Exploring the Interactions Between Psychotic Symptoms, Cognition, and Environmental Risk Factors: A Bayesian Analysis of Networks

Abstract

Background and Hypothesis

Psychotic disorders (PDs) have huge personal and societal impact, and efforts to improve outcomes in patients are continuously needed. Environmental risk factors (ERFs), especially modifiable risk factors, are important to study because they pose a target for intervention and prevention. No studies have investigated ERFs, cognition, and psychotic symptoms together in a network approach.

Study Design

We explored interactions between 3 important ERFs (tobacco smoking, cannabis use, and childhood trauma), 6 cognitive domains, and 3 dimensions of symptoms in psychosis. From the Genetic Risk and Outcome of Psychosis (GROUP) cohort, we used data from patients, siblings, and healthy controls to construct networks using Bayesian analyses of all 12 variables. We constructed networks of the combined sample and of patients and siblings separately.

Study Results

We found that tobacco smoking was directly associated with cognition and psychotic symptoms. The cognitive variable processing speed was the most central node, connecting clusters of psychotic symptoms and substance use through the variables of positive symptoms and tobacco smoking. Comparing the networks of patients and siblings, we found that networks were relatively similar between patients and siblings.

Conclusions

Our results support a potential central role of processing speed deficits in PDs. Findings highlight the importance of integrating tobacco smoking as potential ERFs in the context of PDs and to broaden the perspective from cannabis discontinuation to smoking cessation programs in patients or people at risk of PDs.

Introduction

Psychotic disorders (PDs) have a huge personal and societal impact.^1–3 They can consist of a combination of symptoms, including positive (eg, hallucinations or delusions) and negative symptoms (eg, blunted affect). Additionally, people with PDs often experience more general symptoms, like affective disturbances.⁴ Interventions that improve the life of patients with PDs are essential, and efforts to improve treatment and outcomes in patients are clearly needed.

One relevant line of research is to study environmental risk factors (ERFs) for the onset and course of PDs across the psychosis liability spectrum. For PDs, 2 of the most replicated ERFs are childhood trauma and cannabis use.^5,6 They have been found to be associated with the onset of PDs as well as with the severity of positive symptoms,⁷ negative symptoms,^8,9 and poorer (social) functioning.^6,10,11 An ERF that received less scientific attention is tobacco smoking.¹² However, recent studies have shown that tobacco smoking is associated with the onset of PDs and with the severity of positive symptoms and negative symptoms.^12–14

How ERFs exactly relate to PDs is unknown. One theory is that ERFs have a detrimental effect on neurodevelopment,^15,16 which leads to cognitive deficits^17–19 that contribute to the emergence and maintenance of psychotic symptoms.^20,21 Despite these assumptions, most studies only investigated associations between a small selection of factors,^{17,18,22–25} and have not accounted for potential interactions between multiple ERFs, cognitive domains, and different symptom domains.

We use the network approach, in which variables (like risk factors or symptoms) are represented as nodes, and associations between variables are represented as edges^26,27 to model the interplay between these multiple factors. Because the network approach allows us to look at different associations simultaneously, it provides useful information regarding potentially meaningful interactions between variables. Furthermore, network analyses allow us to identify the most central and densely connected variables.²⁸ This method thereby provides information beyond investigating multiple associations separately and is thus especially well-suited for our exploratory research.²⁹

To date, few studies on ERFs and PDs have taken a network approach. However, some studies have shown the usefulness of the network method, eg, by showing that cumulative trauma is an important bridge node between symptoms of depression and symptoms of psychosis,³⁰ by showing that childhood trauma is associated with psychotic symptoms only through general psychopathology³¹ and by elucidating the importance of ERFs for multiple aspects of population mental health.³² Others have focused on the associations between cognitive and psychopathological variables.^33–36 While the variables within the cognition and psychopathology clusters were densely connected, the connections between the clusters were sparse and specific. The edges that connected the clusters, called bridge edges, differed across studies, likely due to the inclusion of different variables.

