Brain mediators of the effects of noxious heat on pain w w w . e l s e v i e r . c o m / l o c a t e / p a i n PAIN � 155 (2014) 1632–1648 Brain mediators of the effects of noxious heat on pain http://dx.doi.org/10.1016/j.pain.2014.05.015 0304-3959/� 2014 International Association for the Study of Pain. Published by Elsevier B.V. All rights reserved. ⇑ Corresponding author. Address: National Institutes of Health, Building 10, CRC Rm 4-1741, MSC 1302, Bethesda, MD 20892. Tel.: 301 435 4541. E-mail address: lauren.atlas@nih.gov (L.Y. Atlas). Lauren Y. Atlas a,⇑, Martin A. Lindquist b, Niall Bolger c, Tor D. Wager d a Section on Affective Neuroscience and Pain, National Center for Complementary and Alternative Medicine, US National Institutes of Health, Bethesda, MD, USA b Department of Biostatistics, Johns Hopkins University, Baltimore, MD, USA c Department of Psychology, Columbia University, New York, NY, USA d Department of Psychology and Neuroscience, University of Colorado–Boulder, Boulder, CO, USA Sponsorships or competing interests that may be relevant to content are disclosed at the end of this article. a r t i c l e i n f o a b s t r a c t Article history: Received 1 October 2013 Received in revised form 1 May 2014 Accepted 14 May 2014 Keywords: fMRI Mediation Neuroimaging Nociception Pain Connectivity Human Recent human neuroimaging studies have investigated the neural correlates of either noxious stimulus intensity or reported pain. Although useful, analyzing brain relationships with stimulus intensity and behavior separately does not address how sensation and pain are linked in the central nervous system. In this study, we used multi-level mediation analysis to identify brain mediators of pain—regions in which trial-by-trial responses to heat explained variability in the relationship between noxious stimulus intensity (across 4 levels) and pain. This approach has the potential to identify multiple circuits with complementary roles in pain genesis. Brain mediators of noxious heat effects on pain included targets of ascending nociceptive pathways (anterior cingulate, insula, SII, and medial thalamus) and also prefron- tal and subcortical regions not associated with nociceptive pathways per se. Cluster analysis revealed that mediators were grouped into several distinct functional networks, including the following: somatosen- sory, paralimbic, and striatal-cerebellar networks that increased with stimulus intensity; and 2 networks co-localized with ‘‘default mode’’ regions in which stimulus intensity-related decreases mediated increased pain. We also identified ‘‘thermosensory’’ regions that responded to increasing noxious heat but did not predict pain reports. Finally, several regions did not respond to noxious input, but their activity predicted pain; these included ventromedial prefrontal cortex, dorsolateral prefrontal cortex, cerebellar regions, and supplementary motor cortices. These regions likely underlie both nociceptive and non-nociceptive processes that contribute to pain, such as attention and decision-making processes. Overall, these results elucidate how multiple distinct brain systems jointly contribute to the central generation of pain. � 2014 International Association for the Study of Pain. Published by Elsevier B.V. All rights reserved. 1. Introduction The relationship between stimulus intensity and perception is lawful and robust in all perceptual domains, including pain [2,72]. However, although higher stimulus intensities usually lead to greater pain, there is nearly always variability in the stimulus– response relationship. Pain perception is strongly influenced by spontaneous fluctuations in arousal and attention [14,46,62], stim- ulus history [12,37,64], and other factors. Thus, a given stimulus intensity can be perceived or reported as painful or nonpainful depending on brain activity before [14,62], during [63], or after noxious stimulation [9,44]. The purpose of the present study was to examine how variations in noxious stimulus intensity are transformed into variations in pain, focusing specifically on responses during noxious stimulation itself. In particular, we sought to identify regions that mediate stimulus effects on pain and those that do not respond strongly to noxious stimuli but nonetheless play supporting roles in pain genesis. Targets of spino-thalamo-cortical nociceptive pathways [32] and other nociceptive pathways (eg, spino-parabrachial and spi- no-reticular [83]) reliably track the stimulus intensity of painful events in human neuroimaging studies, including somatosensory (SI/SII), dorsal posterior [dpINS], anterior insular [aIns], and ante- rior cingulate [aCC] cortices and thalamus [5,35,60]. There is broad consensus that these ‘‘intensity coding’’ regions also generally cor- relate with pain [25,76], although the stimulus–response function between brain response and pain report may differ depending on the region [15,20,38,48,63]. http://crossmark.crossref.org/dialog/?doi=10.1016/j.pain.2014.05.015&domain=pdf http://dx.doi.org/10.1016/j.pain.2014.05.015 mailto:lauren.atlas@nih.gov http://www.elsevier.com/locate/pain L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1633 However, only a few studies have directly compared brain activity related to stimulus processing with activity related to pain, and findings are mixed on which areas are most strongly associ- ated with each. One seminal study found that dpINS correlated preferentially with stimulus intensity and that aIns correlated pref- erentially with perceived pain [29], whereas another study found the opposite [6]. In addition, these previous studies do not provide models of how stimulus-related brain activity and pain-related brain activity are linked. Although these studies identify correlates of either noxious stimulus intensity or pain perception, we know little about the brain processes that transform stimulus processing into pain, and which processes might contribute to pain indepen- dent of stimulus processing. We used whole-brain multi-level mediation analysis [7,80,81], a linear multivariate approach that relates stimuli, brain responses, and behavior in a single model, to understand the pathways that mediate the effects of noxious input on pain per- ception. We identified 3 classes of relevant brain processes: (1) mediator regions that link stimulus intensity with pain; (2) ther- mosensory regions that respond specifically to noxious input; and (3) pain-related regions that contribute to decisions about pain above and beyond the linear and nonlinear effects of noxious stimulus intensity and thus may reflect endogenous decision- making processes that contribute to variations in pain, such as arousal, attention, and magnitude estimation. We identify net- works with distinct functional properties related to pain genesis, which could help create a clearer picture of the multiple systems involved in creating pain. This approach could also serve as a model for understanding sensory decision making in other per- ceptual modalities. 2. Methods 2.1. Participants and procedure 2.1.1. Participants Thirty healthy, right-handed participants were enrolled in the study. Participants were recruited from the New York metropolitan area through posted flyers, advertisements on Craigslist, and if they had previously participated in studies in our laboratory and volun- teered to be contacted for future research. All participants provided informed consent in accordance with the Declaration of Helsinki, as approved by the Columbia University Institutional Review Board. Preliminary eligibility was assessed with a general health questionnaire, a pain safety screening form, and a functional mag- netic resonance imaging (fMRI) safety screening form. Participants reported no history of psychiatric, neurological, or pain disorders. Three participants completed calibration but did not undergo scanning because of technical problems with the heat equipment (2 participants) or discomfort with the MR environment (1 partic- ipant). The fMRI imaging sequence was incorrect for 1 additional participant, leaving a final sample of 26 participants (9 female and 17 male, mean age = 27.8 years, range: 20–50 years). 2.1.2. Thermal stimulation and pain ratings Thermal stimulation was delivered to the volar surface of the left (nondominant) inner forearm using a 16 � 16-mm Peltier ther- mode (Medoc, Inc.). Each stimulus lasted 10 seconds, with 1.5-sec- ond ramp-up and ramp-down periods and 7 seconds at target temperature. Temperatures were individually calibrated for each participant using an adaptive staircase procedure. During calibra- tion and during the fMRI portion of the experiment, participants rated stimulation on a continuous scale from 0 to 8 (0 = no sensa- tion; 1 = nonpainful warmth; 2 = low pain; 5 = moderate pain; 8 = maximum tolerable pain). This scale has been used in previous studies in our laboratory [7,8] and provides measures of pain threshold and tolerance. It is similar to the 0 to 5 scale used by Bornhovd et al. [15] and Buchel et al. [20] but provides a broader range to increase sensitivities to subtle variations in perception. We used a continuous visual analogue scale (VAS) during fMRI scanning, which provided further sensitivity to small fluctuations in pain. The calibration procedure allowed us to derive each partici- pant’s stimulus–response curve for the relationship between applied thermal stimulation and reported pain, and to identify sites on the forearm with similar nociceptive profiles (ie, the 3 with the lowest average residuals based on the predicted stimulus– response function). During the fMRI experiment, heat was applied to the 3 sites that responded most similarly to changes in temper- ature, and temperatures were selected for each individual based his or her dose–response curve. Institutional review board restric- tions precluded us from applying temperatures higher than 48�C, so all participants were required to have maximum tolerable pain levels fall within the range of 42�C to 48�C. One participant exceeded this range (maximum predicted temperature based on calibration, 50�C) but was included in the experiment and received a maximum stimulus of 48�C. No participants reported maximum tolerable pain that fell below 42�C. 2.1.3. fMRI task design fMRI images were acquired during 6 functional runs (8 trials per run, 48 trials). The thermode was placed on a different skin site for each run, with 2 total runs per skin site. The task design is shown in Fig. 1. At the start of each trial, a square appeared in the center of the screen for 50 milliseconds, followed by a pair of faces [34]. An emotional expression (Happy or Fearful) was presented for 33 milliseconds, masked by a neutral face pre- sented for 1467 milliseconds. Face cues were evenly crossed with temperature. As the conceptual focus of the present article concerns the mechanisms that link changes in temperature with changes in pain, our mediation analyses collapse across the face cues to examine pain-evoked responses during noxious stimulation period as a function of temperature. To test whether face primes were ignorable here, we controlled for the effects of face cues (main effects and interactions with temperature) on regions identified in our mediation analysis. This assessed the possibility that masked emotional faces induced variability in the temperature– pain relationship. No main effects of face primes were found on the regions that we report here, and all results reported were sig- nificant after controlling for face prime identity; thus, we do not report priming effects in detail. A full analysis of the face primes is awaiting replication and extension in future experiments, and is not the main focus of this article. Cue presentation was followed by a 6-second anticipatory inter- val during which a fixation cross was presented on the screen. Thermal stimulation was then delivered via the thermode for 10 seconds (1.5-second ramp up from baseline [32�C], 7 seconds at peak destination temperature, 1.5-second return to baseline) at levels calibrated to elicit ratings of nonpainful warmth (VAS rat- ing = 1; mean = 40.8�C, standard deviation [SD] = 2.03), low pain (VAS rating = 3; mean = 43.1�C, SD = 2.10), medium pain (VAS rat- ing = 5; mean = 45.1�C, SD = 1.79), or high pain (VAS rating = 7; mean = 47.0�C, SD = 1.14). After thermal stimulation, a fixation cross was presented for a 14-second fixed interstimulus interval (ISI). The words ‘‘How painful?’’ then appeared on the screen for 4 seconds above an 8-point VAS. Participants rated the pain evoked by the preceding stimulus using an fMRI-compatible track-ball (Resonance Technologies, Inc.) with resolution equivalent to the screen resolution (ie, approximately 600 discrete values between Fig. 1. Task and analysis framework. (A) Task design. On each trial, participants saw a cue, followed by a 6-second anticipatory delay. They then experienced 10 seconds of noxious thermal stimulation at temperatures calibrated to elicit ratings of nonpainful warmth, low pain, moderate pain, or high pain. After a delay, participants rated the pain that they had experienced. Analyses examine brain responses during noxious stimulation, using area under the curve (AUC) as a measure of evoked response. (B) Multi-level mediation analysis. Our mediation analyses use linear mixed models to examine the dynamic, trial-by-trial relationship between objective stimulus (left, X; temperature), voxelwise stimulus-evoked responses (bottom, M; AUC estimates from single trial analysis), and pain experience (right, Y; pain report). The 4 components of multilevel mediation analysis address the current study’s key questions. Top, path c/c0 ’: Does temperature affect perceived pain as measured by trial-by-trial pain reports? Path c reflects the total relationship between predictive cue and reported pain on medium trials, and path c0 reflects the direct behavioral relationship, controlling for activity in the mediator—in this case, a brain voxel or region. Left: Path a (‘‘Temperature effects’’) provides inferences on whether brain voxels are modulated by noxious input (Fig. 2A). Right: Path b (‘‘report-related responses’’) provides inferences on whether brain activity in each voxel predicts trial-by-trial pain reports, controlling for temperature (Fig. 2B). To supplement the standard linear model, we controlled for nonlinear effects of temperature when assessing path b. Middle: The mediation effect (c – c0) provides inferences on whether brain voxels explain a significant amount of the covariance between temperature and perceived pain (Figs. 3 and 4). (C) Pain-processing brain network. We were most interested in testing for effects within the a priori pain-processing brain network. We created an independent localizer using a mega-analytic approach across 4 different studies of noxious thermal stimulation (n = 114; see Methods) to identify regions associated with pain and nociception. 