Exploring interrelationships between cognition, functioning and quality of life in schizophrenia spectrum disorders: a Bayesian analysis of networks

Abstract

Cognitive impairment is a prevalent feature in schizophrenia spectrum disorders (SSDs), significantly impacting functional ability and quality of life. Although network analysis has been used in recent research, prior studies have frequently overlooked integrating existing knowledge, hindering the understanding of these complex associations. Using a Bayesian analysis of networks, we propose an innovative method incorporating prior knowledge through the utilization of multiple informed prior distributions. We analyzed data from 1150 individuals diagnosed with SSDs. Seven nodes, including cognitive variables, functional capacity, and subjective quality of life (SQL) indicators, were examined. Within a Bayesian framework, we estimated a network of partial associations and constructed a network to quantify the evidence of edge presence and absence, employing multiple informed priors derived from previous network studies. Our analysis uncovered robust associations between cognition (specifically verbal memory and processing speed) and functioning, as well as between functioning and SQL, supported by substantial evidence. While the absence of relationship between cognition and SQL was uncertain with a uniform prior, evidence of absence was observed with the use of informed priors from previous studies. This study underscores the intricate interplay among cognition, functioning, and quality of life within SSDs. Specifically, our results reveal associations between verbal memory and processing speed with functioning, whereas no association was found between cognition and quality of life. Integrating prior knowledge through a Bayesian framework facilitates nuanced insights and may contribute to more reliable inferences.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00406-025-02084-y.

Introduction

Cognitive deficits represent a significant aspect of schizophrenia spectrum disorders (SSDs), alongside positive and negative symptoms [1]. These cognitive impairments profoundly impact the functional outcomes and quality of life of individuals affected by SSDs [2,3]. Individuals with SSDs demonstrate cognitive alterations in various domains such as processing speed, attention, memory, and executive functioning [4,5], with some cognitive domains being more strongly linked to functional outcomes and quality of life than others [6,7]. For example, previous research have indicated robust connections between memory functions and functional outcomes, such as activities of daily living, whereas the impact of executive functions in this context appears to be less important [7]. Furthermore, some studies have shown that executive functioning and working memory are more closely associated with subjective quality of life (SQL) [8–10] than other cognitive domains. The intricate interplay among cognitive deficits, functional impairments, and SQL in SSDs highlights the complexity of such relationships and underscores the importance of gaining a comprehensive understanding.

In the past years, the network approach has gained increased interest in psychopathology research for modeling complex systems [11]. Network approach allows researchers to construct a partial correlation network, highlighting the relationships and interactions among various variables.

Recent research on SSDs has increasingly incorporated cognitive variables into psychological network. While some studies focused on cognition [12–16], others have incorporated clinically relevant factors including symptoms [17–22] and functional outcomes [23–27]. These studies have mainly shown that cognitive variables were closely related to each other and often appear as the most influential nodes in the network, displaying stronger association with functional measures as compared to clinical symptoms [24,25,27]. To date, only one study has explored the link between cognitive variables and quality of life in the context of SSDs [28], finding that cognition was more closely associated with quality of life than with functional outcomes.

Hence, current literature reveals a multitude of empirical findings on the intricate interplay between cognition, functioning, and quality of life. However, previous studies have overlooked insights from available research in their analyses, due to methodological limitations hindering the integration of prior knowledge. To address this limitation, a Bayesian analysis of network emerges as a promising approach [29–31]. The Bayesian framework allows for the integration of prior knowledge into analyses by using prior distributions [32,33]. As new data becomes available, Bayesian inference updates these priors to posterior distributions, reflecting the combined information from the prior and the observed data. This iterative process forms a coherent cycle of inference, wherein prior beliefs are continuously refined and revised based on empirical evidence.

The aim of the present study focuses on two key components, encompassing both clinical and methodological dimensions. As a first step, we estimated a network analysis in a Bayesian framework to identify the unique pairwise associations between cognition, functioning and quality of life in individuals with SSDs. As a second step, we introduced a method to incorporate prior knowledge about cognition, functioning and quality of life in individuals with SSDs. To this end, multiple informed prior distributions were defined for each edge within the network, taking in account the presence or absence of edges reported in previous network studies. To date and to our knowledge, no study has employed such priors in the context of psychometric network analysis.

