Environmental Exposures and Computational Mechanisms in Psychosis Risk: Insights from a Multi-Site Cohort

Abstract

This study leverages computational psychiatry to examine weighting of perceptual priors and reinforcement learning (RL) as potential pathways connecting environmental stressors to psychosis risk. 506 participants with psychosis-like experiences (PLE) and 328 controls performed behavioral tasks measuring these mechanisms, in addition to symptom assessments. Participants’ childhood area social fragmentation and childhood opportunities were also extracted. The association between environmental, computational and clinical measures was assessed, as were mediation pathways. Higher social fragmentation was associated with deficits in RL (r(410)=0.19, p<.001) and lower opportunities were associated with overweighting perceptual priors (r(290)= −0.16, p=.02) as well as issues in RL (r(307)=-0.14, p=.02) for PLE participants. Less available childhood opportunities were also related to more severe positive symptoms (r(354)=-0.14, p=.03). We found a potential pathway connecting available childhood opportunities to positive symptomatology through the mediating effect of overweighting perceptual priors (ACME=-4.91, p=.04). Environmental characteristics could shape symptom severity through computational mechanisms.

1. Introduction

Harsher environmental conditions increase the odds of developing psychotic disorders and symptoms (Beyer et al., 2024; Binbay et al., 2024; Karcher et al., 2021; Ku et al., 2020; Varese et al., 2012; Wahbeh & Avramopoulos, 2021). Several models incorporate the idea that stressors interact with individuals’ predisposition to psychotic disorders (Pruessner et al., 2017; T. Vargas et al., 2020; Walker & Diforio, 1997). The present study describes computational pathways connecting environmental characteristics to clinical presentation in a sample of emerging adults with psychosis-like experiences (PLE).

1.1. Area-Level Characteristics

Social fragmentation (Allardyce et al., 2005; Ku et al., 2021, 2024; Veling et al., 2015) and social and environmental deprivation (Allardyce & Boydell, 2006; Beyer et al., 2024; Bhavsar et al., 2014; Ku et al., 2023; O’Donoghue et al., 2015; Sariaslan et al., 2015) during childhood have been consistently associated to psychosis risk. Social fragmentation is a construct that measures the lack of stable social relationships among people living in the area (Congdon, 1996). Social instability could promote behaviors and cognitions, such as hypervigilance, that interact with psychosis liability (Howes & Murray, 2014; Murray, 2011). Environmental deprivation has been related to psychosis risk through a variety of pathways, including exposure to pollutants, higher stress due to material instability or greater exposure to crime (Allardyce et al., 2005; Allardyce & Boydell, 2006; Bhavsar et al., 2014; Czepielewski et al., 2022; Fond et al., 2023; Hastings et al., 2020; Kirkbride et al., 2016; Ku et al., 2020, 2023; O’Donoghue et al., 2015; Sariaslan et al., 2015). Despite these well-documented associations, the mechanisms connecting detrimental environmental characteristics to psychosis remain poorly understood.

1.2. Computational Mechanisms

Computational mechanisms refer to the algorithmic processes through which the brain transforms experience into perception, learning, and decision-making, and which can be formally modeled via behavioral tasks. The Computerized Assessment of Psychosis Risk (CAPR) project uses this framework to identify young adults and teens at risk for psychosis through a task battery designed to reflect the neurobiological and cognitive mechanisms underlying symptoms (Mittal et al., 2021). Results from this group have shown that perceptual propensity towards social constructs in ambiguous stimuli and anomalies in reinforcement learning have been preferentially associated with psychosis risk and severity (Pokorny et al., 2025, 2026; Tran et al., 2025; Williams et al., 2025). In addition to their clinical relevance, these mechanisms are sensitive to environmental influence and could modulate or compound detrimental environmental influences.