To date, no network study has directly examined the interactions between ERFs, cognition, and symptoms along the continuum of psychosis liability, including patients, siblings, and healthy controls. Therefore, we aimed to investigate the interactions between multiple ERFs (cannabis use, smoking, and childhood trauma), cognitive variables, and positive, negative, and depressive symptoms using advanced Bayesian methodology. The Bayesian approach improves upon common methodology³⁷ as it provides a principled way to quantify the uncertainty of one’s network, such as evidence for edge presence or absence.³⁸ First, we explored the interactions between variables in a combined sample of patients, siblings, and healthy controls. Specifically, we focused on identifying the bridge edges that connect the ERFs with different symptoms of psychosis. Secondly, we investigated these associations along the continuum of psychosis liability by evaluating the interactions in different subsamples of patients and siblings. Although siblings are not affected by illness-related factors, they share a (genetic) risk with their affected relatives. Therefore, similar findings in patients and siblings would suggest associations independent of illness-related confounding effects, such as antipsychotic medication and side effects. The findings of this exploratory study might provide hypotheses for further research into the relationship between risk factors and PDs.

Methods

Subjects

This study was performed on data from the Genetic Risk and Outcome of Psychosis (GROUP) cohort, which consists of 1119 patients diagnosed with PD, 1059 unaffected siblings, and 586 healthy controls. The study protocol was approved centrally by the Ethical Review Board of the University Medical Center Utrecht and also locally by all the participating institutes. Patients and siblings in the cohort were recruited from 4 university psychiatry departments in the Netherlands (Amsterdam, Groningen, Maastricht, and Utrecht) and their associated mental healthcare institutions. Healthy controls were recruited by random mailings to addresses in the vicinity of recruited patients. We analyzed data from the first follow-up assessment (with the exception of the CTQ, which was administered at different assessment times across departments), 3 years after enrollment because this is the only timepoint where all variables of interest were measured. Inclusion criteria for all participants were an age range of 16-50 years, good command of the Dutch language, and ability and willingness to give written informed consent. An additional inclusion criterion for patients was a diagnosis of a PD, and additional inclusion criteria for healthy controls were that they did not ever have a PD and that they had no first-degree family member with a lifetime PD. A detailed description of recruitment, sample characteristics, and procedures of the cohort can be found in Korver et al. For the current study, data release 8.0 was used.

Materials

ERFs

To assess current tobacco and cannabis use, part of the composite international diagnostic interview (World Health Organization (WHO), 1994) was used. A special module assesses the quality and severity of drug dependence and its course, with good test–retest and interrater reliability.³⁹ For cannabis, the most intensive period of use over the last 12 months was scored as zero (none), 1 (less than weekly),2 (weekly) or 3 (daily). Tobacco smoking was scored zero (absent) or 1 (present, at least 1 month daily in the last 12 months).

The Childhood Trauma Questionnaire Short Form⁴⁰ (CTQ-SF) was used to assess childhood trauma. The CTQ-SF is a retrospective self-report questionnaire that originally consisted of 25 items. However, in our study, the item “molestation” was removed due to improper translation into Dutch. The Dutch version of the CTQ-SF thus consists of 24 items that are scored on a Likert scale from 1 (never true) to 5 (always true). It shows good validity and adequate reliability with an α between .63 and .91.⁴¹ The CTQ-SF measures 5 different types of childhood maltreatment: physical neglect, physical abuse, emotional neglect, emotional abuse, and sexual abuse. All of these were used as indicators for our latent variable, which we refer to as childhood trauma. Further details on these measures can be found in the Supplementary Appendix 1.

Cognitive Variables

In the GROUP cohort, all neurocognitive domains in accordance with the Measurement and Treatment Research to Improve Cognition in Schizophrenia (MATRICS) consensus, except visual memory, were included.⁴² The test for each domain was chosen based on reliability, validity, and feasibility for large multisite studies.⁴³ We included all neurocognitive domains in our study: verbal memory, attention/vigilance, processing speed, working memory, and reasoning/problem-solving. Finally, as an addition to the MATRICS domains, we included visuoperception. Details on tests used to measure these domains can be found in Supplementary Appendix 1.