Materials and methods

Study design and participants

Our dataset consists of patients enrolled in the REHABase cohort. The cohort has been extensively described elsewhere [34] and consists of patients living with a severe mental illness referred to one of the 15 French clinics for psychosocial rehabilitation. Clinically stable patients are referred to these centers by general practitioners, private psychiatrists, public mental health agencies, or they may self-refer. Patients enrolled in the REHABase cohort undergo a standardized clinical, functional, and cognitive evaluation by a multidisciplinary team made up of psychiatrists, nurses, neuropsychologists, occupational therapists, and social workers. They subsequently receive an individualized rehabilitation care plan over the course of a year after undergoing a standardized clinical, functional, and cognitive evaluation by a multidisciplinary team made up of psychiatrists, nurses, neuropsychologists, occupational therapists, and social workers (including various psychosocial interventions such as cognitive remediation, social skills training, vocational rehabilitation, etc.).

The study obtained the authorizations required under French legislation (French National Advisory Committee for the Treatment of Information in Health Research, 16.060bis), including information processing (French National Computing and Freedom Committee, DR-2017-268).

Inclusion/exclusion criteria

The present study was restricted to patients included up to November 2022 and fulfilling the following inclusion criteria: (1) being diagnosed with SSDs according to the Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition criteria) and based upon clinical interview performed by a psychiatrist (completed by information collected from medical records), and (2) having benefited from a neurocognitive assessment. The current study used the baseline data and excluded patients with a mother tongue other than French and those with illiteracy.

Measures

Cognition

We used data collected in the REHABase on the following cognitive measures:

• Verbal episodic memory: the French adaptation of the Free and Cued Selective Reminding (FCSR) test [35] was used to assess episodic memory. The FCSRT includes a learning list of 16 written words presented with a semantic cue to control for memory encoding. Participants were asked to retrieve the words spontaneously, then with the help of the semantic cue during three trials. Finally, free and cued delayed recall were examined after 20 min. Total delayed Free Recall scores were used in this study, ranging from 0 to 16, (high score indicating better performance).

• Working memory: the digit span of the Wechsler Adult Intelligence scale 4rd edition (WAIS-IV) [36] was used to measure verbal working memory. The participant hears a sequence of digits (numbers) of increasing length and is asked to recall each sequence in the reverse order (backward digit recall) (score 1–19, high score indicates better working memory performance).

• Executive function: the Multiple Errands Test [37] modified version [38] was used to assess executive abilities in everyday functioning through a number of real-world tasks. Using a neighborhood map, the participant is asked to find a route for shopping while respecting instructions and rules regarding transport (using logical routes to save time, for instance) and schedules of the places to visit. Two scores are obtained: “completion time” assessing the time (mn) taken to complete the itinerary; “total error score” assessing the numbers of errors (logical errors, useless detours and schedule respect). The total error score was used in this study (lower score indicates better performance).

• Attention and processing speed: the D2-Revised (D2-R) [39] was used to measure selective attention and processing speed. The test consists of 14 rows (trials), each with 60 randomly mixed ‘p’ and ‘d’ letters marked with one, two, three, or four small dashes. The target symbol is a ‘d’ with two dashes (hence ‘d2’). The participant is instructed to cancel out as many target symbols as possible, moving from left to right, with a time limit of 20s/trial. Two scores were used: CCT measures processing speed (total number of processed items; high score indicates better performance); number of errors + omissions, measures precision in data processing; lower score indicates better attentional performance).

Functioning and quality of life measures

We used data collected in the REHABase on the following scales:

• Global Assessment of Functioning (GAF) scale [40]: a clinician-rated global measure of psychological, social, and occupational functioning. Scores can range from 1 to 10 (Persistent danger of severely hurting self or other, or persistent inability to maintain minimal personal hygiene, or serious suicidal act with clear expression of death), to 91–100 (Superior functioning in a wide range of activities, life’s problems never seem to get out of hand, is sought out by others because of his or her many positive qualities. No symptoms).

• Schizophrenia Quality of Life-18 (S-QoL18) [41]: 18-item self-administered questionnaire describing 8 dimensions: Psychological Well-being, Self-Esteem, Family Relationships, Relationships with Friends, Resilience, Physical Well-being, Autonomy, and Sentimental Life. It also includes a total score, which was used in the current study (score 0–100; high score indicates better Quality of Life).

Statistical analyses

The statistical software R (version 4.2.2) was used to carry out the statistical analysis [42]. The R-package BGGM [43], qgraph [44] and easybgm [45] were used to estimate and construct the networks.

