The Symptom and Domain Structure of Psychotic Disorders: A Network Analysis Approach

Abstract

Little is understood about the symptom network structure of psychotic disorders. In the current study, we aimed to examine the network structure of psychotic symptoms in a broad and transdiagnostic sample of subjects with psychotic disorders (n = 2240) and to determine whether network structure parameters vary across demographic, sampling method and clinical variables. Gaussian graphical models were estimated for 73 psychotic symptoms assessed using the Comprehensive Assessment of Symptoms and History. A 7-cluster solution (reality distortion, disorganization, catatonia, diminished expressivity, avolition/anhedonia, mania, and depression) best explained the underlying symptom structure of the network. Symptoms with the highest centrality estimates pertained to the disorganization and, to a lesser extent, negative domains. Most bridge symptoms pertained to the disorganization domain, which had a central position within the network and widespread connections with other psychopathological domains. A comparison of networks in subgroups of subjects defined by premorbid adjustment levels, treatment response, and course pattern significantly influenced both network global strength and network structure. The sampling method and diagnostic class influenced network structure but not network global strength. Subgroups of subjects with less densely connected networks had poorer outcomes or more illness severity than those with more densely connected networks. The network structure of psychotic features emphasizes the importance of disorganization symptoms as a central domain of psychopathology and raises the possibility that interventions that target these symptoms may prove of broad use across psychopathology. The network structure of psychotic disorders is dependent on the sampling method and important clinical variables.

Introduction

Psychotic disorders are serious psychiatric conditions that share certain psychotic symptoms, such as delusions or hallucinations, but present with a varied phenomenology and clinical course, which is at the heart of their nosological complexity.^1,2 Traditional approaches to the classification of psychopathology assume that symptoms of a psychiatric disorder reflect an underlying latent construct: the disorder itself in the case of categorical approaches and the latent variables in the case of dimensional classifications.^3,4 These approaches, however, do not take into account the relative non-specificity of psychotic symptoms or the fuzzy boundaries among disorders. In contrast, the network approach is an atheoretical, data-driven methodology that requires no a priori model of causality among variables. It enables the examination of the contemporaneous relationships among a given set of symptoms, controlling for all other symptoms in the network, thus allowing for a global representation of the structure of a disorder at the symptom level.^3,5,6

Network analysis allows for the identification of the most central symptoms in the network, which may be pivotal in influencing other symptoms and illness-related variables and hence represent potential therapeutic targets. This approach may be particularly advantageous in understanding comorbidity and relationships among symptoms in disorders that lack clear boundaries, as is the case with psychotic disorders.^7,8 Network models were implemented only recently in the field of psychotic disorders research and, to date, only 3 previous studies have examined the whole symptom network structure of psychotic disorders using comprehensive measures of symptoms.^9–11 These studies revealed a complex symptom structure and came to the conclusion that the resulting networks showed strong connectivity both within and between symptom domains, but beyond this general finding many questions remain open.

Overall, we identified 5 specific challenges in the prior literature of psychosis symptom networks that we aim to address in this article. First, network studies estimated networks in one sample only, and it is unclear how the results generalize across sampling methods. Second, the communalities (clusters or domains) within the symptom networks are usually predetermined on theoretical grounds or by the predefined structure of assessment instruments; thus, the number and composition of empirically derived clusters in the network remain unknown. Third, within the network approach framework the detection of “bridge” symptoms between different symptom clusters is of particular relevance because of their potential in explaining comorbidity and as target symptoms for treatment^6,12; however, no formal bridge centrality analysis within psychosis exists to date. Fourth, given the vast clinical heterogeneity of psychotic disorders, it is unknown whether the symptom network structure varies as a function of important clinical variables. Five, existing studies featured relatively small samples (typically < 560 subjects), and, given that comprehensive network models of psychotic symptoms require the estimation of many parameters and considerable power to reliably detect small coefficients,¹³ investigations of larger data sets are needed.

