (Table 3 Figure 2) showedbilateral connectivity with the superior lateral occipital cortex(sLOC) IPL and anterior intra-parietal sulcus Results did notsurvive thresholding among high-risk subjects Similarly signif-icant results were not observed for the STG seed for either risk-group Lastly we were unable to replicate previous findings(Ramsay et al 2013 Dinh-Williams et al 2014) as significantresults were not observed when seeding from the IFG seed foreither the high- or low-drug risk groups

Encoding-disruption hypothesis When seeding from the iLOC(Table 4 Figure 3) for the MSVPHYSgtC1PHYS contrastboth high- and low-drug risk subjects showed connections withMTG superior lateral occipital cortex OFC and superior frontalgyrus In general these iLOC connections with temporal corticesdo not conform to the encoding-disruption hypothesis (Seeliget al 2014) which predicts that high MSV should lead to lowerconnectivity between occipital and temporal cortices

Furthermore significant group differences were observed forthe high-gt low-drug risk contrast High-drug risk subjectsshowed iLOC connectivity with the central operculum amyg-dala dorsal striatum (both putamen and caudate nucleus) andMFG

Network-as-predictor results

Of the video-wise seed-target connectivities (PEs) evaluated inthis study four were significantly correlated with PME in ourindependent samples (Table 5) In the stepwise regressionswhich adjusted each PE for both the effect of the self-reportbaseline predictors (pMSV pAS and pMSV x pAS) and each ofthe other PEs in the model only MFG-SPL PEs significantlyimproved out-of-sample prediction accuracy Therefore weonly report these results below

The baseline models using self-reported PSA evaluations inour MRIS sample to predict PSA PME in two independent largesamples (IS1 and IS2) were significant For IS1 the predictionaccuracies were R2frac14 239 (Ffrac14 424 P frac14 0014 all accuracies arereported as adjusted R2) for the high-drug risk group andR2frac14 499 (Ffrac14 1131 Plt 0001) for the low-drug risk group ForIS2 the accuracies were R2frac14 366 (Ffrac14 695 Plt 0001) for thehigh-drug risk group and R2frac14 627 (Ffrac14 1835 Plt 0001) for thelow-drug risk group None of the models violated any centralassumptions of the general linear model

As described above adding the MFG-SPL PEs to the baselinemodel in a stepwise regression increased out-of-sample

prediction accuracy Specifically for the high risk-drug groupfunctional connectivity between MFG and the SPL increased theprediction accuracy in IS1 from R2frac14 239 to R2frac14 351(Ffrac14 518 Pfrac14 0003) an increase of 112 points in this tradition-ally difficult-to-predict target group In IS2 the prediction accu-racy increased from R2frac14 366 to R2frac14 433 (Ffrac14 693 P lt 0001)which is an increase of 67 points In contrast in the low-drugrisk group none of the functional connectivity PEs was able to

increase the out-of-sample prediction accuracy significantlybeyond self-report data for either IS1 or for IS2 These resultsare summarized in Figure 4

Discussion

In order to advance the persuasion neuroscience literature theanalyses reported in the present study were designed to identify

Table 2 Psychophysiological interaction results when seeding from the medial prefrontal cortex

Structure Laterality Cluster size Max Z-score Coordinates

ASxPhys gt CtrlxPhysHigh-risk subjects

Dorsoanterior insula Right 1930 373 38 8 4Temporal pole Right 365 56 8 0Putamen (dorsal striatum) Right 346 28 6 4Frontal operculum Right 326 42 18 0Frontal operculum Right 326 42 24 0Dorsoanterior insula Right 325 44 8 4Putamen Left 1781 346 32 2 2Orbitofrontal cortex Left 345 44 20 8Central operculum Left 332 46 6 0Central operculum Left 317 48 0 4

Low-risk subjectsInferior parietal lobe Right 2451 368 58 34 44Primary somatosensory cortex Right 343 36 34 42Supramarginal gyrus Right 338 38 38 42Supramarginal gyrus Right 337 62 24 30Superior frontal gyrus Right 317 20 6 68Premotor cortex Right 302 30 20 66Inferior parietal lobe Left 1096 387 60 40 36Inferior parietal lobe Left 386 62 30 32Inferior parietal lobe Left 386 62 36 34Supramarginal gyrus Left 358 68 28 26Supramarginal gyrus Left 343 66 30 40

MSVxPhys gt CtrlxPhysHigh-risk subjects

Central operculum Left 1017 367 50 2 8Precentral gyrus Left 361 52 4 10Dorsoanterior insula Left 339 32 20 4Precentral gyrus Left 305 60 4 4Central operculum Left 299 48 6 4Putamen (dorsal striatum) Left 296 32 0 0

Low-risk subjectsInferior parietal lobe Right 3063 461 58 36 42Supramarginal gyrus Right 401 62 24 32Inferior parietal lobe Right 386 54 50 44Inferior parietal lobe Right 384 48 46 50Primary somatosensory cortex Right 343 46 32 48Anterior intra-parietal sulcus Right 335 40 42 42Ventrolateral prefrontal cortex Right 1579 399 40 58 2Dorsolateral prefrontal cortex Right 384 38 58 12Dorsolateral prefrontal cortex Right 366 40 52 18Ventrolateral prefrontal cortex Right 338 46 54 4Orbitofrontal cortex Right 337 40 48 10Orbitofrontal cortex Right 334 22 36 18Inferior parietal lobe Left 1480 406 62 36 32Supramarginal gyrus Left 402 62 30 40Anterior intra-parietal sulcus Left 294 42 44 44

Notes Results are cluster corrected for multiple comparisons with the cluster defining threshold set at Zfrac1423 and a cluster extent threshold of Plt005 and coordinates are

shown in MNI152 space Structures with a value in the cluster size column represent the peak Z-score while all others correspond to local maxima within the cluster

the message-feature by issue-involvement interactions thatcontribute to persuasion-related neural processes particularlyamong high-drug subjects This is the critical advancenecessary if we wish to design effective messages thatmotivate attitude change among this difficult-to-reach audi-ence To that end this discussion section primarily focuses onnetwork connectivity patterns among our high-drug risk

subjects However important differences among low-drug risksubjects particularly as they relate to earlier findings are alsodiscussed

