Introduction
Psychiatry has undergone a significant paradigm shift over the years, with debates emerging particularly in the field of psychosis. Schizophrenia is currently described in the Diagnostic and Statistical Manual of Mental Disorders, fifth edition (DSM-5) as a distinct disorder that combines positive and negative symptoms [1]. While it remains the primary diagnostic tool in clinical settings, the categorical approach faces significant challenges, particularly given the difficulty of categorizing mental disorders, such as schizophrenia, a heterogeneous disorder with overlapping symptoms and comorbidities. Individuals diagnosed with schizophrenia can present with different symptom profiles, illness trajectories, and treatment responses [Reference Tandon, Gaebel, Barch, Bustillo, Gur and Heckers2], making it difficult to identify a singular underlying pathology. Additionally, key symptoms like hallucinations, delusions, and cognitive deficits overlap with other psychiatric disorders and comorbidities, which blur diagnostic boundaries [Reference Yamada, Matsumoto, Iijima and Sumiyoshi3]. Further questioning the categorical approach, research has shown that subclinical psychotic experiences (i.e., occasional and temporary unusual thoughts or sensory experiences) are found in the general population in individuals who do not meet the criteria for psychotic disorders [Reference Rössler, Ajdacic-Gross, Haker, Rodgers, Müller and Hengartner4].
The value of the categorical approach has been debated, with some researchers arguing that the DSM’s rigid classification system may oversimplify the complexity of schizophrenia and fail to capture the full spectrum of psychotic experiences. In response, the dimensional perspective emerged, proposing that psychosis exists on a continuum rather than as a discrete condition. This view has been incorporated in alternative models of psychopathology, such as the Research Domain Criteria (RDoC) and the Hierarchical Taxonomy of Psychopathology (HiTOP) [Reference Insel, Cuthbert, Garvey, Heinssen, Pine and Quinn5, Reference Kotov, Krueger, Watson, Achenbach, Althoff and Bagby6]. The RDoC conceptualizes schizophrenia as a continuum of symptoms that exist across the population and overlap with other psychiatric disorders, and aims to identify shared neurobiological markers [Reference Gordon, Morris and Avenevoli7, Reference Heinssen, Morris and Sherrill8]. However, critics argue that in severe mental disorders, such as schizophrenia, pathogenic processes are more complex and disrupt fundamentally normal neurobiological function, challenging the applicability of a purely dimensional framework [Reference Ross and Margolis9]. The HiTOP addresses the arbitrary boundaries between normality and psychopathology by defining psychopathology as a combination of observable signs, symptoms, and maladaptive traits ranging from subclinical experiences (vulnerability) to established psychosis [Reference Kotov, Jonas, Carpenter, Dretsch, Eaton and Forbes10]. Both the RDoC and HiTOP models conceptualize schizophrenia as an extreme along a continuum, aligning with staging models and clinical high risk approaches that frame schizophrenia as the final stage of illness progression [Reference McGorry, Hickie, Yung, Pantelis and Jackson11, Reference McGorry, Hickie, Kotov, Schmaal, Wood and Allan12]. A key argument supporting this view is that subclinical psychotic symptoms are quite frequent in the general population, with a meta-analysis of 61 cohorts reporting a median prevalence of 7.2% (interquartile range of 2.5–15.5%) [Reference Rössler, Ajdacic-Gross, Haker, Rodgers, Müller and Hengartner4]. Another key argument is that cognitive deficits, considered as core endophenotypes of schizophrenia, are also observed in individuals at-risk for psychosis [Reference Sheffield, Karcher and Barch13]. While their presence may support the psychosis continuum notion, others argue that cognitive impairments in at-risk individuals may differ qualitatively from those in established psychosis, suggesting a possible nonlinear discontinuity within the spectrum [Reference Linscott and van Os14]. The examination of cognitive deficits is particularly relevant to this debate, considering that they are more predictive of social and functional outcomes than positive symptoms [Reference Catalan, McCutcheon, Aymerich, Pedruzo, Radua and Rodríguez15].