1.2.1. Overweighting of Perceptual Priors

The predictive coding framework suggests that people’s brains and minds are continually inferring the nature of sensory input; contrasting existing perceptual priors with incoming sensory experience (Griffin & Fletcher, 2017). Assessing the relevance of expectations and incoming sensory input is sensitive to the nature of the perceptual information, is sensitive to experience and will vary at the individual level. Overweighting of perceptual priors, the disproportionate reliance on prior expectations over incoming sensory evidence, is thought to underlie perceptual abnormalities across the psychosis spectrum (Adams et al., 2013; Corlett et al., 2019; Frith & Friston, 2013). For instance, an overreliance on sensory priors related to faces or speech may contribute to the increased tendency to see faces (Silverstein et al., 2021) and hear voices (Alderson-Day et al., 2017, 2022; Daalman et al., 2012) in ambiguous stimuli and lead to heightened hypervigilance and paranoia in addition to perceptual abnormalities. In turn, heightened perceptions of threat could also relate to limited activities and present as secondary negative symptoms. These perceptual processes could amplify or dampen contextual influences as well be shaped by environmental conditions.

1.2.2. Anomalies in Reinforcement Learning

Translating information regarding rewards and punishments into effective decision-making is thought to be impaired in individuals across the psychosis spectrum (Strauss et al., 2014). Issues in the representation of current and potential rewards, as well as failing to adequately update these representations in response to feedback is thought to underlie positive (Reed et al., 2020) and negative symptoms (Gold et al., 2008; Waltz et al., 2013). These challenges are even more pronounced when rewards are delayed in time (Heerey et al., 2007; Horan et al., 2017) or when reward values need to be updated in response to changing conditions (Ceaser et al., 2008; Waltz et al., 2013; Waltz & Gold, 2007). Reward processing is sensitive to deprivation in the physical environment and to social instability and isolation, which can interfere with effective regulation (Tost et al., 2015). Similarly to the overweighting of perceptual priors, individual differences in reinforcement learning could mediate the influence of the environment as well as shift in response to repeated contextual experience.

1.3. Potential Pathways

Overlapping physiological and cognitive mechanisms suggest that biological and psychological individual vulnerabilities may interact with the effects of adverse environments compounding psychosis risk (DeRosse & Barber, 2021).

Deprivation and stress affect processes underlying perception, including more basic processes, like sensory gating, and higher order processes (August et al., 2012; Bakshi et al., 2012; Hetrick et al., 2012; Li et al., 2022). These alterations can increase psychosis proneness through their interaction higher order processes (Daalman et al., 2012; Vercammen & Aleman, 2010). The neuromodulatory systems engaged in prediction coding are also susceptible to the effect of adversity (Linson et al., 2020). Recent conceptualizations propose that early adversity could fundamentally alter predictive coding processes, including through the overweighting of threat-relevant priors -- namely leading to interpret ambitious stimuli as threatening (Herzog et al., 2022; Krupnik, 2020; Lyndon & Corlett, 2020; Putica & Agathos, 2024). More socially fragmented environments, with their heightened social instability, and more deprived areas, which are more likely to have exposure to threat and material instability, could magnify the weighting of perceptual priors potentially related to threat.

Reinforcement learning patterns are also affected by environmental contingencies. For example, people with higher childhood stress exposures can show apparently maladaptive behavioral and neural patterns in tasks assessing the anticipation and response to rewards and losses which can align with the patterns associated with psychosis risk (Birn et al., 2017). Likewise, early social instability and scarcity can influence implicit and explicit expectations about the environment, shaping people’s responses to rewards and losses towards assumptions that settings will be more volatile, reward contingencies are likely to shift, and that rewards are comparatively less likely than losses (Shafir, 2017; Sheehy-Skeffington, 2020; Wang & Hanson, 2024). These patterns also align with the computational processing of rewards and losses for learning associated with psychosis proneness.

Early environmental adversity can affect the biological and cognitive underpinnings of perceptual prior weighting and reinforcement learning. These mechanisms are also believed to relate to psychosis risk. This shared vulnerability makes them potential bridges between the effects of environmental and illness outcomes.