Psychopathological Symptoms

Besides psychotic symptoms, we decided to also assess depressive symptoms because these occur frequently in people with PDs. Psychotic experiences and depressive symptoms over the last 3 years were assessed with frequency scores of the Community Assessment of Psychic Experiences (CAPE, http://cape42.homestead.com). The CAPE is a self-report questionnaire that assesses positive, negative, and depressive symptoms, with good discriminant validity⁴⁴ and reliability.⁴⁵ It consists of 42 items that are scored on a Likert scale from zero (never) to 3 (almost always). The CAPE was used to derive the latent variables of depressive, positive, and negative symptoms.

Data Analysis

The data analysis of our study consists of 3 steps: (1) calculation of latent variable scores, (2) imputation of missing data values, and (3) Bayesian estimation of network models. Because these techniques are relatively new and not widely understood, we will explain why and how we took these steps in the following sections. A schematic overview of our methods can be found in Supplementary Figure 1. Our R-scripts can be found on OSF (link: https://osf.io/fgzh5/).

Calculation of Latent Variable Scores

Why? In this study, we include childhood trauma and psychopathological symptoms by using responses to items from the CAPE and CTQ-SF. Historically, many studies have done this by calculating the sum of the scores on each item of the questionnaire. This approach is problematic because the item labels are meaningless; they are merely ordered values, and it is unclear what their summed values convey. In addition, the use of a sum score implies that there is a particular psychological construct underlying the individual item scores, a latent variable, but it does not consider the uncertainty in estimating this latent variable. Thus, using sum scores may lead to inaccurate estimates of network relations. For this reason, we decided to use a latent variable model to determine how item scores load on the latent construct and then use this latent variable model to generate plausible values,⁴⁶ a value for an individual’s unknown latent variable based on observed item responses. We use these plausible values to impute the latent variables as a node in the network model.

How? We estimated a latent variable model for ordinal response data, called the generalized partial credit model,⁴⁷ using the R package mirt⁴⁸ to determine how each item score relates to the latent construct. We then used the estimated parameters of the latent variable model to subsequently estimate a value for the latent variable for each participant. Specifically, we simulate a plausible value for the latent variable from its posterior distribution,^49,50 which captures the uncertainty that we have about the latent variable. To account for this uncertainty, we generated 100 plausible values for each latent variable per person. More information on the details of our plausible value imputation procedure can be found in the Supplementary Material, Appendix 2.

Imputation of Missing Values

We used the R package mice,⁵¹ to impute missing values, using the 12 study variables, participant status (patient, sibling, or control), age, and gender as predictors in the imputation algorithm. We generated 100 completed datasets (ie, a dataset with imputed values for the missing data points and latent variables). The variability across the 100 completed datasets is thus due to the uncertainty in imputing the missing data points and latent variables. More information on the missing data and the imputation process can be found in the Supplementary Material, Appendix 3, and Tables 2 and 3.

Bayesian Network Estimation

Why? The Bayesian approach to network analysis allows us to quantify and assess our uncertainty in inferring the structure of the network by estimating the probability that an edge should be included in the network. While the extremes of zero and one for the estimated probability would provide complete certainty of excluding or including the edge in the network (Because these probabilities are estimates, we are uncertain as to their true value. The extreme values of 1 and 0 are most likely numerical bounds in the estimation procedure. While we can take them as strong support for inclusion or exclusion, they are unlikely to be exactly 1 or 0, given finite data), any value between these 2 extremes signals uncertainty about the estimated edge. Two major advantages of the Bayesian approach are thus that (1) it allows us to quantify the evidence of absence, i.e., the exclusion of a network relationship, and (2) it allows us to separate the evidence of absence from the absence of evidence.