Two functional perspectives currently exist on the patternsof network connectivity that underpin persuasive messageprocessing The first argues that persuasive messages rely onaffective-executive processing where network connectivity may

Fig 1 Psychophysiological interaction results when seeding from the medial prefrontal cortex The figure shows the AS x PHYSgtScrambled Control x PHYS contrast

for (A) high-risk and (B) low-risk subjects Results are cluster corrected for multiple comparisons with the cluster defining threshold set at Zfrac14 23 and a cluster extent

threshold of P lt 005 The red circle represents the seed ROI

Table 3 Psychophysiological interaction results when seeding from the middle frontal gyrus

ASxPhys gt CtrlxPhysLow-risk subjects

Superior lateral occipital cortex Left 1678 361 26 64 38Inferior parietal lobe Left 353 54 28 38Superior lateral occipital cortex Left 334 26 72 60Primary somatosensory cortex Left 332 44 42 60Anterior intra-parietal sulcus Left 330 32 44 44Anterior intra-parietal sulcus Left 329 40 46 46Anterior intra-parietal sulcus Right 999 353 44 46 36Inferior parietal lobe Right 344 64 46 54Superior lateral occipital cortex Right 312 32 62 40Superior lateral occipital cortex Right 309 40 62 60Superior lateral occipital cortex Right 308 32 64 54Inferior parietal lobe Right 308 42 56 58

MSVxPhys gt CtrlxPhysLow-risk subjects

Superior lateral occipital cortex Left 2142 377 38 62 42Superior lateral occipital cortex Left 373 26 64 38Inferior parietal lobe Left 360 44 48 52Anterior intra-parietal sulcus Left 360 38 50 48Inferior parietal lobe Left 347 44 54 52Inferior parietal lobe Left 340 56 28 40Anterior intra-parietal sulcus Right 1803 385 46 44 38Anterior intra-parietal sulcus Right 370 40 50 34Inferior parietal lobe Right 351 46 46 54Central precuneous Right 326 8 78 58Superior lateral occipital cortex Right 323 28 74 54

Notes Results are cluster corrected for multiple comparisons with the cluster defining threshold set at Zfrac1423 and a cluster extent threshold of P lt 005 and coordinates

are shown in MNI152 space Structures with a value in the cluster size column represent the peak Z-score while all others correspond to local maxima within the cluster

Fig 2 Psychophysiological interaction results when seeding from the middle frontal gyrus The figure shows the AS x PHYSgtScrambled Control x PHYS contrast for

(A) high-risk and (B) low-risk subjects Results are cluster corrected for multiple comparisons with the cluster defining threshold set at Zfrac1423 and a cluster extent

threshold of Plt005 Note No activations survived thresholding for high-risk subjects The red circle represents the seed ROI

Table 4 Psychophysiological interaction results when seeding from the inferior lateral occipital cortex

Inferior temporal gyrus Left 8636 416 46 38 10Caudate Nucleus (Dosral Striatum) Left 410 8 4 10Superior lateral occipital cortex Left 398 36 64 18Inferior temporal gyrus Left 378 44 0 38Dorsomedial prefrontal cortex Left 359 2 46 40Superior lateral occipital cortex Left 359 50 66 16Subcallosal cortex Left 2012 386 10 16 20Subcallosal cortex Right 386 6 14 18Orbitofrontal cortex Left 364 8 60 14Orbitofrontal cortex Left 359 32 30 8Ventromedial prefrontal cortex Left 345 8 70 6Cerebellum Right 1203 406 24 72 34Cerebellum Right 364 42 68 40Cerebellum Right 317 40 70 32

Low-risk subjectsAngular gyrus Left 3275 407 44 -52 26Superior lateral occipital cortex Left 402 46 66 36Middle temporal gyrus Left 397 62 34 10Middle temporal gyrus Left 391 60 52 4Middle temporal gyrus Left 375 56 38 4Superior lateral occipital cortex Left 66 58 62 28Orbitofrontal cortex Left 2596 364 40 42 14Ventrolateral prefrontal cortex Left 409 52 34 10Orbitofrontal cortex Left 403 42 32 16Ventrolateral prefrontal cortex Left 377 28 64 0Middle frontal gyrus Left 2427 408 32 12 52Dorsomedial prefrontal cortex Left 339 6 42 36Superior frontal gyrus Left 331 14 32 50Middle frontal gyrus Left 2427 331 38 24 38Superior frontal gyrus Left 329 10 30 52Superior frontal gyrus Left 323 20 22 56Cerebellum Right 1762 446 24 74 42

MSVxPhys gt CtrlxPhysHigh-risk gt low-risk subjects

Central operculum Left 1178 315 34 8 16Amygdala Left 309 22 8 8Putamen (dorsal striatum) Left 289 24 10 8Central operculum Left 273 48 0 8Caudate nucleus (dorsal striatum) Right 1097 320 12 12 20Middle frontal gyrus Right 295 36 16 28

(Continued)

Table 4 (Continued)

High-risk subjectsInferior temporal gyrus Left 3905 405 46 38 10Middle temporal gyrus Left 401 50 44 6Middle temporal gyrus Left 391 58 6 28Superior lateral occipital cortex Left 374 38 62 18Inferior lateral occipital cortex Left 374 40 62 4Superior lateral occipital cortex Left 362 52 68 18Cerebellum Right 3035 433 20 76 36Cerebellum Right 430 38 72 42Superior lateral occipital cortex Right 408 44 60 22Cerebellum Right 386 42 66 24Temporal occipital fusiform cortex Right 372 42 54 12Caudate nucleus (dorsal striatum) Left 2926 410 6 2 10Orbitofrontal cortex Left 389 32 30 8Orbitofrontal cortex Left 345 44 48 16Orbitofrontal cortex Left 336 26 52 6Temporal pole Left 332 20 12 32Superior frontal gyrus Left 1108 317 4 24 54Superior frontal gyrus Left 303 6 46 36Paracingulate cortex Left 301 6 36 34Paracingulate cortex Left 301 4 40 32Dorsomedial prefrontal cortex Left 287 12 52 44