Our study attempts to address all these points by investigating the symptom network structure in a large and transdiagnostic sample of subjects with psychotic disorders with the aims of empirically determining the number and composition of symptom domains within the network, ascertaining the most central and bridge symptoms and domains, and comparing networks in subgroups of subjects stratified by relevant demographic and clinical variables.

Methods

Study Sample

The study population consisted of 2240 subjects with a lifetime diagnosis of a functional psychotic disorder according to the Diagnostic and Statistical Manual of Mental Disorders, 4th edition (DSM-IV) criteria.¹⁴ The global study population was composed of samples from 3 primary studies: a first-episode sample (FES, n = 486) recruited between 1988 and 2012,¹⁵ a multi-episode sample (MES, n = 660) recruited between 1988 and 2005,¹⁶ and a community sample (CS, n = 1094) recruited between 2012 and 2014.¹⁷ Subjects from the FES and MES were enrolled from consecutive admissions to the hospital and those from the CS were enrolled thorough the psychiatric case register for the same catchment area of Navarra (Spain). The projects of the 3 primary studies were approved by the ethics committee of the Regional Health Service of Navarra and written informed consent was obtained from all study participants or their legal representatives. The study sample comprised individuals with all types of functional psychotic disorders with varying degrees of illness duration, severity, and impairment (table 1). Subjects with poor quality information or missing data were excluded from the study.

Table 1.

Sociodemographic and Clinical Characteristics of the Overall Study Population (n = 2240)

	Mean (SD)	No. (%)
Age, y	38.7 (15.3)
Gender, male		1225 (54.7)
Civil status, single		1384 (61.8)
Education, y	9.76 (3.47)
DSM-IV diagnosis
Schizophrenia		908 (40.5)
Schizophreniform disorder		180 (8.0)
Delusional disorder		120 (5.4)
Brief psychotic disorder		179 (8.0)
Schizoaffective disorder		124 (5.5)
Bipolar disorder		345 (15.4)
Major depression		245 (10.9)
Psychotic disorder not otherwise specified		139 (6.2)
Premorbid adjustmenta	4.95 (2.91)
Good premorbid adjustment		1118 (49.9)
Poor premorbid adjustment		1122 (50.1)
Age at illness onset	27.5 (11.2)
Duration of untreated psychosis, y	1.84 (3.58)
Duration of illness, y	11.2 (11.8)
Number of hospitalizations	2.74 (3.72)
Treatment responseb	1.83 (0.94)
Good treatment response		1639 (73.2)
Poor treatment response		537 (24.0)
Course pattern
Full remission		993 (44.3)
Partial remission		774 (34.6)
Chronic/continuous		473 (21.1)
Global Assessment of Functioning	65.4 (20.8)

Note: ^aVariable dichotomized at the median value.

^bGood and poor treatment response were respectively defined as a symptom reduction ≥ 50% and <50% according to the Clinical Global Impression scale. Sixty-four subjects from the community sample had never been treated for the psychotic disorder, and thus they were not assessed for treatment response.

Assessment of Symptoms and Clinical Features

Subjects underwent an extensive clinical examination by means of the Comprehensive Assessment of Symptoms and History (CASH).¹⁸ The CASH is a semi-structured interview that served to assess demographics, illness-related variables, psychopathology, and diagnosis. The current condition module of the CASH includes 73 symptoms that were rated at their worst over the past month on a 6-point Likert-type scale. Ratings were performed by experienced psychiatrists or clinical psychologists who conducted face-to-face interviews with the subjects and collected information provided by treating physicians, clinical records, and significant others. Interviewers showed good inter-rater reliability (intraclass correlation coefficient > 0.80) on most CASH items.¹⁵