How? All statistical analyses were performed in R, version 4.1.2.⁵² First, we estimated partial correlation networks of our 3 ERF variables, 6 cognitive variables, and 3 psychopathological symptoms on the sample of patients, siblings, and controls using the package BDgraph.⁵³ The Gaussian Copula Graphical Model (GCGM) used here models the continuous variables using a Gaussian Graphical Model and the binary and ordinal variables in a way that is closely related to the way they are modeled with polychoric correlations, assuming that the observed binary or ordinal data are truncated versions of an underlying continuous variable. We modeled tobacco smoking as a binary variable and cannabis use as an ordinal variable. All other variables were modeled as continuous variables. To estimate our GCGMs, we used reversible jump Markov chain Monte Carlo (100 000 iterations) to estimate the Bayesian model, including default prior specifications: A uniform prior with a prior inclusion probability of .5 for the network relations and a G-Wishart prior with 3 degrees of freedom on the precision matrix (unnormalized partial correlations). After estimation, BDgraph returns the weight and posterior inclusion probability for each edge in the network. We decided to include only edges with a posterior probability over .5.^54,55 We averaged the edge weight estimates and inclusion probabilities across the 100 datasets to keep a single network. The stability and accuracy of all estimated networks were examined by calculating the 95% Bayesian highest posterior density intervals (HDI) of each edge weight and inclusion probability. The 95% HDI is the shortest interval that contains 95% of the posterior mass.⁵⁶

Centrality Analyses

We calculated strength centrality for each node. Strength centrality measures how many connections a node has to other nodes in the network, taking into account the weight of each connection.²⁸ It indicates how strongly the node is connected to the other nodes in the network.⁵⁷

Network Comparison Between Patients and Siblings

Because the data of only 381 healthy controls was available, we did not have enough power to separately analyze the network of healthy controls. However, we were able to compare the network structure between patients and siblings. Because symptom severity may be a confounding factor for network connectivity,⁵⁸ we decided to estimate networks of patients and siblings after controlling for the general severity of illness. We did this by regressing (or partialing) out an indication of general functioning, as measured by the social functioning scale (SFS). By regressing each variable on the SFS sum score, the variance of the variables that are not explained by the severity of illness is contained in the residuals.⁵⁸ The residuals were then used to estimate separate networks of patients and siblings, as described above. We then compared the HDIs of the edge weights and inclusion probabilities averaged across the completed datasets. An overlap between the HDIs of patients and siblings indicates that the edges were present and comparably strong in both patients and siblings, whereas nonoverlapping HDIs indicate a likely difference in the estimated edge. Because the BDgraph package does not allow testing for differences in global strength and overall network structure, we also estimated networks using the Bootnet package. We estimated GGMs for patients and siblings, using EBICglasso,⁵⁹ with a default hyper tuning parameter (γ) of .5. We then used the estimated GGMs to perform a network comparison test (NCT)⁶⁰; a permutation-based hypothesis test that assesses whether networks differ on structural invariance or global strength invariance. The median P-value across all tests (one for each completed dataset) was used as an indication of significance.^61–63

Sensitivity Analyses

We set a uniform prior of .5 on the network structure. To find out how sensitive our estimated networks were to this prior, we estimated the combined network with a prior of .25 and .75 on the network structure. We also estimated the networks of siblings and patients without controlling for SFS.

Results

Sample Characteristics

We removed 23 participants with a diagnosis of bipolar disorder. No outliers were removed; see the Supplementary Appendix 4 for the rationale for this decision. We decided to retain only the participants with no more than 50% missing values per variable cluster. This resulted in a sample of 1748 participants, with data on at least 2 ERFs, 2 symptom dimensions, and 3 cognitive measures. Of this final sample, 653 were patients, 714 were siblings, and 381 were healthy controls. Table 1 shows the demographic characteristics and mean scores of the study variables for all the subgroups.

Table 1.