Low-risk subjectsSuperior frontal gyrus Left 3388 411 26 20 54Superior frontal gyrus Left 408 12 32 52Orbitofrontal cortex Left 397 32 64 8Orbitofrontal cortex Left 394 24 66 0Middle frontal gyrus Left 383 32 20 50Middle frontal gyrus Left 377 34 8 54Middle temporal gyrus Left 2928 408 60 32 8Angular gyrus Left 403 44 54 26Middle temporal gyrus Left 390 66 32 6Middle temporal gyrus Left 382 50 38 4Middle temporal gyrus Left 381 46 46 2Superior lateral occipital cortex Left 372 48 70 36Cerebellum Right 1782 434 38 78 42Orbitofrontal cortex Left 1310 478 42 30 20Orbitofrontal cortex Left 462 40 34 16Orbitofrontal cortex Left 448 50 30 12Ventrolateral prefrontal cortex Left 406 48 44 12Orbitofrontal cortex Left 387 36 46 18Ventrolateral prefrontal cortex Left 385 48 50 2

Fig 3 Psychophysiological interaction results when seeding from the inferior lateral occipital cortex The figure shows the MSV x PHYSgtScrambled Control x PHYS

contrast for (A) high-risk and (B) low-risk subjects Results are cluster corrected for multiple comparisons with the cluster defining threshold set at Zfrac1423 and a cluster

extent threshold of Plt005 The red circle represents the seed ROI

be up- or down-regulated depending on whether audiences findthe message convincing (Ramsay et al 2013 Cooper et al 2017)or choose to engage in defensive processing (Dinh-Williamset al 2014) Another (but not necessarily competing) viewargues that persuasive messages that are too arousing disruptsensory-encoding network connectivity thereby inhibitingmessage processing (Seelig et al 2014) While informative thesestudies manipulate just one of the many factors known to playa role in message processing (ie issue involvement MSV AS)In the present study we draw on the ELM which characterizespersuasion as a multi-determined process where features of amessage (ie MSV AS) are modulated by issue involvement Bybetter specifying the interactions between these factors we findcontinued support for the affective-executive hypothesis par-ticularly among high-risk subjects By comparison we fail toreplicate earlier findings related to the encoding-disruptionhypotheses Furthermore we show that the strength of thesenetwork connections is predictive of PME in two independentsamples We turn our attention to these findings and theirimplications below

The affective-executive hypothesis

For nearly all seed ROIs we see qualitative differences in neuralprocessing by risk-group Previous work has observed that ASand issue involvement together modulate functional connec-tions between regions commonly activated in affective andexecutive processing (Ramsay et al 2013 Dinh-Williams et al2014) While we are unable to replicate earlier findings using anIFG seed other results observed in the ASPHYSgtC1PHYScontrast provide support for this hypothesis For instance whenseeding from the MPFC among high-drug risk subjects we seeconnectivity with structures implicated in salience (dorsoante-rior insula) and consummatory reward (putamen orbitofrontalcortex) processing It is also interesting to note that while theiLOC seed was evaluated to test the encoding-disruptionhypothesis connectivity patterns with the amygdala and dorsalstriatum (unique to high-drug risk subjects) provide additionalsupport for the role of affect in message evaluation (we returnto this point below)

One question that remains is why our results in this contrastdo not perfectly map onto findings from earlier studies Two

possible explanations exist First our task was more of a con-ceptual rather than a direct replication One difference is thatour subjects continuously rated each PSA using a CRM whereasparticipants passively watched messages in earlier studies Thisdifference may have subtly influenced the way audiencesengaged with the message This is particularly true for theDinh-Williams et al study (2014) which provided evidence thatat-risk subjects cognitively disengaged from the messages Bycomparison our task required slightly more message engage-ment compared to passive watching which according to theELM should result in counterarguing (a form of defensive proc-essing) among issue-involved subjects

A second difference may be attributed to neural develop-ment Ramsay et al conducted their study among adolescentswho have important neuropsychological differences fromadults in the amygdala prefrontal cortex and striatum(Tottenham and Galvan 2016) Evidence shows that thestrength of functional connections between affective and exec-utive regions while watching videos featuring smokers (Do andGalvan 2016) or when viewing graphic warning labels on ciga-rette packages (Do and Galvan 2015) influences short-termcraving impulses differently among adolescents compared toadults

At the same time we also know that many of these develop-mental changes persist into early adulthood While our partici-pant group was older than those reported by Do and Galvan(2015 2016) and Ramsay et al (2013) the brains (and underlyingnetwork connections) of college students might still be under-going developmental changes That our network-as-predictorresults were more accurate among the adolescent sample (com-pared to the college sample) provides some evidence for thisview Teasing apart these developmental differences in process-ing and exactly when they occur is an important future direc-tion as persuasion strategies that are effective among adultsmay have different impacts among adolescents

More broadly a number of studies implicate MPFC activity insuccessful persuasion (for a review see Falk and Scholz 2018)The MPFC seed ROI we used was centered on a sub-region ofthe MPFC implicated in positive subjective valuation duringpersuasive message processing (Cooper et al 2017) Whenseeding from this region among high-risk subjects in theASPHYSgtC1 PHYS contrast we see connectivity with

Table 5 Pearson correlations between PEs and PME in IS1 (collegestudents) and IS2 (nationally representative adolescents)

1 2 3 4 5 6

High-risk subjects1 IS1 PME 12 IS2 PME 0935 13 MPFCTP 0257 0323 14 MFGAG 0365 0286 0204 15 MFGSPL 0385 0337 0270 0695 16 STGPSC 0146 0184 0062 0254 0376 1

Low-risk subjects1 IS1 PME 12 IS2 PME 0896 13 MPFCTP 0170 0024 14 MFGAG 0009 0036 0043 15 MFGSPL 0049 0042 0083 0922 16 STGPSC 0230 0374 0207 0176 0184 1

Correlation is significant at the Pfrac14 005 level (two-tailed) and Pfrac14 001 level

(two-tailed)

Fig 4 Network-as-predictor results for high-risk subjects A stepwise regression

model was constructed where self-reported pMSV pAS and their interaction

was entered in the first step and MFG-SPL connectivity parameter estimates

(PEs) were entered in the second step This analysis shows the unique contribu-

tion of self-report and network-connectivity PEs to overall prediction accuracy

of perceived message effectiveness in two independent samples (IS1 and IS2)