A lifetime diagnosis of psychotic disorder, according to the DSM-IV criteria,¹⁴ was reached by consensus among the senior authors (V.P. and M.J.C.) using all available information. For the purpose of the present study, specific diagnoses were collapsed into 3 groups: schizophrenia (n = 908, 40.5%), nonschizophrenic non-affective psychosis (NSNAP, n = 742, 33.2%) and affective psychosis (n = 590, 26.3%). Subjects were treated according to clinical choice, and treatment response was assessed by means of the Efficacy Index from the Clinical Global Impression Scale (CGI-EI)¹⁹ either at the end of the index hospitalization (FES and MES) or at the last admission or the month following the last treatment trial (CS). All other demographic and clinical variables were rated according to the CASH, including premorbid adjustment, age at illness onset, duration of untreated psychosis (DUP), illness duration, course pattern, and functioning.

Statistical Analysis

Procedure. We first estimated the domain network structure of symptoms in the whole sample. We began by estimating a predefined 6-domain structure of CASH symptoms (ie, the theoretical model),²⁰ which well represents the current consensus about the main symptom domains of psychotic disorders, including reality distortion, disorganization, negative, catatonia, mania, and depression.^21–23 Then, we empirically estimated the cluster structure underlying the data using statistical algorithms and compared it with the a priori 6-domain model by means of several fit indices. Second, we estimated the centrality parameters of the symptom network structure, such as their stability and reliability. Third, we conducted a bridge symptom analysis to determine the most important symptoms in this regard. Finally, we compared networks across subgroups of subjects defined by demographic and clinical variables.

Symptom scores were not normally distributed, and as such, the data were z-transformed to relax the normality assumption prior to constructing the networks. To examine the reliability of this procedure, the network based on the transformed variables was compared to a network based on Spearman correlations.²⁴

Network Estimation. Networks are represented through nodes (variables) and edges (relationships between variables). Those relationships between nodes are partial correlations. The partial correlations reveal the relationship between 2 nodes controlling the effect of the rest of the nodes in the network; ie, controlling the spurious correlations that may arise due to the multiple comparisons as it can happen with zero-order correlations. Partial correlations are later regularized to ensure a more interpretable model. The regularization method employed is EBICglasso, graphical Lasso based on an extended Bayesian information criterion.²⁵ For the network estimation, we used the R²⁶ package EGAnet²⁷ with the glasso option, which is implemented via qgraph package.²⁸

Empirical Community Detection. The EGAnet package estimates the number and composition of symptom domains underlying multivariate data by using the walktrap algorithm.²⁹ This algorithm empirically estimates the number of dimensions and which dimension each node belongs to, determined by its ratio of many thickly connected edges to few loosely connected edges. To compare the theoretical and the empirical model we made use of entropy fit indices of both models. A lower Entropy Fit Index value is indicative of a better fit of the structure to the data (correct number of factors and correct positioning of items by factor). The fit of the empirically derived model was qualitatively compared with that of the theoretical model using Global Entropy (GE) and von Neumann’s Entropy Fit Index (EFI) as implemented in R,³⁰ which have been shown to be at least as accurate as traditional fit indices in identifying the number of latent clusters.³⁰ Furthermore, the comparative fit of the theoretical and empirical models was statistically examined using the Satorra-Bentler scaled chi-squared test.³¹

Centrality Estimation. Using the qgraph package, we quantified the importance of each node to the network by estimating 4 centrality indices: strength, closeness, betweenness, and expected influence (EI) (supplementary appendix). Node strength is usually considered the main centrality parameter,⁵ however, because EI also takes into account negative associations among nodes, it represents a more accurate approach to node centrality than strength^32,33; thus, in the present article, EI will be considered the main centrality measure. Edge stability and centrality stability of the sample were estimated via the bootnet package¹³ with nonparametric and case bootstrapping (1000 bootstraps), respectively. To quantify the stability of centrality estimates using subsets bootstraps we, employed the correlation stability coefficient, with coefficients above 0.50 indicating high stability.¹³

Bridge Analyses

Bridge centrality analysis was carried out with the networktools package.¹² It includes the same centrality measures as node centrality analysis described above plus 2-step EI. We obtained the typical centrality measures using the domains of psychopathology derived from the best fitting model.