1.4. Goals and Aims

The present study aims to investigate associations between childhood environmental exposures and symptom presentation in emerging adults experiencing PLE. We focus on the overweighting of perceptual priors and anomalies in reinforcement learning at emerging adulthood as potential mediating mechanisms given their clinical relevance for psychosis symptoms and their susceptibility to the influence of established environmental risk factor for psychosis – social fragmentation and deprivation. Our first aim is determining the association between early environmental exposures and task performance and clinical symptoms during emerging adulthood. We hypothesize that PLE participants from more vulnerable environments will overweight perceptual priors and adopt strategies reflecting aberrant learning from rewards. We also expect that PLE participants from more vulnerable areas will report greater symptom severity. Secondly, we aim to test whether our target mechanisms mediate associations between environmental exposures and clinical outcomes. We hypothesize that the association between detrimental environmental characteristics and clinical outcomes will be partially mediated by task performance.

2. Transparency and Openness

2.1. Preregistration

This study was not preregistered.

2.2. Data, materials, code, and online resources

Given the clinical and sensitive nature of the study sample and collected data we cannot make the data publicly available. De-identified group data is available upon request, please contact the corresponding author (Luz Maria Alliende).

2.3. Reporting

We report all data exclusions, all manipulations, and all measures in the study. We did not perform a prior power analysis. This study is a secondary analysis for available CAPR consortia data, all participants with the available data were included in the respective analyses.

2.4. Ethical approval

Northwestern University’s Institutional Review Board approved all protocols and procedures (Protocol: STU00211351), and they were acknowledged by the Institutional Review Boards of all participating sites. All participants underwent a consent process with a research coordinator or assessor that thoroughly explained the study’s aims, procedures, risks, limits and measures to achieve confidentiality, as well as their rights as a research participant. This study was carried out in accordance with the provisions of the 2013^th Declaration of Helsinki.

3. Materials and Methods

3.1. Participants

Participants were recruited between March 2020 and August 2024 as part of the CAPR study, a United States-based multi-site consortium that aims to develop a computerized battery for the assessment of psychosis risk. Participants were recruited through mass advertisement as well as through referrals from clinical sources and each laboratories’ existing recruiting pools. While CAPR is a longitudinal study, only data from participants baseline visit was used for this study. Further details about this study’s implementation can be found in the published protocol (Mittal et al., 2021).

Participants’ clinical group was determined using the Structured Interview for Psychosis-Risk Syndromes (SIPS) (Miller et al., 2003) and the Structured Clinical Interview for DSM-5 (SCID) (First et al., 2016). Participants who met criteria for a psychosis-risk syndrome according to the SIPS (i.e. Brief Intermittent Psychosis Syndrome, Attenuated Positive Symptom State and/or Genetic Risk and Deterioration) were included in the PLE group. Additionally, the PLE group also included participants experiencing mild positive symptoms as measured by the SIPS positive symptom subscale (score = 2), or experiencing symptoms at the severity associated with clinical high risk (≥ 3) but at either a decreased frequency or duration or better explained by a different disorder (see Supplemental Methods 1 for details). Participants with no subthreshold positive symptoms according to the SIPS, were included as part of the Control group regardless of past or current non-psychosis spectrum diagnoses. Participants with past or current psychotic disorder diagnoses, as determined by the SIPS or SCID, were excluded from the study.

A total of 506 PLE participants were included. The PLE and Control groups were comparable in age, sex, ethnicity and premorbid IQ; but had a significantly different racial composition. The PLE group had more White and Black participants and the Control group had more Asian participants. Expectedly, PLE participants had significantly more severe positive and negative symptoms. Likewise, the PLE group showed significantly higher weighting of perceptual priors and anomalies in reinforcement learning in behavioral tasks than Controls. PLE and Control participants grew-up in neighborhoods comparable in terms of social fragmentation; but PLE participants came from neighborhoods with less available opportunities for children. Table 1 describes the PLE and Control groups; see Supplementary Results 1 for a comparison between the subsample of CHR participants in the PLE group and controls.

Table 1. Demographic and Clinical Sample Characteristics.

This table describes the demographic, cognitive and clinical characteristics of study participants, as well as their zip-codes’ measures of social fragmentation and available childhood opportunities, and their aggregate performance in tasks measuring weighting of perceptual priors and reinforcement learning.