Variable selection

We followed two steps to combine the evidence from previous research and our current data. First, we individually specified and derived priors for each edge from previously published studies (see Sect."Posterior odds network"). Subsequently, to closely align with previous research, we only selected the most comparable variables within our dataset. Notably, psychopathological variables (e.g., negative or positive symptoms) were not systematically collected within the cohort, therefore they were not included in our study. As a result, our study focused on five cognitive variables, alongside total scores for functioning and SQL.

Data preparation

Missing data accounted for 22.9% of the total observations. In line with previous network analyses studies that had equivalent (e.g., 17.4%) [46] or higher amount of missing data (e.g., 39%) [47], we imputed missing data. Within our analysis using the R-package BGGM [48] missing values were addressed by imputing them using the respective posterior predictive distribution. This technique is deemed appropriate based on the underlying assumption that the missingness occurred at random [49]. In order to assess the potential impact of imputing data, we conducted a comprehensive sensitivity analysis of different missing data handling approaches and compared the estimated parameters. In particular, we used Multiple Imputation by Chained Equations (MICE), listwise deletion, and multiple imputation models with auxiliary variables (i.e., age, gender, education) [50]. Minor differences were observed among these three models, yielding highly similar parameter estimates, thereby reinforcing our assumption that our data were missing at random. Detailed information regarding these comparisons is provided in the Supplementary Materials (Supplementary Figs. S1 and S2).

Network estimation

Bayesian analysis of networks

The Bayesian approach specifies prior beliefs and updates them with the data using Bayes’ rule to obtain a posterior distribution of a network’s structure and parameters. Therefore, the network analysis in a Bayesian framework allows quantifying the evidence of edge presence and absence and expresses the precision of the parameters through posterior parameter distributions [29,30]. To estimate and construct the network, we adopted the Bayesian analysis of networks as outlined by Williams [43] and its associated BGGM package. The package stipulates an F-distribution matrix as a prior distribution on the precision matrix, with a standard deviation (SD) set at 0.5 [31,43]. In addition, a sensitivity analysis of the prior standard deviation was conducted, with values set to 0.25 and 0.7, to assess the impact of the prior on posterior estimates. These values were chosen to represent different levels of uncertainty incorporated into the prior beliefs. Full results and methodological details are provided in the Supplementary Materials.

Partial association networks

First, we illustrated the results as networks, where edges represented the partial association between two nodes. Edges represent the strength of association between two variables after having controlled for the influence of all other variables in the network.

Posterior odds network

Prior elicitation approach for the informed edge-specific prior: To determine eligible studies and identify consistent patterns of associations between nodes across various studies, we conducted a systematic search of published network analyses on SSDs. To be eligible, studies had to fulfill the following criteria: (a) publication date up to March 2024, (b) providing networks including at least two distinct neurocognitive measures, (c) inclusion of both male and female participants. Studies that stratified samples into multiple groups based on a particular criterion or incorporating covariates such as age or sex were excluded. In total, 9 out of 19 studies on SSDs met the inclusion criteria (further details are provided in the Supplementary Materials).

From the eligible studies, we extracted the binary information whether the edge between node i (e.g., memory) and j (e.g., functional outcomes) was assessed and present in study s or absent (see Supplementary Tables, 1 and 2). We assumed that the inclusion probability of each edge θ_ij follows a beta-distribution with parameters α and β. If a previous study assessed and found an association for the edge, we updated α. Each time a previous study assessed but did not find an association for the edge, we updated β. As such, α_ij = α’ + ∑ α₀δ_s and β_ij = β’ + ∑ β₀δ_s, where the sum extends over all the studies that assessed the link θ_ij. α’ and β’ indicate a general prior information of edge inclusion if we had not observed any previous studies. α₀ and β₀ indicate how much weight we assign to updating our prior beliefs (the higher their values the stronger our prior beliefs). Since the frequentist approach cannot test for edge absence and as such we cannot be certain whether an edge was excluded because it was too unstable or truly not there, we specified β₀ = α₀/2; thus, the prior is updated more strongly for edge presence. δ_s represents a quality penalization term to address the variability in quality among the selected studies. Our criteria for assessing study quality were based on two factors: the number of subjects and the number of nodes within each study. A higher-quality study was characterized by a larger number of subjects and a lower number of nodes, indicating more robust inference. The informed prior edge inclusion probability is determined as the expected value of the beta-distribution which is θ_ij = α_ij/(α_ij + β_ij). α’ and β’ were set to 2. To perform a sensitivity analysis on this informed prior, we varied the strength of updating factor α₀ and β₀ by 1, 4, 8, and 2, 4 respectively (see Supplementary Fig. S6).