Network Comparison

Networks were compared with the Network Comparison Test (NCT).³⁴ The NCT tests analyses whether 2 networks differ in global strength (overall connectivity, the weighted absolute sum of all edges in the network) and network structure (distribution of edge weights). Furthermore, we examined a similarity index³⁵ and the number of edges that significantly differ across networks. Paired tests were 2-sided with an alpha threshold set at 0.05, and a Bonferroni-Holm correction was applied to control for multiple testing.

Data Availability

The analyses carried out were not formally recorded. The analysis code for all analyses conducted in this study is provided in the supplementary material, as well as the supplementary figures, tables, correlation matrices, and other R-objects that permit the researchers to replicate our findings. Additional information about the data can be obtained from the corresponding author upon request.

Results

Clustering of Symptoms in the Network

The theoretical 6-domain model of psychotic symptoms revealed a rather complex network structure with many positive edges both within and between domains and many negative edges between domains; furthermore, 11 of the 73 symptoms (15%) did not fit the a priori domain structure (supplementary figure 1).

The walktrap community detection algorithm indicated that a 7-cluster solution best explained the underlying symptom structure of the network (figure 1). The comparative fit indices for the theoretical (GE = 1.80, EFI = 4.59, χ ²_(df) = 28472₍₂₄₆₉₎) and empirical (GE = 1.65, EFI = 4,21, χ ²_(df) = 28183₍₂₄₆₃₎) models revealed a better fit for the latter (Satorra-Bentler test P < .001). The major difference among models was the splitting of the negative domain in the theoretical model into 2 domains (diminished expressivity and avolition/anhedonia) in the empirical model. Within the empirical model, the disorganization domain occupied a central position within the network and showed widespread connections with other domains. The symptom connectivity within domains was much higher than that between domains (table 2). The percentage of connections within domains ranged from 40.5% (disorganization) to 100% (catatonia), and the percentage of connections between domains ranged between null (catatonia and mania) and 40.8% (avolition/anhedonia and diminished expressivity). The least interconnected domain was reality distortion, which exhibited between 3.1% and 12.9% of all possible connections with other domains.

Fig. 1.

The empirical network structure of 73 symptoms of psychotic disorders. Each node (circle) represents a symptom, and each edge (line) represents the association between 2 symptoms. Positive edges appear blue, negative red, and stronger and saturated represent regularized partial correlations.

Table 2.

Number of Connections Between and Within Empirically Defined Symptom Domains

	Reality Distortion	Disorganization	Mania	Depression	Avolition/ anhedonia	Catatonia	Diminished Expressivity
Reality distortion, No. (%)	52/91 (57.1)
Disorganization, No. (%)	33/294 (11.2)	85/210 (40.5)
Mania, No. (%)	16/140 (11.4)	40/210 (19)	26/45 (57.8)
Depression, No. (%)	18/140 (12.9)	46/210 (21.9)	11/100 (11)	29/45 (64.4)
Avolition/anhedonia, No. (%)	10/98 (10.2)	29/147 (19.7)	11/70 (15.7)	12/70 (17.1)	18/21 (85.7)
Catatonia, No. (%)	5/42 (11.9)	15/63 (23.8)	0/30 (0)	4/30 (13.3)	1/21 (4.8)	3/3 (100)
Diminished expressivity, No. (%)	3/98 (3.1)	37/147 (25.2)	12/70 (17.1)	10/70 (14.3)	20/49 (40.8)	8/21 (38.1)	19/21 (90.5)

Values are the number (and percentage) of connections between and within (in bold in the diagonal) symptom domains relative to the maximum number of possible connections. The maximum number of connections within domains was computed as follows: given a domain d having Sd number of symptoms, then the maximum number of connections within this domain equals Sd (Sd–1)/2. The maximum number of connections between domains can be calculated by multiplying the number of symptoms from one domain to the number of symptoms of the other domain.

