Widespread and interrelated gray matter reductions in child sexual offenders with and without pedophilia: Evidence from a multivariate structural MRI study

To further investigate the neuroanatomical correlates of child sexual offending and disentangle them from the neural correlates of pedophilia, using a multivariate analytical approach in order to minimize loss of statistical power.

Child sexual offending (CSO) is one of the gravest disturbances to a child's life, often followed by severe disruptions to victims' well-being and development. 1,2 Approximately 11.8% of children become victims of CSO according to a 2011 international meta-analysis (N = 9,911,748). 2 One central risk factor for CSO is pedophilia (P), a persistent or dominating sexual preference for prepubescent children, marked by recurrent sexual urges, fantasies, or behaviors involving sexual activity with a prepubescent child or children. According to the International Classification of Diseases (10th revision; ICD-10 3 ), this preference must prevail for at least 6 months and must be present in a person who is at least 16 years old and at least 5 years older than the child/children. Importantly, the person affected must either suffer from or have enacted their sexual preference (F65.4). These criteria are fairly similar to a 'pedophilic disorder' in the 5th Diagnostic Statistical Manual (DSM-5 (302.2)). 4 The DSM-5 however differentiates between this 'pedophilic disorder' and a 'pedophilic sexual orientation', i.e. the deviant preference in absence of its enactment, feelings of guilt, shame, or anxiety, and functional limitations due to the preference. This highlights that preference and behavior are not synonymous since, contrary to public belief, a pedophilic preference is present in only about 40-50% of child sexual offenders (CSOs). 5 Various motivations for non-pedophiles to commit CSO have been proposed, including a lack of alternative partners, and psychopathology such as other paraphilias or antisociality. Conversely, their offenses have been found to be more opportunistic, violent, and non-exclusive to children. [6][7][8] Both pedophilia and CSO have inspired considerable research, in which possible neural bases have been of special interest.
Until recently, the two phenomena have often remained confounded, either by not clearly differentiating groups or by the use of only data from pedophilic offenders (+CSO+P), leaving out nonpedophilic offenders (+CSO-P) and pedophilic non-offenders (-CSO +P). In this literature, +CSO+P have been found to exhibit abnormalities regarding (prenatal) neurodevelopment markers, [9][10][11][12][13][14] neuropsychological performance, [15][16][17][18] and their neurofunction regarding various paradigms and connectivity. 15,16,[19][20][21][22][23][24][25][26][27] Structural gray matter reductions have been reported for regions in the frontal, 28,29 parietal, 28 and temporal lobe, 28 including for limbic loci, 29,30 as well as for the cingulate cortex, 28 the insula, 28,29 structures of the basal ganglia, 28 and cerebellar loci. 28 Therein, some of these exhibited correlations of gray matter with pedophilia-related markers, leading to the hypothesis that neuroanatomical disturbances in pedophilia may be dimensional rather than categorical. 29 Simultaneously, multiple studies have reported no significant group differences in gray [31][32][33] and some instead in white matter. 31,32 This is in line with findings of resting-state functional connectivity, 24,25 and not necessarily contradictory to former gray matter findings, as the structures connected by significantly reduced white matter tracts overlap with structures found to be abnormal in gray matter 32 and previously identified abnormal gray matter structures have been found to exhibit substantial interconnectivity. 34 However, the literature is heterogeneous regarding analytical approaches (thresholds, whole-brain or ROI analyses), control groups (non-offenders or non-sexual offenders), and inclusion criteria for +CSO+P groups (diagnoses, phallometric results, varying criminological properties, admission to preference, or a combination).