Results are not shown for the low-risk group as the inclusion of network-con-

nectivity PEs did not significantly improve prediction accuracies

dorsal striatum structures particularly the putamen This con-nectivity pattern is worth further consideration as it might onfirst glance seem to contradict findings from Cooper et al whocharacterized MPFC connectivity with the ventral striatum asbeing associated with positive subjective valuation Here wereturn to meta-analytic results to help clarify this pattern Theventral striatum region used by Cooper et al (2017) corre-sponded to the peak meta-analytic voxel by Bartra et al (2013)However these meta-analytic results also indicate that subjec-tive valuation extends more broadly within the striatumincluding into more dorsal regions such as the putamenAccordingly our MPFC connectivity patterns with the putamenare not inconsistent with the view that persuasion requiresmessage receivers to evaluate the positive outcomes associatedwith behavior change These findings also extend this view toimplicate the striatum more broadly in the persuasion process

Still we do not directly replicate MPFC connectivity with ven-tral striatum among either our high- or low-risk subjects asobserved in Cooper et al (2017) This may be driven at least inpart by our choice of stimuli and experimental task Cooper et alemployed a positive intervention where subjects were encouragedto mentalize ways in which they might enact lifestyle changesthat would lead to positive health outcomes Accordingly it is notunexpected that their task engaged the ventral striatum as thisstructure is heavily implicated in reward anticipation (OrsquoDohertyet al 2004) particularly as related to the calculation or rewardeffort tradeoffs associated with goal-directed behavior (Botvinicket al 2009 Kool et al 2013) By comparison our stimuli were con-sistent with more typical anti-drug PSAs which underscore thenegative consequences associated with drug use

The encoding-disruption hypothesis

One surprising area where we see similarities between risk-groups is in network connectivity patterns between occipitaland temporal cortices Contrary to previous findings (Seeliget al 2014) when seeding from the iLOC in theMSVPHYSgtC1PHYS contrast we see connectivity withMTG for both risk groups with high-risk subjects also showingconnectivity with inferior temporal gyrus (ITG) As has beenargued elsewhere (Weber et al 2013) even though MSV may bequite useful for capturing the attention of high sensation-seeking high drug-risk audiences this can backfire if theseaudiences are then required to process a counter-attitudinalmessage

When seeding from the iLOC we see an important group dif-ference where high-risk but not low-risk subjects show func-tional connectivity with regions implicated in affective andsalience processing including AMY dorsal striatum (both puta-men and caudate) and central operculum We interpret theseresults as further evidence that message strategies that relysolely on MSV may not be effective among counter-attitudinaltarget audiences This interpretation is further supported by ourself-reported pMSV evaluations In prediction models that useour self-reported PSA evaluations in our MRI sample to predictPME in two independent large samples of adolescents andfreshman college students pMSV was consistently the weakest(and most non-significant) predictor

The video-specific network connections that underpinmessage effectiveness

To recapitulate our network-as-predictor results we found thatMFG-SPL connectivity significantly improved predictions of per-

video ratings in two independent samples exclusively withinour high-drug risk group Next we consider how this resultrelates to the results of our ELM-based whole-brain PPI analy-ses in which we did not observe any significant modulation inMFG-SPL connectivity by either MSV or AS for either risk group

Practical implications The degree to which MSV and AS interactwith subject specific drug use risk to modulate brain networkconnectivity patterns in response to anti-drug messages helpsadvance our mechanistic understanding of the persuasion proc-ess but practitioners public-health officials and messagedesigners have a different need They are looking for principledyet data-driven ways to identify which anti-drug PSAs are mosteffective among difficult to reach high-risk target audiencesUnfortunately traditional self-report measures either misiden-tify which messages are most effective (eg Falk et al 2012) orgenerate so little variability as to make identifying effectivePSAs all but impossible (eg Weber et al 2013) The brain-as-predictor analysis and our network-as-predictor extension wasdesigned to help solve this problem by asking if there are con-nectivity patterns in the brain that explain additional variancein message-effectiveness evaluations above and beyond self-report data Such an analysis relies on the following logic (i)identify network connectivity patterns that systematically varyby risk group (ii) identify what message features drive this vari-ability (iii) use this information to improve out-of-sample pre-dictions in high-risk target audiences and (iv) use thisinformation for message selection

While our study and results conform to this general logicpoints 2 and 3 are worth considering further This studyhypothesized (and found) that MSV and AS interact with issueinvolvement to modulate persuasion network connectivity pat-terns However we did not observe MFG-SPL connectivity ineither the MSVPHYSgtC1PHYS or ASPHYSgtC1PHYScontrasts This suggests that even though the MFG and SPL arecommonly implicated in the persuasion neuroscience literature(for a synthesis see Kaye et al 2016) regardless of risk-groupconnectivity patterns between these structures do not seem tobe systematically modulated by MSV or AS It seems that someother feature of the PSAs used in this study influences MFG-SPLconnectivity and this feature appears to matter We return tothis idea again in the ELM section below

Point 3 is focused on improving out-of-sample predictionaccuracy While MFG-SPL PEs allow for better PME predictionamong high-risk subjects for both IS1 and IS2 they fall short ofwhat has been achieved in other brain-as-predictor analysesWeber et al (2014) show that measures of within-ROI signalchange in the MFG and SPL contribute to prediction accuraciesas high as 59 Falk et al have shown similar accuracies for theMPFC (eg 2012 2016) If we believe that persuasion is anetwork-level process then we should expect that capturingnetwork information should improve prediction accuraciesabove and beyond self-report and signal change (as has beenshown in Cooper et al 2017) Indeed exploratory analyses onour dataset show that in models including both signal changesand connectivity PEs the inclusion of connectivity PEs no longersignificantly improved adjusted R2 One possible methodologi-cal reason for this finding could be that our functional connec-tivity parameter estimates were drawn from relatively smallROIs that were based on previous studies in the literature andnot from meta-analyses such as Neurosynth (Yarkoni et al2011) or even an independent localizer task as is also commonin the persuasion neuroscience literature (Falk et al 2015)Combined with the facts that PPI analyses are quite sensitive to

the accurate selection of seed regions and that the increasednumber of parameters in a model reduces statistical power(Friston et al 1997 OrsquoReilly et al 2012) it is possible that our PEssimply contain higher levels of noise This coupled with theadditional analytical effort required to conduct a PPI analysissuggests that for now message designers may still be bettersuited by using the traditional brain-as-predictor approachwhich is based on simple BOLD signal change parameters