Inference of network edges via Bayesian methods: To assess edge inclusion in the network, we computed the Bayes Factor (BF) to quantify the relative likelihood of edge inclusion versus exclusion (see Supplementary Fig. S4). Using this BF, we then calculated posterior odds by integrating prior information into our analysis. The posterior odds quantify the relative probability of edge inclusion after accounting for both the prior odds and the evidence provided by the data. This approach extends beyond the BF by incorporating informed prior distributions for edge inclusion, allowing us to integrate existing findings and refine our inference based on cumulative evidence. This methodology enhances the robustness of network estimation by balancing prior knowledge with empirical support from the current dataset.

The interpretation of posterior odds follows the same convention as BF. A posterior odds larger than 1 indicates some evidence for edge inclusion, while a posterior odds larger than 3 is interpreted as weak evidence and a posterior odds larger than 10 as strong evidence for edge inclusion [51]. Conversely, a posterior odds ratio smaller than 1 provides some evidence for edge exclusion, with values below 1/3 indicating weak evidence and those below 1/10 constituting strong evidence for exclusion. The posterior odds will then be visualized as the edge evidence network, in which edges represent the corresponding posterior odds. Blue edges indicate strong evidence for inclusion (i.e., Odds > 10), grey edges reflect inconclusive evidence (i.e., 0.1 < Odds < 10), and red edges denote strong evidence for edge absence (i.e., Odds < 0.1).

Results

Sample characteristics

After applying exclusion criteria (i.e., non-native French speakers, n = 85; individuals with illiteracy, n = 3), a total of 1,150 individuals diagnosed with SSDs were included in the study. The majority of the participants were male (76.7%). The mean age was 32.7 years (SD = 9.81). The sociodemographic and clinical characteristics of the sample are reported in Table 1.

Table 1.

Clinical and socio-demographic characteristics of participants

	Total (N = 1150)	Male (N = 882)	Female (N = 268)
Age, mean (SD)	32.7 (9.81)	32.2 (9.41)	34.3 (10.9)
Education
< 12 years	619 (53.8%)	454 (51.5%)	165 (61.6%)
≥ 12 years	531 (46.2%)	428 (48.5%)	103 (38.4%)
Illness duration
Mean (SD)	11.4 (9.00)	11.0 (8.49)	12.6 (10.4)
Current antipsychotic medication
First generation	75 (6.5%)	57 (6.47%)	18 (6.69%)
Second generation	736 (64%)	551 (62.54%)	185 (68.77%)
Both	283 (24.6%)	224 (25.43%)	59 (21.93%)
None	56 (4.9%)	49 (5.56%)	7 (2.60%)
Diagnosis
Schizophrenia	766 (66.6%)	612 (69.4%)	154 (57.5%)
Schizoaffective disorder	211 (18.3%)	135 (15.3%)	76 (28.4%)
Brief Psychotic Disorder	48 (4.2%)	37 (4.2%)	11 (4.1%)
Schizophreniform Disorder	48 (4.2%)	41 (4.6%)	7 (2.6%)
Other	77 (6.7%)	57 (6.5%)	20 (7.5%)

Network output

Partial associations network and inclusion Bayes factor

The partial association network is shown in Fig. 1 (the Highest Density Intervals (HDI) for all the parameters are detailed in Supplementary Fig. S3). Strongest associations in the network were found between cognitive nodes, notably between Attention and processing speed (θ = 0.387, HDI [0.310,0.424]), and between processing speed and working memory (θ = 0.368, HDI [0.287,0.412]). GAF was positively related to verbal episodic memory (θ = 0.189, HDI [0.111,0.262]), processing speed (θ = 0.123, HDI [0.050,0.194]) and to SQL (θ = 0.141, HDI [0.073,0.206]), all of which demonstrated strong evidence for inclusion based on the Bayes Factor (inclusion Bayes Factor > 10) (Supplementary Fig. 4). SQL was weakly connected to executive function (θ = 0.084, HDI [0.016,0.151]); however, there was insufficient evidence to conclude the inclusion or exclusion of this association (0.1 < inclusion Bayes Factor < 10). In total, the inclusion Bayes Factor network showed evidence for exclusion for 5 edges (inclusion Bayes Factor < 0.1), including edges linking cognition with functioning, as well as those connecting cognition with SQL. Importantly, Attention and executive function nodes are quantified by the number of errors. In our partial association network, an increased occurrence of errors in the assessment of executive function is thus associated with a decreased score in working memory.

Fig. 1.