Demographic Characteristics and Mean Scores on Study Variables in Full Sample, Patients, Siblings, and Controls

Variables	Unit	Combined	Controls	Siblings	Patients
Age	Years (SD)	31.63 (8.51)	34.75 (10.57)	30.98 (8.01)	30.524 (7.16)
Gender	Female	0.43	0.53	0.55	0.24
Positive symptoms	Mean (SD)	4.95 (3.95)	3.48 (2.73)	4.05 (3.20)	6.83 (4.55)
Negative symptoms	Mean (SD)	5.00 (7.50)	1.82 (2.42)	2.13 (2.90)	10.18 (9.93)
Depressive symptoms	Mean (SD)	8.78 (7.17)	5.61 (4.39)	6.56 (5.63)	13.09 (7.90)
CPT sensitivity	Mean (SD)	95.44 (12.14)	96.71 (10.71)	97.15 (9.34)	92.86 (14.88)
WAIS digit symbol	Mean (SD)	10.18 (3.52)	12.06 (3.24)	11.14 (2.99)	8.05 (3.10)
WAIS arithmetic	Mean (SD)	10.76 (3.32)	11.75 (3.18)	11.10 (3.17)	9.80 (3.32)
WAIS block design	Mean (SD)	11.26 (3.21)	11.94 (3.07)	11.60 (3.03)	10.50 (3.34)
15WT total delayed	Mean (SD)	9.49 (2.89)	10.22 (2.59)	10.17 (2.59)	8.32 (2.99)
BFR	Mean (SD)	23.27 (2.20)	23.43 (1.99)	23.52 (2.14)	22.91 (2.32)
CTQ total	Mean (SD)	11.09 (10.39)	8.48 (8.83)	9.90 (9.52)	14.15 (11.49)
Cannabis use	Mean (SD)	1.32 (0.77)	1.19 (0.57)	1.26 (0.68)	1.45 (0.94)
Tobacco smoking	Proportion users	0.41	0.21	0.34	0.60

Abbreviations: BFR, Benton facial recognition, CTQ, childhood trauma questionnaire; CPT, continuous performance test; WAIS, Wechsler adult intelligence scale; 15WT: 15 Word Test.

For all the cognitive variables (CPT sensitivity, WAIS, 15WT total delayed, BFT), a higher score indicates a better performance. For the variables of positive symptoms, negative symptoms, and depressive symptoms, a higher score indicates the presence of more symptoms. For cannabis use, a higher score indicates a higher frequency of cannabis use.

Edge Inclusion

The posterior edge inclusion probabilities aggregated over the 100 completed data sets, ie, estimated networks, are shown in Figure 1. This figure shows that some edges were estimated very reliably, with an inclusion probability of (almost) 0 or 1 in all the samples. However, there are also edges whose inclusion probability varies greatly from sample to sample. This means that the data do not provide enough information to conclude whether the edge should be included in the network. Most of the uncertain edges had a mean inclusion probability below .5, so they were not included in the visualized networks. The few uncertain edges that were included in the network had very low edge weights. The visualization of the HDIs of the estimated edge weights can be found in Supplementary Figure 2.

Figure 1.

Posterior Edge Inclusion Probabilities of All Edges, Including HDI. Edges With an Average Inclusion Probability of Greater Than 0.5 (eg, a Dot on the Right Side of the Line in This Figure) Were Included in the Network. A Wider Interval Means More Uncertainty in Whether or Not an Edge Should Be Included in the Network. Abbreviations: BFR, Benton facial recognition, CTQ, childhood trauma questionnaire; CPT, continuous performance test; WAIS, Wechsler adult intelligence scale

Edge Estimation

All networks were plotted in qgraph,⁶⁴ using the layout = spring option. Figure 2 shows the average network of the combined sample of patients, siblings, and controls. The weakest edge in the network represents a partial correlation of r = .05 (WAIS block design—negative symptoms), and the strongest edge represents a partial correlation of r = .57 (cannabis—cigarette smoking).

Figure 2.