	PLEs N = 506	Controls N = 328	Statistic
Demographics
Age	23.12 (4.01) N = 506	23.49 (3.95) N = 327	t (8313) = −1.32, p = .19
% Female (N)	62.85% (318) N = 506	62.80% (206) N = 328	χ2 (1) = 0.00, p = .99
% Race (N)	White: 56.52% (286) Black: 16.60% (84) Asian: 14.43% (73) Indigenous: 1.19% (6) Other: 11.26% (57) N = 506	White:50.61% (166) Black: 12.80% (42) Asian:26.83% (88) Indigenous: 0.91% (3) Other:8.84% (29) N = 328	Fisher’s exact test, p < .001
% Hispanic (N)	16.40% (83) N = 506	12.80% (42) N = 328	χ2 (1) = 1.75, p = .186
WRAT	112.06 (13.87) N = 416	112.54 (13.70) N = 283	t (697) = −0.46 p = .647
Clinical
SIPS-P	9.18 (4.14) N = 505	1.66 (1.78) N = 319	t (822) = 30.70, p < .001
NSI-PR	13.38 (8.57) N = 506	4.80 (7.20) N = 328	t (832) = 15.03, p < .001
Area Characteristics
Social Fragmentationa	0.02 (0.82) N = 465	−0.07 (0.76) N = 316	t (779) = 1.46, p = .14)
Childhood Opportunitiesa	0.01 (0.03) N = 356	0.01 (0.03) N = 208	t (562) = −2.33, p = .02)
Computational Measures
Weighting of Perceptual Priorsa	0.10 (0.89) N = 412	−0.14 (0.76) N = 280	t (690) = 3.65, p < .001
Anomalies in Reinforcement Learninga	0.05 (0.64) N = 440	−0.06 (0.65) N = 296	t (609) = −2.29, p = .022

Abbreviations: WRAT, Wide Range Achievement Test-4 Reading subtest; SIPS-P, Structured Interview for Psychosis-Risk Syndromes, Positive Scale; NSI-PR, Negative Symptom Inventory-Psychosis Risk. Mean and standard deviation are reported unless otherwise specified.

^aData was scaled across all participants.

3.2. Measures

3.2.1. Area Characteristics

During their baseline visit participants were asked to retrospectively recall the zip code they grew up in. For reference, the average zip code in our sample corresponded to an area with an average of 34490.81 inhabitants (sd=19465.66). 78.46% of potential participants reported growing up in a valid United States zip code. There were no significant differences in the proportion of PLE or Control participants reporting a valid US zip code (χ²(1)=0.16, p=.693). More so, PLE participants with missing zip code data did not significantly differ in positive or negative symptoms. However, PLE participants with missing zip code data had significantly lower premorbid IQ (t (158.31)=-3.24, p=.001), were less likely to be White and not Hispanic (χ²(1)=8.59, p=.003), and were slightly but significantly older (t (171.55)=2.96, p=.004). See the Supplementary Methods 2 for a full description of participants with missing zip code data. Participants in the PLE group reported growing up in 441 different zip codes across the U.S. 90.93% of these zip codes were unique to a single participant, 7.71% were shared by two participants, and with the remaining 1.36% were shared by three or four participants.

3.2.1.1. Social Fragmentation:

social fragmentation has been operationalized as the transiency and non-nuclear families in the community (Allardyce & Boydell, 2006; Ku et al., 2021). To measure social fragmentation, we extracted the percent of non-owner residents, percent of single-parent households, and percent of residents in participants’ zip code living in a different house than they were a year ago from the 2011 American Community Survey as a reflection of this transiency. We chose to use data from the 2011 survey as it was most likely to coincide with participants’ middle childhood and reflect the state of their neighborhood growing up; this year was also close to the best available measure of environmental deprivation. We normalized each percent score across the full sample and then constructed a composite measure of social fragmentation using the average z-scores across measures.

