Introduction

Schizophrenia is a chronic and heterogeneous mental disorder, ranked among the top 20 leading causes of disability worldwide1,2,3,4. It is associated with impairment in neurocognitive and social cognition performance, which, together with a broad range of symptoms (positive symptoms, disorganization, and negative symptoms), concurs in determining poor functional outcomes5,6,7,8,9,10,11,12,13. Cognitive impairment has long been recognized as a fundamental aspect of schizophrenia14,15,16. Indeed, Kraepelin documented disturbances in memory, attention, motor function, and perception, which align with contemporary research findings17. As a matter of fact, an impairment in multiple domains of neurocognition, including processing speed, verbal memory, and working memory, as well as of social cognition, compared to healthy controls and people with affective disorders, is commonly observed in people with schizophrenia. Moreover, in people with schizophrenia, cognitive impairment is associated with poor functional outcome and is increasingly viewed as a treatment target to improve recovery18,19. A recent meta-analysis focusing on individuals with early psychosis revealed that impairments in all domains of neurocognition and social cognition are correlated with lower levels of psychosocial functioning, both in a cross-sectional and longitudinal context20.

The complex associations of illness-related aspects, such as cognitive impairment and psychopathological variables, with psychosocial functioning have been investigated in cross-sectional and follow-up studies, through linear regressions and structural equation models (SEM). In a study by Mucci et al.21, SEM identified 5 variables (ie, neurocognition, social cognition, positive symptoms, the experiential domain of negative symptoms, and availability of incentives) that were directly associated with real-life functioning at four-year follow-up. The results of this study also showed that neurocognition was the variable with the strongest association with everyday life and work skills, as well as with interpersonal relationships through the mediation of social cognition. Other SEM studies suggested that neurocognitive impairment, specifically in processing speed and verbal memory, contributed directly and indirectly to impaired real-life functioning22,23,24. A limitation of the structural equation models is the need of predetermined assumptions concerning the selection of predictors, mediators, precursors, and outcomes. Therefore, to investigate the complex patterns of relationships of these variables, network analysis represents a robust methodological approach25,26,27. This data-driven technique operates without the necessity of preconceived assumptions about variable relationships28. A few studies used a network-based approach to identify complex associations among psychopathological variables, personal resources, context-related factors, and real-life functioning in schizophrenia. The first study using network analysis29 proved that social cognition, neurocognition, resilience, and real-life functioning represent robust and independent constructs, with functional capacity and everyday life skills playing a key role in the network, partly bridging the effects of cognitive impairment and symptoms on other domains of functioning. These relationships were substantially stable in a four-year follow-up study conducted by the same group30. The results of these studies have been independently replicated by Moura et al.31, confirming the interplay of functional capacity, everyday life skills, interpersonal relationships, neurocognition, and avolition in schizophrenia-spectrum disorders. However, despite the capacity to capture the connectedness of multiple variables and to identify those with the highest number of associations with functioning, the network analysis does not provide information on causal interactions between variables and the direction of effects. Understanding the causal relationships among various variables could prove beneficial in identifying therapeutic targets which might have the highest impact on the other interconnected variables.

Recently, research has taken advantage of probabilistic graphical models based on directed acyclic graphs (DAGs) to get improved insights into the complexity of the mechanisms linking variables to each other32. A Directed Acyclic Graph is a graph with nodes only connected by arrows and does not include directed cycles. The directed edges in these graphs are frequently interpreted causally, such that the edge A → B is interpreted to mean that the variable A causally influences the variable B. However, the consequence of using a DAG is that directionality is forced, and the possible influence of latent variables is not accounted for. To overcome these limitations, researchers have implemented Partial Ancestral Graphs (PAGs). Partial Ancestral Graphs are a methodological framework used to represent and analyze causal relationships when the presence of latent variables or unmeasured confounders is possible. PAGs extend the standard DAG model by allowing for a more flexible depiction of causal relationships, particularly in cases where not all variables are observable. In a PAG, edges between nodes can take different forms—directed, undirected, or bidirected—each indicating different potential relationships, such as direct causality, correlation due to latent confounding, or ancestral relationships33. These qualities make PAGs particularly useful in observational studies where hidden confounders might bias causal inference. Like DAGs, also PAGs do not contain cycles.