Mechanistic implications We know from previous research thatthe MFG is an important structure implicated in counterarguingprocesses particularly among high-drug risk subjects (Weberet al 2014) Further the fact that activity in this structure iscommonly observed across a number of studies suggests that itis a core component of the persuasion network (eg Chua et al2009 Falk et al 2010 Ramsay et al 2013) We also know that theMFG is implicated in processing audiovisual narratives (Wilsonet al 2008) like the PSAs used in this study However a recentstudy investigating intersubject correlations (ISC) among bothhigh- and low-risk subjects viewing health-message PSAs didnot observe ISCs in this region (Imhof et al 2017) To be clearwe are not arguing that the absence of a statistically significantresult implies no effect But it does suggest that there is some-thing important about the MFG that emerges in group-levelGLMs and at the individual video level but does not systemati-cally vary for ISCs or group-level PPIs Identifying what messagefeatures contribute to MFG activity in future studies is impor-tant both for our mechanistic understanding of the persuasionprocess as well as for message designers looking for practicalways of improving anti-drug campaigns

More broadly while this study replicates and extends ourunderstanding of connectivity patterns between structureswithin the persuasion network the PPI analyses employedherein and in previous studies are not without limitation Mostnotably PPI only considers the way in which a task modulatesconnectivity between one seed ROI and all other regions in thebrain Such a one-to-many connectivity analysis necessarilyoversimplifies the true network architecture In fact it isentirely likely that dynamic connections between multiple ROIswithin the network better characterize the persuasion processRecent advances in the application of graph theory to neuroi-maging data (eg Bassett and Sporns 2017) provide an analyti-cal approach for interrogating these complex dynamics Twoimmediate next-steps can help resolve ambiguity about net-work structure and dynamics First hypothesis driven analysesshould construct sparse graphs where nodes in the graph repre-sent recurrent structures within the persuasion network At thesame time more data-driven approaches based on a whole-brain parcellation (eg Glasser et al 2016) may illuminate as ofyet undiscovered network typologies New toolboxes particu-larly as implemented in both Python Abraham et al (2014) andMatlab (Rubinov and Sporns 2010 Sizemore and Bassett 2017)should help advance research in this area

Returning to the ELM

The ELM (Petty and Cacioppo 1986) makes the case that mes-sage features interact with individual differences to shape moti-vation and ability to process a message Different message-by-audience interactions determine if audiences (i) choose to proc-ess a message deeply or superficially and (ii) if this processing isbiased (eg counterarguing message disengagement) or notWith two exceptions (Weber et al 2014 Imhof et al 2017) tests

using an ELM framework are largely absent from the persuasionneuroscience literature

Accordingly much of the language used to describe the neu-ral processes that underpin persuasion are not framed usingELM-consistent language in-fact many test single- and notdual-process models (such as the ELM) of persuasion (for moredetail see Vezich et al 2016) In an effort to draw linkagesbetween these literatures we might re-frame the cognitiveprocesses identified with the ELM to be more consonant withthe persuasion neuroscience literature Explained in theseterms message features and issue-involvement modulateaffective responses that motivate executive processing strat-egies The results reported within this manuscript as well asacross a number of studies (Kaye et al 2016) largely support thisview

For instance affective processing associated with subjectivevalue is well-known to motivate executive processing andbehavior (for a review see Braver et al 2014) and studies impli-cating subjective value in persuasion (eg Cooper et al 2017Falk and Scholz 2018) provide additional clarity as to whatmight be driving this motivational response At the same timeour network-as-predictor analysis identified variation in indi-vidual videos that was not captured by ELM-relevant variables(ie MSV AS) but still predicted message perceptions in inde-pendent samples This suggests possible extensions to the ELM

Falk et al (2009) implicated social processing in persuasionand the literature to-date extends this into additional areasincluding reward processing self-reflection social pain andsalience detection (Kaye et al 2016) Another possible explana-tion lies in the ways audiences process narratives (Slater andRouner 2002) Recent work shows that audiences with differentinitial beliefs interpret narratives in rather different ways andthat these differences are driven by similar or dissimilar activa-tion in a variety of structures including the right MFG (the seedROI we used in our network-as-predictor analysis for furtherdetails see Yeshurun et al 2017) These studies hint at factorsthat underpin individual differences in processing and suggestfuture directions for ELM-based research

On a more critical note one limitation on much of the per-suasion neuroscience research including the present study isthat it largely relies on a lsquomagic bulletrsquo logic where a singleexposure to a single message should result in persuasion Intodaylsquos media landscape individuals are often exposed to thesame (or even competing) messages multiple times across dif-ferent channels And even with these multiple exposures weknow that changing attitudes and behaviors is nontrivialAddressing these challenges requires longitudinal work focusedon how plasticity in attitudes and behaviors corresponds withneurobiological changes There is some work on this particu-larly among smoker populations (Berkman et al 2011 Falk et al2011) but these studies still utilize a one-shot fMRI sessionAdmittedly longitudinal fMRI work is both costly and slow butrepeated measures are necessary if we wish to understand howmessage dynamics modulate neural dynamics and behaviorover time

Conclusion

Although no meta-analysis currently exists the persuasionneuroscience literature has grown to a point that the sameregions consistently emerge across a number of studies (Kayeet al 2016) However our study shows important qualitative dif-ferences in connectivity patterns by risk group These resultsare bolstered by their predictive-validity Specifically these

network connections are predictive of how independent audi-ences process the message This suggests that careful work isneeded to identify and tailor (Noar et al 2007) messages for spe-cific and narrowly defined target audiences This point is partic-ularly important given that recent evidence demonstrates thatindividual differences in message encoding and social networkstructure predict persuasive message outcomes (Pegors et al2017) In the past traditional message channels (eg broadcasttelevision) made such fine-grained tailoring difficult Howeverwith the advent of social media and other Internet channelsmicro-targeting strategies have become increasingly feasibleFuture work should seek to identify how specific message com-ponents interact with theoretically relevant individual differen-ces in order to increase message effectiveness

Acknowledgement

We thank our colleagues at the Annenberg School forCommunication at The University of Pennsylvania for gen-erously sharing the stimuli and one of the data sets used inthis study