More recently, neuroscientific works have begun to disentangle P and CSO, regarding executive functioning, [35][36][37][38][39] IQ, 40 behavioral control, 23 affect recognition, 41 resting-state functional connectivity, 42 and neurostructure. 40,43 Schiffer et al., 43 (as part of the NeMUP research initiative) conducted a whole-brain comparison of 58 +CSO +P, 60 -CSO+P, and 101 non-pedophilic, non-offending controls (please note that some studies, due to their psychiatric rather than forensic focus, use the notation of P+CSO to describe pedophilic CSOs, thereby putting P first; inclusion criteria may vary slightly due to prioritizing one before the other when building the sample). No difference emerged between pedophiles and controls overall, but +CSO+P exhibited lower gray matter volume relative to -CSO+P in the right temporal pole. Lett et al., 40 also examined pedophilic men with and without a history of CSO and again found (gray and white) matter reductions in +CSO+P but not -CSO+P, in the right motor cortex, the frontal and temporal lobe, and the corpus callosum, among others.
Overall, it remains uncertain whether +CSO+P exhibit neurostructural abnormalities and which features any differences may be related to, not least because data from +CSO-P are still missing. First neuroanatomical data from -CSO+P and +CSO+P may indicate CSO, rather than P, to be associated with neuronal anomalies, though it remains possible that both phenomena exhibit their own (overlapping) neuropathology. 40,43 Concurrently, both +CSO+P and +CSO-P have been found to exhibit neuropsychological impairments, though with stronger weaknesses in +CSO-P and differing executive profiles. 36,37,39 The present study contributes to ongoing efforts in neuroscience to disentangle the phenomenology of CSO and P by presenting gray matter data from incarcerated +CSO+P, +CSO-P, and non-sexual violent offenders (NSOs). It is, to our knowledge, the first neuroanatomical study on CSOs to include +CSO-P subjects. Given the mixed evidence thus far and hints of connectivity-related abnormalities, we conducted a whole-brain analysis using a multivariate, data-driven approach that allows for the identification of networks of neuroanatomical loci exhibiting interrelated abnormal gray matter volume. 44 Methods Data collection and participants All data were taken from from the Kiehl lab data pool at the Mind Research Network (MRN), collected between 2010 and 2014. [45][46][47] The lab uses a Siemens mobile MRI scanner to collect neuroimaging data in maximum-security detention facilities across the United States, which has allowed for the analysis of large samples in the field of forensic psychiatry. Subjects provide informed consent to the use of their data in continued research prior to participation. All assessments were ethically approved by The University of New Mexico Institutional Review Board and are continually approved by Ethical and Independent Review (E&I); the present study conforms to the Declaration of Helsinki.
The data pool was searched for subjects who fit inclusion criteria for this particular study. As the most reliable information was criminological, inclusion and subject allocation was primarily based on this info supplemented by psychiatric information where possible and applicable, and kept as conservative as possible to avoid misallocation of (CSO) subjects, resulting in a moderate final sample size despite the initially large data pool.
General inclusion criteria were an age between 18 and 65, a fullscale IQ of or above 75, fluency in English, a 4th grade reading level or higher, no history of seizures in self or of psychotic disorder in self or first-degree relative, as well as no mood disorder and no alcohol or drug use at the time of assessment. All CSOs had offended against at least one victim under 11 in a sexual offense. To be sorted into the +CSO+P group, subjects had to have had no hands-on sexual victims above the age of 16 and a DSM-IV-TR pedophilia diagnosis, for which the criteria are the same as those for a pedophilic disorder in the DSM-5. 48 +CSO-P either had victims over 16 or no pedophilia diagnosis. NSOs had committed at least one homicide, to hold constant the factors of having committed a violent offense among subjects, and no sexual offenses.
Information on subjects' demographics, psychological, psychiatric (including neurobiologically relevant medication at time of assessment, i.e. neuroleptics, SSRIs, amphetamines, and antihistamines), and criminological background was extracted from legal documents, inmate files and, if available, psychiatric assessments. This information had been assessed beforehand by MRN staff or by the incarceration institutions. IQ had been assessed using the Wechsler Adult Intelligence Scale III 49 and psychopathy had been assessed using the Psychopathy Checklist-Revised 50 (see former Kiehl lab publications for more info). Crimes were summarized into the total number of all, non-violent, and violent (sexual and non-sexual) crimes, both from convictions and from self-reports.