To our knowledge, very few studies used a data-driven analysis based on partial ancestral graphs to identify causal links among multiple variables in psychiatry and related disciplines34,35,36,37,38. In one of these studies, using data on first-episode schizophrenia, variables were neurocognition, socio-affective capacity (as a proxy of social cognition), negative symptoms and social/occupational functioning34. Results demonstrated a complex interplay among the included variables: social functioning impairment caused motivational impairment (a measure that overlaps with the negative symptom domain of avolition), which in turn caused reduced occupational functioning. There was no causal link between neurocognition and real-life functioning, and only a causal link between negative symptoms and impairment in social cognition, which in turn caused motivational impairment. However, the authors did not use a specific and validated assessment instrument for social cognition, because their measure of impairment in this domain was derived from items of the Heinrichs-Carpenter Quality of Life Scale. Moreover, a composite score rather than individual domains was used to assess the neurocognitive impairment of patients with schizophrenia. Two studies, carried out by a research group on alcohol use disorders, highlighted, using the Greedy Fast Causal Inference (GFCI) algorithm, the unconfounded causal chain from social anxiety to drinking, not captured by other analytic methods35 and the intricate interplay between multiple constructs and pathways involved in the maintenance of drinking among those with internalizing pathology comorbidity36. Another study compared different methodologies (Structural Equation Models SEM; Fast Casual Inference, FCI; Fast Greedy Equivalence Search, FGES) to recreate the known ground-truth causal structure of Alzheimer’s disease data using different degrees of background “knowledge” and concluded that SEM was unable to recover the true graph37. The last is a recent methodological paper showing the challenges and opportunities of applying causal models to chronic pain data to guide improvements to existing treatments and develop new treatment strategies38.

To contribute further evidence of the opportunities offered by causal models in psychiatry, the aim of the present study was to investigate the causal relationships among impairment in neurocognition and social cognition, psychopathology, functional capacity and real-life functioning in a large sample of patients with schizophrenia, using state-of-the-art assessment instruments. The data-driven analysis based on partial ancestral graphs was used to identify causal links among these variables in order to provide additional insights for the identification of possible treatment targets to improve the outcome of people with schizophrenia. We also examined whether the relationships among variables held over time by replicating the analysis in the same sample after a 4-year follow-up period. We decided not to use follow-up data for implementing a longitudinal design because of the long interval between the two assessments and the overall stability of the sample characteristics.

Results

Characteristics of the study sample

Of the 921 patients who participated in the Italian Network for Research on Psychoses study at baseline, 618 patients provided follow-up data. No clinically and statistically significant differences were found in the baseline characteristics of participants and non-participants at the follow-up30. Patients participating in both waves were selected for the analyses. PAGs were obtained from patients with complete data on all selected variables (N = 612 at baseline and N = 602 at follow-up). The characteristics of patients at baseline and the summary statistics of the variables in the PAGs are shown in Table 1. At baseline and follow-up, virtually all patients were treated with antipsychotics (97% and 98%). The prevalence of anticholinergics use was recorded at follow-up: current use was 9.2%, and, in a subset of patients with information available, the use over the past 4 years was 6.5%.

Table 1 Characteristics of the study sample at baseline and follow-up.
Full size table

Baseline PAG

Table 2 lists the types of connections in the final PAG resulting from the bootstrap analysis. The table reports the probability that an edge is present (regardless of the orientation) for pairs of variables (Node 1 and Node 2), which in most cases was greater than 0.90. Furthermore, Table 2 reports in bold the most likely type of connection among those tested and the most likely orientation of the arrow: the final choice of arrow set in a PAG is related to the probability of an arrow being present and to the most likely type of arrow, among the types considered. Blue-and-bold directed arrows between a pair of variables indicate that the first variable can be modeled as a direct cause of the second; blue-and-not-bold directed arrows, indicate that the first variable can be modeled as a direct or indirect cause of the second variable; and edges with two circle marks, indicate an undetermined causal relationship between variables with the possible presence of a latent confounder. Detailed information on edge types and specialization can be found in Fig. 1, which graphically depicts the types of connections in the PAGs. For example, the blue-and-bold directed arrow from node 1 (Disorganization) to node 2 (Positive symptoms) indicates that a direct causal link is the most frequently occurring type of connection (with a probability of 0.7982 highlighted in bold) over the 1000 bootstrap replications, thus disorganization can be modeled as a direct cause of positive symptoms. Similarly, for the pair Working memory-Functional capacity the light blue arrow indicates that Working memory is estimated to be a direct or an indirect cause of functional capacity, with a probability of 0.355.