3.2.1.2. Childhood Opportunities as a proxy for Deprivation:

We used the Childhood Opportunity Index as a proxy for participants’ childhood areas’ material and social deprivation. This index measures features in neighborhoods across the United States established to aid children’s healthy development (Noelke et al., 2020). We used the z-score nationally normed composite index as a proxy for available opportunities for children. This index includes measures of education, health and environment, and social and economic opportunities. All index indicators and domains are briefly described in the Supplementary Methods 3 and thoroughly described in the dataset’s technical documentation (Noelke et al., 2020). We used the 2010 COI 2.0 dataset (diversitydatakids.org, 2023) as it was closest to participant’s middle childhood. For this dataset we cross-walked census tract level information to participants’ corresponding zip codes.

3.2.2. Tasks

All tasks were completed by participants during their baseline assessments.

3.2.2.1. Weighting of Perceptual Priors

To determine participants’ weighing of perceptual priors, potentially related to threat, we constructed a composite index based on tasks that determined the likelihood participants had to interpret ambiguous stimuli as human voices and faces respectively. This was done using the following two tasks:

Filtered Speech Task (Alderson-Day et al., 2017):

participants are tasked with determining whether sound clips contain intelligible speech. First, participants are exposed to, often unintelligible to most people, filtered recordings of short phrases. Then the participant listens to the unfiltered, clear version of the audio. Lastly, they re-listen to the distorted versions. The variable of interest for this task is the number of audio clips identified as speech across all trials - distorted or clear.

Mooney Faces (Mooney, 1957):

participants are presented with degraded pictures of faces and tasked with determining whether a face is present in the image or not. Faces are presented both upright and inverted, with the latter condition being more ambiguous. While this task has been associated with alterations in perceptual organization in people with schizophrenia (Uhlhaas & Silverstein, 2005), it is believed that in clinical high risk populations it could operate as a measure of overweighting perceptual priors for faces (Mittal et al., 2021; Silverstein et al., 2021).

3.2.2.2. Reinforcement Learning

This variable consisted of the average z score for measures from the following three tasks:

Delay Discounting (Kirby et al., 1999):

in this task participants are asked to choose between a smaller reward closer in time and a larger reward later. The magnitude of the rewards, the difference between rewards and the delay changes across trials allowing us to model participants’ rate of discounting. The parameter of interest for this task is the rate of discounting (K) for large rewards.

Gain vs Loss Avoidance (Gold et al., 2012):

participants are faced with pairs of pictures with different associated probabilities of, in some trials, providing a reward or, in others, leading to the loss of previously earned rewards. Gain and loss trials both have one picture that gives the desired outcome 80% and one where this happens 90% of the time. The measure of interest is the difference between the proportion of correct choices in gain trials and the proportion of correct choices on loss avoidance trials. This measure was reverse coded so higher scores related more relative learning from losses than gains.

Probabilistic Reversal Learning Task (Mittal et al., 2021):

the task follows a three-armed bandit paradigm in which participants aim to maximize their final score by selecting from three digital decks of cards. The decks have different probabilities of yielding rewards. However, these probabilities change over time. Participants must adapt to these shifts by updating their choices based on past outcomes. The key measure of interest is the win-switch rate - the number of times a participant switches decks after receiving a reward on the previous trial.

3.2.3. Symptoms and Cognition

Attenuated positive symptoms were assessed using the Structured Interview for Prodromal Symptoms (SIPS) (Miller et al., 2003). Negative symptoms were assessed using the Negative Symptom Inventory-Psychosis Risk (NSI-PR) (Strauss et al., 2024). Premorbid IQ was determined using the WRAT-4 (Wilkinson & Robertson, 2006).

3.3. Analyses

The relationship between environmental variables and computational or clinical measures was calculated using pairwise Pearson correlations. Analyses were performed on the full sample as well as the subset of PLE participants. Given the reduced sample size of the PLE group, results might be underpowered and need to be considered as exploratory in nature.

