1.1 Participants

The MRI sample consisted of 37 adolescents who varied in conduct problem severity and CU traits. Of the sample, 20 were assessed as exhibiting clinically significant CU traits; the remaining 17 adolescents did not exhibit significant levels of either CU traits or disruptive behavior. The sample included males and females aged 10 to 17 years old who were recruited from Washington, D. C. and surrounding areas through referrals, advertisements, fliers seeking both healthy children and children with conduct problems. All youths and a parent or guardian completed an initial screening visit to determine eligibility for the scanning portion of the study. During this visit, demographic and clinical assessments were conducted, including a test of cognitive intelligence (K-BIT, Kaufman & Kaufman, Reference Kaufman and Kaufman2004), and assessments of clinical symptomology using the Strengths and Difficulties Questionnaire (Goodman & Scott, Reference Goodman and Scott1999), Child Behavior Checklist (CBCL; Achenbach, Reference Achenbach1991), and the Inventory of Callous-Unemotional Traits (ICU; Frick & Ray, Reference Frick and Ray2015; Kimonis et al., Reference Kimonis, Frick, Skeem, Marsee, Cruise, Munoz and Morris2008) completed separately by parent and child. Participants were excluded based on the following criteria: full-scale IQ scores <80, history of head trauma, neurological disorder, parent report of pervasive developmental disorder, or magnetic resonance imaging (MRI) contraindications. In addition, no siblings were permitted to participate. Additional exclusion criteria for healthy controls included any history of mood, anxiety, or disruptive behavior disorders. Written informed assent and consent were obtained from children and parents before testing. Approval for all procedures was obtained from the Georgetown University Institutional Review Board.

Of this sample, four participants were excluded from analyses due to excessive motion during the MRI scanning and three participants were excluded from analyses for failing the attention checks during the task (>50% errors). The resulting sample consisted of 30 adolescents (M _age = 13.33, SD = 2.29; 14 females; see Tables 1a and 1b). All participants were native English speakers.

Table 1a. Demographic and behavioral characteristics

Table 1b. Spearman ρ correlations among demographic and CU variables (N = 30)

**p < 0.001, *p < 0.05; ^{^}imputed one question item as average of remaining responses for one subject due to blank data

1.2 Inventory on callous-unemotional traits

In accordance with standard procedures, ICU scores were calculated using the highest item response from either the child or parent for each item, which reduces susceptibility to social desirability biases and optimizes accuracy across multiple contexts (Frick et al., Reference Frick, Cornell, Bodin, Dane, Barry and Loney2003). Two subjects did not report an answer for a question item but their parents did, and one parent did not provide an answer for a question item but the subject did. For these instances, we used the only available item response. We also assigned “somewhat true” for two question items where one subject’s parent selected both “not at all true” and “somewhat true.” We then calculated a summary score of all item responses (ICU total) and each subscale (callous, uncaring, and unemotional) for each participant (Kimonis et al., Reference Kimonis, Frick, Skeem, Marsee, Cruise, Munoz and Morris2008). Internal consistency was high for the ICU total scale (α = .92), callousness (α = .82), and uncaring (α = .90) subscales, but low-to-moderate for the unemotional subscale (α = .49), which is commonly observed among studies (Cardinale & Marsh, Reference Cardinale and Marsh2017).

1.3 fMRI task

We adapted an experimental task from Palmer and colleagues (Reference Palmer2013), which comprised six runs. In the first two runs, participants viewed four 18-s blocks of rapidly presented fearful faces interleaved with four 18-s blocks of fixation for 3 min. During the fearful face blocks, faces were presented for 200 ms followed by a 300-ms fixation cross. The purpose of these baseline runs was to orient the participants to the fearful faces before presenting them in three experimental contexts. During the next four runs, participants again viewed 18-s blocks of rapidly presented fearful faces interleaved with 18-s blocks of fixation. During these runs, a 2000-ms cue appeared before each block that placed the faces in one of three contexts: “afraid of you,” “afraid for you,” or “afraid for themselves.” During this prompt, participants pressed one of three buttons corresponding to the condition to indicate that they had read and understood the context for the upcoming block. The interval between the offset of the cue and the onset of a face block pseudorandomly varied between 2500 and 3600 ms. Each contextualized block was presented twice per run for a total of six blocks of contextualized fearful faces interleaved with six blocks of fixation in each run. These final four runs of the task were 4 min 13 s each. The blocked design was selected to reduce cognitive load resulting from changing contexts (Figure 1).