Table 2 Baseline edge probabilities obtained with 1000 bootstraps.
Full size table
Fig. 1
Fig. 1
Full size image

Causal inference: edge types and meaning in the PAGs.

Overall, the pathways of the baseline PAG (Fig. 2) suggested that working memory can be modeled as the first ancestor of the causal links among the nodes of the graph, as it directly affects the other 5 neurocognitive variables and through its direct effect on Facial Emotion Identification (FEI) the other 3 social cognitive variables. In addition, it affects (either directly or indirectly) functional capacity, which in turn directly affects daily functioning. Three pathways started from this domain of functioning, one to disorganization and positive symptoms and the other two to work skills and interpersonal relationships; the latter functioning domain in turn was directly linked to asociality and the other domains of negative symptoms.

Fig. 2: Baseline partial ancestral graph that best models the causal relationships among included variables.
Fig. 2: Baseline partial ancestral graph that best models the causal relationships among included variables.
Full size image

Blue-and-bold directed arrow: direct causal path; blue-and-not-bold directed arrow: indirect causal path; black edges with two circle marks: undetermined causal relationship between variables with the possible presence of a latent confounder. More details on edge types and meaning in Fig. 1. Alo Brief Negative Symptom Scale (BNSS) alogia, Anh BNSS anhedonia, Apa BNSS apathy, Aso BNSS asociality, Att Measurement and Treatment Research to Improve Cognition in Schizophrenia Consensus Cognitive Battery (MCCB) attention/vigilance, BAf BNSS blunted affect, Dep Calgary Depression Scale for Schizophrenia total score for depression, Dis PANSS disorganization, ELS Specific Level of Functioning Scale (SLOF) everyday life skills, FC functional capacity, FEI Facial Emotion Identification Test, Int SLOF interpersonal relationships, MSC MCCB Mayer-Salovey-Caruso Emotional Intelligence Test managing emotion section, Pos PANSS positive factor, PrS MCCB problem solving, PSp MCCB processing speed, SLe MCCB visuospatial learning, Stg stigma, Ta1 The Awareness of Social Inference Test (TASIT) Section 1, Ta2 TASIT Section 2, Ta3 TASIT Section 3, VLe MCCB verbal learning, WMe MCCB working memory; and Wrk SLOF work skills.

The Mayer-Salovey-Caruso Emotional Intelligence Test (MSCEIT) domain was isolated and disconnected from the rest of the graph, and depression (CDSS) has only one direct link to stigma.

Follow-up PAG

Table 3 and Fig. 2 show the type of connections at follow-up. Most of the baseline relationships linking variables belonging to the same and different conceptual domains held over time: working memory was again modeled as a key variable which, directly and indirectly affects the other cognitive domains and functional capacity. Note that the blue-and-not-bold directed arrow between “ToM – minimal social inference” and “ToM – enriched social inference” suggests that “ToM – minimal social inference” most likely can be modeled as directly or indirectly causing “ToM – enriched social inference” (Fig. 3), and the black-and-bold arrow between “ToM – emotion evaluation” and “ToM – minimal social inference” suggests that their relationship could be explained by an unmeasured latent confounder (Fig. 3). The network of links is richer and more varied than at baseline: attention/memory appeared to mediate the relationship of working memory with social cognition and stigma. Functional capacity, in turn, mediated the relationships of cognitive function ing with disorganization, and everyday life functioning, as well as of social cognition and everyday life functioning. Negative symptoms were again modeled as directly caused by impaired interpersonal relationships. Moreover, some flipped edges were found between variables belonging to the same domain (i.e. FEIT to TASIT-1 at baseline, TASIT-1 to FEIT at follow-up).

Table 3 Follow-up edge probabilities obtained with 1000 bootstraps.
Full size table
Fig. 3: Follow-up partial ancestral graph that best models the causal relationships among included variables.
Fig. 3: Follow-up partial ancestral graph that best models the causal relationships among included variables.
Full size image