Empirical evidence consistently identifies cognitive deficits as a core feature of schizophrenia [Reference Catalan, McCutcheon, Aymerich, Pedruzo, Radua and Rodríguez15]. Meta-analyses in individuals experiencing a first episode of psychosis (FEP) have reported significant impairments across all cognitive domains compared with controls, with large effect sizes (Cohen’s d) around 1.00, especially in processing speed and verbal learning [Reference Catalan, McCutcheon, Aymerich, Pedruzo, Radua and Rodríguez15–Reference Watson, Harrison, Preti, Wykes and Cella17]. A meta-analysis of longitudinal studies indicated that these deficits persist at follow-up, with verbal learning and memory being the most affected [Reference Catalan, McCutcheon, Aymerich, Pedruzo, Radua and Rodríguez15]. Beyond FEP, a large-scale meta-analysis, combining 240 studies involving schizophrenia in- and outpatients, reported cognitive deficits ranging from 0.99 to 1.22, with memory functions being the most impaired [Reference Fioravanti, Bianchi and Cinti18]. Several reviews and meta-analyses have identified verbal memory and learning as the most impaired [Reference Guo, Ragland and Carter19–Reference Bora, Binnur Akdede and Alptekin21]. Cognitive deficits are also observed in individuals at-risk for psychosis, albeit to a lesser extent. A meta-analysis of familial high risk (FHR) and ultra-high risk (UHR) individuals reported impairments across all domains, with effect sizes around 0.36 and 0.48, respectively [Reference Bora, Lin, Wood, Yung, McGorry and Pantelis22]. Other reviews estimated a global cognitive impairment of 0.34 in at-risk individuals [Reference Fusar-Poli, Deste, Smieskova, Barlati, Yung and Howes23] and task-specific deficits around 0.50 [Reference Catalan, Salazar de Pablo, Aymerich, Damiani, Sordi and Radua24]. However, greater variability exists across studies in at-risk populations compared to clinical populations, partly due to the heterogeneous definitions of at-risk conditions. Taken together, these findings suggest a continuity in cognitive deficits across the psychosis spectrum, but the disproportionate level of impairment might reflect a nonlinear discontinuity rather than a linear progression between at-risk states and established psychosis.
To date, no meta-analysis has directly compared cognitive deficits across multiple risk groups (FHR, clinical high risk [CHR], schizotypy/schizotypal personality [ST], and UHR) alongside individuals with FEP and chronic schizophrenia. While some meta-analyses have examined at-risk individuals relative to FEP, they included heterogeneous definitions of risk and did not distinguish between different forms of vulnerability [Reference Catalan, Salazar de Pablo, Aymerich, Damiani, Sordi and Radua24]. Moreover, they did not use statistical approaches that allow for testing categorical, linear, and nonlinear models. A direct comparison across these groups with relevant statistical approaches would help determine if cognitive deficits in the psychosis spectrum are better explained by a categorical model (deficits emerge only at psychosis onset), a linear continuum (deficits increase proportionally), or a nonlinear continuum (with a discontinuity between at-risk and clinical states), and provide crucial insights into the nature of this psychopathology. Given the extensive literature on cognitive impairments, this question can be effectively addressed by focusing our effort on verbal learning, one of the most profoundly affected domains in schizophrenia, and by selecting standardized tasks to ensure comparability across studies. Indeed, the selection of homogenous tasks and trials is necessary because verbal memory engages multiple cognitive processes, including the initial encoding, the maintenance of information over a short delay, the consolidation, and the retrieval after a longer delay [Reference Bedwell, Horner, Yamanaka, Li, Myrick and Nahas25].
Therefore, the first objective of this meta-analysis is to compare verbal learning impairments across at-risk groups (FHR, CHR, ST, and UHR) and clinical populations (FEP and chronic schizophrenia) using commonly employed verbal learning tests in the psychosis literature. The second objective is to evaluate whether a categorical, linear, or nonlinear model best explains the distribution of these deficits, using polynomial regression analyses. We hypothesized that verbal learning deficits would generally increase with greater clinical severity across the psychosis spectrum, but anticipated that this relationship might be nonlinear, reflecting disproportionate increases in impairment at later stages.
Methods
Literature search
The protocol for this meta-analysis was registered in the International Prospective Register of Systematic Reviews (PROSPERO; registration number: 1042287). The systematic search was done on September 21, 2024, and records were queried from three different databases: Web of Sciences, PubMed, and EMBASE. Together, they host all the records from Web of Science Core Collection, MEDLINE, KCI-Korean Journal Database, Current Contents Connect, SciELO Citation Index, and Preprint Citation Index. The search query was as follows: [(psychosis, at-risk-mental-states, self-report, psychotic experiences, schizotypy, schizotypal personality disorder, biological high risk, familial high risk, clinical high risk, ultra-high risk, first episode of psychosis, schizophrenia, schizoaffective, schizophreniform) AND (cognit* OR neuropsychologi* OR memory OR episodic OR verbal learning OR CVLT OR HVLT OR RAVLT)]. No restrictions on study design were applied to ensure comprehensive retrieval of studies. The literature search was done by one author (FP).
Selection criteria
To be included, records had to meet the following general criteria: (i) investigated individuals within the psychosis spectrum, such as those at FHR, CHR, ST, UHR, and individuals with a FEP and chronic schizophrenia, (ii) included a healthy group for comparison, and (iii) reported the data from at least one validated list learning test (i.e., California Verbal Learning Test (CVLT), Hopkins Verbal Learning Test (HVLT), or the Rey Auditory Verbal Learning Test (RAVLT)). These tasks were selected based on a preliminary search in at-risk populations to ensure that the selected tasks are widely used. Moreover, their psychometric properties (e.g., reliability) have been rigorously validated [Reference Shapiro, Benedict, Schretlen and Brandt26–Reference Woods, Delis, Scott, Kramer and Holdnack28]. The selection of these tasks ensures consistency in administration and scoring, promoting a high level of homogeneity and standardization across studies. More specifically, studies were included in the FHR group if participants had at least one first-degree relative with schizophrenia, schizophreniform, or schizoaffective disorder, excluding those with second-degree relatives or other psychotic diagnoses. The CHR group included studies using semi-structured interviews to identify CHR individuals. The ST group comprised studies of individuals with ST traits or schizotypal personality. The UHR group included studies with a mix of clinical and familial risk. For the FEP group, studies involved individuals experiencing an acute psychotic episode, with illness duration lasting no longer than 3 years, or who had been in treatment for no more than 1 year before the study. The chronic group included studies with individuals diagnosed with chronic schizophrenia. Studies with both FEP and chronic participants were included if data were reported separately. More details are presented in the Supplementary Materials.