Figure 1. Visualization of the fMRI Task. During the first two runs of fMRI, participants viewed 18-s blocks of noncontextualized fearful facial expressions (200 ms) and fixation (300 ms) interleaved with 18-s blocks of fixation (not depicted). During the final four runs, participants again viewed 18-s blocks of fearful facial expressions (200 ms) and fixation (300 ms) followed by 18-s blocks of fixation. Prior to each face block, a sentence appeared for 2000 s indicating that the “following people are all afraid… ‘FOR YOU’, ‘FOR THEMSELVES’, or ‘OF YOU’.” Participants were asked to press one of three buttons that corresponded to the instruction as an attentional check.

Face stimuli included fearful facial expressions from 24 different exemplar individuals from the Karolinska face database (Lundqvist, Flykt, & Ohman, Reference Lundqvist, Flykt and Ohman1998), selected because the actors in this database were between the ages of 20 and 30, which is younger than most facial expression sets. Faces presented in grayscale were normalized for size and luminance. Each individual face was presented in only one of the three context condition blocks (i.e., six identities per context). The other six identities were presented within the first two (baseline) runs.

1.4 fMRI data acquisition

MRI data were acquired on a 3.0 T Siemens TIM Trio MRI scanner (Erlangen, Germany), located at the Georgetown University Center for Functional and Molecular Imaging. The first two functional scans consisted of 72 contiguous T2*-weighted echo planar imaging (EPI) whole-brain functional volumes while the final four functional scans consisted of 102 T2*-weighted EPI volumes. All functional scan contained the following parameters: repetition time (TR) = 2500 ms; echo time (TE) = 35 ms; flip angle =90°, 43 slices, matrix = 64 × 64; field of view (FOV) = 240 × 240 × 129 mm³; acquisition voxel size = 3.75 × 3.75 × 3.00 mm³. A T1-weighted high-resolution anatomical image was acquired for coregistration and normalization of functional images with the following parameters: TR = 1900 ms; TE = 2.52 ms; flip angle = 9°; 176 slices; FOV = 176 × 250 × 250 mm³; acquisition voxel size = 1.00 × 0.98 × 0.98 mm³.

1.5 fMRI preprocessing and first-level analysis

Preprocessing was performed using fMRIPrep 1.3.2 (Esteban et al., Reference Esteban, Blair, Markiewicz, Berleant, Moodie, Ma and Gorgolewski2017, Reference Esteban, Markiewicz, Blair, Moodie, Ayse, Erramuzpe and Gorgolewski2018), which is based on Nipype 1.1.9 (Gorgolewski et al., Reference Gorgolewski, Esteban, Markiewicz, Ziegler, Ellis, Jarecka and Ghosh2019; Gorgolewski et al., Reference Gorgolewski, Madison, Burns, Clark, Halchenko, Waskom and Ghosh2011) and included anatomical T1-weighted brain extraction, head-motion estimation and correction, slice-timing correction, intrasubject registration, and spatial normalization to the Montreal Neurological Institute 152 Nonlinear Asymmetric 2009c Template. The data were smoothed using a 6 mm³ full-width half-maximum Gaussian kernel prior to first-level beta parameter estimation. Using Statistical Parametric Mapping (SPM12; https://www.fil.ion.ucl.ac.uk/spm/), a general lineal model (GLM) was constructed for each participant across each of the six runs using boxcar regressors for each of the three instruction prompts, three fearful faces conditions with contexts, six motion parameters, one fearful faces condition without context, and a constant regressor for each run. The resulting GLM was convolved using a canonical hemodynamic response function and corrected for temporal autocorrelations using a first-order autoregressive model. Lastly, a standard high-pass filter (cutoff at 128 s) was used to exclude low-frequency drifts. Subject-level beta maps are available in Supplementary Data.