1634 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1 and 8). The delay between pain offset and rating was included to maximize our ability to isolate pain-period responses using single- trial analysis. A 10-second ISI separated the rating period from the start of the next trial. In this report, we focus on brain responses evoked during the 10-second thermal stimulation period. Anticipatory responses and responses during the pain rating period were analyzed and reported for a subset of participants, along with heat-evoked responses, in a separate study [78] that used qualitatively different methods (pattern classification, rather than mediation analysis) and thus provides complementary findings. 2.1.4. Behavioral analyses Behavioral data were analyzed using custom code implement- ing a linear mixed-effects model in Matlab software (MathWorks, Natick, MA), and verified using the Mixed procedure in SAS 9.1 (SAS Institute, Cary, NC). We modeled within-subject effects of Temperature and Masked Emotion prime on pain report, as well as the Temperature �Masked Emotion interaction. To test for non- linear effects of temperature on pain, we log-transformed temper- atures and ratings and tested a linear mixed-effects model in log space. 2.2. fMRI acquisition and preprocessing 2.2.1. Imaging acquisition Whole-brain fMRI data were acquired on a 1.5-T GE Signa Twin Speed Excite HD scanner (GE Medical Systems) at Columbia Uni- versity’s Program for Imaging in Cognitive Science. Functional images were acquired with an echo-planar imaging sequence (TR = 2000 ms, TE = 34 ms, field of view = 224 mm, 64 � 64 matrix, 3.5 � 3.5 � 4.0-mm voxels, 29 slices). Each run lasted 6 minutes and 18 seconds (189 TRs). Stimulus presentation and data acquisi- tion were controlled using E-Prime software (PST Inc.). 2.2.2. Preprocessing Before preprocessing, we identified global outlier time points by computing both the mean and the SD of values in each image for each slice. Mahalanobis distances for the matrix of mean values (1 per slice) � functional volumes were computed, and images with a value greater than 3 SDs were considered to be outliers. The same procedure was used for SD values. Outlier time points were modeled as indicator vectors at the level of single trial esti- mation (see below, ‘‘Single trial analysis’’). We also used principal components analysis (PCA) as an approach to noise reduction, L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1635 using methods similar to previous work [74,75]. In brief, standard PCA was used to define 10 components expressed across the brain. A task-related design matrix was created using the timing of the fMRI paradigm. Cue, heat, and rating onsets were modeled with boxcars representing event duration and convolved with SPM’s canonical hemodynamic response function plus time and disper- sion derivatives. A nuisance-related design matrix was also con- structed from the time series of global outlier time points and motion parameters. For each run, task- and nuisance-related predictors were regressed on each component time series, which provided correlation coefficients for how nuisance related or task related each component was. The ratio between correlation coeffi- cients was used as a first step in determining whether each compo- nent was artifactual and should be removed: No components were removed that had a task-related R2 value greater than 0.1, or whose ratio of nuisance-to-task was less than 2.0. Components that fit these criteria were then visualized both spatially (pattern of weights across voxels) and temporally (average activity across time). Trained research assistants carefully inspected the compo- nents, and removed only those components that were clearly artifactual (eg, those expressed only at the edge of the brain, those that included an obvious single spike, etc.). An average of 1.09 components per run was removed per subject (SD = 0.69). After PCA denoising, functional images were slice-acquisition- timing and motion corrected using FSL (FMRIB’s Software Library; http://www.fmrib.ox.ac.uk/fsl/). Structural T1-weighted images were coregistered to the first functional image for each subject using an iterative procedure of automated registration using mutual information coregistration in SPM5 (Wellcome Trust Centre for Neuroimaging, London, UK) and manual adjustment of the automated algorithm’s starting point until the automated pro- cedure provided satisfactory alignment. Data were smoothed with an 8-mm–full width at half maximum Gaussian smoothing kernel using SPM5. Structural images were normalized to MNI space using a genetic algorithm (GA)–based normalization. This approach is a refine- ment of the standard SPM5 warping and improves intersubject registration, and is described in detail elsewhere [7]. We note that because we started with a group mean registered to MNI space, the resulting warped images were still registered with MNI space. Normalized functional images were interpolated to 2 � 2 � 2-mm voxels. 2.3. fMRI analysis 2.3.1. Single trial analysis We used single trial analysis to elicit a measure of brain response (eg., area under the curve) to each trial, in each voxel, for each subject. Our method is described in detail elsewhere [7]. In brief, we quantified single-trial response magnitudes by con- structing a GLM design matrix with separate regressors for each trial, as in the ‘‘beta series’’ approach of Rissman et al. [67]. How- ever, we used a flexible basis set to model each trial, thus allowing the shape of the modeled hemodynamic response function (HRF) to vary across trials and voxels. The basis set consisted of 3 curves shifted in time and was customized for thermal pain responses based on previous studies [47]. The positions of the 3 basis func- tions were fixed in time relative to trial onset, and we allowed amplitude to vary to construct a fitted response on each trial. We evaluated the height, duration, delay-to-peak, and area under the curve (AUC) of each trialwise fitted response. We chose to use the AUC of each fitted response (for each trial within each voxel) as a summary estimate of pain-period activity, as noxious heat has been shown to influence not only amplitude but also the dura- tion of the evoked HRF [47], and thus the AUC captures these joint effects better than standard beta regressor amplitudes. When using a single-trial approach, it is common for some trials to be contam- inated by movement or other artifacts. To mitigate such artifacts, we calculated trial-by-trial variance inflation factors (a measure of design-induced uncertainty due, in this case, to collinearity with nuisance regressors). Any trials with variance inflation factors that exceeded 2 or whose whole-brain AUC estimates exceeded 3 SDs from the mean were excluded from subsequent analyses (mean = 2.69, SD = 2.07, range = 0–7 trials per participant). 2.3.2. Mediation analysis We used whole-brain multi-level mediation to examine the relationship between trial-by-trial changes in temperature (X), heat-evoked responses within a voxel (M), and reported pain (Y), as illustrated in Fig. 1. Our methods have been described in detail in previous work [7,80,81]. The mediation approach is complemen- tary to recent attempts to develop pain-predictive brain patterns using machine learning [16,18,52,66,78]. Whereas machine learn- ing integrates activity across many regions into patterns that predict a single outcome variable (eg, pain reports), the present multivariate approach relates responses within voxels to both noxious input and pain. We then address the complexity of these relationships by identifying multiple networks that make separate contributions to pain (see ‘‘Network analysis,’’ below). The mediation analysis presented here includes the following tests: (1) whether changes in temperature influence activation within a voxel (measured by AUC) during heat (path a); (2) whether heat-evoked responses are correlated with pain reports, controlling for temperature (path b); and (3) whether the overall relationship between temperature and pain report (path c) decreases when controlling for heat-evoked responses within a voxel [c-c0, equivalent to a⁄b+cov(a,b)]. This final test is a test of mediation. The interpretation of a mediator is that there is an indi- rect pathway through this region that links changes in external stimuli with subjective responses, and that if a mediator were dis- rupted, the stimulus–response relationship would be diminished or abolished entirely. We used bootstrap tests to test the significance of the mediation effect and each individual path. Bootstrap testing provides a more sensitive test of mediation than the Sobel test [13,33,71,73]. We tested the significance of all effects (a, b, and a�b) using a bias- corrected, accelerated bootstrap [33]. We estimated distributions of subject level path coefficients by randomly sampling with replacement 10,000 observations (rows) from the matrix of [a b a�b] path coefficients for each voxel. Two-tailed, uncorrected P values were calculated from the bootstrap confidence interval. Multi-level mediation analysis incorporates a critical compo- nent, not present in single-level mediation, that offers a novel perspective on the relationship between within- and between- subject effects. The mediation effect (the decrease in the behavioral relationship when controlling for a mediator) is driven by a combi- nation of consistent effects (paths a and b) across a group of sub- jects and by covariance between paths a and b across individuals [39]—that is, the individuals that show strongly positive path a effects also show strongly positive path b effects, and vice versa. This implies that a region can mediate the relationship between independent and dependent variables either because it is used in similar ways by all subjects (eg, the region shows increases with temperature [path a] and greater activation predicts higher pain controlling for temperature [path b]) or because individuals vary in consistent ways across paths a and b (eg, for subjects who show increases with temperature, greater activation predicts higher pain, and for subjects who show decreases with temperature, less activation predicts higher pain). The notion of mediation driven by covariance is relevant because individuals vary in the magnitude of pain that they report with a given stimulus temperature in mean- ingful ways [24,56,84]. These person-level differences are often http://www.fmrib.ox.ac.uk/fsl/ 1636 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 likely to be psychologically meaningful; for instance, individuals may use different strategies to cope with pain. To examine whether mediation was driven by consistent effects across the group or by covariance between temperature effects and pain-related responses, we identified clusters that showed evi- dence of mediation, and extracted trialwise data from each subject within each cluster. We then tested whether there were significant effects across the group in paths a and b (ie, whether there was a consistent group effect, essentially identical to what would be found with a standard 1-sample t test on path a or path b coeffi- cients), or whether there were significant correlations between a and b coefficients across individuals. The former are reported as mediators that show consistent effects, whereas the latter are reported as driven by covariance and are likely to reflect psycho- logically meaningful individual differences, although identifying the sources and psychological correlates of these individual differ- ences likely requires large-sample studies and is beyond the scope of the present report. 2.3.3. Thresholding All results are false discovery rate (FDR) corrected at q < 0.05, which corresponds to a voxelwise threshold of P < .001 for the mediation effect, and a threshold of P < .003 for the conjunction across all 3 maps (a, b, a�b). We imposed a cluster extent threshold of 3 voxels at the lowest threshold. Network analyses and cluster- wise tests were performed on data extracted from these voxels along with any contiguous voxels at uncorrected thresholds of P < .005 and P < .01. We were specifically interested in characterizing responses within regions classically associated with pain-related processing [5,35,60]. We used an independent localizer to identify pain- processing brain network (PPBN) regions of interest (ROIs), using methods similar to those previously reported [7]. In brief, we used a mega-analytic approach to compare high- versus low-intensity thermal stimulation across 4 different studies from our laboratory (n = 114). Voxels that showed a significant [high – low intensity] effect on the basis of familywise error correction (P < .05) using Gaussian random fields as implemented in SPM5 (Fig. 1C) were defined as the PPBN. Anatomical localization was determined based on the LONI Probabilistic Brain Atlas [70]. This revealed the following regions as comprising the PPBN (Fig. 1C): bilateral superior, inferior, and middle frontal gyrus; bilateral precentral gyrus; right middle and bilateral lateral orbitofrontal gyrus; bilat- eral postcentral gyrus; bilateral supramarginal gyrus; bilateral superior and middle temporal gyrus; right hippocampus; right parahippocampal gyrus; left lingual gyrus; bilateral insular cortex; bilateral cingulate gyrus; bilateral caudate; bilateral putamen; brainstem; cerebellum, bilateral thalamus, and medial thalamus. For each path, we report results from ROIs that fell within the PPBN localizer, as well as regions outside the a priori PPBN. 2.3.4. Relationship to stimulus versus report We tested for selectivity to temperature effects or variations in pain report by identifying voxels that were activated by tempera- ture but unrelated to variations in report (ie, showed path a effects but neither path b effects nor mediation), and voxels that showed the reverse relationship (path b effects but neither path a effects nor mediation). To ensure specificity to one process over the other, we identified voxels that showed significant effects in one path at P < .001, and whose P values for the other path effect and the medi- ation effect exceeded P > .05 uncorrected. Within path b regions, we also tested for specificity to pain report and independence from nociceptive responses by testing whether brain regions predicted pain controlling for temperature, as well as brain responses in mediators that fell within the bound- aries of the PPBN. Trial-by-trial AUC estimates were averaged across PPBN mediator regions and included as covariates in multi- level mediation analyses for each path b region (activation averaged across all voxels). This in turn allowed us to test the rela- tionship between brain response and pain report controlling for both temperature and PPBN mediator response. Results are reported in Table 2. 