Data analyses
Pre-processing of imaging data and SBM Imaging data were pre-processed using Statistical Parametric Mapping 12 (SPM12, Wellcome Department of Cognitive Neurology, London, UK; http://www.fil.ion.ucl.ac.uk/spm) and analyzed using Source-Based Morphometry (SBM, performed using GIFT software, available from TReNDS; https://trendscenter.org/software/), 44 a multivariate alternative to Voxel-Based Morphometry (VBM 51 ). Preprocessing for SBM was done in the same manner as has been done for VBM 52 , details of the procedure used with the present data can be found in the Supplement. Using Independent Component Analysis (ICA), 24 SBM identifies patterns of voxels that show maximal covariation regarding gray matter in the sample and maximal independence from other networks of covariation. Its qualities and comparison to VBM have been elaborated on before. 44 Each network is captured in one component. For the present work, the number of components was set to 30, as in other studies using SBM. 53 ICA is independent from subjects' group affiliation. It produces individual loading values that represent the extent to which a given component contributes to each subject's T1 image. Loading values can therefore subsequently be used for statistical group analyses. The final interpretation of group differences is conditional on a component's spatial image. Higher loadings of predominantly positive components should be interpreted as higher matter volume or density, whereas higher loadings of negative components stand for lower matter volume or density. tests for normally and non-normally distributed interval-scaled numeric data, respectively. For pairwise comparisons, independent samples t-tests and Mann-Whitney-U-tests were performed. For count data, χ 2 -tests were performed when expected frequencies exceeded five in all cells and Fisher's exact tests were used when they did not.

SBM component loadings
Individual component loadings were entered into a multivariate analysis of variance (MANOVA) with group as the sole predictor and all 30 components as dependent variables (step 1). A significant MANOVA result can be interpreted as indication for a true, non-noise difference between groups in at least one dependent variable (in this case a component), but should be followed-up with an analysis of variance (ANOVA) for each dependent variable to determine in which one(s) the difference lies.
We therefore determined that following a significant MANOVA, we would compute a follow-up one-way ANOVA for every component, again with group as the sole predictor (step 2). If the ANOVA for a component exhibited a significant result (Holm-corrected P < 0.05), it would be followed up with (i) a linear model with relevant control variables as additional predictors; and (ii) pair-wise comparisons of the three groups using independent samples t-tests (step 3). If a component did not exhibit a significant group difference in step 2, we would refrain from further investigating them.

Average gray matter in significant component clusters
Since one component comprises multiple morphological structures, if any components were to exhibit a significant P-corrected ANOVA result (as indicated in step 2), they were decomposed into the clusters they encompassed, identified using probabilistic cytoarchitectonic maps in the Anatomy Toolbox 2.1 implemented in SPM8 (height threshold: z > 2, extend threshold: k = 2 voxels, http://www.fz-juelich.de/inm/ inm-1/DE/Forschung/_docs/SPMAnatomyToolbox/ SPMAnatomyToolbox_node.html, [55][56][57] ) and the Brainnetome Atlas, a morphological parcellation atlas based on meta-analytic structural and functional connectivity data 58 (please see Table S1 in the Supplement for Brainnetome regions and subregions). Then, individual average gray matter was extracted for 10 mm radius spheres surrounding the peak voxels of clusters exceeding 300 vx. On this data, steps 1, 2, and 3 were run again, to determine whether one or more clusters within components that were significantly different between groups drove this significance finding particularly.

Anatomically defined ROIs
Lastly, steps 1 through 3 were again used with average gray in a set of 11 regions of interest (ROIs) chosen for their theoretical or empirical relevance to the neurobiology of P/CSO, namely the prefrontal (PFC) and orbitofrontal cortex (OFC), the anterior (ACC) and posterior cingulate (PCC), insula, amygdala, hippocampus, striatum, nucleus accumbens, hypothalamus, and thalamus. Masks for volume/ density extraction were created using the WFU PickAtlas (https:// www.nitrc.org/projects/wfu_pickatlas/). For the OFC, ACC, PCC, the amygdala, and the nucleus accumbens, additional ROIs were defined for each hemispheric half of the structure, in addition to the ROIs covering the whole structures, amounting to 21 ROIs total.

