Pathways from threat exposure to psychotic symptoms in youth: The role of emotion recognition bias and brain structure

Background: Research supports an association between threatening experiences in childhood and psychosis. It is possible that early threat exposure disrupts the development of emotion recognition (specifically, producing a bias for facial expressions relating to threat) and the brain structures subserving it, contributing to psychosis development. Methods: Using data from the Philadelphia Neurodevelopmental Cohort, we examined associations between threat exposure and both the misattribution of facial expressions to fear/anger in an emotion recognition task, and gray matter volumes in key emotion processing regions. Our sample comprised youth with psychosis spectrum symptoms (N = 304), control youth (N = 787), and to evaluate specificity, youth with internalizing symptoms (N = 92). The moderating effects of group and sex were examined. Results: Both the psychosis spectrum and internalizing groups had higher levels of threat exposure than controls. In the total sample, threat exposure was associated with lower left medial prefrontal cortex (mPFC) volume but not misattributions to fear/anger. The effects of threat exposure did not significantly differ by group or sex. Conclusions: The findings of this study provide evidence for an effect of threat exposure on mPFC morphology, but do not support an association between threat exposure and a recognition bias for threat-related expressions, that is particularly pronounced in psychosis. Future research should investigate factors linking transdiagnostic alterations related to threat exposure with psychotic symptoms, and attempt to clarify the mechanisms underpinning emotion recognition mis-attributions in threat-exposed youth