2.3.5. Network analysis We used dimension reduction with nonmetric multidimen- sional scaling (NMDS) and clustering to test whether mediator regions grouped into separate networks, using a multisubject extension of methods described in previous work [40,79]. We extracted the entire series of trial AUC estimates for each subject for each voxel within each mediator region. The goal was to pro- vide a reasonable estimate of which voxels were grouped into coherent parcels (contiguous regions) and how these regions grouped into functional networks. A detailed explanation of the steps is provided below; in addition, the code for the analysis is available on our Web site (http://wagerlab.colorado.edu/tools). The function parcel_clusters.m in the CANlab Core Tools package implements the 3 major steps described below and visualized in Fig. 2: 1. Dimension reduction: Clustering of multivariate data is most stable when the data are not sparse, that is, the dimensionality is low relative to the number of observations. To limit the dimensionality of the data, a spatio-temporal dimension-reduction step is first performed on the [n � v � N] data matrix of AUC data for n trials � v vox- els � N participants (here, n = 48 trials [usually], v = 17,112 voxels, and N = 26 participants). A temporal data reduction is first performed to identify components with correlated AUC trial time series within each participant, followed by a spatial reduction to identify components with correlated spatial pat- terns across subjects. First, the [n � v] matrix of AUC data for each participant was subjected to PCA, using the [v � v] corre- lation matrices. Based on the scree plots across subjects, we saved the first 7 eigenvectors (spatial maps). These eigenvectors explained 74 ± 2.5% (SD across subjects) of the variance in the full dataset. These eigenvectors were scaled by their variances (eigenvalues) and concatenated across subjects to form an [v � N⁄7] matrix of eigenvectors. This matrix was subjected to a second (across participant) PCA step to identify components with similar spatial maps across participants. We retained 12 eigenvectors (maps) based on the scree plot, which explained 65% of the variance across individuals. Component scores in this space were used for clustering. This is a data reduction step, and the results are not expected to depend strongly on the number of eigenvectors retained at either step, as long as most of the variance in the data is explained. 2. Parcellation: Hierarchical agglomerative clustering with average linkage was used to group voxels into parcels—sets of contiguous voxels with similar profiles—in the [v � 12] matrix of component maps. The goal of parcellation was to reduce the space from voxels to parcels (regions) for nonmetric multidimensional scaling (NMDS)–based clustering of regions, so that NMDS is computationally tractable. Voxels whose trial AUC time series did not correlate with that of other voxels in the same parcel at P < .001 (in a random effects analysis across participants) were pruned from each parcel. Of 140 parcels total, 127 parcels with more than 3 voxels after pruning were retained for subse- quent analysis. 3. NMDS and clustering: Parcels were now treated as the unit of analysis, and the AUC trial time series data were extracted from each parcel for each Table 1 Path a: Regions correlated with stimulation temperature. Brain region (Brodmann area) Coordinates (mm) No. of voxels Volume (mm) Max z score Mean x y z P < .001 Path a coeff (SE) Mean p value path a Temperature-related increases Bilateral cerebellum –14 –66 –34 6081 48,648 15.13 0.9010 (0.17) .0000 Right temporal pole 48 –22 –36 6 48 5.86 0.0774 (0.03) 0.0026 Right insula, contiguous with bilateral thalamus, midbrain, right inferior frontal gyrus, Right lateral PFC, right premotor cortex, right lateral OFC, Right hippocampus 36 2 8 12180 97,440 17.22 0.9217 (0.12) .0001 Right inferior frontal gyrus p. orbitalis, contiguous with right anterior insula and Right VLPFC 26 14 –24 12 96 5.45 0.4488 (0.17) .0087 Left middle orbital Gyrus –26 38 –20 84 672 11.18 0.6833 (0.18) .0002 Left middle orbital Gyrus –22 58 –16 38 304 6.26 0.8102 (0.34) .0050 Right middle temporal gyrus 58 –40 –10 43 344 8.87 0.4104 (0.12) .0006 Left rolandic operculum –46 –4 10 3654 29,232 13.63 0.8228 (0.10) .0002 Left calcarine gyrus (BA17) 2 –84 0 416 3328 10.26 0.9811 (0.25) .0003 Right calcarine Gyrus (BA17) 16 –102 –4 18 144 5.9 0.3819 (0.15) .0043 Cerebellar vermis 4/5 2 –46 –2 106 848 13.82 1.2624 (0.42) .0000 Left insula lobe (OP2) –38 –18 14 54 432 12.09 0.4722 (0.13) .0001 Right middle cingulate cortex (BA6) 6 6 44 5699 45,592 19.94 0.8975 (0.12) .0002 Left middle frontal gyrus –38 42 28 320 2560 10.1 0.7566 (0.17) .0016 Right middle frontal gyrus 38 50 28 18 144 5.62 0.8758 (0.30) .0037 Left caudal cingulate –18 –28 34 55 440 8.06 0.4528 (0.13) .0003 Left inferior parietal lobule (hIP2) –44 –42 36 6 48 6.27 0.3704 (0.15) .0019 Left precuneus –6 –80 54 4 32 5.21 0.5817 (0.25) .0100 Left superior frontal gyrus (BA6) –18 4 66 238 1904 12.39 0.2750 (0.07) .0000 Right superior Parietal lobule 24 –46 70 367 2936 13.42 0.2867 (0.07) .0001 Temperature-related decreases Right medial Temporal pole 44 12 –42 30 240 6.62 –0.5364 (0.18) .0006 Left hippocampus, contiguous with parahippocampal gyrus left medial temporal lobe, left occipital cortex –34 –54 –8 7155 57,240 10.95 –0.7752 (0.10) .0006 Right hippocampus, contiguous with parahippocampal gyrus and right occipital cortex 40 –64 2 5973 47,784 11.59 –0.8269 (0.14) .0005 Right medial Temporal pole 54 14 –32 2 16 4.87 –0.6793 (0.27) .0084 Right middle temporal gyrus 58 –8 –16 429 3432 10.36 –0.9178 (0.21) .0001 Right amygdala 14 2 –28 4 32 5.71 –0.4142 (0.14) .0016 Left rectal gyrus –6 14 –24 22 176 7.02 –0.3120 (0.09) .0005 Right inferior Temporal gyrus 66 –24 –26 3 24 5.02 –0.1715 (0.06) .0033 Left superior medial gyrus –4 54 4 1634 13,072 9.44 –0.9172 (0.20) .0047 Left rectal gyrus –8 24 –14 11 88 5.86 –0.4067 (0.14) .0034 Right cerebelum IV–V 10 –56 –12 4 32 4.86 –0.5879 (0.22) .0075 Left inferior frontal gyrus p. orbitalis –54 24 –4 42 336 6.96 –0.7147 (0.23) .0013 Right precuneus 2 –44 44 2458 19,664 11.26 –0.7598 (0.13) .0001 Left superior frontal gyrus –18 62 2 6 48 5.28 –0.4488 (0.17) .0078 Right superior temporal gyrus 44 –28 4 33 264 7.57 –0.3953 (0.11) .0002 Right calcarine gyrus (BA18) 22 –58 8 13 104 6.52 –0.6114 (0.21) .0052 Right precentral gyrus (BA1) 52 –14 46 1377 11,016 10.94 –0.7646 (0.15) .0004 Left postcentral gyrus (BA3b) –40 –32 52 3869 30,952 11.1 –0.7422 (0.13) .0014 Right superior frontal gyrus 22 36 38 350 2800 9.41 –0.5574 (0.13) .0003 Left middle frontal gyrus –40 10 50 122 976 7.41 –0.5397 (0.16) .0011 Left middle frontal gyrus –26 24 46 138 1104 8.49 –0.8461 (0.25) .0009 Coeff, coefficient; Max, maximum; OFC, orbitofrontal cortex; PFC, prefrontal cortex; SE, standard error; VLPFC, This table presents results from path a of the multilevel mediation analysis presented as a path diagram in Fig. 1. Mediation effect parametric mapping results are presented in Fig. 3. L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1637 subject. Values in the [n trials � parcels � N subjects] data matrix were z scored within participant to remove intersubject differ- ences in scaling, then concatenated into an [n⁄N � parcels] data matrix. Correlations among parcels were converted into a [parcels � parcels] matrix of distances using the formula dis- tance = (1�r)/2. The NMDS stress plot (which operates on ranked distances and is therefore more robust to outliers and does not require a strictly Euclidean distance space) was exam- ined, and 7 dimensions (which explained 79% of the variance in distances) were retained for the final cluster analysis. Hierarchical agglomerative clustering with average linkage was used to cluster the parcels in this space into interconnected networks with correlated AUC trial time series. To choose the number of clusters in the final solution (k), for every possible choice of clusters between k = 2 and 20, we compared the clus- ter solution to the average and SD of 1000 clustering iterations with permuted parcel time series. This yielded a z score ([actual solution – mean permuted solution]/standard deviation) of per- muted solution for each value of k. The best solution was k = 13, with a value of Z = 5.57 compared to the null-hypothesis single- cluster solution (P < .0001). We report results from the solution that divided parcels into 13 networks. Two networks consisted solely of parcels in white mat- ter and were excluded from the results reported below. For each network, we averaged over in-network voxels for each subject and used multilevel mediation to test whether average network responses mediated temperature effects on pain. Table 2 Path b: Regions correlated with subjective pain, controlling for temperature. Brain region (Brodmann area) Coordinates (mm) No. of voxels Volume (mm) Max z score Linear temperature effects model: Mean path b coeff (SE) Nonparametric temperature effects model: Mean path b coeff (SE) Linear temperature effects, controlling for faces: Mean path b coeff (SE) PPBN mediator independence model: Mean path b coeff. (SE) x y z P < .001 Report-related increases Cerebellar vermis 2 –66 –32 3119 24,952 36.04 0.0381 (0.01)⁄⁄⁄ 0.0283 (0.01)⁄⁄⁄ 0.0371 (0.01)⁄⁄⁄ 0.0289 (0.01)⁄⁄⁄ Right superior orbital gyrus 18 44 –24 310 2480 12.52 0.0163 (0.00)⁄⁄⁄ 0.0129(0.00)⁄⁄⁄ 0.0157 (0.00)⁄⁄⁄ 0.0153(0.00)⁄⁄⁄ Left lateral OFC –36 44 –20 58 464 12.19 0.0065 (0.00)⁄⁄⁄ 0.0064 (0.00)⁄⁄⁄ 0.0062(0.00)⁄⁄⁄ 0.0067 (0.00)⁄⁄⁄ Left middle temporal gyrus –62 –26 –16 33 264 9.77 0.0051 (0.00)⁄⁄⁄ 0.0034 (0.00)⁄⁄ 0.0047(0.00)⁄⁄⁄ 0.0077 (0.00)⁄⁄ Right middle orbital gyrus 32 52 –18 32 256 9.48 0.0036 (0.00)⁄⁄⁄ 0.0032 (0.00)⁄⁄ 0.0030(0.00)⁄⁄⁄ 0.0025 (0.00)⁄⁄ Right insula lobe 46 10 2 2042 16,336 11.07 0.0179 (0.00)⁄⁄⁄ 0.0150 (0.00)⁄⁄⁄ 0.0170(0.00)⁄⁄⁄ 0.0068 (0.01)NS Left rolandic operculum, contiguous with insula –44 –2 2 2001 16,008 14.48 0.0185 (0.00)⁄⁄⁄ 0.0158 (0.00)⁄⁄⁄ 0.0178 (0.00)⁄⁄⁄ 0.0092 (0.00)⁄ Right globus pallidus 14 –16 –8 41 328 7.65 0.0095 (0.00)⁄⁄⁄ 0.0073 (0.00)⁄⁄ 0.0092 (0.00)⁄⁄⁄ 0.0037 (0.00)NS Left superior occipital gyrus (BA17) –6 104 4 72 576 9.14 0.2459 (0.21)⁄⁄⁄ 0.0785 (0.07)⁄ 0.1983 (0.17)⁄⁄ 0.0566 (0.04)⁄⁄ Left superior temporal gyrus –58 –39 20 428 3424 10.24 0.0117 (0.00)⁄⁄⁄ 0.0091 (0.00)⁄⁄ 0.0110 (0.00)⁄⁄⁄ 0.0014 (0.00)NS Right supramarginal gyrus 56 -32 32 1136 9088 13.44 0.0171 (0.00) 0.0147 (0.00)⁄⁄⁄ 0.0165 (0.00)⁄⁄⁄ 0.0096 (0.00)⁄ Right middle frontal gyrus 40 48 30 294 2352 11.21 0.0077 (0.00)⁄⁄⁄ 0.0074 (0.00)⁄⁄⁄ 0.0068 (0.00)⁄⁄⁄ 0.0039 (0.00)NS Left middle frontal gyrus –36 48 26 111 888 11.42 0.0074 (0.00)⁄⁄⁄ 0.0074 (0.00)⁄⁄⁄ 0.0075(0.00)⁄⁄⁄ 0.0046 (0.00)NS Right SMA (BA6), contiguous with dACC 8 –14 60 4191 33,528 30.8 0.0346 (0.01)⁄⁄⁄ 0.0244 (0.01)⁄⁄⁄ 0.0350 (0.01)⁄⁄⁄ 0.0227 (0.01)⁄⁄ Left superior parietal lobule –24 –50 70 336 2688 36.04 0.0173 (0.01)⁄⁄⁄ 0.0140 (0.00)⁄⁄ 0.0185 (0.01)⁄⁄⁄ 0.0081 (0.00)NS Right superior parietal lobule 40 –54 64 42 336 11.44 0.0080 (0.00)⁄⁄⁄ 0.0076 (0.00)⁄⁄⁄ 0.0073(0.00)⁄⁄⁄ 0.0040 (0.00)NS Report-related decreases Left parahippocampal gyrus –26 –18 –26 724 5792 10.36 –.0229 (0.00)⁄⁄⁄ –.0172 (0.00)⁄⁄⁄ –0.5456 (0.13)⁄⁄⁄ –.0194 (0.00)⁄⁄⁄ Left temporal pole –42 18 –42 54 432 9.29 –.0092 (0.00)⁄⁄⁄ –.0091 (0.00)⁄⁄ –0.2334 (0.13)NS –.0060 (0.00)⁄ Right medial temporal pole 26 16 –38 67 536 7.25 –.0132 (0.00)⁄⁄ –.0128 (0.00)⁄⁄⁄ 0.0643 (0.12)NS –.0137 (0.00)⁄⁄ Right temporalpole 42 10 –28 169 1352 8.74 –.0090 (0.00)⁄⁄⁄ –.0093 (0.00)⁄⁄⁄ –0.1961 (0.24)NS –.0086 (0.00)⁄⁄ Right parahippocampal gyrus (Hipp CA) 26 –12 –24 263 2104 9.33 –.0121 (0.00)⁄⁄⁄ –.0090 (0.00)⁄⁄⁄ –0.6034 (0.16)⁄⁄⁄ –.0099 (0.00)⁄⁄⁄ Right fusiform gyrus 30 –36 –14 168 1344 8.85 –.0116 (0.00)⁄⁄⁄ –.0087 (0.00)⁄⁄⁄ –0.7728 (0.18)⁄⁄⁄ –.0087 (0.00)⁄⁄ Right middle temporal gyrus 56 –6 –18 228 1824 9.16 –.0080 (0.00)⁄⁄⁄ –.0066 (0.00)⁄⁄⁄ –0.8675 (0.24)⁄⁄⁄ –.0068 (0.00)⁄⁄⁄ Left mid orbital gyrus, contiguous with MPFC (VMPFC) –2 54 –10 684 5472 10.62 –.0083 (0.00)⁄⁄⁄ –.0073 (0.00)⁄⁄⁄ –1.1621 (0.32)⁄⁄ –.0057 (0.00)⁄ Right middle 54 –66 16 649 5192 9.11 –.0181 (0.00)⁄⁄⁄ –.0165 (0.00)⁄⁄⁄ –0.5814 (0.15)⁄⁄⁄ –.0140 (0.00)⁄⁄ Temporal gyrus left inferior frontal gyrus p. orbitalis –28 32 –12 49 392 8.4 –.0068 (0.00)⁄⁄⁄ –.0044 (0.00)⁄⁄ –0.1378 (0.16) NS –.0073 (0.00)⁄⁄⁄ Left middle occipital gyrus –40 –80 22 752 6016 10.35 –.0135 (0.00)⁄⁄⁄ –.0110 (0.00)⁄⁄⁄ –0.6046 (0.16)⁄⁄⁄ –.0113 (0.00)⁄⁄⁄ Right precuneus 4 –56 26 985 7880 11.02 –.0092 (0.00)⁄⁄⁄ –.0089 (0.00)⁄⁄⁄ –0.7416 (.22)⁄⁄ –.0057 (0.00)⁄ Left rostral ACC –16 32 26 111 888 8.88 –.0120 (0.00)⁄⁄⁄ –.0012 (0.00)⁄⁄⁄ 0.1080 (0.15) NS –.0160 (0.00) ⁄⁄⁄ Left inferior frontal gyrus p. triangularis –38 18 26 127 1016 8.13 –.0122 (0.00)⁄⁄ –.0092 (0.00)⁄⁄ –0.1411 (0.18) NS –.0121 (0.00)⁄⁄ Left middle frontal gyrus –24 20 44 162 1296 9.12 –.0095 (0.00)⁄⁄⁄ –.0077 (0.00)⁄⁄⁄ –0.4318 (0.19)⁄ –.0063 (0.00) NS ACC, anterior cingulate cortex; coeff, coefficient; Max, maximum; MPFC, medial prefrontal cortex; NS, not significant; OFC, orbitofrontal cortex; PFC, prefrontal cortex; PPBN, pain-processing brain network; SE, standard error; VMPFC, ventromedial prefrontal cortex. This table presents results from path b of the multilevel mediation analysis presented as a path diagram in Fig. 1. Mediation effect parametric mapping results are presented in Fig. 3. ⁄⁄⁄P < .001; ⁄⁄P < .01; ⁄P < .05. 