Correlational analyses
In addition, we explored dimensional associations of gray matter volume with pedophilic interest as measured by the Screening Scale for Pedophilic Interests (SSPI 59,60 ), an association which was found in previous works for the left DLPFC and the left transition zone between insula and parietal operculum. 29 Due to the component loadings being relative and standardized values, this was only done with average gray matter values extracted from cluster spheres and ROIs.

Sample characteristics
Out of approximately 300 screened subjects, a total of N = 63 was included in the final sample. Out of these, 22 subjects were entered as +CSO+P, 21 as +CSO-P, and 20 as NSOs. Available demographic, psychological, criminological, and global morphological information and group comparisons can be found in Table 1.
Group comparisons revealed significant differences in age, IQ, total number of crimes, violent crimes, non-sexual crimes, and in PCL-R scores (Table 1). More specifically, NSOs were younger and had committed less crimes overall than both CSO groups (though, per definition, at least one homicide). Among CSOs, +CSO-P exhibited a lower IQ and more non-sexual crimes than +CSO+P, and, on average, the highest PCL-R scores in the sample. 40.9% of +CSO+P and 19% of +CSO-P were predominantly homosexually oriented (χ 2 = 7.28, P = 0.0263) (this information was missing for 1 +CSO +P, 3 +CSO-P, and all NSOs), and median scores on the SSPI were the maximum of 5 for +CSO+P (min = 3, max = 5) and 3 for +CSO-P (min = 2, max = 5) (χ 2 = 7.48, P = 0.0293). Among +CSO+P, pedophilia diagnoses were marked as 'exclusive' for 17 (77.3%) subjects and as 'non-exclusive' for 2 (9.1%) (information was missing for the remaining 3 subjects). Among +CSO-P, 10 subjects (47.6%) had victims above 16; 5 of those had a pedophilia diagnosis (23.8% of all +CSO-P) in their file. Note, that one additional subject who had a pedophilia diagnosis and no convictions for sexual assault against an adult was sorted into the +CSO-P group due to a) his generally deviant behavior as expressed by a high number of (violent) crimes, b) the violence of the CSO, and c) allegations of rape by other inmates against him. Groups did not differ with regard to total brain , overall gray matter, white matter, or cerebrospinal fluid volume. In addition to the instances of missing information named above, information was missing in three NSOs for axis I diagnoses, in two for alcohol and substance use, and in all for paraphilia diagnoses .
Overall, no subject met criteria for any axis I disorder except for one NSO, who had been diagnosed with posttraumatic stress and anxiety. Substance use diagnoses, if present, referred to Cannabis use in all groups and to Cocaine and polydrug use in some +CSO-P. If subjects met criteria for any personality disorder, it was of antisocial personality disorder (see Table 1).

Neuroanatomical group differences
Before each MANOVA analysis, general (multivariate) data distribution assumptions were examined. If data did not entirely fulfill these assumptions, an additional, rank-based MANOVA 61,62 was computed.
A subsequent series of 30 one-way ANOVAs with only group as predictor, revealed a significant difference between groups in compo-  Crimes comprise all crimes, including those admitted to but not necessarily prosecuted for. Percentages are indicated as proportion of total n in group, regardless of missing data. predictors (F model (6, 54) = 5.883, P < 0.001, R 2 adj = 0.328), though this model did not perform significantly better than the full model (F (5, 49) = 1.058, P = 0.3948).