Introduction
Evidence suggests that childhood adversity is a risk factor for psychosis development (Varese et al., 2012).However, the mechanisms involved in this relationship remain elusive.Non-clinical psychotic-like experiences are common in childhood and adolescence (Kelleher et al., 2012), are more prevalent in people with a history of childhood adversity (Arseneault et al., 2011;Croft et al., 2019), and are associated with an increased risk for psychotic illness (Poulton et al., 2000;Kaymaz et al., 2012).Investigating these experiences alongside more advanced symptoms on the psychosis spectrum in youth may help to characterize adversity-related deviations from normative development and illuminate the processes linking them to psychotic illness.
One possible mechanism involves disruption to socio-cognitive development and its underlying neurobiology.Impaired emotion recognition (the ability to identify facial expressions) has been observed in individuals with a history of adversity (Koizumi and Takagishi, 2014) and individuals on the psychosis spectrum (Green et al., 2015;van Donkersgoed et al., 2015).Deficits in this domain may also predict transition from a clinically high risk (CHR) state to a psychotic disorder (Corcoran et al., 2015).Both adversity (Teicher and Samson, 2016) and psychosis (Fornito et al., 2009) tend to correlate with altered brain structure in regions involved in socio-cognitive and emotion processing.The amygdala, medial prefrontal cortex (mPFC) and hippocampus are key to these functions (Amodio and Frith, 2006;Kober et al., 2008;Montagrin et al., 2018), and due partly to their high density of glucocorticoid receptors, may be particularly vulnerable to the effects of stress related to early adversity (Teicher and Samson, 2016).
Research investigating this pathway, however, has revealed conflicting findings.For example, while adversity has mostly predicted worse performance in emotion recognition tasks in psychosis (Lysaker et al., 2011;Rokita et al., 2020;Tognin et al., 2020), positive associations have also been reported (Tognin et al., 2020).Such inconsistencies may have arisen due to several limitations.First, a focus on speed and accuracy does not consider an individual's degree of response bias or misattribution, whereby the emotions of certain facial expressions are systematically mistaken as another.Second, in much of the adversity literature, several adversity types (e.g., abuse and neglect) are combined under one broad grouping, which may conceal impacts unique to specific exposures.It has been proposed, for example, that experiences characterized by "deprivation" (e.g., neglect, poverty) and "threat" (e.g., abuse, bullying), have distinct impacts on neurodevelopment (Sheridan and McLaughlin, 2014); likewise, distinctions in the nature of emotion recognition alterations have been reported following adversities in these dimensions (e.g., Pollak et al., 2000).Importantly, threat exposure has been specifically implicated in a type of emotion recognition impairment particularly prevalent in psychosis; a recognition bias for threat-related emotions.
Several studies in children and adolescents have reported associations between threat exposure and both attentional and interpretational biases for facial expressions that relate to threat, such as anger (Pollak et al., 2000;Pollak and Sinha, 2002;Pollak and Tolley-Schell, 2003;Shackman et al., 2007;Gibb et al., 2009;Pfaltz et al., 2019;Saarinen et al., 2021), possibly reflecting hypervigilance adapted for protection in unsafe surroundings (McLaughlin and Lambert, 2017).Similarly, a recent meta-analysis found that people with psychosis tend to overperceive threat in the behavior of others, and over-attend to potentially threatening cues (Livet et al., 2020).Neutral and positive facial expressions are also more frequently mislabeled as anger and fear by individuals on the psychosis spectrum, relative to controls (Pinkham et al., 2011;van Rijn et al., 2011;Catalan et al., 2016;Seo et al., 2020).
Given the high prevalence of childhood threat exposure in psychosis, it is possible that the threat recognition bias observed in this population is (partly) driven by these experiences.In the only study to examine this, Catalan et al. (2020) found, in a sample comprising adults with first episode psychosis (FEP), adults with borderline personality disorder, and healthy controls, that abuse in childhood was associated with the mislabeling of happy and neutral expressions as fear or anger, regardless of diagnostic group.These findings have not yet received replication efforts, nor have they been examined in youth with psychotic-like experiences, at a time when emotion recognition continues to develop (Durand et al., 2007;Rodger et al., 2015).
At a neural level, exposure to threat, but not deprivation, in children and adolescents, is associated with lower gray matter volume in the amygdala, mPFC and hippocampus (McLaughlin et al., 2019), suggesting that adversity characterized by threat specifically may disrupt the development of these regions.It is possible that such disruption contributes to psychosis, but only three studies to our knowledge have investigated threat-specific effects on these structures in the psychosis spectrum, and findings were mixed.Sheffield et al. (2013) found an association between threat exposure and volumetric reductions in the mPFC in a whole brain voxel-wise analysis of adult psychosis patients.Vargas et al. (2019) reported threat-related mPFC reductions in both CHR and healthy adolescents, and amygdala reductions across both combined; no effects on the hippocampus were observed.LoPilato et al. (2019) found no association between threat exposure and either amygdala or hippocampus volume in CHR or control youth.Given the importance of these regions to threat processing, and the convincing evidence for their alteration in threat-exposed children and adolescents (McLaughlin et al., 2019), further investigation of these associations in youth with psychosis is warranted.
Using the Philadelphia Neurodevelopmental Cohort (PNC) dataset, the present study investigated whether threat exposure is associated with changes to threat recognition (specifically, the tendency to overclassify facial expressions as anger and fear in an emotion recognition task) and gray matter volume in emotion processing brain regions (specifically, the mPFC, amygdala and anterior hippocampus), and whether increased susceptibility to such changes leads certain youth to experience psychotic spectrum symptoms (PSS).Additionally, despite reports of adversity-related alterations to threat processing in people with anxious (Shackman et al., 2007;Briggs-Gowan et al., 2015) and depressive (Suzuki et al., 2015;Flechsenhar et al., 2022) symptoms, no study to our knowledge has yet compared these alterations with those in the psychosis spectrum; it is thus unclear whether effects found in the research discussed are psychosis-specific or generalize to other psychopathologies.We therefore compared all associations between youth with PSS, control youth, and youth with internalizing symptoms, to contextualize our findings in the broader psychiatric literature.
Although previous studies using the PNC have broadly examined associations of threat exposure with cognition (Barzilay et al., 2018(Barzilay et al., , 2019) ) and brain structure/function (Gur et al., 2019), none have specifically examined threat exposure, threat recognition bias, and underlying brain structure in youth with PSS.We hypothesized that greater threat exposure would be associated with more frequent misattribution to fear/anger, and lower amygdala, anterior hippocampus, and mPFC volumes, and that these effects would be stronger in youth with PSS than controls.We made no hypotheses regarding the relative effects of youth with PSS and youth with internalizing symptoms, given the lack of prior research comparing these groups.