Average Estimated Network on the Sample of Patients, Siblings, and Controls. Positive Edges Are Shown in Blue, and Negative Edges Are Shown in Red. A Stronger Partial Correlation Is Represented by a Thicker, More Visible Edge. To Increase Readability, We Have Renamed Some of the Variables. Attention/Vigilance Represents the Measure of the Continuous Performance Test, Processing Speed Is Measured With the WAIS-III Digit Symbol Substitution, Working Memory Is Measured with the WAIS-III Arithmetic, Reasoning/Problem Solving Is Measured with the WAIS-III Block Design Test, Verbal Memory Is Measured With the Auditory Verbal Learning Test, and Visuoperception is measured With the Benton Facial Recognition test

The WAIS digit symbol substitution is the most central node in the network (see Supplementary Figure 4). It connects the clusters of substance use, cognition, and psychopathological symptoms through the edge with positive symptoms (r = −0.18) and the edge with smoking (r = −0.17). It is also noteworthy that smoking is highly connected to cognitive variables and positive symptoms in the combined and separate networks (Figure 3), whereas cannabis use is only directly connected with WAIS block design (r = 0.16). Because this correlation is in an unexpected direction (cannabis use is associated with better performance on WAIS block design test), we performed a post hoc analysis. This showed that the bivariate correlation between cannabis and WAIS block design (Spearman’s ρ = 0.045) was substantially lower than the partial correlation, suggesting that this partial correlation may be spurious. Finally, we see that childhood trauma is associated with negative and depressive symptoms in the combined and separate networks.

Figure 3.

Network Visualization of (A) Patients and (B) Siblings. Positive Edges Are Shown in Blue, and Negative Edges Are Shown in Red. A Stronger Partial Correlation Is Represented by a Thicker, More Visible Edge. To Increase Readability, We Have Renamed Some of the Variables. Attention/vigilance Represents the Measure of the Continuous Performance Test, Processing Speed Is Measured With the WAIS-III Digit symbol Substitution, Working Memory Is Measured With the WAIS-III Arithmetic, Reasoning/Problem Solving Is Measured With the WAIS-III Block Design Test, Verbal Memory Is Measured With the Auditory Verbal Learning Test, and Visuoperception Is Measured With the Benton Facial Recognition Test

Network Comparison Patients and Siblings

Figure 3 shows the separate networks of patients and siblings. A visual comparison of the networks shows that, overall, there are many similarities. Both networks show that the cognitive variables are very interconnected, that smoking is associated with processing speed and positive symptoms, and that childhood trauma is only connected to negative and depressive symptoms. The similarity of the networks is in accordance with the results of the NCT test, which rejects the null hypothesis that the networks of patients and siblings differ in general network structure (P = .217) or global strength (P = .155). Supplementary Figures 5, 6, and 7 show the centrality plot and the the HDIs of edge weights and inclusion probabilities in the networks of patients and siblings.

Sensitivity Analyses

The networks with a prior of .25 and .75 on the network structure showed approximately the same structure as the network with a prior of .5, with only small differences in the number of included connections with small edge weights. Further results of the sensitivity analyses can be found in Supplementary Appendix 5.

Discussion

To our knowledge, this is the first study that directly investigated interactions between ERFs, cognition, and psychotic symptoms across the psychosis liability spectrum using a network approach. Furthermore, our study is one of the first to apply Bayesian network modeling combined with plausible value imputation in a clinical framework. Processing speed was the most central node in the estimated networks, connecting the clusters of substance use, cognition, and psychotic symptoms. Among the ERFs examined, tobacco smoking was highly connected to cognition and psychotic symptoms, both in the combined and separate networks of patients and siblings. Contrary to expectations, we found no evidence of a direct association between cannabis and cognition or symptoms. We did find a positive relationship between experienced childhood trauma and negative/depressive symptoms in the combined and separate networks. These findings add to the current knowledge in several ways, which we will outline below.