Network displaying the partial associations between cognition, functioning and subjective quality of life. Blue edges indicate positive partial correlations while dashed red edges indicate negative partial correlations. The thickness of the line indicates the strength of the relationship. Attention and executive function variables are quantified by the number of errors (e.g., a high number of errors in the task measuring executive function is associated with a smaller working memory span). s_qual_life, subjective quality of life; verb_mem, verbal episodic memory; work_mem, working memory; proc_speed, processing speed; exec_func, executive function.

Posterior odds network

While the Inclusion Bayes Factor provides initial evidence for edge inclusion based solely on observed data, it does not account for prior knowledge regarding network structure. To address this limitation, we incorporated informed prior distributions for edge inclusion, refining the inference process. As a result, discrepancies were identified between the inclusion Bayes factor network (see Supplementary Fig. S4) and the posterior-odds network (Fig. 2). As compared to the inclusion Bayes factor network, establishing informed prior distributions for edge inclusion provided evidence of exclusion for two connections (between SQL and processing speed, and between GAF and Attention), which showed insufficient evidence for exclusion in the inclusion Bayes factor network.

Fig. 2.

Networks displaying the evidence of edges between cognition, functioning and subjective quality of life. Blue edges indicate evidence supporting inclusion (Odds > 10), red edges indicate evidence supporting exclusion (Odds < 0.1), and grey edges indicate a lack of evidence for either inclusion or exclusion (0.1 < Odds < 10). For easier reading each edge evidence network was divided in two plots: the left plot represents edges with some degree of evidence for inclusion (Odds > 1), while the right one represents edges with some degree of evidence for exclusion (Odds < 1). s_qual_life, subjective quality of life; verb_mem, verbal episodic memory; work_mem, working memory; proc_speed, processing speed; exec_func, executive function.

Discussion

To date and to our knowledge, this study provides a novel perspective on the interrelations between cognition, functioning, and quality of life in individuals with SSDs by employing a network analysis within a Bayesian framework. This approach enabled us to quantify the stability of the network and assess the presence or absence of edges with greater nuance. Methodologically, the use of informed prior distributions enabled the integration of previous findings, resulting in a posterior-odds network that synthesizes both our data and prior evidence. The main findings may be summarized as follows: among cognitive measures, only verbal episodic memory and processing speed were related to functional outcomes, SQL was not associated with any cognitive node, only with functioning [52].

Indeed, according to the posterior-odds network, only verbal episodic memory and processing speed showed strong evidence supporting inclusion of a relation with functional outcomes. This is in line with previous research in individuals with SSDs indicating that verbal memory and processing speed are among of the most consistent correlates or predictors of functioning [6,7,53,54], especially when assessed with the GAF scale [55–58]. Moreover, our results are consistent with recent network analyses that found a direct association between processing speed and functioning [18,27]. For example, Galderisi et al., [24,25] found a relationship between processing speed and daily living skills, which can be seen as a more focused extension of global functioning. Similarly, Hajduk [18] observed that processing speed was associated with work skills and activities. Overall, our study reinforces the central role of processing speed in functional outcomes, regardless of the specific assessment tool used. However, other network studies failed to observe any direct association between processing speed and functioning [24,28], nor between verbal episodic memory and functioning [18,24,27,28]. The discrepancies may be explained by the inclusion of other variables within the network that bridged these relationships. Indeed, in most studies, verbal episodic memory was indirectly connected to functional outcomes through functional capacities [18,24,27], while processing speed was indirectly related to functioning through psychopathological measures such as disorganization [28] or negative symptoms [18]. Although previous network studies did not find any direct association between verbal episodic memory and functioning (which led to a weak prior in favor of the presence of such a link), the association was so robust in the present research that the prior had no influence on the subsequent estimation.

The present study identified evidence supporting the inclusion of the edge between GAF and SQL, suggesting that patients with low level of functioning are subject to a poorer self-reported quality of life. This is consistent with previous research [59] and with a previous network study where SQL was directly connected to functional outcomes [28].

Although literature has shown consistent findings regarding relationships between cognition and objective quality of life (i.e., clinician-rated) in individuals with SSDs, inconsistent results have been reported regarding the relationship between cognition and SQL. Some authors have reported positive associations [8–10] whereas others reported negative ones [2,60,61]. Others, in line with our results, failed to identify any relationship [62–64]. The discrepant results may stem from the use of different instruments, some of which were developed for the general population and might be less sensitive to variations in individuals with a specific illness, such as schizophrenia, potentially reflecting a more comprehensive assessment of quality of Life [2,65].