Centrality Measures Analysis

Overall, the most central symptoms pertained to the disorganization domain and to a lesser extent to the negative domain. More specifically, items with EI > 1 included, in decreasing order of importance, inattentiveness during mental testing, increase in activity, inability to enjoy recreational interest/activities, derailment, unchanging facial expression, inappropriate social/sexual behavior, impersistence at work/school and paucity of expressive gestures (figure 2). Notably, inappropriate social/sexual behavior had high values on all 4 centrality measures. In contrast, the less central symptoms (ie, the less connected ones) pertained to the reality distortion and catatonia domains.

Fig. 2.

Centrality estimates of 73 symptoms of psychotic disorders depicting strength, betweenness, closeness, and expected influence of each node. See figure 1 for description of symptoms’ abbreviations.

The networks based on transformed and Spearman correlations were virtually the same with a difference of 0.2% in the total number of significant connections, this indicating that results derived from the z-transformed correlation matrix were trustworthy (supplementary figure 2).

The case-dropping subset bootstrap procedure showed that the centrality values remained stable even after dropping large proportions of the sample (supplementary figure 3). Bootstrapped 95% CIs for the estimated edge weights were narrow, suggesting that the estimates were reliable (supplementary figure 4).

The correlation stability coefficients of centrality measures ranged between 0.60 and 0.75, indicating robustness of findings. Furthermore, the EI centrality across subgroups was also highly stable with correlations stability coefficients >0.70 for each subgroup.

To examine the stability of networks across regularized and unregularized models, we performed a graphical comparison between EBICglasso, thresholded EBICglasso, and unregularized GGM models (gamma 0 and 0.5). The 4 models resulted in very similar centrality parameters and network structures that were hardly impacted by the estimation method (supplementary figures 5a and 5b).

Bridge Symptoms Across Domains

According to the 2-step EI, most bridge symptoms (9 of 10 symptoms with z score > 1) pertained to the disorganization domain (supplementary figure 6). Psychomotor agitation was the symptom that most consistently connected the different domains (z score > 1 in all 5 bridge parameters).

Effects of Demographic and Clinical Variables on the Network

Table 3 shows pair-wise comparisons among subgroups of subjects with psychotic disorders. Supplementary figure 7 shows the EI parameter across subgroups and supplementary figure 8 depicts the specific network structure of each subgroup. Network comparison in subgroups of subjects defined by gender did not yield significant differences in network global strength or network structure. Sample subtypes exhibited similar network global strength but highly differed in network structure (all P < .001). Diagnostic groupings did differ in network global strength; however, the network structure of schizophrenia significantly differed from those of NSNAP (P < .001) and affective psychosis (P = .04). Subjects with good vs poor premorbid adjustment significantly differed in terms of both network global strength (P = .03) and network structure (P = .04). Regarding the course pattern, subjects with remitting psychosis significantly differed from those with partially remitting psychosis in network global strength (P = .02), and from those with chronic psychosis in both network global strength (P < .001) and network structure (P < .001); furthermore, subjects with partially remitting psychosis also significantly differed from those with chronic psychoses in network structure (P = .024). Subgroups of subjects with less densely connected networks had poorer outcomes than those with more densely connected networks (supplementary figure 8).

Table 3.