Blue-and-bold directed arrow: direct causal path; blue-and-not-bold directed arrow: indirect causal path; black-and-bold arrow: the causal relationship could be explained by an unmeasured latent confounder. More details on edge types and meaning in Fig. 1. Alo Brief Negative Symptom Scale (BNSS) alogia, Anh BNSS anhedonia, Apa BNSS apathy, Aso BNSS asociality, Att Measurement and Treatment Research to Improve Cognition in Schizophrenia Consensus Cognitive Battery (MCCB) attention/vigilance, BAf BNSS blunted affect, Dep Calgary Depression Scale for Schizophrenia total score for depression, Dis PANSS disorganization, ELS Specific Level of Functioning Scale (SLOF) everyday life skills, FC functional capacity, FEI Facial Emotion Identification Test, Int SLOF interpersonal relationships, MSC MCCB Mayer-Salovey-Caruso Emotional Intelligence Test managing emotion section, Pos PANSS positive factor, PrS MCCB problem solving, PSp MCCB processing speed, SLe MCCB visuospatial learning, Stg stigma, Ta1 The Awareness of Social Inference Test (TASIT) Section 1, Ta2 TASIT Section 2, Ta3 TASIT Section 3, VLe MCCB verbal learning, WMe MCCB working memory and Wrk SLOF work skills.

The comparison of the follow-up and the baseline PAGs showed 75.9% agreement on edge presence, 74.2% on edge absence, and 81.8% on edge orientation.

Effect size estimation and model fit statistics

Fit statistics indicated an acceptable fit to the data of the baseline and follow-up models (baseline CFI = 0.878; baseline RMSEA = 0.083; follow-up CFI = 0.871; follow-up RMSEA = 0.094).

Raw and standardized effect sizes obtained using SEM models are provided in Supplementary Tables S1, S2. The baseline model contains 11 direct edges standardized effect sizes (between 0.5 and 0.8), such as those connecting asociality with apathy (ES = 0.532) and asociality with anhedonia (ES = 0.695).

The follow-up model showed 11 edges with medium standardized effect sizes, such as those connecting with each other variables in the social cognition and in the negative symptom domains. For example, ToM – minimal social inference was connected with ToM – enriched social inference (ES = 0.785) and blunted affect connected with alogia (ES = 0.760).

Notably, at baseline and follow-up, working memory was positively connected with functional capacity with a moderate effect size (baseline ES = 0.549; follow-up ES = 0.533).

Sensitivity and bootstrap analyses

At baseline, the stability analysis based on a random 90% sample showed 100% agreement on edge presence, 97.7% on edge absence, and 96.6% on edge orientation (see Supplementary Table S3). At follow-up, the stability analysis showed 98.8% agreement on edge presence, 93.8% on edge absence, and 100% on edge orientation (see Supplementary Table S4).

At baseline, the PAGs derived from the cohort not taking anticholinergics showed 87.5% agreement on edge presence, 92.6% on edge absence, and 100% on edge orientation (see Supplementary Table S5). At follow-up, the sensitivity analysis showed 93.3% agreement on edge presence, 96.6% on edge absence, and 100% on edge orientation (see Supplementary Table S6).

In a sensitivity analysis in which functional capacity was excluded from the model, there were no direct edges between working memory and the functional outcomes of everyday life skills, work skills, or interpersonal relationships at baseline and follow-up.

Discussion

The present study provided, through a data-driven discovery algorithm, a plausible model of the causal relationships linking symptoms, cognition and functioning in schizophrenia. This model showed that in patients with schizophrenia at baseline: 1) working memory impairment played a key role in the PAG, directly affecting other cognitive domains and everyday functioning through the mediation of functional capacity; 2) the impairment in everyday functioning influenced the severity of conceptual disorganization and, thus, of positive symptoms; 3) impairment in everyday life skills influenced work skills and interpersonal relationships; 4) impairment in interpersonal relationships directly affected asociality, which in turn influenced the severity of other negative symptoms.

After 4 years, most of the baseline relationships held over time, as did the mediation of functional capacity, but the PAG included additional relationships. The main differences between baseline and follow-up relations were 1) attention/vigilance mediated the relationship between working memory, social cognition and stigma; 2) visual learning mediated the relationship between working memory and social cognition; 3) functional capacity mediated not only the relationships between working memory and everyday functioning, but also the relationships between cognitive functioning and disorganization, and between social cognition and everyday functioning.

The finding of a key role in the PAGs model for working memory impairment is in line with the literature supporting the centrality of neurocognitive functioning in schizophrenia, particularly in the working memory domain18,39. Neurocognitive impairments are considered a trait marker in patients with schizophrenia, as they are present before the onset of psychotic symptoms and tend to remain fairly stable over the course of the illness, independent of illness phase, symptom severity and antipsychotic medication40,41,42,43. The centrality of working memory supports the idea that this domain represents a critical function for complex cognitive processes and the executive control of behavior41. Consistent with this hypothesis, our findings suggest that in schizophrenia working memory impairment may underlie a substantial part of the other neurocognitive as well as social cognition impairments.