1 6 3 8 L.Y . A tla s et a l./P A IN � 1 5 5 (2 0 1 4 ) 1 6 3 2 – 1 6 4 8 Fig. 2. Connectivity analysis. We used a 3-step approach to define networks of mediators. (A) Step 1: Dimension reduction with principal components analysis. (B) Step 2: Parcellation to group voxels into regions. (C) Step 3: NMDS and clustering to identify 13 networks of parcels. L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1639 2.3.6. Accounting for potential nonlinearities Recent work demonstrates that noxious stimulus effects on brain responses may be nonlinear with respect to temperature [38] and that some regions show nonlinear relationships with respect to pain [15,19,48]. Nonlinearities in the relationship between temperature and pain are one of the effects that decouple stimulus intensity from reported pain, along with others (sequence and learning effects, attention, and variance related to decision- making processes). Our primary analysis controlled for linear effects of stimulus intensity when assessing the brain relationship with pain (path b), which allows nonlinear effects of temperature to drive some path b effects. In additional analyses, we ruled out potential nonlinear influences of temperature: We assessed path b (reported pain) and mediation effects controlling for temperature in a non- parametric fashion by including covariates for all possible differ- ences among the 4 discrete levels of heat. To do this, we evaluated a second mediation in which the input variable (‘‘X’’) reflected the difference between the top 2 temperatures and the bottom 2 temperatures (eg, [11�1�1] across the vector of [High pain, Med- ium pain, Low pain, Warmth] conditions), and we also modeled covariates for the difference between High and Medium conditions and Medium and Low conditions. This controls for any pairwise differences, because any systematic linear or nonlinear differences among intensity levels can be modeled as a linear combination across these 3 regressors. The results of these tests are reported in Tables 2 and 3. Path b effects on all regions remained significant after controlling for nonlinear effects of stimulus intensity. 2.3.7. Relationship to masked emotion face cues The primary focus of this report was to introduce a method to isolate the brain regions that link changes in noxious input to changes in pain perception. Although our design included masked presentations of face stimuli with happy or fearful expressions, iso- lating the effects of the faces themselves on brain responses was outside the scope of this article. However, mediation analysis takes advantage of the variability in the temperature-to-pain relation- ship, and, given our experimental design, some proportion of this variance might relate to the emotion face primes. To account for this possibility, Masked Emotion and Emotion � Temperature interactions were included as covariates in follow-up analyses for regions that showed significant path b and mediation effects in our primary linear mediation model (similar to the nonparametric analysis above). Results are reported below and included in Tables 2 and 3. 3. Results 3.1. Behavioral analysis 3.1.1. Behavioral relationship between stimulus and report Pain reports were highly correlated with noxious stimulus intensity. The linear mixed model revealed a significant effect of temperature on pain (t25 = 14.45, P < .001), such that pain reports increased by 0.73 points on the VAS scale for each unit change in temperature. The linear effect of temperature on pain is modeled as path c in the multilevel mediation analysis (Fig. 1A). As shown in Fig. 3, there was also substantial variability in this relationship (variance in linear effect of temperature: mean = 1.93, SD = 0.75). We searched for mediators that explained this variance. We also tested nonlinear effects of temperature on pain and found that the relationship was captured by a power law with exponent 2.49, consistent with previous work [65]. We therefore additionally modeled nonparametric effects of temperature for path b and mediator regions below. Table 3 Path a � b: Brain mediators of temperature effects on pain. Brain region (Brodmann area) Coordinates (mm) No. of voxels Vol (mm) Max a � b (SE) Cov (a, b) Path a and path b effects, if consistent across individuals Nonpara metric temperature effects model: Mean path ab coeff (SE) Mediation effect controlling for faces: Mean path ab coeff (SE) x y z P < .001 Positive mediators Right cerebelum crus 2 52 –52 –44 108 864 0.004 (0.001)⁄⁄⁄ 0.014 0.006 (0.002)⁄ 0.0042 (0.00)⁄⁄⁄ Right cerebelum VIII 28 –60 –48 61 488 0.003 (0.001)⁄⁄⁄ –0.104 0.003 (0.001)⁄ 0.0041 (0.00)⁄⁄⁄ Left parahippocamp al gyrus, contiguous with left cerebellum –24 –52 –26 4422 35376 0.005 (0.001)⁄⁄⁄ 0.023 0.007 (0.002)⁄ 0.0064 (0.00)⁄⁄⁄ Left medial temporal pole –36 14 –36 85 680 0.006 (0.002)⁄⁄⁄ 0.023 0.014 (0.003)⁄⁄ 0.0067 (0.00)⁄⁄⁄ Right temporal pole 48 –12 –46 15 120 0.002 (0.000)⁄⁄ 0.012 0.004 (0.001)⁄⁄ 0.0026 (0.00)⁄⁄⁄ Left temporal pole –34 –20 –42 20 160 0.004 (0.001)⁄⁄ 0.017 0.005 (0.002)⁄ 0.0044 (0.00)⁄⁄⁄ Right fusiform gyrus 36 –12 –40 16 128 0.002 (0.001)⁄⁄ 0.011 0.005 (0.002)⁄ 0.0044 (0.00)⁄⁄⁄ Right medial temporal pole 50 12 –26 461 3688 0.003 (0.001)⁄ 0.008 Negative a, negative b 0.007 (0.002)⁄⁄ 0.0041 (0.00)⁄ Right hippocampus, parahippocamp us, fusiform gyrus 28 –40 –16 749 5992 0.005 (0.002)⁄⁄⁄ 0.005 Negative a, negative b 0.007 (0.003)⁄ 0.0079 (0.00)⁄⁄⁄ Right inferior occipital gyrus 40 –74 –16 319 2552 0.007 (0.001)⁄⁄ 0.025 0.006 (0.002)⁄ 0.0083 (0.00)⁄⁄⁄ Left superior orbital gyrus –12 30 –22 14 112 0.003 (0.001)⁄⁄⁄ 0.015 0.004 (0.001)⁄⁄ 0.0040 (0.00)⁄⁄⁄ Left inferior frontal gyrus p. orbitalis, contiguous with left MPFC, rACC –36 40 –4 805 6440 0.007 (0.001)⁄⁄⁄ 0.027 0.009 (0.002)⁄⁄⁄ 0.0089 (0.00)⁄⁄⁄ Right rectal gyrus 6 32 –20 29 232 0.003 (0.001)⁄⁄⁄ 0.014 0.007 (0.002)⁄ 0.0034 (0.00)⁄⁄⁄ Left calcarinegyrus (BA18) –6 –94 –10 129 1032 0.006 (0.002)⁄⁄⁄ 0.013 0.005 (0.002)⁄ 0.0078 (0.00)⁄⁄⁄ Right cerebelum III 14 –32 –16 30 240 0.002 (0.001)⁄⁄ 0.011 0.006 (0.002)⁄⁄⁄ 0.0043 (0.00)⁄⁄⁄ Left middle orbital gyrus –20 54 –16 94 752 0.006 (0.001)⁄⁄⁄ 0.02 0.006 (0.002)⁄⁄ 0.0072 (0.00)⁄⁄⁄ Midbrain, hypothalamus 4 –12 –16 29 232 0.003 (0.001)⁄⁄⁄ 0.004 0.003 (0.002) NS 0.0050 (0.00)⁄⁄⁄ Left insulalobe, contiguous with caudate, putamen, left globus pallidus –36 8 2 1702 13616 0.013 (0.002)⁄⁄⁄ 0.011 Positive a, positive b 0.023 (0.003)⁄⁄ 0.0150 (0.00)⁄⁄⁄ Right insula Lobe 42 14 2 1971 15768 0.017 (0.03)⁄⁄⁄ 0.012 Positive a, positive b 0.030 (0.004)⁄⁄⁄ 0.0178 (0.00)⁄⁄⁄ Right middle occipital gyrus 36 –74 16 1686 13488 0.009 (0.002)⁄⁄⁄ 0.009 Negative a, negative b 0.015 (0.003)⁄⁄⁄ 0.0103 (0.00)⁄⁄⁄ Right middle orbital gyrus 46 52 –8 30 240 0.006 (0.002)⁄⁄⁄ 0.022 0.006 (0.002)⁄⁄ 0.0063 (0.00)⁄⁄⁄ Right thalamus 12 –14 0 220 1760 0.006 (0.002)⁄⁄⁄ 0.011 0.010 (0.004)⁄ 0.0081 (0.00)⁄⁄⁄ Right lateral PFC 38 42 –2 122 976 0.005 (0.001)⁄⁄⁄ 0.014 0.004 (0.001)⁄⁄ 0.0056 (0.00)⁄⁄⁄ Right middle temporal gyrus 58 –36 2 113 904 0.002 (0.001)⁄ 0.012 0.004 (0.001)⁄⁄ 0.0030 (0.00)⁄ Left superior medial gyrus –10 70 6 42 336 0.003 (0.001)⁄⁄ 0.009 Negative a, negative b 0.006 (0.002)⁄⁄ 0.0055 (0.00)⁄⁄⁄ Left superior temporal gyrus –62 –24 4 11 88 0.003 (0.001)⁄⁄⁄ 0.005 0.006 (0.002)⁄⁄⁄ 0.0046 (0.00)⁄⁄⁄ Area 17 –26 –72 6 19 152 0.003 (0.001)⁄⁄ 0.011 0.005 (0.002)⁄⁄ 0.0044 (0.00)⁄⁄⁄ Left middle temporal gyrus –56 –38 8 22 176 0.002 (0.001)⁄⁄⁄ 0.004 0.003 (0.001)⁄ 0.0031 (0.00)⁄ Right middle frontal gyrus 44 42 20 119 952 0.004 (0.001)⁄ 0.008 Positive a, positive b 0.004 (0.001)⁄⁄ 0.0056 (0.00)⁄⁄⁄ Right precuneus 12 –52 18 57 456 0.002 (0.001)⁄⁄⁄ 0.005 Negative a, negative b 0.004 (0.002) NS 0.0035 (0.00)⁄⁄⁄ Right SII, supramarginal gyrus 56 –24 24 139 1112 0.005 (0.001)⁄⁄⁄ 0.005 Positive a, positive b 0.005 (0.001)⁄⁄⁄ 0.0069 (0.00)⁄⁄⁄ Posterior cingulate cortex 0 –40 18 29 232 0.002 (0.000)⁄⁄⁄ 0.008 0.004 (0.001)⁄ 0.0038 (0.00)⁄⁄⁄ Left cuneus –4 –84 34 328 2624 0.004 (0.001)⁄⁄⁄ 0.017 0.002 (0.002) NS 0.0061 (0.00)⁄⁄⁄ Left cuneus –14 –56 22 75 600 0.002 (0.001)⁄⁄⁄ 0.008 0.008 (0.001) NS 0.0040 (0.00)⁄⁄⁄ Left SII, supramarginal Gyrus –58 –30 24 127 1016 (0.002)⁄⁄⁄ 0.009 Positive a, positive b 0.008 (0.002)⁄⁄ 0.0078 (0.00)⁄⁄⁄ Left postcentral gyrus (area 3b) –60 –4 28 65 520 0.005 (0.001)⁄⁄⁄ 0.014 0.007 (0.003)⁄⁄ 0.0063 (0.00)⁄⁄⁄ Left precentral gyrus –48 –2 24 14 112 0.003 (0.001)⁄⁄⁄ 0.008 0.003 (0.001)⁄ 0.0041 (0.00)⁄⁄ Left middle frontal gyrus –30 22 30 20 160 0.003 (0.001)⁄⁄⁄ 0.013 0.001 (0.001) NS 0.0029 (0.00)⁄ Right middlecingulate Cortex 4 10 40 434 3472 0.013 (0.003)⁄⁄⁄ 0.006 Positive a, positive b 0.016 (0.005)⁄⁄ 0.0156 (0.00)⁄⁄⁄ Right superior frontal gyrus 20 22 38 29 232 0.003 (0.001)⁄⁄⁄ 0.009 0.004 (0.001)⁄ 0.0040 (0.00)⁄⁄⁄ Left superior medial gyrus –4 26 52 198 1584 0.004 (0.001)⁄⁄⁄ 0.014 0.013 (0.003)⁄⁄⁄ 0.0062 (0.00)⁄⁄⁄ Left precentral gyrus (BA6) –54 –2 46 69 552 0.004 (0.001)⁄⁄⁄ 0.007 0.005 (0.002)⁄⁄ 0.0048 (0.00)⁄⁄⁄ Right precentralgyrus (BA6) 48 –10 52 226 1808 0.006 (0.002)⁄⁄⁄ 0.012 0.008 (0.004) NS 0.0072 (0.00)⁄⁄ Right precentral gyrus 52 8 50 70 560 0.005 (0.001)⁄⁄⁄ 0.013 0.007 (0.002)⁄⁄ 0.0082 (0.00)⁄⁄⁄ Right middle frontal gyrus 32 22 44 40 320 0.002 (0.001)⁄⁄ 0.006 0.001 (0.001) NS 0.0035 (0.00)⁄⁄ Left superior parietal lobule –34 –70 52 70 560 0.003 (0.001)⁄⁄⁄⁄⁄⁄ 0.009 Negative a, negative b 0.002 (0.002) NS 0.0046 (0.00)⁄⁄⁄ Left paracentral lobule (BA4a) 0 –24 66 598 4784 0.008 (0.002)⁄⁄⁄ 0.016 Negative a, negative b 0.009 (0.005)⁄ 0.0106 (0.00)⁄⁄⁄ Left precuneus –2 –56 60 141 1128 0.003 (0.001)⁄⁄⁄ 0.013 0.005 (0.002)⁄ 0.0050 (0.00)⁄⁄⁄ Right superior parietal lobule 38 –58 64 74 592 0.002 (0.001)⁄⁄⁄ 0.006 0.004 (0.002)⁄ 0.0047 (0.00)⁄⁄⁄ Left precuneus –14 –72 62 9 72 0.002 (0.000)⁄⁄⁄ 0.006 0.001 (0.001) NS 0.0029 (0.00)⁄⁄ Right SMA (BA6) 14 8 68 297 2376 0.006 (0.002)⁄⁄⁄ 0.038 Positive a, positive b 0.004 (0.002)⁄ 0.0081 (0.00)⁄⁄⁄ Right postcentral gyrus (BA1) 24 –42 68 70 560 0.005 (0.001)⁄⁄⁄ 0.003 Positive a, positive b 0.010 (0.003)⁄⁄⁄ 0.0072 (0.00)⁄⁄⁄ Left precuneus –10 –66 66 6 48 0.002 (0.001)⁄⁄⁄ 0.008 0.003 (0.002) NS 0.0030 (0.00)⁄⁄ Right precentral gyrus 26 –14 70 68 544 0.005 (0.001)⁄⁄⁄ 0.013 Positive a, positive b 0.002 (0.002) NS 0.0065 (0.00)⁄⁄⁄ Negative mediators Left rectal gyrus 65 –2 18 –24 15 –0.002 (0.001)⁄⁄⁄ 0.000 –0.003 (0.002)⁄ –0.0028 (0.00)⁄⁄ 1640 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 Table 3 (continued) Brain region (Brodmann area) Coordinates (mm) No. of voxels Vol (mm) Max a � b (SE) Cov (a, b) Path a and path b effects, if consistent across individuals Nonpara metric temperature effects model: Mean path ab coeff (SE) Mediation effect controlling for faces: Mean path ab coeff (SE) x y z P < .001 Left midbrain, hippocampus 66 –16 –16 –14 10 –0.001 (0.000)⁄⁄⁄ –0.006 –0.003 (0.001)⁄⁄⁄ –0.003 (0.00)⁄⁄⁄ Left superior frontal gyrus 68 –22 32 30 21 –0.002 (0.001)⁄⁄⁄ –0.001 –0.004 (0.001)⁄⁄ –0.004 (0.00)⁄⁄⁄ Right middle frontal gyrus 69 26 48 36 22 –0.009 (0.001)⁄⁄⁄ –0.008 –0.004 (0.002)⁄⁄ –0.005 (0.00)⁄⁄⁄ Coeff, coefficient; SE, standard error. This table presents results from path a⁄b of the multilevel mediation analysis presented as a path diagram in Fig. 1. Mediation effect parametric mapping results are presented in Fig. 4. ⁄⁄⁄P < .001; ⁄⁄P < .01; ⁄P < .05. Fig. 3. Variability in temperature effects on pain. The relationship between temperature and pain is reliable. However, there is also substantial variability in this relationship both within and between individuals, as demonstrated by plots of each participant’s pain ratings as a function of temperature. Mediators help to explain the within-subject variability. (A) Variability across subjects. Each participant is represented in a different color. (B) Variability within subjects. Green = nonpainful warmth (calibrated level 1); yellow = low pain (calibrated level 3); orange = medium pain (calibrated level 5); red = high pain (calibrated level 7). L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1641 Finally, we tested whether the masked emotion faces influenced pain reports. There was no main effect of Masked Emotion (Fear vs Happy) on pain (P > .3), however there was a significant Tempera- ture �Masked Emotion interaction (t25 = 3.79, P < .001), such that there was a steeper slope in temperature effects on pain when stimuli were preceded by subliminal fear faces, relative to sublim- inal happy faces. Because this comparison lacks a neutral control condition and requires replication before we interpret it strongly, we cannot make claims about these effects without replication and extension in additional experiments. Thus, we do not focus on these effects here, but we evaluate whether results reported below remain significant controlling for priming effects due to faces. 3.2. Multilevel mediation analysis The mediation analysis (Fig. 1b) included 3 tests of the links between the objective stimulus and the pain response: (1) path a, stimulus-related responses; (2) path b, pain-related responses, controlling for temperature; and (3) path a � b, mediators of the relationship between stimulus and response, that is, a formal test of whether the brain region explains a significant amount of the covariance between temperature and pain. 3.2.1. Path a, stimulus-related activity Increasing temperature produced increases in activation throughout the PPBN, including bilateral insula (with contiguous activation covering anterior, middle, and dorsal posterior insula), bilateral secondary somatosensory cortex (SII), right primary somatosensory cortex (SI), bilateral medial thalamus, dorsal anterior cingulate cortex (dACC; with coverage extending to sup- plementary motor area (SMA) and middle cingulate regions), in the midbrain surrounding the periaqueductal gray (PAG), and bilateral and medial cerebellum (Fig. 4A and Table 1). Tempera- ture-related increases were also found in regions outside of the PPBN, including bilateral lateral orbitofrontal cortex (latOFC), right Fig. 4. Mediation analysis results. (A) Path a identifies brain regions that show linear effects of temperature. (B) Path b identifies brain regions that correlate with pain reports, controlling for temperature. All identified regions were significantly related to trial-by-trial variations in pain using both linear and nonlinear models of temperature-related changes. 