Network Comparison Across Subgroups of Psychotic Disorders

	Group Comparison	S (P)	M (P)	SI	N (%)a
Gender	Male (n = 125) vs female (n = 1015)	0.60 (.50)	0.19 (.08)	0.73	3 (0.11)
Sample type	First-episode sample (n = 486) vs multi-episode sample (n = 660)	0.38 (.79)	0.56 (<.001)	0.63	64 (2.44)
	First-episode sample (n = 486) vs community sample (n = 1094)	0.41 (.85)	0.47 (<.001)	0.54	78 (2.97)
	Multi-episode sample (n = 660) vs community sample (n = 1094)	0.79 (.64)	0.55 (<.001)	0.48	126 (4.79)
Diagnostic groupings	Schizophrenia (n = 908) vs NSNAP (n = 742)	1.89 (.1)	0.31 (<.001)	0.58	86 (3.27)
	Schizophrenia (n = 908) vs affective psychoses (n = 590)	1.28 (.21)	0.24 (.04)	0.68	32 (1.21)
	NSNAP (n = 742) vs affective psychoses (n = 590)	0.60 (.49)	0.22 (.31)	0.59	32 (1.21)
Premorbid adjustment	Good (n = 1118) vs poor (n = 1122)	1.76 (.03)	0.21 (.04)	0.72	16 (0.61)
Response to treatment	Good (n = 1639) vs poor (n = 537)	5.02 (<.001)	0.28 (.010)	0.67	12 (0.46)
Course pattern	Remitting (n = 993) vs partial remitting (n = 774)	2.34 (.02)	0.19 (.18)	0.67	16 (0.61)
	Remitting (n = 993) vs chronic/continuous (n = 473)	4.79 (<.001)	0.34 (<.001)	0.60	20 (0.76)
	Partial remitting (n = 774) vs chronic/continuous (n = 743)	2.45 (.2)	0.30 (.024)	0.74	11 (0.42)

Note: S = network global strength, M = network structure (maximum difference value in any of the edge weights of the compared networks); SI = similarity index; NSNAP = nonschizophrenia non-affective psychosis.

^aNumber (and %) of edges that significantly differ across a pair of networks over the total of edges (n = 2628).

NCT centrality test of EI values across subgroups showed that the major differences occurred between first/multi-episode vs community samples, schizophrenia vs affective psychoses diagnoses and remission vs chronic courses (supplementary table 1), which is in line with the above-mentioned finding that the degree of symptoms’ connectivity is highly dependent on the severity of the clinical picture.

Discussion

Main Findings

In this study, we sought to examine the network structure of psychopathology in psychotic disorders according to several network parameters, such as the variability of networks across subgroups of subjects defined on the basis of meaningful variables. Five major findings may be considered in this study. First, the network structure of psychotic symptoms could be reliably estimated, since precision of network parameter estimates was high and the findings were robust; thus, the obtained network structure appears efficiently represent the “true” model underlying the data. Second, the symptoms of the psychoses were part of a single “connected component” in that the global network structure was formed by densely connected symptoms. Against this globally connected web of symptoms, a 7-domain symptom model structure best explained the data. This model included the domains of reality distortion, disorganization, catatonia, diminished expressivity, anhedonia/asociality, mania, and depression. Third, centrality and bridge analyses showed that the most interconnected symptoms pertained to the disorganization and, to a lesser extent, negative domains. Fourth, in congruence with the above-mentioned findings, the disorganization domain occupied a central position within the whole network and showed widespread connections with other psychopathological domains, which suggests that disorganization represents a true “bridge domain.” The core character of disorganization symptoms within psychotic disorders was already outlined by relevant authors in the field^36–38 and current research confirms its importance regarding etiopathogenic factors and real-world functioning.^39,40 This domain had a complex structure in that it was the domain comprising more and more diverse symptoms such as formal thought disorders, disorganized behavior, inappropriate affectivity, motor abnormalities, and attentional problems. Fifth, the network structure of psychotic symptoms differed according to important variables such as sampling method, diagnosis, premorbid adjustment, illness duration, treatment response, and course pattern. These differences were due to overall connectivity (remitting vs partially remitting psychosis), distribution of edge weights (sample type, schizophrenia vs affective psychosis and partially remitting vs chronic psychosis) or both (good vs poor premorbid adjustment, good vs poor treatment response and remitting vs continuous psychosis). Furthermore, more densely connected networks were indicative of less severe illness or better outcomes.