Funding

This work was supported by the University of CaliforniaSanta Barbara Brain Imaging Center RH is an awardee ofthe George D McCune Dissertation Fellowship

Conflict of interest None declared

ReferencesAbraham A Pedregosa F Eickenberg M et al (2014) Machine

learning for neuroimaging with scikit-learn Frontiers in

Neuroinformatics 8 14Bartra O McGuire JT Kable JW (2013) The valuation system

a coordinate-based meta-analysis of BOLD fMRI experimentsexamining neural correlates of subjective value NeuroImage76 412ndash27

Bassett DS Gazzaniga MS (2011) Understanding complexityin the human brain Trends in Cognitive Sciences 15 (5)200ndash9

Bassett DS Sporns O (2017) Network neuroscience Nature

Neuroscience 20 (3) 353ndash64Beckmann CF Smith SM (2004) Probabilistic independent

component analysis for functional magnetic resonance imag-ing IEEE Transactions on Medical Imaging 23 (2) 137ndash52

Berkman ET Falk EB (2013) Beyond brain mapping usingneural measures to predict real-world outcomes CurrentDirections in Psychological Science 22(1) 45ndash50

Berkman ET Falk EB Lieberman MD (2011) In the trenchesof real-world self-control neural correlates of breaking thelink between craving and smoking Psychological Science22(4)498ndash506

Bigsby E Cappella JN Seitz HH (2013) Efficiently andeffectively evaluating public service announcements addi-tional evidence for the utility of perceived effectivenessCommunication Monographs 80(1) 1ndash23

Botvinick MM Huffstetler S McGuire JT (2009) Effort dis-counting in human nucleus accumbens Cognitive Affective ampBehavioral Neuroscience 9(1) 16ndash27

Braver TS Krug MK Chiew KS et al (2014) Mechanisms ofmotivation-cognition interaction Challenges and opportuni-ties Cognitive Affective amp Behavioral Neuroscience 14(2) 443ndash72

Bullmore E Sporns O (2012) The economy of brain networkorganization Nature Reviews Neuroscience 13(5) 336ndash49

Cappella JN Yzer MC Fishbein M (2003) Using beliefs aboutpositive and negative consequences as the basis for designingmessage interventions for lowering risky behavior In RomerD editors Reducing Adolescent Risk Toward an IntegratedApproach pp 210ndash220 Thousand Oaks CA Sage Publications

Chua HF Liberzon I Welsh RC Strecher VJ (2009) Neuralcorrelates of message tailoring and self-relatedness in smok-ing cessation programming Biological Psychiatry 65(2) 165ndash8

Cohen J Cohen P West SG Aiken LS (2003) Applied MultipleRegressionCorrelation Analysis for the Behavioral Sciences 3rdedn Mahwah NJ Lawrence Erlbaum Associates

Cooper N Bassett DS Falk EB (2017) Coherent activitybetween brain regions that code for value is linked to the mal-leability of human behavior Scientific Reports 7(43250) 4325

Dillard JP Shen L Vail RG (2007a) Does perceived messageeffectiveness cause persuasion or vice versa 17 consistentanswers Human Communication Research 33(4) 467ndash88

Dillard JP Weber KM Vail RG (2007b) The relationshipbetween the perceived and actual effectiveness of persuasivemessages a meta-analysis with implications for formativecampaign research Journal of Communication 57(4) 613ndash31

Dinh-Williams L Mendrek A Dumais A Bourque J PotvinS (2014) Executive-affective connectivity in smokers viewinganti-smoking images an fMRI study Psychiatry ResearchNeuroimaging 224(3) 262ndash8

Do KT Galvan A (2015) FDA cigarette warning labels lowercraving and elicit frontoinsular activation in adolescent smok-ers Social Cognitive and Affective Neuroscience 10 (11)1484ndash96

Do KT Galvan A (2016) Neural sensitivity to smoking stimuliis associated with cigarette craving in adolescent smokersJournal of Adolescent Health 58(2) 186ndash94

Falk EB Berkman ET Lieberman MD (2012) From neuralresponses to population behavior neural focus group predictspopulation-level media effects Psychological Science 23(5)439ndash45

Falk EB Berkman ET Mann T Harrison B Lieberman MD(2010) Predicting persuasion-induced behavior change fromthe brain Journal of Neuroscience 30(25) 8421ndash4

Falk EB Berkman ET Whalen D Lieberman MD (2011)Neural activity during health messaging predicts reductions insmoking above and beyond self-report Health Psychology 30(2)177ndash85

Falk EB Cascio CN Coronel JC (2015) Neural prediction ofcommunication-relevant outcomes Communication Methodsand Measures 9(1ndash2) 30ndash54

Falk EB OrsquoDonnell MB Tompson S et al (2016) Functionalbrain imaging predicts public health campaign success SocialCognitive and Affective Neuroscience 11(2) 204ndash14

Falk E B Rameson L Berkman E T et al (2009) The neuralcorrelates of persuasion A common network across culturesand media Journal of Cognitive Neuroscience 22(11) 2447ndash2459

Falk EB Rameson L Berkman ET et al (2010) The neuralcorrelates of persuasion a common network across culturesand media Journal of Cognitive Neuroscience 22(11) 2447ndash59

Falk EB Scholz C (2018) Persuasion influence and value per-spectives from communication and social neuroscienceAnnual Review of Psychology 69(1) doi 101146annurev-psych-122216-011821

Friston KJ (1994) Functional and effective connectivity in neu-roimaging a synthesis Human Brain Mapping 2(1ndash2) 56ndash78

Friston KJ (2011) Functional and effective connectivity areview Brain Connectivity 1(1) 13ndash36

Friston KJ Buechel C Fink GR Morris J Rolls E Dolan RJ(1997) Psychophysiological and modulatory interactions inneuroimaging NeuroImage 6(3) 218ndash29

Glasser MF Coalson TS Robinson EC et al (2016) Amulti-modal parcellation of human cerebral cortex Nature7615(563) 171ndash8

Imhof MA Schmalzle R Renner B Schupp HT (2017) Howreal-life health messages engage our brains shared processingof effective anti-alcohol videos Social Cognitive and AffectiveNeuroscience 12(7) 1188ndash96