1642 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 lateral prefrontal cortex (latPFC), bilateral dorsolateral prefrontal cortex (DLPFC), and right caudate. Temperature was inversely related to blood oxygen level– dependent (BOLD) responses in the right dorsomedial prefrontal cortex (DMPFC; superior frontal gyrus), ventromedial prefrontal cortex (VMPFC), bilateral parahippocampal gyrus, bilateral hippo- campus, bilateral primary motor cortex (M1), bilateral fusiform gyrus, precuneus, cuneus, subgenual and posterior cingulate cor- tex, bilateral middle temporal cortex, and bilateral occipital cortex (Table 1). As we describe below, some of these regions were also related to pain report, and others were not. Path a effects on all regions remained significant when Masked Emotion type and the Emotion � Temperature interaction were included as covariates. 3.2.2. Path b, pain-related responses fMRI activity in a number of regions in the a priori PPBN mask predicted pain reports, controlling for temperature (Fig. 4B and Table 2), including bilateral anterior and middle insula, bilateral SII, right S1, dACC, SMA, right globus pallidus, and bilateral cerebel- lum. Other regions outside of the PPBN also positively predicted pain, including bilateral DLPFC (middle frontal gyrus), bilateral latOFC (middle orbital gyrus), and bilateral superior parietal lobule (SPL). BOLD responses were inversely related to pain reports in another set of other areas, including bilateral parahippocampal cortex, right hippocampus, left VMPFC, MPFC, left/lateral OFC (inferior frontal gyrus pars orbitalis), rostral anterior cingulate cor- tex (rACC), pons, cuneus, bilateral temporal pole, right middle and inferior temporal gyrus, posterior cingulate, and bilateral occipital cortex (Table 2). Every region that showed significant path b effects remained significant when we used a nonparametric model of temperature, suggesting that these effects were not driven by nonlinearities with respect to temperature (Table 2). We also evaluated effects control- ling for the face primes by modeling Masked Emotion and the Emotion � Temperature interaction as covariates (Table 2). All posi- tive path b effects remained significant when we controlled for the effects of face primes. Although most negative path b regions remained significant, path b effects on bilateral temporal pole, rACC, and 1 cluster in left VLPFC were no longer significant (Table 2). Thus, trial-by-trial variations in activity in these regions predicted trial-by-trial variations in report, even when tempera- ture was held constant. As we describe below, some of these regions covaried with stimulus temperature and others did not; in addition, some predicted pain report even when controlling for activity in PPBN mediators. 3.2.3. Path a � b, mediation effect A number of regions positively mediated the within-subject relationship between stimulus intensity and reported pain (Fig. 5A and Table 3). Mediation effects in some regions were dri- ven by consistent effects—that is, significant stimulus-related (path a) and pain-related effects (path b) across all participants—whereas others were driven by covariance in paths (ie, correlated individual differences in path a and path b effects). In either case, positive mediation implies that activity in this region explains some of the effects of noxious heat on reported pain. Mediator regions that showed consistent effects are illustrated in Fig. 5B. In most mediators that fell within the a priori PPBN mask, activation both increased with temperature and predicted pain (ie, positive path a and b; red in Fig. 5B, Table 3). We observed this pattern of effects in bilateral anterior/mid-insula, bilateral SII, dACC, left cerebellum, right lateral PFC, right SMA (BA6), right superior parietal lobule (SPL), and right premotor cortex. In other mediator regions (blue in Fig. 5B, Table 3), temperature increases produced deactivation, and greater deactivation predicted greater pain. Because both path a and path b are negative, the overall mediation effect is positive. Such effects were found in included bilateral hippocampus and parahippocampus, bilateral inferior temporal gyrus, left medial prefrontal cortex (MPFC), bilateral occipital cortex, right posterior cingulate cortex, left SPL, and precuneus. Interestingly, most mediators were driven by covariance between paths (Fig. 5A and Table 3), including the right parahippo- campus, left latOFC, bilateral latPFC, left caudate, left temporal pole, right fusiform, left VLPFC (inferior frontal gyrus), right middle temporal gyrus, bilateral retrosplenial cortex, posterior cingulate, cuneus, left premotor/lateral PFC, left DLPFC, right DMPFC, right SPL, and precuneus. Significant covariance implies systematic het- erogeneity across individuals; that is, some participants show Fig. 5. Brain mediators of the relationship between stimulus and response. (A) Mediators of the relationship between temperature and pain. Brain regions whose responses to noxious stimulation formally mediate the relationship between temperature and pain include those that were activated consistently across the group, as well as those that show covariance between paths. (B) Consistent mediators (significant path a and path b effects across subjects). Regions that were activated consistently across the group include both regions that showed temperature- and pain-related increases (red), as well as those that showed temperature- and pain-related decreases (blue). L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1643 strong path a and path b responses, whereas others are weak on both. Alternatively, there could be interindividual differences in the qualitative nature of the relationship between fMRI activity and pain: Some subjects show positive relationships with both temperature and pain, whereas others show negative relationships with both, in the same region. This is possible if, for example, brain activity represents appraisal or attention processes that serve to increase pain in some individuals and mitigate it in others. Thus, these results raise the possibility that different brain models may be appropriate for different groups of individuals. Finally, we observed negative mediation effects (blue in Fig. 5A; Table 3) in the subgenual anterior cingulate/medial orbitofrontal cortex (mOFC), left midbrain, and bilateral DMPFC (superior frontal gyrus). These findings suggest that some regions contribute to the stimulus–response relationship but show responses that go against the overall behavioral effect. These patterns have been described as suppressive effects in the mediation literature [51], and the inter- pretation is that the temperature–pain relationship would be stronger if activity in these regions were held constant. Regions in which activation is triggered by increases in temperature, but that serve an endogenous regulatory function, would fit this pat- tern; however, we do not interpret these results strongly here, as they were not predicted a priori. Finally, mediation effects in all regions remained significant when we controlled for cue effects by including Masked Emotion and Emotion � Temperature as covariates in the mediation analysis (Table 3). 3.3. Relationship to stimulus versus response Among the regions with significant temperature-related effects, some did not show any substantial evidence for correlations with pain report (‘‘a not b regions’’). Regions positively and uniquely related to temperature (yellow in Fig. 6) included the right dorsal posterior insula, bilateral thalamus, midbrain surrounding the PAG, right lateral prefrontal cortex, rostral ACC, bilateral lateral OFC, bilateral inferior frontal gyrus, caudal anterior cingulate cor- tex, and bilateral cerebellum. Regions negatively and uniquely related to temperature (green in Fig. 6) included bilateral medial prefrontal cortex, right dorsomedial prefrontal cortex (DMPFC), bilateral primary motor cortex, left S1, precuneus, left hippocam- pus, bilateral posterior hippocampus/fusiform gyrus, and bilateral occipital lobe. Activity in these regions is likely to reflect early pro- cessing of the noxious input. Regions that predict variations in report but are not affected by changes in noxious input (‘‘b not a regions’’, red/purple in Fig. 6) included the right mOFC, bilateral DLPFC (BA46), right DMPFC (superior frontal sulcus), bilateral SMA, right pre-SMA, left poster- ior insula, left posterior parahippocampus, bilateral superior parie- tal lobule, and bilateral cerebellum. To ensure specificity to pain reports, we also tested whether these ‘‘b not a’’ regions predicted pain report controlling for evoked responses within PPBN mediator regions. All regions remained significant controlling for PPBN mediator activation (Table 2), and remained significant controlling for nonlinear effects of temperature. These regions are likely to represent magnitude estimation, attention, decision making, and other endogenous processes that relate to variation in report but are unaffected by changes in temperature. Interestingly, we observed qualitative evidence for anterior- to-posterior gradients in several regions, pictured in Fig. 6A. In the left insula and ACC, anterior portions were temperature- related, middle regions showed both temperature and pain effects, and posterior regions were pain-related. We found the opposite Fig. 6. Relationship to stimulus versus response. Regions illustrated in yellow and green were affected by temperature (P < .001) but were unrelated to pain reports, controlling for temperature (P > .05) and showed no evidence of mediation (P > .05). Red and purple regions were associated with trial-by-trial pain reports, controlling for temperature (P < .001), but showed neither temperature-related changes nor mediation (P > .05), and each was significantly related to variations in pain controlling for nonlinear effects of temperature and for activation in the pain-processing brain network (PPBN) mediators. Mediators that showed consistent effects within subjects, with significant temperature- and pain-related effects, are depicted in orange (all paths: P < .001). (A) We observed anterior-to-posterior gradients in several regions. In the anterior cingulate cortex (ACC; left inset) and left insula (right inset), anterior portions were stimulus related; middle regions showed activation that was correlated with both increases in stimulus intensity and increases in pain; and posterior regions were related to pain reports. The opposite pattern was observed in the left parahippocampus (middle inset): More posterior portions were related to the stimulus, whereas anterior portions were related to pain reports. (B) Regions related to temperature but not pain included bilateral medial prefrontal cortex, right dorsomedial prefrontal cortex (DMPFC), bilateral primary motor cortex, left S1, precuneus, left hippocampus, bilateral posterior hippocampus/fusiform gyrus, bilateral occipital lobe, right ventromedial prefrontal cortex (VMPFC), bilateral dorsolateral prefrontal cortex (DLPFC; BA46), right supplementary motor area (SMA), right pre-SMA, left posterior insula, left posterior parahippocampus, and bilateral cerebellum. (C) Regions related to pain but not temperature included right medial orbitofrontal cortex (mOFC), bilateral DLPFC (BA46), right DMPFC, right SMA, right pre-SMA, left posterior insula, left posterior parahippocampus, and bilateral cerebellum. (D) We extracted responses from the regions that showed unique relationships to temperature or pain and averaged across regions and subjects. Loess curves depict responses as a function of applied temperature (top row), and pain as a function of brain response to heat (area-under-the-curve [AUC] estimate), controlling for temperature (bottom row). Shaded regions reflect standard error across subjects and therefore should not be used to make inferences on random effects. 1644 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 gradient in left parahippocampus: more posterior portions were temperature-related, whereas anterior parahippocampus was pain-related. Such gradients should be subjected to quantative spatial tests in future studies. 3.4. Mediation by network averages Finally, we used nonmetric multidimensional scaling (NMDS) to identify networks of mediators expressed similarly across trials and across individuals. We averaged across clusters within each network and tested for mediation to understand the network’s functional significance. Three networks showed consistent media- tion effects with positive paths a and b (Fig. 7, top panel). The first network included bilateral SII; the second network was made up of bilateral anterior and middle insula, dorsal anterior cingulate, right latPFC, right premotor cortex, and right VLPFC; and the third net- work included bilateral thalamus, left ventral striatum, midbrain/ PAG, and bilateral cerebellum. Two other networks showed posi- tive mediation effects with negative paths a and b. One network included bilateral parahippocampal cortex, right hippocampus, bilateral middle temporal gyrus/temporoparietal junction (TPJ), bilateral temporal pole, and right posterior hippocampus (Fig. 7, bottom), and the other network included bilateral retrosplenial cortex/posterior cingulate and inferior parietal lobule (IPL). Networks illustrated in Fig. 7 showed evidence of mediation even controlling for all other mediators, suggesting that they are the most proximal predictors of temperature effects on pain. Finally, seven more networks were identified with inconsistent positive mediation effects (significant a⁄b, but nonsignificant path a and/ or path b effects), likely driven by covariance (individual differ- ences) and/or heterogeneity of the behavior of the constituent regions. All effects are reported in Table 4. 