Participants
Participant data was obtained from the PNC dataset.Participants in this cohort were recruited from the greater Philadelphia region and ranged from age 8-21 years.Information regarding recruitment procedures can be found in previous work (Calkins et al., 2014;Satterthwaite et al., 2014).Inclusion and exclusion criteria are provided in the Supplementary Material.
Within the PNC sample, 1598 participants underwent a T1-weighted MRI scan.Of these, 415 were excluded due to poor T1 image quality, a lack of usable emotion recognition data, or the presence of major medical, neurological or developmental conditions.This left a total of 1183 participants in the final analytic sample.Excluded participants were significantly younger [t(703.64)= 12.31, p < 0.001] and more likely to be male [53 % vs 46 % male, X 2 (1, N = 1598) = 4.62, p = 0.032] than those included.Exclusion steps are provided in more detail in the Supplementary Material (Section 1.2 and Fig. S1).

Assignment to diagnostic group
Participants were identified as endorsing PSS (PSS group, N = 304), internalizing symptoms and not PSS (INT group, N = 92), or neither internalizing symptoms nor PSS (control group, N = 787).
Following Calkins et al. (2015), PSS endorsement was defined as having either: 1) a total PRIME Screen-Revised (PS-R) score (Kobayashi et al., 2008) or summed total of 6 items on the Scale of Prodromal Symptoms (SOPS) (Miller et al., 2003), at least two standard deviations above age-matched peers; 2) at least one "somewhat agree" or three "definitely agree" responses to any PS-R item; or 3) definite or possible hallucinations/delusions on the adapted Kiddie-Schedule for Affective Disorders and Schizophrenia (K-SADS - Kaufman et al., 1997).
INT group status was based on current or past endorsement of K-SADS items corresponding to DSM-V criteria for either a major depressive episode or generalized anxiety disorder (GAD), and no endorsement of PSS criteria.Participants not classified as PSS or INT were assigned to the control group.Further details are provided in Tables S1-S3.

Exposure measures
Threat exposure comprised 8 questions relating to traumatic experiences of actual or threatened death or serious injury to oneself or others (see Table S4).Participants were given a score of 0-8 denoting the number of items endorsed, in accordance with Gur et al. (2019).When participants had incomplete data for 5 or fewer items (N = 16; 4 PSS, 1 INT, 11 CTRL), missing items were replaced with the mean score for that participants' other (completed) items.When participants had incomplete data for 6 or more constituent items (N = 2; 1 PSS, 1 CTRL), their total threat exposure score did not contribute to analyses.
Given our focus on threat-specific adversity and the typically observed correlation between threat and deprivation exposure (McLaughlin et al., 2014), we controlled for maternal education, a proxy measure of deprivation, in all threat exposure-related analyses.While low maternal education does not directly reflect deprivation, it has been shown to predict a lack of other family resources (Jackson et al., 2017).

Neuropsychological measures 2.4.1. Threat recognition bias
Emotion recognition was measured using behavioral data acquired during an emotion-identification functional MRI task (functional MRI methods and findings in this dataset are described elsewhere [Satterthwaite et al., 2014;Wolf et al., 2015;Zhang et al., 2019]).In the task, a series of images depicting human faces displaying either a happy, sad, fearful, angry or neutral expression were presented on a screen.During each trial, a face was presented for 5.5 s, and the participant was required to choose which of the 5 emotions was presented.A variable inter-stimulus interval of 0.5-18.5 s was used, during which a fixation crosshair matched to the perceptual qualities of faces was displayed.There were 60 trials in total (12 for each emotion), presented in a pseudorandom order.
To quantify the mislabeling of emotions as threat, a "false alarm" rate was calculated for fear and anger (separately), by dividing the number of trials in which a stimulus was falsely attributed to the target emotion (i.e., anger or fear), by all valid trials where that emotion was not the target emotion (Macmillan and Creelman, 1990;Stanislaw and Todorov, 1999;Goos and Silverman, 2002).Trials were deemed valid if the participant responded 0.2 s or more after the presentation of the face, and before the onset of the following trial.Mean trial validity for each emotion within each group ranged from 95 to 98 %.When a participant responded more than once, the first response in this time window was used.False alarm rates for anger and fear were then averaged to create a combined measure of threat bias.