Interestingly, tobacco smoking was more diversely associated with symptoms and cognition than cannabis use, which was only indirectly associated with smoking. Based on previous meta-analytical evidence,⁶⁵ one would expect to find a direct association between cannabis use and psychotic symptoms. Notably, only recent studies investigating this association have explicitly controlled for tobacco smoking as a potential confounder or correlate. Quattrone and colleagues⁶⁶ found that current cannabis use was associated with positive symptoms only in patients who used high-potency cannabis daily. Another study found a significant attenuation of the association between cannabis and psychotic symptoms when controlling for tobacco and other drug use.⁶⁷ Consistent with these findings, our results suggest that the association between cannabis use and psychotic symptoms could, at least partially, be explained by the association between tobacco use and psychotic symptoms. However, other authors report a stronger association between cannabis compared to tobacco use and psychotic experiences when investigating confounding effects of combined use.⁶⁸ Thus, the exact relationship between cannabis use and psychotic symptoms remains unclear, which highlights the importance of disentangling the relationships between these variables. As the combined use of cannabis and tobacco appears to be consistently associated with an increased risk of psychotic symptoms,⁶⁹ future research should not only focus on cannabis use but also include tobacco smoking as an important, as yet under-recognized, potential risk factor.

Analyses further revealed that the clusters of substance use, cognition, and psychotic symptoms were all connected through processing speed as the most central node in the networks. This finding is in line with recent network analyses on PDs, which also found processing speed to be a central node.^70,71 It is important to note that high centrality does not necessarily mean that a node is the most (clinically) relevant variable. It could also mean that the node is simply a causal endpoint to the pathways in the data, which would imply that intervening in this variable would not lead to any changes in the network.⁷² However, previous studies into the importance of processing speed have shown that it might be a starting point of the causal chain,^73,74 meaning that processing speed may represent a core cognitive deficit underlying generalized neurocognitive impairment in PDs.⁷⁵ More practically, this could imply that a slower pace of information processing could underlie the other cognitive impairments that people with PDs encounter, like difficulty with planning, concentration, or memory.

Interestingly, tobacco smoking and positive symptoms, in particular, seem to be connected through processing speed. This raises the hypothesis that smoking-related cognitive impairment may, in turn, be associated with more severe positive symptoms. However, based on our study, the opposite (people use smoking as a coping mechanism for cognitive impairments) could also be true. We encourage future confirmatory research to test these hypotheses.

The direct positive relationship between childhood trauma and negative/depressive symptoms in the estimated networks is consistent with previous research^31,76,77 and supports the theory of an affective pathway to psychosis. This theory suggests that childhood trauma may lead to emotional distress (eg, depression), which increases the risk of psychotic symptoms.⁷⁸

Comparison of Patients and Siblings

Our results show that the networks of patients and siblings are relatively similar, which means that many of the associations we found are likely independent of illness-related confounding effects, such as antipsychotic medication.

Limitations

Although our study is both clinically and methodologically innovative, there are important limitations that must be considered. First, as network modeling in a Bayesian framework is still largely in development, certain features available for frequentist network modeling, such as comparing networks, are not yet implemented for Bayesian network modeling. Therefore, we used the frequentist analysis to compare networks between patients and siblings by using the HDIs over the 100 completed datasets and by using the NCT. While the method allowed us to detect group differences, it does not allow for the obtaining of evidence for group equivalence, which is a benefit that could have been obtained with the Bayesian approach.

A further consideration is the influence of Berkson’s bias on our results. Berkson’s bias is the estimation error that arises from selecting a subpopulation for analyses, and it is an often encountered issue when using network analyses to study a clinical population.⁷⁹ However, in the main analyses of our study, we have explicitly included people experiencing the full possible range of psychotic symptoms (from healthy controls to patients with psychoses). We have thus avoided only analyzing a subpopulation, therefore minimizing the chance of Berkson’s bias influencing our results. In the additional analyses where we did select a subpopulation (patients and siblings), we have tried to reduce the influence of Berkson’s bias by controlling for the general severity of illness, as described in the methods section. Importantly, the additional analyses of subgroups showed similar results to the analyses of the total group, and we haven’t drawn any conclusions solely based on the analyses of subgroups. Altogether, we do not believe that the main conclusions of our study are heavily influenced by Berkson’s bias.