1.6 fMRI univariate analysis

We first conducted a series of whole-brain univariate multiple regression analyses in SPM12 to examine how neural activation to fearful facial expressions varied as a function of CU traits. For each model, we controlled for age, gender, and IQ. A family-wise error (FWE) corrected threshold of p < 0.05 with >10 voxels per cluster at the whole-brain level was applied to all analyses. To address a priori hypotheses regarding the role of the amygdala in fearful face processing, we also applied small volume correction using an anatomically-defined bilateral amygdala parcel from the Harvard Oxford Subcortical Atlas (Desikan et al., Reference Desikan, Ségonne, Fischl, Quinn, Dickerson, Blacker and Killiany2006; Goldstein et al., Reference Goldstein, Seidman, Makris, Ahern, O’Brien, Caviness and Tsuang2007; Makris et al., Reference Makris, Goldstein, Kennedy, Hodge, Caviness, Faraone and Seidman2006) and conducted region-of-interest (ROI) analyses that averaged across voxels in the left and right amygdala.

Our first regression model examined the relationship between CU traits and activation during the first two baseline runs. The second regression analysis modeled how neural activation during the final four runs to fearful facial expressions varied across contexts (“afraid of you,” “afraid of you,” and “afraid for self”) and as a function of CU traits. We conducted each test for ICU total scores and subscale scores.

1.7 Searchlight intersubject representational similarity analysis (IS-RSA)

To test where multivoxel neural activity patterns associated with viewing contextualized fearful facial expressions corresponded to each of three models (high-CU alike, low-CU alike, and nearest neighbors), we employed IS-RSA (Chen et al., Reference Chen, Jolly, Cheong and Chang2020; Finn et al., Reference Finn, Glerean, Khojandi, Nielson, Molfese, Handwerker and Bandettini2020; Nguyen et al., Reference Nguyen, Vanderwal and Hasson2019; van Baar et al., Reference van Baar, Chang and Sanfey2019) using the NLTools package version 0.3.11 (Chang et al., Reference Chang, Sam, Jolly, Cheong, Burnashev, Chen and Frey2019) in Python 3.5.2 and the CoSMoMVPA Toolbox version 1.1.0 in MATLAB 2017a. We first constructed three intersubject models (30 subjects × 30 subjects symmetrical matrices) based on ICU total scores, with each cell corresponding to the models’ prediction of a subject pairs’ neural pattern dissimilarity based on their CU scores. The first two models were generated using the Anna Karenina principle that “all low (or high) CU scorers are alike; each high (or low) CU scorer is different in his or her own way” (Finn et al., Reference Finn, Glerean, Khojandi, Nielson, Molfese, Handwerker and Bandettini2020) according to the following formulas: the low-CU alike model was constructed using the maximum score for each subject pair; the high-CU alike model was constructed using 1 minus the minimum score for each subject pair. The third nearest neighbors model was constructed using the absolute value of the difference between subject pairs’ scores. Prior to constructing each model, all subjects’ summary scores were normalized relative to the maximum CU score in the sample. To investigate finer-grained patterns of CU traits, we also constructed a nearest neighbors model by calculating the Euclidean distance between subject pairs’ item-wise responses on the ICU. Each matrix consisted of 435 unique combinations of subject pairs (30!/2! × (30 − 2)!).