Furthermore, our results show that working memory impairment might be a key determinant of poor functioning in people with schizophrenia, as working memory is the only variable directly or indirectly influencing functional capacity and then real-life functioning in the graph. These results suggest that difficulties in the maintenance and manipulation of stored information may influence the ability to perform tasks relevant to daily life, thus contributing to the patients’ disability. The pivotal role of working memory in influencing functional capacity in both chronic44,45,46 and first-episode47 psychosis patients has been widely described. At odds with our results, the only other study that used a data-driven analysis based on partial ancestral graphs in patients with schizophrenia did not find any causal relationship between cognition and functioning, either directly or indirectly35. However, the set of variables used is different between the two studies, and scoring or frequency distribution of measures can differ and could impact results. Moreover, the differences in sample size, characteristics (acute vs. chronic schizophrenia), and model parameters between the studies may account for differences in the results. In interpreting our results, it is important to consider that working memory was estimated to have a direct or indirect effect on functional capacity. One possible explanation is that contextual factors, such as environmental demands or personal attributes (e.g., motivation), may modulate the impact of working memory on real-life functioning, making its role more complex and multidimensional.

Our findings suggest a plausible pathway in which working memory impairment appears as an ancestor of disability in schizophrenia. Within this pathway, functional capacity acts as a bridge between cognitive performance and real-life functioning. The central role of functional capacity in patients with schizophrenia has already been largely reported in the literature21,23,30,48,49,50 with only a few studies suggesting that the inclusion of functional capacity does not improve the power of models predicting real-life functioning51. The results of our sensitivity analysis, where functional capacity variables were excluded, confirmed the role of functional capacity as a mediator between cognition and functional outcomes.

Overall, our model shows that working memory impairment affects the ability of people with schizophrenia to perform everyday tasks, but not the other domains of real-life functioning. This result is consistent with our previous findings that functional capacity is strongly related to the domain of daily living skills, which in turn is related to work skills and interpersonal relationships29,30.

Furthermore, the results of the present study, in line with previous literature, highlight the linkage of everyday life skills with disorganization and, through this, with positive symptoms, as well as the link between interpersonal relationships and negative symptoms21,23,29,30,52. In particular, in our model the everyday life skills impairment determines the severity of disorganization, and through this of positive symptoms, while the interpersonal relationships domain causes asociality, which, in turn, influences the severity of the other negative symptoms. This is a surprising result because literature has provided evidence that the relationship is from psychopathology to functioning impairment11,23,53. However, in most research, functioning has been considered an outcome, while our findings showed that the patients’ functional impairment might affect their experience, thus influencing symptomatology. For instance, the impairment in skills necessary to manage everyday life may manifest in specific behaviors and particular patterns of thinking that constitute the construct of disorganization. In addition, the impairment of social functioning might decrease the patients’ motivation for social interactions, which, in turn, might reduce pleasure during other activities. The inverse pathway was observed by Miley et al. at baseline34. However, the authors found that negative symptoms had an indirect effect on functioning through the mediation of socio-affective capacity and motivation. In contrast, after 6 months of follow-up, the direction of these relationships reversed, where social functioning was a cause of motivation, which in turn affected socio-affective capacity. This seems to be in line with the findings of our study, which was conducted on patients with chronic schizophrenia. Overall, the results of the two studies suggest either 1) a possible inversion of the relationships between negative symptoms, motivation, and social functioning from first-episode patients to later stages of the illness, or 2) a recursive mechanism where negative symptoms lead to a lack of motivation and impaired social functioning, which, over time, further diminishes motivation and worsens negative symptoms. Supporting this hypothesis, we note that in our results, the peripheral position of psychopathology in the graph does not exclude possible effects of symptoms on real-life functioning. Indeed, the algorithm used in the analyses does not include feedback loops. Therefore, it does not evaluate the potential reciprocal effects of symptoms on functioning. Overall, the results of the present study suggest that difficulties encountered in facing daily life activities and then with social interactions in clinically stable subjects with chronic schizophrenia may influence the clinical presentation of the disease54. The impact of real-life functioning on symptom severity might shed new light on the manifestation of positive and negative symptoms. Indeed, it is well known that psychotic onset is preceded by a long premorbid phase, characterized by the presence of cognitive impairment and a progressive worsening in real-life functioning. Therefore, according to our results, this temporal sequence could imply a cause-effect relationship. However, longitudinal studies including subjects at high-risk for psychosis are needed to confirm this hypothesis.