To test mediation effects, we conducted a series of mediation analyses using R’s mediation package (Tingley et al., 2025). All mediation models included PLE participants’ sex, age and premorbid IQ as covariates. For each variable combination we created subsets of complete cases and estimated two linear regression models. A first model predicts the computational outcome based on neighborhood characteristics and all included covariates. Then, a second model predicts symptoms including the neighborhood exposure variable and computational index. Lastly, we included these models in the mediate function in R’s mediation package to test for mediated effects of the neighborhood index on the symptom variable through the computational mechanism. All analyses were performed using 1,000 nonparametric bootstrap simulations and robust standard errors. Results for our metric of interest, average causal mediation effect, were corrected for multiple comparisons (2 environmental variables x 2 computational mechanisms x 2 clinical outcomes) using the Benjamini and Hochberg method, using the p.adjust function in R (Benjamini & Hochberg, 1995). As for prior analyzes, the reduced size of the PLE subsample makes these results more likely to be underpowered and they need to be considered as exploratory.

4. Results

4.1. Environmental Exposures and Computational Mechanisms

We found that some, but not all, of our area measures were significantly associated with performance in computational tasks. Contrary to our hypothesis we did not find a significant correlation between social fragmentation and participants’ weighing of perceptual priors when looking at our full sample (r (664) = 0.00, p = 0.98) or subset of PLE participants (r (385) = 0.01, p = .89). However, we did find a significant association between social fragmentation and our reinforcement learning metric in the full sample (r (704) = 0.17, p < .001) and PLE subsample (r (410)= 0.19, p < .001), with participants from more fragmented areas showing more aberrant performance in reinforcement learning tasks. Secondary analyses on the relationship between social fragmentation and computational mechanisms exclusively on CHR participants found the same pattern of results (see Supplementary Results 2).

Also in line with our hypotheses, participants from less deprived areas, operationalized as greater availability of childhood opportunities, had significantly lower weighing of perceptual priors (full sample: r (488) = −0.10, p = .04; PLE: r (290) = - 0.16, p = .02) and less reinforcement learning abnormalities in both our full sample (r (515) = −0.13) and PLE group (r (307) = −0.14, p = .02). See Figure 1 for a visualization of these results. These findings were replicated in secondary analyzes looking at the CHR subset of participants (see Supplementary Results 2).

Figure 1. Associations Between Environmental Exposures and Computational Task Performance.

This figure represents the association between our chosen environmental variables: social fragmentation and deprivation, operationalized as an areas’ available opportunities for children, and computational variables: weighting of perceptual priors and reinforcement learning aberrancies. For the purposes of visualization all environmental and task values have been scaled and the childhood opportunity index has been reversed for ease of comparison across metrics. Controls are represented in orange and the PLE group is represented in blue. Asterisks indicate whether a correlation was significant for the full sample.

4.2. Environmental Exposures and Clinical Outcomes

In line with our hypotheses, higher social fragmentation significantly correlated with higher positive (r (769) = 0.08, p = .03) and negative symptoms (r (783) = 0.09, p < .001) in the full study sample. However, the association between social fragmentation and symptoms was not significant for the PLE subsample (positive symptoms: r (462) = 0.05, p = .35; negative symptoms: r (463) = 0.07, p = .24). Secondary analyses showed that this association was not present among CHR participants (see Supplementary Results 3).

As expected, we found a significant association between participants’ childhood area deprivation and positive symptoms. Higher available childhood opportunities were associated to lower positive symptoms for the full sample (r (555) = −.16, p < .001) and PLE subsample (r (354) = −0.14, p = .03). This association was not present for negative symptoms in either the full sample (r (603) = −0.06, p =.13) or PLE participants (r (354) = −0.04, p = .44). See Figure 2 for a visualization of these results. The significant association between area deprivation and positive, but not negative, symptoms was replicated in secondary analyses in the CHR sub-set of participants (see Supplementary Results 3).

Figure 2. Associations Between Environmental Exposures and Symptomatology.

This figure represents the association between our chosen environmental variables: social fragmentation and area deprivation, operationalized as available childhood opportunities, and clinical variables: positive and negative symptoms. For the purposes of visualization all environmental and clinical measures have been scaled and the childhood opportunity index has been reversed for ease of comparison across metrics. Controls are represented in orange and the PLE group is represented in blue. Asterisks indicate whether a correlation was significant for the full sample.

4.3. Mediation Effects

For our mediation analyses our metric of interest was the average causal mediation effect (ACME) for PLE participants. We found one significant mediation pathway but did not find significant mediation effects in most models.