Additional emotion recognition measures
As a specificity check, we calculated the false alarm rate for all other facial expressions in addition to fear/anger (happy, sad, neutral, and fear and anger separately), as well as the percentage of correct trials (hit rate) for all emotions combined, fear/anger combined, and each emotion separately.

Brain measures
T1-weighted images were acquired using a Siemens Tim Trio 3 Tesla MRI scanner and processed using the FreeSurfer software package (version 6.0, http://surfer.nmr.mgh.harvard.edu/).Scans were screened for quality using set thresholds on FreeSurfer's Euler number and the Mahalanobis distance of cortical thickness measurements (see supplementary material for details).FreeSurfer's automated volumebased stream was used to extract total intracranial volume (ICV), and mean volume estimates within each hemisphere for the amygdala and the cortical areas constituting the mPFC.The mPFC was defined in accordance with Ding et al. (2015), and comprised the medial orbitofrontal cortex, frontal pole, and rostral and caudal anterior cingulate cortex according to the Desikan-Killiany atlas (Desikan et al., 2006).
Volumes for the left and right anterior hippocampus were obtained using a hippocampal subfield segmentation tool implemented in Freesurfer v7.1 (Iglesias et al., 2015), which delineates the hippocampus along its longitudinal segments (i.e., head, body and tail).We constrained hippocampal analyses to the anterior portion due to the putative specialized role of this region in emotion processing (Fanselow and Dong, 2010;Poppenk et al., 2013;Strange et al., 2014;Vogel et al., 2020).Segments of the head within each hemisphere were extracted for each participant as a measure of the anterior hippocampus (Poppenk et al., 2013).Further information on image acquisition, quality control and processing is provided in Section 1.2 of the supplementary material.

Statistical analysis
All analyses were conducted in R version 4.1.1.The packages and functions used are provided in Supplementary Table S5.

Descriptive analyses
Analysis of variance (ANOVA), Kruskal Wallis and chi-squared tests were conducted as appropriate to examine group (control, PSS and INT) differences in age, maternal education, threat exposure, medication use and sex.When significant effects were found, post-hoc comparisons were performed using Tukey's method (age, maternal education), Fisher's exact test (sex, medication use), and Siegel and Castellan Jr. 's (1988) approach for non-parametric analyses (threat exposure).
Due to positive skew in the distribution of fear/anger false alarm rate, and positive and negative skew in the additional emotion recognition measures, all were transformed using ordered quantile normalization (Peterson and Cavanaugh, 2020).Between-group analyses for false alarm rate and brain measures were performed using analysis of covariance (ANCOVA), controlling for age and sex, as well as ICV for brain measures.Post-hoc pairwise comparisons were conducted using Tukey's method.

Associations of threat exposure with cognitive and brain measures
A series of multiple linear regression models were performed to examine the effect of threat exposure on fear/anger false alarm rate, as well as left and right amygdala, left and right mPFC, and left and right anterior hippocampus volume, separately (one model for each dependent variable, N = 7).All models simultaneously included threat exposure as the independent variable, and age, sex, maternal education, and ICV (for brain measures) as covariates of no interest.To investigate whether associations were moderated by group, as well as sex given prior evidence for sex-specific associations between threat exposure and both emotion recognition (Barzilay et al., 2018) and corticolimbic morphology (Samplin et al., 2013;Paquola et al., 2016;Helpman et al., 2017), two (threat exposure by group, threat exposure by sex) and three (threat exposure by group by sex)-way interaction terms were then added to each model hierarchically.Age by threat exposure interaction effects were also examined to evaluate whether associations varied with age.Benjamini and Hochberg's (1995) false discovery rate (FDR) correction was used to adjust for multiple comparisons across the 7 models.