The main strength of the current study lies in the use of the Bayesian framework, which allowed us to incorporate several well-informed prior from previous studies. This approach reduced inconclusive evidence regarding the absence of edges as compared to solely relying on the Bayes factor. Indeed, when solely relying on the Bayes factor, the edges between processing speed and SQL and between attention and GAF were inconclusive for absence (see Supplementary Fig. S4 for a visual comparison). However, when taking previous knowledge into account, the posterior odds provided evidence for the exclusion of these relationships. This highlights the critical importance of determining a Bayes Factor within the framework of existing prior knowledge. Furthermore, the Bayesian approach facilitated the precise quantification of evidence regarding the presence or absence of edges, providing a more robust analysis. In contrast, previous network analyses relied on frequentist inference, lacking the ability to effectively quantify such uncertainties, and consequently leading to potentially risky conclusions about the network structure. In the present study, numerous partial associations have been identified as uncertain when quantifying the evidence regarding the absence of edges, emphasizing the need for caution in their interpretation.

Interpretation of the present results is affected by limiting factors. Firstly, our study is of cross-sectional nature, which does not allow us to establish the direction of the effect. For instance, it is not possible to determine the hierarchical nature of cognitive processes. Additional research is necessary to better understand the underlying causal mechanisms, thereby informing more effective approaches to facilitate recovery. Secondly, despite a substantial sample size, our study demonstrated a marked male predominance, which may limit the generalizability of the findings to predominantly female populations. Additionally, as our sample was drawn from specialized psychosocial rehabilitation clinics, it may not fully represent individuals with schizophrenia spectrum disorders in acute or general psychiatric settings. Therefore, future research should prioritize more gender-balanced and heterogeneous samples to better clarify the observed relationships and enhance the generalizability of results. Thirdly, the considerable heterogeneity in tests used to assess cognition, functioning or quality of life across studies is challenging. Indeed, using distinct assessments for similar concepts make comparisons within studies more complex, which in turn complicates our analyses. For instance, the instrument used in the current study to assess functioning (GAF) also measures psychopathological domains which may influence the association. This limitation poses significant challenges in accurately identifying and inferring the intricate relationship between cognitive processes, functioning and psychopathological manifestations. More specific functional scales, such as the Personal and Social Performance Scale, the Specific Levels of Functioning Scale (SLOF), or the University of California, San Diego Performance-Based Skills Assessment (UPSA), could have provided more nuanced insights. Regarding cognition, we derived our prior for executive function from studies that relied on problem solving and flexibility measures. In contrast, the test used in the present study did not purely assess executive functioning which did not allow us to precisely specify the impaired executive function(s) [38]. Similarly, while we used a free recall measure for VEM, our prior was derived from both verbal learning and free recall measures. Although part of the same cognitive domain, they do not strictly measure the same process. Standardized and homogeneous assessment measures are critical to enhance the robustness of future network studies exploring interrelationships between cognitive function, functioning, and SQL. Fourthly, while the focus of our study was to investigate the relationships between cognition, functional outcomes and SQL, all previous network studies that included these variables also encompassed variables across psychopathological domains. The unavailability of psychopathological variables in our dataset may have constrained the comprehensiveness of our analysis. Moreover, certain factors, such as treatment variables, are known to influence cognitive functioning. As previously highlighted, the presence or absence of specific variables in a network significantly impacts its structure and affects relationships between variables [66,67]. Selecting which variables should be included in a network is a recurring topic in the literature, raising concerns about the replicability of networks [67–69]. However, to date, no comprehensive guidelines addressing this specific concern are available which might contribute to some disparities between previous results and our present findings.

Conclusion

In the current study, we used a Bayesian analysis of network to explore the interplay between cognition, functioning, and quality of life in individuals with SSD. Our findings underscore noteworthy associations between cognition (specifically verbal episodic memory and processing speed) and functional outcomes, whereas no association between cognition and SQL were observed. Furthermore, the study highlights the strengths of the Bayesian framework in network analysis. In particular, integrating informed prior knowledge for each edge allowed us to incorporate relevant contextual information into the analysis, which may help reduce uncertainty in the estimated network and enhance the robustness of the inferences drawn.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

TB, KH, MM, LFL, RR: conception and design of the study. TB, KH, MM, RR: data analysis/interpretation. TB, KH: drafting the article. All authors have revised the manuscript. All authors contributed to the article and approved the submitted version.

Funding

The Authors have declared that there are no conflicts of interest in relation to the subject of this study.