We did not find a significant mediation effect between social fragmentation and weighing of perceptual priors on either positive (ACME = 0.02, CI 95% = [-0.07 – 0.12], N = 329, p = .87) or negative (ACME = - 0.02, CI 95% = [-0.12 – 0.06], N = 329, p = .87) symptoms. This was also the case in secondary analyses including only the subset of CHR participants (see Supplementary Results 4).

In line with our hypothesis, we found a significant mediation effect between participants’ childhood area deprivation and positive symptoms through the weighting of perceptual priors (ACME = −4.91, CI 95% = [-10.18 – 1.07], N = 247, p = .04). In this model there was a significant total effect of deprivation, operationalized as available childhood opportunities, on positive symptoms (-24.03, CI 95% = [-44.18 - - 2.15], p = .04). Participants from less deprived areas had lower positive symptoms overall. Notably, the average direct effect of deprivation on positive symptoms was only present at the trend level (-19.12, CI 95% = [-38.81 – 3.10], p = .09) with a significant proportion of this effect coming from mediation (proportion mediated = 0.19, CI 95% = [0.01– 1.04], p = .04). These results were replicated in the subset of CHR participants (see Supplementary Results 4) but did not remain statistically significant after correcting for multiple comparisons. The mediation effect was not significant in the case of negative symptoms among PLE participants (ACME = 3.38, CI 95% = [-3.88 – 11.85], N = 247, p = .81) or secondary CHR analysis (see Supplementary Results 4).

We did not find a significant mediation effect between social fragmentation and positive (ACME = 0.01, CI 95% = [-0.08 – 0.10], N = 351, p = .87) or negative symptoms (ACME = - 0.09, CI 95% = [-0.35 – 0.07], N = 351, p = .81) through reinforcement learning. Nor did we find a significant mediation effect between area deprivation and positive (ACME = 0.18, CI 95% = [-1.90 – 2.76], N = 261, p = .87) or negative (ACME = 1.92, CI 95% = [-2.20 – 8.43], N = 261, p = .81) symptoms through reinforcement learning. We did not find significant moderation effects between any of the environmental variables and any of the symptom variables through reinforcement learning in secondary analyses on the CHR subgroup (see Supplementary Results 4).

5. Discussion

Establishing mechanisms connecting environmental conditions and individuals’ psychosis proneness has remained elusive. To our knowledge, this is the first study to use tools from georeferencing and computational psychiatry to shed light on this relationship.

The first relationships highlighted by our findings are associations between some of our targeted childhood environmental exposures and computational mechanisms. We mostly found that participants from more vulnerable environmental had tasks outcomes more akin to those found in people with psychosis spectrum disorders; except for social fragmentation and the overweighting of perceptual priors, where we found no significant relation. These associations could reflect the effect of early environmental exposures on information processing mechanisms. One possible explanation is that perceptual processing and reinforcement learning are adaptatively biased to reflect early environmental conditions. For example, people in unstable environments where threats are more likely, can show a perceptual predisposition towards threats. Within this context, this tendency to perceive threats could lead to potentially functional or even protective increase in vigilance (Ellett et al., 2008; Freeman et al., 2002; Garety et al., 2001; Johns et al., 2004). Likewise, reward-based learning that can be considered suboptimal in theory, could be environmentally adaptive in adverse environments. For example, in psychosis studies, opting for immediate rewards over greater delayed rewards can be typically interpreted as an issue in reward representation over time. However, this strategy can be functional in unstable environments where consummating delayed rewards is unlikely or when immediate needs are very high (Shafir, 2017; Sheehy-Skeffington, 2020; Wang & Hanson, 2024). Conversely, certain environmental conditions may worsen pre-existing tendencies in perceptual processing and reinforcement learning abnormalities in individuals with an underlying psychosis vulnerability, compounding their clinical risk (Birn et al., 2017; Sterzer et al., 2018; Strauss et al., 2014). Lastly, these associations could also be responding to a third variable. Some specificity is found in our study, where we found no relationship between participants’ weighing of perceptual priors and social fragmentation. Perhaps, social fragmentation in our sample did not relate as strongly to the environmental contigencies affecting perceptual processing. It might be the case that in our sample social instability related less strongly with increased crime or threat than our deprivation index, which could in turn lead to heightened perception for potential sources of social danger. Future longitudinal studies can elucidate the nature and timing of the associations between more specific environmental conditions, as well as determine under what conditions these associations are adaptive or conducive to psychosis risk.