Average gray matter in component clusters
Two cerebellar clusters (7705 and 911 vx) were located in the anterior (lobules IV and V) and posterior (lobules VIIA, Crus I and II, and VIIIA) vermis, respectively, with the first cluster paravermally also reaching into the right posterior cerebellum, and the second one reaching into the left occipital cortex. Two further cerebellar clusters (1907 vx in lobules VI, VIIA, Crus I, and VIIIB, and 739 vx in lobules VIIIA/B, IX, and X) were posteriorly and laterally situated, with the right cluster reaching into the right inferior temporal and fusiform gyrus. Four clusters (3066, 2001, 587, and 403 vx) were situated in the frontal lobe, comprising the bilateral precentral gyri (premotor area), dorsomedial/dorsolateral prefrontal cortex (DM/DLPFC), inferior frontal junction (IFJ), and the left medial OFC (mOFC) reaching into the right mOFC, including voxels in the frontal pole. In the parietal lobe, in addition to two clusters (401 and 1111 vx) in the bilateral medial precuneus close to the parieto-occipital sulcus and the PCC, a right-hemispheric cluster (393 vx) was located in the superior parietal lobule, while a left cluster (354 vx) lay in the inferior parietal lobule, more specifically the intraparietal sulcus. Finally, one cluster each in the left (665 vx) and right (325 vx) basal ganglia emerged, comprising the bilateral ventromedial putamen and left ventral caudate. Overall, the right hemisphere was slightly more impacted than the left.
Unsurprisingly, a MANOVA of the 10 mm spheres surrounding the 37 peak voxels in the 15 biggest clusters (>300 vx) with group as the sole predictor again yielded a significant result (Λ Roy = 4.14, F(2, 60) = 2.612, P = 0.0076). However, none of the 37 one-way ANOVAs (with group as the only predictor) computed as a follow-up remained significant after correcting for multiple comparisons.
An examination of smaller clusters using the SPM Anatomy Toolbox (<300 vx, T > 2.00; see Table 3) and the Brainnetome Atlas (<300 vx, no height threshold; see Table S2 for assignments) revealed abnormalities mirroring the structures in the bigger clusters in the respective opposite hemisphere (right inferior and left superior parietal lobule, left inferior temporal and fusiform gyri, right occipital lobe, right ventral caudate). Additionally, they included the bilateral postcentral gyri, the right temporal pole, the right superior and middle temporal gyri, the bilateral posterior parahippocampal gyrus, medial cingulate, and hippocampus, and revealed that the basal ganglia clusters spread into the left dorsal caudate and bilaterally into the dorsolateral putamen, nucleus accumbens, and globus pallidus.

Anatomically defined ROIs
A MANOVA for average gray matter volume in the 21 anatomically defined ROIs also yielded a significant result for group (Λ Roy = 1.078, F(2, 60) = 2.1, P = 0.0205). None of the results for the 21 follow-up one-way ANOVAs (group as only predictor) survived correction, however.  General localization of clusters and maxima was identified using the Anatomy Toolbox implemented in SPM. Structure assignment was also done using the SPM Anatomy Toolbox and complemented by assignment via the Brainnetome Atlas; in general, assignment via the two tools did not conflict. Where no probabilistic assignment was possible via the Anatomy Toolbox, only Brainnetome assignments are given (as indicated by the absence of a probability estimate for the assignment). A full table of Brainnetome abbreviations is available in the Supplement (Table S1)

Correlational analyses
In contrast to former findings, no significant correlation of SSPI, PCL-R scores, or number of violent crimes with gray matter volume was identified in our data, neither when examining the decomposed C22 clusters, nor when examining the anatomically defined ROIs. This was the case both for all CSOs and for each CSO group on its own. Correlation values and pertaining significance tests can be found in the Supplement (Tables S3 to S24).