The comparison between baseline and follow-up PAGs indicates that most of the baseline links are consistent over time. Moreover, no bidirectional links are found, suggesting the absence of latent unmeasured or measured confounders. However, the follow-up PAG is richer than the baseline PAG, and there are some differences. First, attention/vigilance and visual learning mediated the relationship between working memory and social cognition. This might suggest that although working memory impairment remains the key change in the pathway to functioning, the relationships between cognitive domains become more complex over the years, potentially due to factors like anticholinergic burden or metabolic syndrome. At follow-up, functional capacity mediated the relationships between social cognition and everyday life functioning and had a direct impact on disorganization, without the mediating role of everyday life functioning. This pattern suggests that over time, functional capacity might have an extended impact on functioning and clinical outcomes. Overall, although the PAG structure remained relatively stable, the complexity of relationships increased at the follow-up, as indicated by the greater number of links among the study variables. The augmented density of the follow-up network may be associated with the alleviation of specific symptoms within an overall stable clinical picture, with cognitive impairment and poor functional capacity being more pronounced than the symptoms themselves. The augmented density might be related to increased variability at the individual level in the main variables with a significant impact on functioning. However, the alterations in some of the variables at the individual level (clinical, cognitive, and functional capacity) did not result in clinically meaningful changes. Thus, despite the augmented density of the follow-up network, its overall stability may be associated with the preservation of the overall clinical presentation characterized by mild positive symptoms, stable negative symptoms, cognitive impairment and poor functional capacity in the whole sample.

Our results should be considered in the light of some limitations. First, the sample consisted of chronic and clinically stable patients. All participants were on antipsychotic medication and had low symptom severity, possibly resulting in restricted variance. Further studies including patients in different stages of the disease are needed to confirm the present findings. Second, it is essential to acknowledge that, given its novelty, our model requires external validation. A further limitation of the method is that GCFI cannot identify synchronous cycles, should any be present in the data as suggested by the almost equal frequency of the causal link from functional capacity to working memory with respect to the one chosen by the algorithm from working memory to functional capacity. Finally, in the absence of a longitudinal analysis, no definitive conclusions can be drawn about the causal relationships among variables.

In conclusion, the key role of working memory impairment in the graph, influencing clinical and functional outcome suggests the importance of large-scale implementation of cognitive remediation interventions focusing on this domain, at least in chronic stable patients living in the community. Moreover, the possible impact of impairment in everyday activities and interpersonal relationships on the clinical presentation of schizophrenia suggests that integrating pharmacotherapy and psychosocial treatments may improve clinical outcomes, in addition to improving patients’ real-life functioning.

Methods

Participants

This cross-sectional study was carried out as part of the project “Factors influencing real-life functioning of people with a diagnosis of schizophrenia: a four-year follow-up multicenter study” (grant number: 2017M7SZM8) and in collaboration with the Italian Network for Research on Psychoses.

Study participants were recruited from patients living in the community and consecutively seen at the outpatient units of 26 Italian university psychiatric clinics and/or mental health departments. The study population consisted of patients with schizophrenia who had been stabilized on antipsychotic treatment. Twenty-four centers also participated in the follow-up study. In these centers, all the patients recruited for the baseline study were asked to join the follow‐up study. Subjects were contacted by phone, e‐mail or during a routine follow‐up visit or rehabilitation session. At baseline, inclusion criteria were a diagnosis of schizophrenia according to DSM-IV, confirmed with the Structured Clinical Interview for DSM IV — Patient version (SCID-I-P), and an age between 18 and 65 years. Exclusion criteria were: (a) history of head trauma with loss of consciousness; (b) history of moderate to severe mental retardation or of neurological diseases; (c) history of alcohol and/or substance abuse in the last six months; (d) current pregnancy or lactation; (e) inability to provide an informed consent; (f) treatment modifications and/or hospitalization due to symptom exacerbation in the last three months. For the follow-up study, additional exclusions applied: (a) head trauma with loss of consciousness during the four-year interval, (b) progressive cognitive deterioration diagnosed in the last four years, and the same criteria as the baseline study.