Secondary analyses
Secondary analyses examined the effect of threat exposure on the additional emotion recognition measures using the same multiple linear regression model described above.FDR correction was applied to the 12 models.

Demographics
304 participants were assigned to the PSS group, 92 to the INT group, and 787 to the control group.Demographic characteristics are shown in Table 1.Descriptive information on psychopathology is provided in Table S6.Threat exposure was greatest in the PSS group, significantly exceeding that of control participants but not those with INT; the INT group also had greater exposure than controls (see Fig. 1 and Table S7).

Group differences in brain measures and threat-recognition bias
Descriptive statistics and between-group comparisons for the fear/ anger false alarm rate and brain measures are provided in Table 2.No group differences were yielded for fear/anger false alarm rate but significantly lower gray matter volumes were found in all regions except the left and right mPFC in PSS participants relative to controls (see Supplementary Figs.S2 & S3).

Associations between threat exposure and threat recognition bias and brain measures
The results for each regression model are presented in Table 3.A significant main effect of threat exposure was found for the left mPFC [β = − 0.063, t = − 2.884, FDR-adjusted p = 0.028, partial r 2 = 0.007], such that higher exposure was associated with lower volumes in the total sample (see Fig. 2).No other threat-related effects for brain measures or the fear/anger false alarm rate were significant, although a trend-level association was found between threat exposure and left anterior hippocampal volume [β = − 0.054, t = − 2.264, FDR-adjusted p = 0.083, partial r 2 = 0.004].There were no significant two-way (threat exposure by group, threat exposure by sex, threat exposure by age) or three-way (threat exposure by group by sex) interactions when these variables were included in the regression models, indicating a lack of significant moderation effects; interaction terms were thus omitted from the models reported in Table 3.
b Answered "yes" to "are you currently taking medication because of your emotions and/or behaviors?"

Secondary analyses
Group differences for the additional emotion recognition measures are presented in Supplementary Table S8, and threat exposure regression models for these measures are in Table S9.Relative to both the control and INT groups, the PSS group had higher false alarm rates for neutral faces, and lower hit rates for all emotions combined and sad faces specifically.The PSS group also had lower hit rates for fearful and happy faces than the control but not the INT group.In the total sample, higher threat exposure was associated with lower hit rates for happy faces.See Supplementary Figs.S4 and S5 for significant effects.