Discussion
The present study is the first to present a structural MRI comparison of +CSO+P, +CSO-P and NSOs. Using a multivariate whole-brain analysis approach, we identified a widespread, interrelated, and robust network of reduced gray matter volume in CSOs relative to NSOs, but no significant difference between +CSO+P and +CSO-P (despite a trend placing +CSO-P between +CSO+P and NSOs, which may indicate that a remainder of variance may be explained by pedophilia per se, though this is slightly at odds with other current works in the field 43  The network character of our findings was additionally underlined by two aspects. First, our findings encompassed important associative structures, including the precuneus and the DLPFC, the basal ganglia, and the cerebellum. The basal ganglia and cerebellum especially have been found to form (complementary) parallel closed loops with cortical structures, 64 mediating the integration of various cognitive and emotional processes and contributing to several higher-level functions including social cognition. 65,66 Interestingly, loci with abnormal gray in CSOs in our results seemed to be more strongly attributable to cognitive than emotional processes in both circuit bases. 67 Second, our non-cerebellar clusters appeared to exhibit connectivity with each other as informally identified using the Brainnetome Atlas. Overall, this may indicate connectivity-related neural impairments in CSOs, affecting higher-level cognitive and inhibitory functioning, in line with neuropsychological findings. 68 Though no study thus far has conducted the same group comparison as we did, our results show overlap with the (sometimes uncorrected) findings of studies comparing +CSO+P to nonoffending controls 28,32 (see the Supplement for regions). They also overlap with results of reduced gray in +CSO+P compared to -CSO +P specifically. 40,43 We are furthermore not the first to identify network-related impairments of brain structure (and function 24,25 ) in CSOs. In fact, Cantor et al., argue that connectivity-based abnormalities are the only findings on brain structure in +CSO+P not attributable to low statistical power, 32 as they identified no abnormal gray but abnormal white matter in CSOs compared to NSOs 31 and healthy controls. 32 Of note, the gray matter structures connected by their white matter findings overlap with the loci we identified (left DLPFC, frontal pole, superior parietal lobule, occipital cortex). Furthermore, in a review examining functional connectivity between gray matter regions previously identified as abnormal in +CSO+P and key areas of sexual processing, Poeppl et al., suggested that altered brain structure in these men may 'affect neural networks for sexual processing through disrupted functional connectivity'. 34 In parallel to this interpretation, our findings overlap with the inhibitory (mOFC, caudate), cognitive (appraisal: inferior temporal cortex; attention: superior and inferior parietal lobule), and autonomic/endocrine (putamen) components of the neurophenomenological model of sexual arousal in men. 69 This network has been found to be activated similarly in pedophiles and teleiophiles when each group is confronted with their preferred stimuli, 22,27 but has thus far not been explicitly studied in sex offenders vs. non(sexual)-offenders. Our results suggest that structural brain abnormalities may be pertaining to CSO rather than to pedophilia, a conclusion also drawn in former studies comparing +CSO+P and -CSO+P. 40,43 Using a multivariate, whole-brain analysis approach, we were able to present an unbiased comparison with a minimal loss of statistical power. Our results therefore are a rather strong testament to the presence of brain structural abnormalities in CSOs. However, despite the non-significant comparison between +CSO+P and +CSO-P, our results cannot be seen as proof for there being no neuroanatomical differences between these groups. As has been noted before, it remains possible that both CSO and P exhibit (overlapping) neural abnormalities, and that, among CSOs, abnormalities may (therefore) be especially pronounced in pedophilic men.
Apart from group, no variable in our data showed significant covariance with gray matter, and the group variable remained significant after the introduction of covariates. This included number of violent crimes and psychopathy scores, pointing toward the interpretation that a general proneness to violence may not be explanatory for the abnormalities at hand. However, CSOs had committed significantly more violent crimes and were older than NSOs, possibly reflecting that the NSOs in our sample were incarcerated earlier, providing less time to further commit violent acts. Though age did covary slightly with C22 based on its non-but marginally significant result in the model, its predictive power was neither responsible for nor greater than that of group.
Our study exhibits several limitations. First, though we maximized statistical power when answering our main research question, our sample size was limited. Our findings therefore need to be considered preliminary and should be confirmed by larger scale studies. Second, allocation to groups on the basis of criminological data was not ideal and led to some subjects with a DSM-IV pedophilia diagnosis being sorted into the non-pedophilic group. Identifying pedophilic interest (among CSOs) is a notorious issue, 7,70-72 and though criminal records can give an indication, 59 a multimodal diagnostic process would be superior. Given that the identification of covarying loci in SBM was blind to group allocation however, this did not impact the identification of the later significant gray matter loci, thereby decreasing the influence of possible misallocations. Third, we did not include -CSO+P subjects, a group needed to fully disentangle P and CSO. This addition should therefore be addressed in future research, despite the well-known difficulties of assembling samples of larger sizes in the field, which also challenged the attainment of the present sample. Finally, though our main result points to network-related abnormalities, we did not follow-up with a formal connectivity analysis of our findings.
Overall, our results are a testament to, possibly network-related, neural impairments in CSOs. These abnormalities are unlikely correlates of general violence but may be pertaining to sexual violence (against children). Future research should seek to provide wellpowered P/CSO-2x2-designs and explore multi-instead of univariate analysis approaches further.