The study was conducted in strict adherence to the ethical standards of the 1964 Declaration of Helsinki. Ethical approval was obtained from the Ethics Committee of the involved institutions. All participants provided written informed consent after receiving a comprehensive explanation of the study’s procedures and objectives. Participants were recruited between March 2012, and September 2013 for baseline assessment, while the follow-up took place after 4 years, from March 2016 to December 2017.

Assessment

At baseline, socio-demographic variables such as age, education and gender were collected.

The Positive and Negative Syndrome Scale (PANSS) was used to rate symptom severity55. We based the positive symptoms score calculation on the consensus 5-factor solution proposed by Wallwork and colleagues56. Disorganization was obtained by the P2 item of the scale to avoid overlap with cognitive impairment57.

Negative symptoms were assessed with the Brief Negative Symptom Scale, validated in Italian by Mucci and colleagues58,59. The scale comprises 13 items, organized into six subscales (five negative symptom subscales: Anhedonia, Asociality, Apathy, Blunted Affect and Alogia, and a control subscale: Lack of distress). All the items are rated on a 7-point (0–6) scale, thus ranging from absent (0) to moderate (3) to extremely severe (6).

The Measurement and Treatment Research to Improve Cognition in Schizophrenia-MATRICS Consensus Cognitive Battery (MCCB)60,61 was used for the assessment of the following neurocognitive domains: speed of processing, verbal memory and learning, visual memory and learning, reasoning and problem solving, attention and vigilance, and working memory. Higher scores on all domains reflect better neurocognitive function in the corresponding domains.

The assessment of social cognition, partly included in the MCCB Mayer-Salovey-Caruso Emotional Intelligence Test (MSCEIT)60,61 managing emotion section, was integrated with the Facial Emotion Identification Task (FEIT)62 and The Awareness of Social Inference Test (TASIT)63. FEIT is a facial emotion recognition test which consists in identifying the correct emotion (joy, anger, fear, disgust, surprise, sadness or neutral) represented in a specific photo. A total of 55 photos are presented randomly62. The total test score was computed as the number of correct answers. TASIT63 consists of seven scales (positive emotions, negative emotions, sincere, simple sarcasm, paradoxical sarcasm, sarcasm enriched and lie), organized into 59 videos divided in three sections (TASIT 1–3): TASIT 1 “The Emotion Evaluation Test”, which explores emotional processing; TASIT 2 “Social Inference-Minimal” and TASIT 3 “Social Inference-Enriched”, which explore theory of mind. Higher scores on TASIT reflect better social cognition skills.

Real-life functioning was evaluated using the Specific Level of Functioning Scale (SLOF), a hybrid instrument which evaluates many aspects of functioning and is based on the key caregiver’s judgment on behavior and functioning of the patient64. It consists of 43 items arranged into the following domains: physical efficiency, skills in self-care, interpersonal relationships, social acceptability, everyday life skills, and work skills. In our study the scale was administered by a trained researcher to a key relative of each patient. Only the domains interpersonal relationships, work skills, and everyday life skills were used, as the other subscales showed ceiling effects65. Each of the items is rated on a 5-point Likert scale (1 = poorest functioning, 5 = best functioning).

Functional capacity was evaluated using the short version of the University of California San Diego (UCSD) Performance-based Skills Assessment Brief (UPSA-B)66, a performance-based instrument that assesses “financial skills” (e.g., counting money and paying bills) and “communication skills” (e.g., to dial a telephone number for emergency or reschedule an appointment by telephone). The total score, ranges from 0 to 100, with higher score reflecting higher functional capacity.

The Internalized Stigma of Mental Illness (ISMI) was used to evaluate the experience of stigma. It includes 29 items and 5 subscales for self-assessment of subjective experience of stigma67. Each item is rated on a 4-level Likert scale, where higher scores indicate greater levels of internalized stigma.

At follow-up, a clinical form was filled with data about the course of the disease and treatment information during the previous 4 years, using every available source of information (patients, relatives, medical records and mental health workers). All baseline assessments were also conducted at follow-up, using the same assessment tools.

Statistical analyses

To investigate the nature of the causal relationships linking symptoms, cognition and functioning, Partial Ancestral Graphs (PAGs) were used33.