Kang Y Cappella JN Fishbein M (2006) The attentionalmechanism of message sensation value interaction betweenmessage sensation value and argument quality on messageeffectiveness Communication Monographs 73(4) 351ndash78

Kaye S-A White MJ Lewis I (2016) The use of neurocogni-tive methods in assessing health communication messages asystematic review Journal of Health Psychology 22(12) 1534ndash55

Kool W McGuire JT Wang GJ Botvinick MM Brosnan SF(2013) Neural and behavioral evidence for an intrinsic cost ofself-control PLoS ONE 8(8) 72626

Lang A (2006) Using the limited capacity model of motivatedmediated message processing to design effective cancer com-munication messages Journal of Communication 56(1) S57ndash80

McLaren DG Ries ML Xu G Johnson SC (2012) A general-ized form of context-dependent psychophysiological interac-tions (gPPI) a comparison to standard approaches NeuroImage61(4) 1277ndash86

Noar SM Benac CN Harris MS (2007) Does tailoring matterMeta-analytic review of tailored print health behavior changeinterventions Psychological Bulletin 133(4) 673ndash93

OrsquoDoherty J Dayan P Schultz J Deichmann R Friston KDolan RJ (2004) Dissociable roles of ventral and dorsal stria-tum in instrumental conditioning Science 304(5669) 452ndash4

OrsquoReilly JX Woolrich MW Behrens TEJ Smith SMJohansen-Berg H (2012) Tools of the trade psychophysiologi-cal interactions and functional connectivity Social Cognitiveand Affective Neuroscience 7(5) 604ndash9

Pegors TK Tompson S OrsquoDonnell MB Falk EB (2017)Predicting behavior change from persuasive messages usingneural representational similarity and social network analy-ses NeuroImage 157 118ndash28

Petersen SE Sporns O (2015) Brain networks and cognitivearchitectures Neuron 88(1) 207ndash19

Petty RE Cacioppo JT (1986) The elaboration likelihoodmodel of persuasion In Berkowitz L editor Advances inExperimental Social Psychology (v19 ed pp 123ndash205) New YorkNY Academic Press

Poldrack RA Farah MJ (2015) Progress and challenges inprobing the human brain Nature 526(7573) 371ndash9

Ramsay IS Yzer MC Luciana M Vohs KD MacdonaldAWI (2013) Affective and executive network processingassociated with persuasive antidrug messages Journal ofCognitive Neuroscience 25(7) 1136ndash47

Rogers EM (1994) A History of Communication Study ABiographical Approach New York NY The Free Press

Rubinov M Sporns O (2010) Complex network measures ofbrain connectivity uses and interpretations NeuroImage 52(3)1059ndash69

Seelig D Wang A-L Jaganathan K Loughead JW Blady SJChildress AR Romer D (2014) Low message sensationhealth promotion videos are better remembered and activate

areas of the brain associated with memory encoding PLoS One9(11) e113256

Sizemore AE Bassett DS (inpress) Dynamic graph metricsTutorial toolbox and tale NeuroImage doi 101016jneuroimage201706081

Slater MD Rouner D (2002) Entertainment-education andelaboration likelihood understanding the processing of narra-tive persuasion Communication Theory 12(2) 173ndash91

Tottenham N Galvan A (2016) Stress and the adolescentbrain amygdala-prefrontal cortex circuitry and ventral stria-tum as developmental targets Neuroscience and BiobehavioralReviews 70 217ndash27

Vezich IS Falk EB Lieberman MD (2016) Persuasion neuro-science new potential to test dual-process theories InHarmon-Jones E Inzlicht M editors Social NeuroscienceBiological Approaches to Social Psychology 1st edn pp 34ndash58New York NY Routledge

Vezich IS Katzman PL Ames DL Falk EB Lieberman MD(2016) Modulating the neural bases of persuasion whyhowgainloss and usersnon-users Social Cognitive and AffectiveNeuroscience 12(2) 283ndash97

Wang Z Vang M Lookadoo K Tchernev JM Cooper C(2014) Engaging high-sensation seekers the dynamic inter-play of sensation seeking message visual-auditory complexityand arousing content Journal of Communication 5(1) 101ndash24

Weber R Eden A Huskey R Mangus JM Falk E (2015)Bridging media psychology and cognitive neuroscience chal-lenges and opportunities Journal of Media Psychology 27(3)146ndash56

Weber R Huskey R Mangus JM et al (2014) Neural predictorsof message effectiveness during counterarguing in anti-drugcampaigns Communication Monographs 82(1) 4ndash30

Weber R Huskey R Mangus JM Westcott-Baker A TurnerBO (2015) Neural predictors of message effectiveness duringcounterarguing in anti-drug campaigns CommunicationMonographs 82(1) 4ndash30

Weber R Mangus JM Huskey R (2015) Brain imaging in com-munication research a practical guide to understanding andevaluating fMRI studies Communication Methods and Measures9(1-2) 5ndash29

Weber R Westcott-Baker A Anderson GL (2013) A multilevelanalysis of antimarijuana public service announcement effec-tiveness Communication Monographs 80(3) 302ndash30

Wilson SM Molnar-Szakacs I Iacoboni M (2008) Beyondsuperior temporal cortex intersubject correlations in narrativespeech comprehension Cerebral Cortex 18(1) 230ndash42

Woolrich MW Behrens TEJ Beckmann CF Jenkinson MSmith SM (2004) Multilevel linear modelling for FMRI groupanalysis using Bayesian inference NeuroImage 21(4) 1732ndash47

Worsley KJ (2001) Statistical analysis of activation images InJezzard P Matthews PM Smith SM editors Functional MriAn Introduction to Methods (pp 251ndash270) Oxford UnitedKingdom Oxford University Press

Yarkoni T Poldrack RA Nichols TE Van Essen DC WagerTD (2011) Large-scale automated synthesis of human func-tional neuroimaging data Nature Methods 8(8) 665ndash70

Yeshurun Y Swanson S Simony E et al (2017) Samestory different story the neural representation of interpretiveframeworks Psychological Science 28(3) 307ndash19

Zhao X Strasser A Cappella JN Lerman C Fishbein M(2011) A measure of perceived argument strength reliabilityand validity Communication Methods and Measures 5(1) 48ndash75