Discussion
This study aimed to investigate whether childhood threat exposure is associated with altered threat processing and underlying brain structure in youth, and whether such effects are particularly pronounced in the psychosis spectrum.Overall, threat exposure was higher in both the PSS and INT groups relative to controls, and was associated with lower left mPFC volume in the total sample.In contrast to our hypothesis, threat exposure was not significantly associated with gray matter volume in any other region of interest, nor was it associated with misattribution of emotional expressions to fear or anger.Moreover, no moderating effects of group, age or sex were found.
The association between threat exposure and lower left mPFC volume aligns with previous research in a range of populations including psychosis (Sheffield et al., 2013;Vargas et al., 2019), depression (Malykhin et al., 2012), and healthy youth (McLaughlin et al., 2019).The absence of moderation effects, however, somewhat opposes the suggestion that varying susceptibility to mPFC changes with threat exposure renders some people more vulnerable to psychopathology than others.Given that threat exposure-related volume reductions were expected in all three groups, it is possible that differences between them were too subtle to yield significant effects in our sample.It is also possible that threat exposure is associated with the mPFC in a comparable manner in all youth, and may relate to psychosis and other psychopathology only when combined with other risk factors, such as certain genetic traits (Alemany et al., 2011;Aas et al., 2013Aas et al., , 2014;;Collip et al., 2013).Additionally, compensatory mechanisms may prevent such reductions from progressing into psychopathology in certain individuals (Teicher et al., 2016).
Interestingly, we found no evidence for associations between threat exposure and amygdala volume, and little evidence for associations with the anterior hippocampus (although a negative association between threat exposure and volume in the left hemisphere was significant prior to correction for multiple comparisons).While previous hippocampal and amygdala findings have been mixed in psychosis spectrum samples, several studies have found lower volumes in both regions with threat exposure in community samples (McLaughlin et al., 2019).However, these have been inconsistent for the amygdala, and effects on the hippocampus may be attenuated in youth (Teicher and Samson, 2016).Further, it is thought that the timing of adversity influences the extent and nature of alterations in these regions (Gee and Casey, 2015).While the age of participation did not moderate associations between threat exposure and brain structure or threat recognition bias, we were unable to consider the age at which exposure occurred.Studies comparing the impact of adversity at different stages of development are needed to effectively characterize these effects.
The absence of an association between threat exposure and misattribution of emotional expressions to fear or anger conflicts with reports of attentional and interpretational biases towards threatening stimuli in threat-exposed clinical (Mitchell et al., 2014;Seitz et al., 2021) and healthy (Pollak et al., 2000;Pollak and Tolley-Schell, 2003;Gibb et al., 2009) populations.In particular, our findings differ from those of Catalan et al. (2020), who found increased misattribution of both happy and neutral faces to anger/fear in a sample of adults with FEP, borderline personality disorder, and controls, who reported abuse in childhood.
Several factors may explain this discrepancy.Firstly, differing sample characteristics may have led to distinct manifestations of threat exposure-related bias.For example, while many studies have reported associations between threat exposure and attentional biases towards threat-related faces (Pollak and Tolley-Schell, 2003;Shackman et al., 2007;Gibb et al., 2009), others have found biases away from such stimuli (Pine et al., 2005;Grossheinrich et al., 2022).Moreover, avoidance of threatening faces may be more pronounced in younger children (Pine et al., 2005), suggesting that these biases evolve over time.In this context, it is plausible that most threat-exposed participants b corrected for multiple comparisons using Tukey's method. in Catalan's sample (aged 18-65 years) had experienced a shift from threat-avoidance to threat-over attending, while our sample (aged 8-21 years) may have varied depending on each participants' stage of emotion processing development.
Secondly, ours and Catalan's study used slightly different emotion recognition paradigms.While both had a similar number of overall trials (64 vs 60), the latter did not examine sad faces (thus including more trials for each other expression) and degraded facial stimuli resolution by 30 % to increase task difficulty.Both factors may have increased sensitivity to variation between participants.Additionally, Catalan did not control for deprivation.Lower maternal education, despite being included in our regression models as a covariate of no interest, yielded an association with higher fear/anger false alarm rate.This was surprising, given prior work suggesting that early exposure to threat, and not deprivation, predicts a bias for threat-related emotions (e.g., Pollak et al., 2000).While this finding could indicate that certain types of deprivation contribute to threat recognition bias instead of or in addition to threat exposure, an association between lower maternal education and worse overall task performance may also have been responsible; a conclusion somewhat supported by additional associations between lower maternal education and higher false alarm rates for neutral faces, and lower hit rates for sad faces and all emotions combined (see Table S9).
We also found that lower maternal education was associated with lower right mPFC, left amygdala, and right and left anterior hippocampus volume.Although we only measured one aspect of deprivation, these findings do not support the proposition that threat exposure primarily disrupts affective mechanisms and supporting brain structure, while deprivation mainly affects complex cognition (Sheridan and McLaughlin, 2014).Moreover, prior work in psychosis and healthy samples have similarly found associations between early threat exposure and poor executive function (Aas et al., 2012;Üçok et al., 2015;Op den Kelder et al., 2018;Barzilay et al., 2019), and between deprivation and alterations to amygdala and hippocampal structure (Hanson et al., 2015;LoPilato et al., 2019).Aspects of deprivation and threat exposure may thus disrupt the development of these cognitive and biological characteristics to some degree.There were several limitations to our study.First, as noted, the emotion recognition task used may have been insufficiently challenging to detect between-group and threat-related effects on emotion misattributions.Researchers examining similar associations may benefit from varying trial difficulty via adjustment to facial stimuli resolution or emotional intensity.They should also consider the use of more ecologically valid paradigms, which incorporate facial movement or provide situational context to the emotions displayed (Straulino et al., 2023).Second, our threat exposure measure, which quantified the number of distinct threatening experiences encountered, did not consider the frequency or severity of such experiences.Some highly exposed participants may have therefore received a score of only one or two; future studies should try to include measures capturing both a wide range of experiences and their frequency and severity.Third, while our proxy measure of deprivation (low maternal education) likely gave some indication of each participants' level of cognitive enrichment during development, it gave little to no information about other types of deprivation exposure, such as physical or emotional neglect (McLaughlin et al., 2014).Such experiences would have thus been largely uncontrolled for in our models, and may have influenced associations with threat exposure.Fourth, our study was cross sectional, and therefore unable to estimate the effect of threat exposure on longitudinal trajectories of emotion recognition and brain structure.Additionally, many youth on the psychosis spectrum never develop a psychotic disorder; often remitting (Kaymaz et al., 2012) or developing other psychiatric illnesses (Werbeloff et al., 2012;Lindgren et al., 2022).Longitudinal data could allow the stratification of participants based on psychopathology endorsed at later timepoints, allowing better understanding of effects observed at this stage.
Given past evidence for dynamic and multidirectional (threat avoidance vs over-attending) attentional biases in people exposed to threat, researchers should consider investigating interpretational and attentional biases in a combined paradigm.Longitudinal studies incorporating techniques used in the latter, such as eye tracking (Grossheinrich et al., 2022), may help us to better understand the mechanisms underpinning emotion recognition misattributions, clarify how these processes interact with threat exposure, and explore how these interactions evolve over time.