A PAG is a causal graph suitable for situations in which latent measured or unmeasured variables may be present. This type of graph consists of three different types of endpoints: (o, >, −), and each edge here represents ancestral relationships between the nodes of the graph (Fig. 1). A → B denotes that A is an ancestor of B, and A ↔ B denotes that there is a latent variable causing both A and B. The circular endpoint denotes uncertainty about the correct causal endpoint, e.g., A o → B could be A → B or A ↔ B in the true graph, and thus the only certain knowledge from this orientation is that B is not an ancestor of A68.

The edge color and thickness denote its specialization (Fig. 1). An edge colored in blue is called definitely visible. In a PAG without selection bias (we assumed that no selection bias is present in the current analysis since the study sample is composed by real-world patients treated by community mental health services), a blue (definitely visible) edge from A to B denotes that A and B do not have a latent confounder. If an edge is not definitely visible (represented as black) then A and B may have a latent confounder68.

Another edge specialization shown as bold is called definitely direct. In a PAG without selection bias, a bold (definitely direct) edge from A to B denotes that A is a direct cause of B, relative to the other measured variables. If an edge is not definitely direct (represented as not bolded) then A may not be a direct cause of B, in which case there may be one or more measured variables on every causal path from A to B.

The PAG represents a Markov equivalence class of Directed Acyclic Graphs (DAGs) with latent and selection variables in the acyclic case33, and a Markov equivalence class of directed graphs with latent variables (but without selection variables) in the cyclic σ-separation case.

We employed the Greedy Fast Causal Inference (GFCI) algorithm69, a robust approach for causal discovery, both at baseline and follow-up. GFCI integrates two advanced algorithms: the Fast Causal Inference (FCI) and the Fast Greedy Equivalence Search (FGES), to construct and refine causal graphs.

Initially, FGES is utilized to establish a preliminary graph or “supergraph” of connections among the variables by iteratively adding edges to an empty graph. This step relies on a score-based method to identify potential causal relationships. After this initial construction, the FCI algorithm is applied to prune the supergraph. FCI identifies and eliminates indirect connections, thereby refining the graph. This process facilitates the discernment of direct relationships and the identification of potential latent confounders.

The final models were obtained using a bootstrap method with 1000 replications. The output of the bootstrap analysis provides the probability of an edge connecting two variables over 1000 replications, which must be greater than 0.5 to declare that the edge is present, and the most likely type of edge among 11 possible edges (i.e. the maximum proportion in the columns headed --- to <->), i.e. the one occurring most frequently over the 1000 bootstrap replications. This approach allowed us to estimate the stability and variance of the causal inferences made by GFCI, providing a robust assessment of the causal relationships identified in our data.

The GFCI algorithm’s implementation, sourced from the Tetrad software package version 7.6.0 (https://sites.google.com/view/tetradcausal), includes a statistical model score module and a conditional independence test module, utilizing BIC and Fisher Z, respectively. For this study, we configured the BIC score with a “penalty discount” parameter set to a value of 2, deviating from the standard BIC value to accommodate our specific data characteristics, i.e. the relatively high dimensionality of the dataset and the potential for noise in the relationships between variables. A higher penalty discount helps prevent model overfitting by reducing the likelihood of including spurious edges. This adjustment ensures that only more robust relationships are retained in the graph, which is particularly important when working with observational data that may contain noise or variability.

The Fisher Z module was employed with a p-value set at 0.05. These settings were chosen to optimize the balance between sensitivity and specificity in identifying causal connections. PAGs were compared using the Tetrad suite automated comparison tools.

Effect size estimation and model fit statistics

To estimate the raw and standardized effect sizes (ES) of the causal relationships learned by PAGs, linear structural equation models (SEM) were applied. Indirect covariations were used to model the edges connected to potential latent variables within the SEM. Model fit was evaluated using the Comparative Fit Index (CFI) and Root Mean Square Error of Approximation (RMSEA). The analysis was performed using the R package Lavaan 0.6-1833.

Sensitivity and bootstrap analyses

The stability of the PAGs at baseline and follow-up was assessed in a bootstrap analysis using GFCI on 1000 datasets, each containing 90% of the original data.

A sensitivity analysis was conducted to investigate the stability of the PAGs after excluding patients treated with anticholinergics.

An additional sensitivity analysis was conducted by excluding the functional capacity variable from the data. This analysis was performed on both the baseline and follow-up datasets to determine whether the pathway between cognition, specifically working memory, and functional outcomes would remain the same without the mediation role of functional capacity.