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

processing are recruited while evaluating persuasive messagesToday the literature has developed such that a structural con-sensus is emerging A synthesis of 20 persuasion-neurosciencestudies implicated structures involved in reward processingself-reflection social pain executive processing and saliencedetection as key components of the so-called persuasion network(Kaye et al 2016)

Despite reference to a persuasion network the majority ofwork in this domain has in fact ignored network structure andhas instead used the traditional mass-univariate approach thataddresses questions of how much activity is elicited within dif-ferent brain regions (but not how those regions interact withone another Friston 1994 2011)

However cognitive neuroscience has increasingly come torecognize that a complete understanding of how psychologicalprocessess (including persuasion) are instantiated in the brainrequires studying the brain at the network level (Bassett andGazzaniga 2011 Bullmore and Sporns 2012 Petersen andSporns 2015 Weber et al 2015 Bassett and Sporns 2017Cooper et al 2017) Accordingly prominent calls have forcefullyadvocated that our understanding of the neural basis of psycho-logical processes can only advance by moving beyondlsquoblobologyrsquo (eg Poldrack and Farah 2015) and adopting a net-worked perspective Applying this perspective to the persuasionnetwork requires analytical procedures that illuminate thetask-modulated relationship between one network structureand another (Friston 2011) To date a small handful of persua-sion neuroscience studies have addressed this issue using psy-chophysiological interaction analyses (PPI Friston et al 1997)Among these two (not necessarily incompatible) functionalviews of the network connectivity patterns that underpin per-suasion are beginning to emerge Each is briefly summarizedbelow

Affective-executive hypothesis

Ramsay et al (2013) proposed a mechanism where evaluationsof persuasive messages are modulated by affective and execu-tive processing In their study among both non-drug and drugusers strong argument strength (AS) persuasive messagesyielded increased functional connectivity between left inferiorfrontal gyrus (lIFG a putatively lsquoexecutiversquo region) and theamygdala (AMY a putatively lsquoaffectiversquo region) and insula com-pared to weak AS messages A related study among smokersobserved a negative relationship between structures commonlyimplicated in affective and executive processing in response toaversive anti-smoking (and therefore self-relevant) imagescompared to aversive but not self-relevant images (Dinh-Williams et al 2014) Specifically a seed region of interest (ROI)in the lIFG showed negative connectivity with the anterior cin-gulate cortex (ACC) medial prefrontal cortex (MPFC) precuneusand insula This connectivity pattern suggests that issueinvolvement modulates the way in which audiences process apersuasive message such that smokers engage in defensiveprocessing in an attempt to down regulate negative feelingsassociated with a threatening message Unfortunately thisstudy did not include a non-smoking comparison group so itis difficult to fully integrate this finding with the study byRamsay et al Nevertheless these two results account for keyvariables known to modulate persuasion specifically AS andaudience issue-involvement (Petty and Cacioppo 1986) andshow that each variable contributes to different neural process-ing patterns

Similarly a study by Cooper et al (2017) examined connectiv-ity patterns in sedentary individuals in response to (individuallyrelevant) messages promoting physical activity and found thatconnectivity strength between a subregion within the MPFCand ventral striatum (VS) and AMY during message exposurepredicted decreases in sedentary behavior Taken togetherthese three studies provide converging evidence that affective-executive processing underlies attitude and behavior change

Encoding-disruption hypothesis

The previous studies evaluated variables such as AS and issueinvolvement At the same time message sensation value (MSVa measure of message arousingness or intensity) has also beenhypothesized as a mechanism that can disrupt cognitive proc-essing thereby limiting the extent to which a persuasive mes-sage is encoded and ultimately recalled (eg Lang 2006)Experimental fMRI work has shown that low-compared to high-MSV persuasive ads are better remembered elicit increasedactivity in temporal and prefrontal cortex and correspond tolower levels of occipital cortex activation (Seelig et al 2014)

The same study also tested the hypothesis that increases inMSV disrupt message encoding and therefore should result indecreased functional connectivity between structures impli-cated in sensory processing and message encoding The resultswere ambiguous in that a negative relationship was observedbetween inferior lateral occipital cortex (iLOC) and the middletemporal (MTG) and inferior frontal gyrus (IFG) regardless ofvariation in MSV although this negative relationship wasweaker when MSV was low One possible explanation for thisfinding is that audience characteristics modulate audience reac-tions to MSV For instance Wang et al (2014) show thatincreases in MSV lead to approach tendency among high-riskaudiences whereas low-risk audiences show avoidanceresponses

The present study

As these studies indicate and a large body of theoretical andempirical evidence demonstrates issue involvement modulatespersuasive message processing (Petty and Cacioppo 1986) Theliterature shows that being at-risk for drug use constitutes aform of issue involvement where those that are at higher-riskfor using a drug hold stronger counter-attitudinal views whenpresented with an anti-drug message compared to those thatare at lower-risk (Cappella et al 2003) This increases the proba-bility that at-risk and therefore issue-involved audiencesengage in defensive message processing (ie counterarguing)when presented with a counter attitudinal message This defen-sive processing makes it difficult to use self-report measures toidentify which messages are likely to result in attitude or behav-ior change (Weber et al 2013 2014) Identifying what messagefeatures contribute to attitude change among this audience iscritically important

Unfortunately our understanding in this area is incompleteOf the four fMRI PPI studies to-date two only evaluate at-riskparticipants (Dinh-Williams et al 2014 Cooper et al 2017) Theremaining two studies (Ramsay et al 2013 Seelig et al 2014)include both not-at-risk and at-risk participants in their sam-ples however they do not consider the different ways each riskgroup processes persuasive messages Furthermore each ofthese studies has focused narrowly on just one aspect of per-suasive messages (eg MSV AS)

Methods

PPI analysis

Results

PPI results

Discussion

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Methods

PPI analysis

Results

PPI results

Discussion

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Results

PPI results

Discussion

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Results

PPI results

Discussion

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Discussion

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

(Continued)

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Table 4 (Continued)

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

1 2 3 4 5 6

(two-tailed)

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Conclusion

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Acknowledgement

Funding

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

nsx126-TF1

nsx126-TF2

nsx126-TF3

nsx126-TF4

nsx126-TF5

Documents

The persuasion network is modulated by drug-use risk and