Comparison With the Literature

Our results align with those of previous research indicating that symptoms of psychotic and other mental disorders are part of a single connected component, the psychopathology web,⁴¹ in that a global continuity was found between most symptoms and syndromic domains of the psychoses, a finding that argues against the categorical view of psychopathological syndromes as discontinuous phenotypes and supports a dimensional view of psychopathology of psychotic disorders. We empirically identified a 7-domain structure of the psychosis network that adds to the widely accepted 6-domain structure the further separation of negative symptoms into diminished expressivity and avolition/anhedonia subdomains. These subdomains have been reported to have factor validity^42,43 and network structure validity,⁴⁴ such as differential aetiopathogenic and functional underpinnings^45,46 including those reported in a network analysis approach.⁴⁷

The previous study most comparable to ours also used the CASH to rate symptoms⁹ and reported an overall network structure highly resembling our 7-domain model, although the clustering model was not statistically tested. Furthermore, this study also showed strong connectivity between individual CASH symptoms both within- and between domains. However, some important differences were observed between the studies regarding centrality parameters, which may be explained by variability among study samples and power to obtain reliable results, since, eg, the previous study included only 408 male subjects with a diagnosis of non-affective psychoses.⁹ In line with previous studies of psychosis^11,48 and schizotypal personality traits,^33,49 we found that there are more within-domain than between-domain connections, which argues for a strong clustering of the symptoms in the network. Interestingly, a network analysis of 74 schizotypal personality traits, many of which are similar, at the subclinical level, to the CASH symptoms, revealed, as in the present study, that some disorganization traits were among the most central features of the network and that the network structures of men and women were similar.³³

It has been suggested that the group-level connectivity between symptoms may be a marker of treatment responsiveness in that more densely connected networks would feature stronger self-reinforcing feedback loops among symptoms, which may play a role in the maintenance of psychopathology and negative outcomes.^50,51 However, the existing evidence is contradictory in this regard both in mood^51,52 and in psychotic disorders,^48,53–56 In line with one previous study,⁵³ our results indicate that densely connected networks predict good treatment response. We also found more densely connected networks in subjects with good premorbid adjustment or remitting psychosis than in their counterparts; thus, our results suggest that densely connected symptom networks may be a marker of good outcome overall. In a similar vein, Strauss et al⁵⁴ found that negative symptom networks were less densely connected in subjects with schizophrenia than in those with bipolar disorder. However, given the uncertainty within this area, the putative association between densely connected networks and better outcomes needs to be confirmed by collecting reliable measurement of symptoms over time.^56,57

Implications

Overall, the symptom network structure of the psychoses was highly dependent on sampling method and important clinical features; hence, it is plausible that an invariant symptom network structure of psychotic disorders does not exist at all. Indeed, the network structure of symptoms needs to be understood within a multilayered network framework,⁸ in which a given structure would be the result of a combination of risk factors operating at several levels of illness development together with some disorder-specific factors.

We found evidence that the disorganization domain occupied a central position in the network, acting as a bridge domain; hence, and according to the network theory, focusing treatment on these symptoms has potential relevance for alleviating the whole clinical picture of psychotic disorders. It is often claimed that antipsychotic medications are particularly helpful for positive symptoms, including the disorganization ones,^58–60 but are less helpful for other psychopathological domains. However, converging evidence indicates that improvement in response to antipsychotic drugs spans across specific diagnoses of psychotic disorders⁵⁸ and that a global effect, rather than domain-specific effects, explains, on average, 80% of the variance in symptoms change.⁶¹ Thus, it can be speculated that this general effect of treatment might be largely mediated by the disorganization domain through its connectedness with other domains. Indeed, the most central symptoms predicting treatment responsiveness in the Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE) study pertained to the disorganization domain as defined in the present study.⁶² This is a testable hypothesis that should be addressed using prospective designs and mediation models examining the effect of antipsychotic drugs on psychopathology controlling for symptomatic domains.