Conclusion
This study suggests that threatening experiences in childhood may affect key corticolimbic brain regions involved in emotion processing, but not social cognitive processes indexing a bias to threat-related facial expressions.Future studies should investigate whether other factors interact with threat exposure-related variation of these brain regions to promote risk or resilience to psychosis.Longitudinal cohorts stratified by subsequent psychopathological outcomes may improve understanding of alterations related to threat exposure, observed in youth with early psychotic symptoms.

Funding sources
VLC was supported by National Health and Medical Research Council (NHMRC) Investigator Grant No. 1177370 and a Dame Kate Campbell Fellowship from the University of Melbourne.TEVR was supported by a Dame Kate Campbell Fellowship from the University of Melbourne.YET was supported by the Mary Lugton Postdoctoral Fellowship.AZ was supported by a research fellowship from the NHMRC (APP1118153).The funding sources were not involved in the study design, analysis, or interpretation of data.

CRediT authorship contribution statement
Megan Thomas: Conceptualization, Methodology, Formal analysis,

Fig. 1 .
Fig. 1.Mean threat exposure by group.CTRL = controls, PSS = psychosis spectrum symptoms group, INT = internalizing group.Error bars show standard deviation; both PSS and INT had significantly higher threat exposure than CTRL.

Fig. 2 .
Fig. 2. Association between threat exposure and left medial prefrontal cortex volume.Data points are adjusted for covariates included in the regression model (age, sex, total intracranial volume, maternal education, group).

Table 1
Demographics of sample.

Table 2
Descriptive statistics for brain and cognitive measures.

Table 3
Multiple regression models examining the effect of threat exposure on fear/anger bias and brain measures.= standardized regression coefficient, ICV = intracranial volume, INT = internalizing group, ME = maternal education, mPFC = medial prefrontal cortex, PSS = psychosis spectrum symptom group, SE = standard error.Covariates = age, sex, maternal education, ICV (for brain measures), group split into dummy variables for PSS and INT (reference group = controls).