Another limitation is that our data and methods do not allow for causal interpretation. The causal mechanism of the association between many of our variables, for example, between smoking and psychotic symptoms,⁸⁰ remains unclear. While smoking could be a risk factor for PDs, it is also possible that experiencing psychotic symptoms leads to smoking, eg, as a coping mechanism. Therefore, we encourage future research to extend the network approach to longitudinal data (ie, panel data) to elucidate the possible causal relationship between ERFs, cognition, and psychotic symptoms.

In addition to the methodological limitations, our study has several conceptual limitations. First, our measures of visual perception and attention showed a ceiling effect. That is, many participants achieved (near) perfect scores on these tasks, reducing the variance and power to detect associations with these variables. Secondly, the task we used to measure processing speed (WAIS digit symbol substitution) likely also measures other abilities, such as fine motor skills. Impairments in these abilities could confound the measure of processing speed. It is also important to note that most of the participants who used cannabis also reported smoking tobacco (78%), which may partially explain why cannabis was related to other variables through smoking. Also, we measured tobacco smoking as a binary variable but cannabis use as an ordinal variable, which might have influenced our results. Future research should measure these variables on an identical scale and should attempt to assess well-separated groups when investigating the differential effect of smoking versus cannabis use.

Notably, the GROUP sample suffers from selection bias in favor of relatively high-functioning patients with mild symptoms. Therefore, our results cannot be easily generalized to the overall population of patients with PDs. In addition, the GROUP sample consists of patients who were diagnosed with a PD at least 3 years prior to the data assessment used in the current study, whereas many studies on PDs have been conducted in patients with a first episode of psychosis. For broader generalizability, future studies should include patients at different stages of illness. Finally, our study did not control for (antipsychotic) medication use, although we partially overcame this shortcoming by showing similar effects in siblings (who do not use antipsychotic medication) and patients.

Conclusion

In conclusion, our study is the first to examine the associations between ERFs, cognition, and psychotic symptoms using (a Bayesian approach to) network analysis. With our innovative methods, we have provided new insights into the association between these factors, showing that processing speed links clusters of substance use and psychotic symptoms. Most importantly, our results suggest that tobacco smoking, in particular, is associated with cognition and psychotic symptoms in both patients and siblings of patients with psychosis. This finding highlights the need to include tobacco smoking when studying cannabis use in the context of PDs. Finally, it underlines the importance of discouraging smoking in patients or people at risk of PDs.

Supplementary Material

Supplementary material is available at https://academic.oup.com/schizophreniabulletin/.

Funding

The infrastructure for the GROUP study is funded through the Geestkracht programme of the Dutch Health Research Council (Zon-Mw, grant number 10-000-1001) and matching funds from participating pharmaceutical companies (Lundbeck, AstraZeneca, Eli Lilly, Janssen Cilag) and universities and mental health care organizations (Amsterdam: Academic Psychiatric Centre of the Academic Medical Center and the mental health institutions: GGZ Ingeest, Arkin, Dijk en Duin, GGZ Rivierduinen, Erasmus Medical Centre, GGZ Noord Holland Noord. Groningen: University Medical Center Groningen and the mental health institutions: Lentis, GGZ Friesland, GGZ Drenthe, Dimence, Mediant, GGNet Warnsveld, Yulius Dordrecht and Parnassia psycho-medical center The Hague. Maastricht: Maastricht University Medical Centre and the mental health institutions: GGzE, GGZ Breburg, GGZ Oost-Brabant, Vincent van Gogh voor Geestelijke Gezondheid, Mondriaan, Virenze riagg, Zuyderland GGZ, MET ggz, Universitair Centrum Sint-Jozef Kortenberg, CAPRI University of Antwerp, PC Ziekeren Sint-Truiden, PZ Sancta Maria Sint-Truiden, GGZ Overpelt, OPZ Rekem. Utrecht: University Medical Center Utrecht and the mental health institutions Altrecht, GGZ Centraal and Delta.)