Our study also showcased a potential direct association between some environmental exposures and psychosis-spectrum symptoms. We found that higher social fragmentation and deprivation significantly correlated with increased positive symptoms in the full sample. This is in line with previous literature connecting harsher environments to increased positive symptomatology (Howes & Murray, 2014; Varese et al., 2012). Higher social fragmentation, but not deprivation, was significantly associated with increased negative symptomatology. The relationship between negative symptoms and social fragmentation, but not neighborhood deprivation is also found in a study on the Avon Longitudinal Study of Parents and Children (Solmi et al., 2020). However, other groups have found increased negative symptoms in PLE or CHR samples from more materially deprived neighborhoods (Strauss, 2021, 2024). Notably, some of our significant findings were present only for the full sample, but not for the subset of PLE or CHR participants. While these findings could suggest differential effects of environmental stressors by underlying psychosis risk, they could also reflect constraints in sample power or symptom range for the PLE and CHR subsamples. Future large studies with high symptom variability across the psychosis spectrum are needed to ascertain the nature of these findings.

Finally, our findings hint at a potential computational mechanism mediating the relationship between environmental factors and symptomatology. The association between higher area deprivation, operationalized as lower available opportunities for children, and higher positive symptoms was partially explained by PLE participants’ increased proclivity to report faces or voices in ambiguous stimuli. These findings are consistent with the hypothesis that environmental and computational risk factors may interact to confer increased psychosis risk; however, the cross-sectional nature of our study precludes causal interpretation. Notably, this mediation effect was not replicated in the CHR subgroup, nor did we find significant mediation effects between our remaining variables of interest. Negative findings could be due to our sample being underpowered to capture small existing mediation effects, conversely, they could reflect specificity in the role of perceptual priors for faces and voices in interfacing the relationship between environmental characteristics and symptoms.

Our study findings need to be interpreted in the context of some limitations. First, our metric of childhood area characteristics relies on self-report, which makes this information vulnerable to misclassifications. Further, our data does not include when or for how long participants lived in this zip code. While this time probably aligned with middle childhood for participants, future studies could characterize these trajectories and determine the timing and “dose” effects of exposures (Gomes et al., 2019; Howes & Murray, 2014; T. Vargas et al., 2020; T. G. Vargas & Mittal, 2022). Additionally, while we collected retrospective data on participants’ childhood area characteristics, our study’s cross-sectional design limits causal or directional interpretations of our results; future longitudinal studies are needed to shed light on the nature of our findings. Since this study partly took place during the COVID-19 pandemic, participant’s task performance and symptomatology could have been influenced by the personal effects of the pandemic at the time of the study. Another limitation in our study is the clinical variability within our PLE sample. We chose to include PLE participants, rather than participants that met formal criteria for CHR, in an attempt to balance heightened psychosis risk and attaining a sufficiently powered sample size. Notably, our secondary analyses on the subset of CHR participants mirrored all results in the PLE sample, except for the multiple comparison corrected mediation effect of perceptual priors on the effect of childhood opportunities on positive symptoms. This could be due to the reduced sample size and respective power, in addition to potential restrictions of range for positive symptoms since the CHR syndrome necessitates that symptoms fall within a narrower range.

Our study’s findings showcase how environments can interact with individuals’ cognitive and computational vulnerability for psychosis and relate to differences in clinical outcomes. Notably, our findings suggest a potential pathophysiological mechanism through the overweighting of perceptual priors modulating the relationship between diminished opportunities in the environment and positive symptoms. Our findings pave the way for a potentially fruitful line of research: characterizing the computational effects of exposure to environmental stressors. More so, future studies, incorporating neural and physiological measures can further describe the biological mechanisms through which environmental conditions modulate the cognitive mechanisms connected with psychosis risk.