4. Discussion The present study identified brain networks that contribute to the generation of pain in response to noxious stimulation. This pro- vides the first direct evidence that regions within an a priori ‘‘pain Fig. 7. Classes of mediators. We performed nonmetric multi-dimensional scaling to identify networks of mediators, based on functional coactivation within-subjects across trials. Eleven functional networks were identified (Table 4). (Top) Three networks consistently showed temperature- and pain-related increases, including a network in bilateral secondary somatosensory cortex (orange), a second network that consisted of bilateral anterior and middle insula, dorsal anterior cingulate, right lateral prefrontal cortex (latPFC), right premotor cortex, and right ventrolateral prefrontal cortex (red), and a third network that included bilateral thalamus, left ventral striatum, midbrain/ periaqueductal gray (PAG), and bilateral cerebellum (yellow). Plots of average activation reveal that networks 2 and 3 were largely linear in relationship to noxious input, whereas network 1 showed nonlinearities that mirror the behavioral effects of temperature on pain. (Bottom) Two networks were inversely related to both noxious stimulus and pain reports: The first included bilateral retrosplenial cortex/posterior cingulate and inferior parietal lobule (green), and the second included bilateral parahippocampal cortex, right hippocampus, bilateral middle temporal gyrus/temporoparietal junction (TPJ), bilateral temporal pole, and right posterior hippocampus (cyan). L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1645 processing’’ brain network (PPBN) formally mediate stimulus intensity effects on pain. Our results also reveal complexity that goes beyond a unitary pain-processing system. First, regions out- side the PPBN also mediate temperature effects on pain. Second, mediators were grouped into several networks that make separa- ble contributions to pain. Finally, some regions responded selec- tively to changes in temperature, whereas others predicted pain without responding to changes in heat. 4.1. Pain-processing brain network mediators Most mediators that showed consistent profiles across individ- uals fell within a priori ROIs, including bilateral SII, aIns, dACC, left cerebellum, IPL, and SPL. These regions showed temperature- related increases and predicted subtle variations in pain, control- ling for temperature. Previous work has linked activation in some of these regions to variations in reports without changes in stimu- lus intensity [7,14,62,82]. Our findings reveal that these regions drive fluctuations in pain even when stimulus intensity varies. These trial-by-trial fluctuations are not reducible to nonlinearities in temperature effects on pain [33], but they could reflect sensory variations due to contact with the thermode (which we attempted to minimize), stimulus history [12,37,64], or non-nociceptive pro- cesses that modify pain, such as fluctuations in attention [14,46,62] and regulatory or preparatory mechanisms that reduce pain [22,77]. Future studies should distinguish among these possibili- ties and perhaps quantify how much variance in the brain activ- ity–pain relationship is due to each factor. Network analyses suggest that PPBN mediators comprise distinct somatosensory, paralimbic cortical (anterior cingulate/ insula/lateral sensorimotor/frontal), and subcortical (striatal/ brainstem/cerebellar) groups. These findings are broadly consis- tent with separation of lateral sensory/discriminative versus med- ial affective/motivational systems [17], but provide an elaborated view. The cortical network that we identified includes lateral and medial areas, and the subcortical regions often included in the ‘‘medial pain system’’ form a separate, coherent network. Thus, the results extend the concept of the unitary ‘‘pain processing’’ system, and lay a foundation for (1) understanding how networks interact during pain genesis, and (2) examining how different fac- tors (stimulus types, psychological contexts, or treatments) affect each network. 4.2. Mediators not targeted by ascending nociceptive pathways Regions outside the a priori PPBN also mediated pain. We observed temperature- and pain-related increases as well as medi- ation in the right latPFC, right SMA, right SPL, and right premotor Table 4 Mediator networks and relationship to temperature effects on pain. Network components Path a Path b Path c Path c’ Mediation (c – c’) Consistent mediator networks: Positive path coefficients Network 1: Bilateral SII, M1 2.148⁄ 2.387⁄ 3.933⁄⁄⁄ 3.953⁄⁄⁄ 3.144⁄⁄ Network 2: Bilateral anterior insula, Bilateral middle insula, dACC, R DMPFC, bilateral precentral gyrus, R latPFC, R VLPFC, R SPL 3.747⁄⁄⁄ 3.635⁄⁄⁄⁄ 3.652⁄⁄⁄ 3.967⁄⁄⁄ 3.586⁄⁄⁄ Network 3: Right thalamus, PAG, midbrain, medial cerebellum, left ventral striatum, ventral striatum, right globus pallidus 3.570⁄⁄⁄ 2.213⁄ 4.040⁄⁄⁄ 3.988⁄⁄⁄ 3.078⁄⁄ Consistent mediator networks: Negative path coefficients Network 4: Bilateral retrosplenial cortex, R – – IPL 3.313⁄⁄⁄ 3.551⁄⁄⁄ 3.979⁄⁄⁄ 3.921⁄⁄⁄ 3.720⁄⁄⁄ Network 5: Bilateral hippocampus, R parahippocampus, bilateral middle temporal gyrus / TPJ, bilateral temporal operculum, L IFG –3.236⁄⁄⁄ –1.784\ 3.937⁄⁄⁄ 3.920⁄⁄⁄ 3.572⁄⁄⁄ Inconsistent mediators: a,b covariance-driven mediation Network 6: Bilateral VMPFC –0.8142 –0.196 3.900⁄⁄⁄ 3.957⁄⁄⁄ 2.826⁄⁄ Network 7: Bilateral VLPFC, L IFG 0.564 0.266 3.984⁄⁄⁄ 3.942⁄⁄⁄ 3.677⁄⁄⁄ Network 8: Bilateral cerebellum, bilateral latOFC, DMPFC, R DMPFC, L SPL 1.930� 0.960 3.859⁄⁄⁄ 3.917⁄⁄⁄ 3.084⁄⁄ Network 9: Bilateral temporal pole, L anterior PFC, L parahippocampal gyrus –2.472⁄ –1.587 3.910⁄⁄⁄ 3.935⁄⁄⁄ 3.143⁄⁄ Network 10: Bilateral occipital cortex, posterior parahippocampal cortex, precuneus, R M1 –3.814⁄⁄⁄ –0.708 3.923⁄⁄⁄ 3.943⁄⁄⁄ 2.457⁄ Network 11: L pons, cuneus –0.706 –0.531 3.935⁄⁄⁄ 3.944⁄⁄⁄ 3.125⁄⁄ dACC, dorsal anterior cingulate cortex; DMPFC, dorsomedial prefrontal cortex; IFG, ; IPL, inferior parietal lobule; L, left; latOFC, lateral orbitofrontal cortex; lat PFC, lateral prefrontal cortex; PAG, periaqueductal gray; PFC, prefrontal cortex; R, right; SPL, superior parietal lobule; TPJ, temporoparietal junction; VLPFC. Results of mediation analysis testing whether average network activation mediates temperature effects on pain reports. Path a: temperature effects. Path b: prediction of pain reports, controlling for temperature. Values reflect z scores for each effect. ⁄⁄⁄P < .001; ⁄⁄P < .01; ⁄P < .05; �P < .10. 1646 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 cortex. Although it is not a direct target of the spino-thalamo-cor- tical pathway, right latPFC (near Brodmann Area 10) has been implicated in previous pain studies [10,49,86]. SMA and premotor activation is also commonly observed in imaging studies of neuro- pathic pain and acute pain in healthy individuals and might be involved in pain-related motor planning [5,57,59]. Two core networks, including the hippocampal and parahippo- campal cortex, anterior pole, TPJ, and the precuneus, mediated stimulus effects on pain by virtue of deactivation with increasing heat, the magnitude of which predicted increased pain. These regions are all associated with the ‘‘default mode network’’ (DMN). They are active at rest and may represent stimulus- independent thought and internally focused mnemonic, viscero- motor, and valuation processes [4,21]. Previous studies have also shown intensity- and pain-related DMN deactivation [48,69], link- ing pain with other processes that capture attention and limit internal thought. Although these temperature- and pain-related decreases may reflect pain-related attention capture, some of these regions might play a more active role in pain inhibition and generation. Existing evidence suggests that the hippocampus and parahippocampal cortices can both enhance or decrease pain. Both receive input from SII via the insula [36]. Noxious stimuli reduce activation in the hippocampus [43,59] and the parahippocampal gyrus [30,43]. Hippocampal stimulation in animals produces antinociception [50,53], and placebo analgesia has been associated with parahippo- campal increases [26,42]. However, noxious stimuli can sometimes increase hippocampal activation [11,30], and hippocampal activity has been associated with anxiety-induced hyperalgesia [61] and nocebo-related increases in pain [41]. Disentangling the factors that lead to anti-nociceptive or pro-nociceptive responses is an important direction for future research. 4.3. Regions preferential to stimulus versus pain report Tests of mediation in single voxels provide an important, but partial, view of systems encoding pain. Areas that represent distinct aspects of the experience—that is, selective relationships with temperature or pain—can provide clues about the dissociabil- ity of nociceptive and decision-making processes and potential neurophysiological targets for the modulation of each. For exam- ple, placebo analgesia is associated with widespread changes in responses to noxious stimuli [78]. It is unclear whether any brain circuits reliably encode thermosensory nociceptive responses without being affected by the treatment context. This study iden- tified potential thermosensory circuits in subcortical (eg, PAG and thalamus), sensory cortex, and fronto-cerebellar circuits. These could be related to early nociception, attentional orienting, noci- fensive responses, or other processes; whatever the case, they are more strongly related to stimulus intensity than to pain report. Conversely, several regions were related to pain but showed no relationship with stimulus intensity. These regions may support the cognitive and affective processes that contribute to pain evalu- ation, such as magnitude estimation, conceptual representation, and fluctuations in arousal and attention. The medial and lateral OFC have been associated with expected value and outcome expec- tancy [45,68]. The DLPFC is broadly associated with decision mak- ing and executive function [54,55], and the right DMPFC has been associated with prospection and mental simulation [1,3]. These systems may be responsible when contextual factors influence pain without affecting nociception (ie, as a function of decision- making processes). Previous studies isolated brain regions whose activation during rating predicts variation in pain [9,44]. An extension of the present work might use mediation to separate regions whose activation during the report period is related to heat, and those whose pain- related responses are related to endogenous variations in report- ing-related activity alone. These would be informative probes of activity related to decision-making biases or memory-related pro- cesses. Our study was not designed to address reporting-period activity, as motor-related processes would correlate with heat- related processes due to the consistent mapping of increased pain to rightward movement on our scale. The pattern of temperature-related versus pain-related regions provides new evidence on the question of whether these processes are encoded in different circuits. One pioneering fMRI study associated thermosensation with dpIns and pain perception with aINS [29]. We also found that contralateral dpIns showed linear increases with temperature but did not predict pain; however, the picture is likely to be more complex, as other portions of dpINS L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 1647 predicted pain but did not respond to temperature, replicating recent results [9], and anterior portions of aINS/operculum tracked temperature without predicting pain. In addition, our data appeared to show posterior–anterior gradients in the insula and dACC from pain report to stimulus encoding. Previous studies have found similar gradients in the parietal cortex [6]. These should be replicated and explicitly tested with spatial statistics in future studies. 4.4. Future directions There are several important outstanding questions, including the need for more information on (1) the specificity of brain responses to pain, (2) connections between experimental thermal pain and other types of pain (eg, inflammatory and neuropathic pain), (3) further characterization of complex responses to stimu- lus intensity and pain (eg, nonlinearities and temporal dissocia- tions [6,15,63], and (4) sources of individuals’ variability in path effects (ie, understanding effects on mediators driven by covari- ance), such as gender differences and relationship to psychological variables (eg, fear of pain, anxiety). Although assessment of these questions is beyond the scope of the current study, the specificity issue deserves discussion. We observed heterogeneity in PPBN responses, suggesting that this network contains both thermosensory regions and regions involved in both thermosensation and magnitude estimation. However, we examined only one modality in this study, and cannot determine whether stimulus intensity effects and relationships with subjective reports would be similar in other modalities, or whether the relationships that we observed are unique to nocicep- tion and pain. AIns and dACC are activated by a range of tasks [9,31,85]. One view is that aIns mediates general interoception [27,28], and dACC and aIns integrate salient external stimuli of all types with internal states [23]. A similar argument has been made for somatosensory regions [58]. Another view is that these regions contain many types of neurons that encode diverse, but specific, processes, including nociception and pain, and indeed, responses within these regions predict pain across individuals with high accuracy [78]. Future research should use mediation to link stimulus intensity with perception across domains and to deter- mine the specificity of pathways identified here. 4.5. Conclusion Pain is generated by a combination of independent networks. Some are involved in nociception and pain in a classic sense: They show graded responses to noxious input, and responses predict the magnitude of pain. Others show decreases with temperature that predict pain increases and may be related to mental processes that antagonize (or sometimes promote) pain. Finally, some regions predict pain without responding to noxious input per se, and may encode decision-making processes that shape pain evaluation. The networks identified here provide a rich neurophysiological basis for understanding diverse pain-generation processes, and this approach can be extended to stimulus effects on conscious percep- tion in other domains. Conflict of interest statement The authors declare no conflicts of interest. Acknowledgements The authors acknowledge Katherine Dahl and Natalie Johnston for assistance with data collection. Funding was provided by National Institutes of Health grants RO1 MH076136 and DA027794. References [1] Abraham A, Schubotz RI, von Cramon DY. Thinking about the future versus the past in personal and non-personal contexts. Brain Res 2008;1233:106–19. [2] Adair ER, Stevens JC, Marks LE. Thermally induced pain, the Dol scale, and the psychophysical power law. Am J Psychol 1968;81:147–64. [3] Addis DR, Pan L, Vu M-A, Laiser N, Schacter DL. Constructive episodic simulation of the future and the past: distinct subsystems of a core brain network mediate imagining and remembering. Neuropsychologia 2009;47:2222–38. [4] Andrews-Hanna JR. The brain’s default network and its adaptive role in internal mentation. Neuroscientist 2011;18:1–21. [5] Apkarian AV, Bushnell MC, Treede RD, Zubieta JK. Human brain mechanisms of pain perception and regulation in health and disease. Eur J Pain 2005;9:463–84. [6] Apkarian AV, Darbar A, Krauss BR, Gelnar PA, Szeverenyi NM. Differentiating cortical areas related to pain perception from stimulus identification: temporal analysis of fMRI activity. J Neurophysiol 1999;81:2956–63. [7] Atlas LY, Bolger N, Lindquist MA, Wager TD. Brain mediators of predictive cue effects on perceived pain. J Neurosci 2010;30:12964–77. [8] Atlas LY, Whittington RA, Lindquist MA, Wielgosz J, Sonty N, Wager TD. Dissociable influences of opiates and expectations on pain. J Neurosci 2012;32:8053–64. [9] Baliki MN, Geha PY, Apkarian AV. Parsing pain perception between nociceptive representation and magnitude estimation. J Neurophysiol 2009;101:875–87. [10] Becerra L, Breiter HC, Stojanovic M, Fishman S, Edwards A, Comite AR, Gonzalez RG, Borsook D. Human brain activation under controlled thermal stimulation and habituation to noxious heat: an fMRI study. Magnet Reson Med 1999;41:1044–57. [11] Bingel U, Quante M, Knab R, Bromm B, Weiller C, Buchel C. Subcortical structures involved in pain processing: Evidence from single-trial fMRI. PAIN� 2002;99:313–21. [12] Bingel U, Schoell ED, Herken W, Buchel C, May A. Habituation to painful stimulation involves the antinociceptive system. PAIN� 2007;131:21–30. [13] Bollen K, Stine R. Direct and indirect effects: classical and bootstrap estimates of variability. Sociol Methodol 1990:115–40. [14] Boly M, Balteau E, Schnakers C, Degueldre C, Moonen G, Luxen A, Phillips C, Peigneux P, Maquet P, Laureys S. Baseline brain activity fluctuations predict somatosensory perception in humans. Proc Natl Acad Sci U S A 2007;104:12187–92. [15] Bornhovd K, Quante M, Glauche V, Bromm B, Weiller C, Buchel C. Painful stimuli evoke different stimulus-response functions in the amygdala, prefrontal, insula and somatosensory cortex: a single-trial fMRI study. Brain 2002;125:1326–36. [16] Brodersen KH, Wiech K, Lomakina EI, Lin C-s, Buhmann JM, Bingel U, Ploner M, Stephan KE, Tracey I. Decoding the perception of pain from fMRI using multivariate pattern analysis. Neuroimage 2012;63:1162–70. [17] Brooks J, Tracey I. From nociception to pain perception: imaging the spinal and supraspinal pathways. J Anat 2005;207:19–33. [18] Brown JE, Chatterjee N, Younger J, Mackey S. Towards a physiology-based measure of pain: patterns of human brain activity distinguish painful from non-painful thermal stimulation. PLoS One 2011;6:e24124. [19] Buchel C, Bornhovd K, Quante M, Glauche V, Bromm B, Weiller C. Dissociable neural responses related to pain intensity, stimulus intensity, and stimulus awareness within the anterior cingulate cortex: a parametric single-trial laser functional magnetic resonance imaging study. J Neurosci 2002;22:970–6. [20] Büchel C, Bornhovd K, Quante M, Glauche V, Bromm B, Weiller C. Dissociable neural responses related to pain intensity, stimulus intensity, and stimulus awareness within the anterior cingulate cortex: a parametric single-trial laser functional magnetic resonance imaging study. J Neurosci 2002;22:970–6. [21] Buckner RL, Andrews-Hanna JR, Schacter DL. The brain’s default network: anatomy, function, and relevance to disease. Ann N Y Acad Sci 2008;1124:1–38. [22] Bushnell MC, Čeko M, Low LA. Cognitive and emotional control of pain and its disruption in chronic pain. Nat Rev Neurosci 2013;14:502–11. [23] Cauda F, Torta DM-E, Sacco K, Geda E, D’agata F, Costa T, Duca S, Geminiani G, Amanzio M. Shared, ‘‘core’’ areas between the pain and other task-related networks. PLoS One 2012;7:e41929. [24] Coghill RC, Eisenach J. Individual differences in pain sensitivity: implications for treatment decisions. Anesthesiology 2003;98:1312–4. [25] Coghill RC, Sang CN, Maisog JM, Iadarola MJ. Pain intensity processing within the human brain: a bilateral, distributed mechanism. J Neurophysiol 1999;82:1934–43. [26] Craggs J, Price D, Perlstein W, Nicholas Verne G, Robinson M. The dynamic mechanisms of placebo induced analgesia: evidence of sustained and transient regional involvement. PAIN� 2008;139:660–9. [27] Craig AD. How do you feel? Interoception: the sense of the physiological condition of the body. Nat Rev Neurosci 2002;3:655–66. [28] Craig AD. How do you feel—now? The anterior insula and human awareness. Nat Rev Neurosci 2009;10:12. [29] Craig AD, Chen K, Bandy D, Reiman EM. Thermosensory activation of insular cortex. Nat Neurosci 2000;3:184–90. http://refhub.elsevier.com/S0304-3959(14)00248-6/h0005 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0005 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0010 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0010 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0015 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0015 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0015 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0015 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0020 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0020 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0025 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0025 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0025 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0030 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0030 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0030 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0035 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0035 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0040 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0040 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0040 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0045 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0045 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0050 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0050 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0050 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0050 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0055 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0055 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0055 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0060 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0060 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0060 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0065 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0065 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0070 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0070 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0070 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0070 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0075 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0075 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0075 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0075 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0080 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0080 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0080 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0085 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0085 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0090 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0090 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0090 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0095 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0095 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0095 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0095 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0100 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0100 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0100 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0100 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0105 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0105 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0105 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0110 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0110 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0115 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0115 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0115 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0120 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0120 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0125 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0125 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0125 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0130 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0130 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0130 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0130 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0135 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0135 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0140 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0140 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0145 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0145 1648 L.Y. Atlas et al. / PAIN � 155 (2014) 1632–1648 [30] Derbyshire SW, Jones AK, Gyulai F, Clark S, Townsend D, Firestone LL. Pain processing during three levels of noxious stimulation produces differential patterns of central activity. PAIN� 1997;73:431–45. [31] Dosenbach NUF, Visscher KM, Palmer ED, Miezin FM, Wenger KK, Kang HC, Burgund ED, Grimes AL, Schlaggar BL, Petersen SE. A core system for the implementation of task sets. Neuron 2006;50:799–812. [32] Dum RP, Levinthal DJ, Strick PL. The spinothalamic system targets motor and sensory areas in the cerebral cortex of monkeys. J Neurosci 2009;29:14223–35. [33] Efron B, Tibshirani RJ. An introduction to the bootstrap. New York: Chapman & Hall; 1993. [34] Ekman P, Oster H. Facial expressions of emotion. Annu Rev Psychol 1979;30:527–54. [35] Friebel U, Eickhoff S, Lotze M. Coordinate-based meta-analysis of experimentally induced and chronic persistent neuropathic pain. Neuroimage 2011;58:1070–80. [36] Friedman DP, Murray EA, O’Neill JB, Mishkin M. Cortical connections of the somatosensory fields of the lateral sulcus of macaques: evidence for a corticolimbic pathway for touch. J Comp Neurol 1986;252:323–47. [37] Jepma M, Jones M, Wager TD. The dynamics of pain: Evidence for simultaneous site-specific habituation and site-nonspecific sensitization in thermal pain. J Pain, in press. http://dx.doi.org/10.1016/j.jpain.2014.02.010. [38] Johnstone T, Salomons TV, Backonja MM, Davidson RJ. Turning on the alarm: the neural mechanisms of the transition from innocuous to painful sensation. Neuroimage 2012;59:1594–601. [39] Kenny D, Korchmaros J, Bolger N. Lower level mediation in multilevel models. Psychol Methods 2003;8:115–28. [40] Kober H, Barrett L, Joseph J, Bliss-Moreau E, Lindquist K, Wager T. Functional grouping and cortical–subcortical interactions in emotion: a meta-analysis of neuroimaging studies. Neuroimage 2008;42:998–1031. [41] Kong J, Gollub RL, Polich G, Kirsch I, Laviolette P, Vangel M, Rosen B, Kaptchuk TJ. A functional magnetic resonance imaging study on the neural mechanisms of hyperalgesic nocebo effect. J Neurosci 2008;28:13354–62. [42] Kong J, Gollub RL, Rosman IS, Webb JM, Vangel MG, Kirsch I, Kaptchuk TJ. Brain activity associated with expectancy-enhanced placebo analgesia as measured by functional magnetic resonance imaging. J Neurosci 2006;26:381–8. [43] Kong J, Loggia ML, Zyloney C, Tu P, Laviolette P, Gollub RL. Exploring the brain in pain: activations, deactivations and their relation. PAIN� 2010;148:257–67. [44] Kong J, White NS, Kwong KK, Vangel MG, Rosman IS, Gracely RH, Gollub RL. Using fMRI to dissociate sensory encoding from cognitive evaluation of heat pain intensity. Hum Brain Mapp 2006;27:715–21. [45] Kringelbach ML. The human orbitofrontal cortex: linking reward to hedonic experience. Nat Rev Neurosci 2005;6:691–702. [46] Kucyi A, Salomons TV, Davis KD. Mind wandering away from pain dynamically engages antinociceptive and default mode brain networks. Proc Natl Acad Sci 2013;110:18692–7. [47] Lindquist MA, Meng Loh J, Atlas LY, Wager TD. Modeling the hemodynamic response function in fMRI: efficiency, bias and mis-modeling. Neuroimage 2009;45(1 Suppl):S187–S198. [48] Loggia ML, Edwards RR, Kim J, Vangel MG, Wasan AD, Gollub RL, Harris RE, Park K, Napadow V. Disentangling linear and nonlinear brain responses to evoked deep tissue pain. PAIN� 2012;153:2140–51. [49] Lorenz J, Cross D, Minoshima S, Morrow T, Paulson P, Casey K. A unique representation of heat allodynia in the human brain. Neuron 2002;35:383–93. [50] Lovick TA. Central nervous system integration of pain control and autonomic function. News Physiol Sci 1991;6:82–6. [51] MacKinnon D, Krull J, Lockwood C. Equivalence of the mediation, confounding and suppression effect. Prev Sci 2000;1:173–81. [52] Marquand AF, Howard M, Brammer MJ, Chu C, Coen S, Mourão-Miranda J. Quantitative prediction of subjective pain intensity from whole-brain fMRI data using Gaussian processes. Neuroimage 2010;49:2178–89. [53] McKenna JE, Melzack R. Analgesia produced by lidocaine microinjection into the dentate gyrus. PAIN� 1992;49:105–12. [54] Miller EK. The prefrontal cortex and cognitive control. Nat Rev Neurosci 2000;1:59–65. [55] Miller EK, Cohen JD. An integrative theory of prefrontal cortex function. Annu Rev Neurosci 2001;24:167–202. [56] Mogil JS. The genetic mediation of individual differences in sensitivity to pain and its inhibition. Proc Natl Acad Sci U S A 1999;96:7744–51. [57] Moisset X, Bouhassira D. Brain imaging of neuropathic pain. Neuroimage 2007;37:S80–8. [58] Mouraux A, Diukova A, Lee MC, Wise RG, Iannetti GD. A multisensory investigation of the functional significance of the ‘‘pain matrix’’. NeuroImage 2011;54:2237–49. [59] Peyron R, García-Larrea L, Grégoire MC, Costes N, Convers P, Lavenne F, Mauguière F, Michel D, Laurent B. Haemodynamic brain responses to acute pain in humans: sensory and attentional networks. Brain 1999;122: 1765–80. [60] Peyron R, Laurent B, Garcia-Larrea L. Functional