With respect to the theoretical underpinning of network psychometrics stating that networks are the result of mutual influence among symptoms rather that one or more underlying factors,^3,50 our results suggest that the 2 approaches are not necessarily antagonistic but complementary. Indeed, the strong cluster structure of symptoms arising within the framework of a highly interconnected web may be indicative of both general and domain-specific factors underlying the networks. Furthermore, this conceptualization is broadly consistent with hierarchical bifactor models emphasizing general and specific dimensions of psychosis,^63,64 and mental disorders overall.⁶⁵ This convergence is not surprising, since factor models appear to be nested within network models.^41,66

Strengths and Limitations

This study was marked by several strengths, including the large, diagnostically heterogeneous sample and the use of a comprehensive and semi-structured interview with strong psychometric properties to characterize the network structure of the psychotic pathology. Moreover, our findings move beyond those from previous research in providing evidence of reliable network parameters, the clustering of symptoms in the network and the influence of sampling method and clinical variables in the network.

Despite these strengths, several limitations warrant consideration. First, our study population was made of samples of convenience; thus, it is not representative of the general population of psychotic disorders. However, the primary samples were ascertained through consecutive admissions or psychiatric case register for a defined catchment area; therefore, our study population may approach a representative sample of treated subjects with psychotic disorders. In support of this, the percentage diagnostic breakdown of our sample was comparable to that reported in an epidemiological study of psychotic disorders.⁶⁷ Second, the prevalence of certain diagnostic categories fell slightly below what has been recommended for network analysis,⁶⁸ which precluded the generation of networks within each diagnosis of psychotic disorder. Understanding how specific diagnoses are similar or distinct in terms of their overall networks and central features might be important in clarifying the nosology of psychotic disorders. Third, findings were cross-sectional, which limits our ability to draw causal inferences, and the direction of influence among symptoms in the network is unclear. Fourth, in comparing networks, subgroups had unequal number of subjects, which may have introduced some bias in the parameter estimation by the NCT; to avoid sample size differences, we performed subsampling from the largest group,³⁴ which resulted in virtually the same results. In line with this, simulation studies have showed that an n = 500 achieves a correlation between the “true” and estimated networks above 0.90 for edge weights and strength and above 0.8 for sensitivity.²⁴ Fifth, several CASH items may reflect similar underlying constructs and future network analyses should consider more parsimonious methods for constructing symptom networks of psychosis (ie, reducing the number of nodes). On the other hand, anxiety symptoms, which have been shown to be important when modeling psychosis from a network perspective,¹⁰ and lack of insight, among other features, need to be incorporated within future studies to fully map the network structure of the whole psychosis phenotype.

Lastly, a mention needs to be made regarding the current debate about the replicability of network structures in psychopathology. Some authors consider that network structures are replicable and generalizable⁶⁹ while others hold that networks have limited replicability.⁷⁰ This debate, however, appears to be more a question about the definition of replicability and its criteria than a problem of replicability itself. Failure to replicate and failure to generalize across samples need to be distinguished,⁷¹ and research should critically interrogate the data to determine if inconsistency between studies is best characterized as a failure to replicate or a failure to generalize across samples, clinical variables and other contextual factors. In fact, our study by providing robust results on the network structure of psychotic symptoms and the factors influencing it may shed some light on the replicability issue. In other words, our data support that lack of replicability across studies may be, at least in part, due to differences in sampling methodology and clinical factors, findings that need to be replicated by other authors using similar study samples and methodology.

Funding

The study was funded by the Carlos III Health Institute (FEDER Funds) from the Spanish Ministry of Economy and Competitiveness (grant number PI16/02148) and the Gobierno de Navarra (grant number 31/17). The funding sources had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.