Next, we searched for brain regions that responded to fearful facial expressions similarly 1) among high-CU subjects, 2) among low-CU subjects, and 3) for subjects who were similar in CU traits in a relative rather than an absolute sense. We conducted this analysis using a spherical searchlight comprised of 100 voxels. For each task condition (baseline, “afraid for you,” “afraid of you,” and “afraid for self”), we created a 30 × 30 dissimilarity matrix in each searchlight in terms of the multivoxel activity pattern correlation distances (1 minus Spearman ρ) between all pairs of participants while they viewed the expressions in each context (again, 435 unique combinations). We assessed correspondence between the lower triangles of these matrices using a Spearman ρ correlation and assigned ρ values to the center voxel of each searchlight. Statistical significance was determined using a Mantel permutation test (Mantel, Reference Mantel1967; Nummenmaa et al., Reference Nummenmaa, Glerean, Viinikainen, Jääskeläinen, Hari and Sams2012), in which both the rows and columns of one subject × subject dissimilarity matrix (e.g., subject labels) were shuffled and the Spearman correlation between both correlation matrices was recomputed 1000 times to generate an empirical null distribution of rank correlations (Figure 2). For each searchlight, we calculated the p-value as the proportion of instances in which the permuted Spearman ρ statistic exceeded the true Spearman ρ statistic, and thresholded the resulting maps using a cluster-forming threshold at p < .005 and cluster-extent threshold at k = 10, and using a false-discovery rate (FDR) at q = .05.

Figure 2. Visualization of Searchlight Intersubject Representational Similarity Analysis. Searchlight Intersubject Representational Similarity Analysis (IS-RSA) consisted of the following steps: (1) We computed three subject × subject disimilarity matrices based on CU summary scores across subjects. The first matrix tested a model in which low scoring CU adolescents’ neural response patterns were more alike while all others’ were different from each other; the second matrix tested a model in which high scoring CU adolescents’ neural response patterns were more alike while all others’ were different from each other; and the third matrix tested a model where adolescents’ neural response patterns were similar to each other in a relative rather than an absolute sense. Depicted trait dissimilarity models are sorted by ICU total scores in ascending order. (2) For each condition (“Afraid for you”, “Afraid for self”, “Afraid of you”), we then computed a subject × subject neural dissimilarity matrix within 100-voxel searchlights across gray matter. (3) Again for each condition, we vectorized the lower triangle of each matrix and performed a Spearman ρ correlation at each searchlight between intersubject behavioral dissimilarity and intersubject neural dissimilarity matrices, and assigned the ρ statistic to the center voxel in the searchlight. (4) Statistical significance was determined using a Mantel permutation test, in which both the rows and columns of the subject × subject model dissimilarity matrix were shuffled and the Spearman correlation between both correlation matrices was recomputed 1000 times to generate an empirical null distribution of rank correlations. (5) At each searchlight, we calculated the p-value as the proportion of instances in which the permuted Spearman ρ statistic exceeded the true Spearman ρ statistic.

For each condition, we assessed the voxel-wise similarity between the nearest neighbors model constructed using subjects’ item-wise responses and the nearest neighbors model based on the absolute difference of subject pairs’ summary scores by conducting a Spearman rank correlation, list-wise excluding nonoverlapping voxels. Because these tests revealed significant similarity across voxels between models for all conditions (ρ _baseline = .78, p < .0001; ρ _ForYou = .76, p < .0001; ρ _OfYou = .68, p < .0001; ρ _ForSelf = .70, p < .0001), we opted to interpret results from the absolute difference nearest neighbors model across all conditions. We report findings using the Euclidean distance model in Table 2.

Table 2. Table of searchlight IS-RSA results meeting significance threshold.

Note. Results are reported using a cluster-forming threshold at p < .005 and cluster-extent threshold at k = 10 with voxel size = 3mm³. ⁺ indicates a cluster surviving false discovery rate (FDR) correction at q = .05. ρ values represent the degree to which the specific cluster corresponds to the inter-subject model with positive values indicating higher correspondence and negative values indicating anticorrespondence. Regions were labeled using the Harvard Oxford Atlas, and organized by condition, intersubject model, positive-to-negative ρ values, and then posterior-to-anterior.