imaging of brain responses to pain. A review and meta-analysis. Neurophysiol Clinique (Clin Neurophysiol) 2000;30:263–88. [61] Ploghaus A, Narain C, Beckmann CF, Clare S, Bantick SJ, Wise RG, Matthews PM, Rawlins JN, Tracey I. Exacerbation of pain by anxiety is associated with activity in a hippocampal network. J Neurosci 2001;21:9896–903. [62] Ploner M, Lee MC, Wiech K, Bingel U, Tracey I. Prestimulus functional connectivity determines pain perception in humans. Proc Natl Acad Sci U S A 2009;107:355–60. [63] Pomares FB, Faillenot I, Barral FG, Peyron R. The ‘where’ and the ‘when’ of the BOLD response to pain in the insular cortex. Discussion on amplitudes and latencies. Neuroimage 2013;64:466–75. [64] Price DD, Hu JW, Dubner R, Gracely RH. Peripheral suppression of first pain and central summation of second pain evoked by noxious heat pulses. PAIN� 1977;3:57–68. [65] Price DD, McHaffie J, Stein BE. The psychophysical attributes of heat-induced pain and their relationships to neural mechanisms. J Cognit Neurosci 1992;4:1–14. [66] Rish I, Cecchi G, Baliki M, Apkarian A. Sparse regression models of pain perception. Brain Informat 2010:212–23. [67] Rissman J, Gazzaley A, D’esposito M. Measuring functional connectivity during distinct stages of a cognitive task. Neuroimage 2004;23:752–63. [68] Schoenbaum G, Roesch MR, Stalnaker TA, Takahashi YK. A new perspective on the role of the orbitofrontal cortex in adaptive behaviour. Nat Rev Neurosci 2009;10:885–92. [69] Seminowicz DA, Davis KD. Pain enhances functional connectivity of a brain network evoked by performance of a cognitive task. J Neurophysiol 2007;97:3651–9. [70] Shattuck DW, Mirza M, Adisetiyo V, Hojatkashani C, Salamon G, Narr KL, Poldrack RA, Bilder RM, Toga AW. Construction of a 3D probabilistic atlas of human cortical structures. Neuroimage 2008;39:1064–80. [71] Shrout P, Bolger N. Mediation in experimental and nonexperimental studies: New procedures and recommendations. Psychol Methods 2002;7:422–45. [72] Stevens SS. To honor Fechner and repeal his law. Science 1961;133:80–6. [73] Stone C, Sobel M. The robustness of estimates of total indirect effects in covariance structure models estimated by maximum. Psychometrika 1990;55:337–52. [74] Thomas CG, Harshman RA, Menon RS. Noise reduction in BOLD-based fMRI using component analysis. Neuroimage 2002;17:1521–37. [75] Tohka J, Foerde K, Aron AR, Tom SM, Toga AW, Poldrack RA. Automatic independent component labeling for artifact removal in fMRI. Neuroimage 2008;39:1227–45. [76] Tölle TR, Kaufmann T, Siessmeier T, Lautenbacher S, Berthele A, Munz F, Zieglgänsberger W, Willoch F, Schwaiger M, Conrad B, Bartenstein P. Region- specific encoding of sensory and affective components of pain in the human brain: a positron emission tomography correlation analysis. Ann Neurol 1999;45:40–7. [77] Tracey I, Mantyh PW. The cerebral signature for pain perception and its modulation. Neuron 2007;55:377–91. [78] Wager T, Atlas L, Lindquist M, Roy M, Woo C-W, Kross E. An fMRI-based neurologic signature for physical pain. N Engl J Med 2013;368:1388–97. [79] Wager TD, Scott DJ, Zubieta JK. Placebo effects on human l-opioid activity during pain. Proc Natl Acad Sci U S A 2007;104:11056–61. [80] Wager TD, van Ast VA, Hughes BL, Davidson ML, Lindquist MA, Ochsner KN. Brain mediators of cardiovascular responses to social threat, part II: prefrontal-subcortical pathways and relationship with anxiety. Neuroimage 2009;47:836–51. [81] Wager TD, Waugh CE, Lindquist M, Noll DC, Fredrickson BL, Taylor SF. Brain mediators of cardiovascular responses to social threat: Part I: reciprocal dorsal and ventral sub-regions of the medial prefrontal cortex and heart-rate reactivity: Supplementary material. Neuroimage 2009;47:821–35. [82] Wiech K, Lin C-s, Brodersen KH, Bingel U, Ploner M, Tracey I. Anterior insula integrates information about salience into perceptual decisions about pain. J Neurosci 2010;30:16324–31. [83] Willis WD, Westlund KN. Neuroanatomy of the pain system and of the pathways that modulate pain. J Clin Neurophysiol 1997;14:2–31. [84] Woodrow KM, Friedman GD, Siegelaub AB, Collen MF. Pain tolerance: differences according to age, sex and race. Psychosom Med 1972;34:548–56. [85] Yarkoni T, Poldrack RA, Nichols TE, Van Essen DC, Wager TD. Large-scale automated synthesis of human functional neuroimaging data. Nat Methods 2011;8:665–70. [86] Zambreanu L, Wise RG, Brooks JCW, Iannetti GD, Tracey I. A role for the brainstem in central sensitisation in humans. Evidence from functional magnetic resonance imaging. PAIN� 2005;114:397–407. http://refhub.elsevier.com/S0304-3959(14)00248-6/h0150 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0150 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0150 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0150 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0155 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0155 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0155 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0160 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0160 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0160 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0165 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0165 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0170 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0170 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0175 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0175 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0175 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0180 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0180 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0180 http://dx.doi.org/10.1016/j.jpain.2014.02.010 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0190 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0190 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0190 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0195 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0195 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0200 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0200 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0200 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0205 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0205 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0205 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0210 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0210 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0210 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0215 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0215 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0215 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0220 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0220 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0220 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0225 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0225 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0230 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0230 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0230 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0240 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0240 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0240 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0240 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0245 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0245 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0250 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0250 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0255 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0255 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0260 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0260 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0260 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0265 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0265 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0265 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0270 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0270 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0275 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0275 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0280 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0280 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0285 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0285 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0290 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0290 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0290 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0295 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0295 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0295 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0295 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0300 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0300 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0300 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0305 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0305 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0305 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0310 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0310 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0310 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0315 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0315 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0315 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0320 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0320 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0320 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0325 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0325 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0325 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0330 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0330 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0335 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0335 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0340 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0340 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0340 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0345 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0345 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0345 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0350 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0350 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0350 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0355 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0355 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0360 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0365 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0365 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0365 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0370 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0370 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0375 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0375 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0375 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0380 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0380 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0380 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0380 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0380 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0385 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0385 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0390 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0390 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0395 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0395 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0400 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0400 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0400 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0400 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0405 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0405 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0405 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0405 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0410 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0410 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0410 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0415 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0415 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0420 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0420 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0425 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0425 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0425 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0430 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0430 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0430 http://refhub.elsevier.com/S0304-3959(14)00248-6/h0430 Brain mediators of the effects of noxious heat on pain 1 Introduction 2 Methods 2.1 Participants and procedure 2.1.1 Participants 2.1.2 Thermal stimulation and pain ratings 2.1.3 fMRI task design 2.1.4 Behavioral analyses 2.2 fMRI acquisition and preprocessing 2.2.1 Imaging acquisition 2.2.2 Preprocessing 2.3 fMRI analysis 2.3.1 Single trial analysis 2.3.2 Mediation analysis 2.3.3 Thresholding 2.3.4 Relationship to stimulus versus report 2.3.5 Network analysis 2.3.6 Accounting for potential nonlinearities 2.3.7 Relationship to masked emotion face cues 3 Results 3.1 Behavioral analysis 3.1.1 Behavioral relationship between stimulus and report 3.2 Multilevel mediation analysis 3.2.1 Path a, stimulus-related activity 3.2.2 Path b, pain-related responses 3.2.3 Path a×b, mediation effect 3.3 Relationship to stimulus versus response 3.4 Mediation by network averages 4 Discussion 4.1 Pain-processing brain network mediators 4.2 Mediators not targeted by ascending nociceptive pathways 4.3 Regions preferential to stimulus versus pain report 4.4 Future directions 4.5 Conclusion Conflict of interest statement Acknowledgements References