Upon general criteria, studies were excluded if: (i) they did not investigate individuals that fall within the psychosis spectrum entities as defined previously, (ii) there was no healthy group, (iii) they did not report data from the CVLT, HVLT, or RAVLT, (iv) there was missing or inadequate cognitive data (e.g., scores derived from a combination of different tasks), and (v) there was an overlap with data from another study (same task, same participants). The selection of reports was reviewed by two researchers (FP and SP) based on the inclusion/exclusion criteria. Moreover, the Preferred Reported Items for Systematic Reviews and Meta-Analysis guidelines were followed (see Supplementary Table S1) [Reference Page, McKenzie, Bossuyt, Boutron, Hoffmann and Mulrow29].
Classification of scores and effect size calculation
Scores (means and SD) from the CVLT, the HVLT, and the RAVLT were manually extracted by two authors (FP and SP). The scores were then categorized into three types: immediate, delayed, and total recall scores. For studies that did not report a global task score, the total score was calculated by averaging the immediate and delayed recall scores, if both were available. Immediate recall was assessed right after encoding and thus focused on the initial storage of memory and rehearsal strategies, whereas delayed recall was assessed at least 20 min after encoding and is known to capture recollection [Reference James, Weiss-Cowie, Hopton, Verhaeghen, Dotson and Duarte30]. Analyses were restricted to immediate and delayed free recall, since cued recall and recognition were measured only in a minority of studies. The categorization of trials was done by two researchers (FP and SP).
Then, the standardized mean differences (Cohen’s d) and their variances were computed in R [Reference Team31] using the metafor package [Reference Viechtbauer32] for each type of score: immediate, delayed, and total recall.
Univariate meta-analyses
Three separate univariate meta-analyses were conducted in R [Reference Team31] using the metafor package [Reference Viechtbauer32] to estimate the effect sizes for immediate, delayed, and total recall. Mixed-effects models were used with a random intercept nested in studies to adjust for intra-study correlation using the Restricted Maximum Likelihood estimation method. The direction of the effect size was considered positive if cognitive performance in individuals on the psychosis spectrum was worse than that of healthy controls. According to Cohen’s convention [Reference Cohen33], effect sizes of 0.2, 0.5, and 0.8 were interpreted as small, medium, and large, respectively. Heterogeneity analysis for mixed-effects models was performed [Reference Higgins and Thompson34] with the Q statistics [Reference Paulson and Bazemore35] and the magnitude evaluated with the I 2 index [Reference Lipsey and Wilson36]. Given the data set’s high levels of heterogeneity (see Section “Univariate meta-analyses”), effect sizes were pooled using random-effects models [Reference DerSimonian and Laird37]. A funnel plot and an Egger’s regression test were performed to assess the potential of publication bias [Reference Egger, Davey Smith, Schneider and Minder38].
Comparison of models
In each model, the primary predictor variable was psychosis spectrum group, and the outcome variable was verbal learning performance. Model comparisons were conducted separately for immediate, delayed, and total recall conditions using polynomial regression analyses [Reference Kroc and Olvera Astivia39]. To enable this, based on effect size estimates from the most reliable univariate model (i.e., immediate recall), which was determined by statistical power (i.e., larger number of studies available), categorical groups were coded into continuous variables from the least to most impaired (i.e., ST = 1; familial high risk (FHR) = 2; CHR = 3; UHR = 4; FEP = 5; chronic = 6). This data-driven transformation allowed us to treat psychosis spectrum groups as a continuous variable to test for linear and nonlinear (i.e., quadratic) models using all individual effect sizes rather than pooled effects. First, a linear model was compared to the intercept. Then, each regression model (
$ y=a+{b}_1x $
) was expanded with a quadratic term (
$ y=a+{b}_1x+{b}_2{x}^2 $
). Likelihood ratio tests (LRTs) were performed to evaluate whether the quadratic model provided a significantly better fit than the linear model after accounting for its additional parameter. Model predictions were generated, and comparisons were conducted using analyses of variance. Finally, categorical group models were contrasted with the intercept and quadratic models. For this purpose, participants were classified into two broader categories using the factor function: risk groups (FHR, CHR, UHR, and ST) and clinical groups (FEP and chronic schizophrenia). Post-hoc comparisons were performed in the categorical models to examine differences in effect sizes between these categories. To identify the most parsimonious model that best explained the data, we used the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC) to evaluate model quality by balancing goodness-of-fit and complexity, while the explained variance (R 2) quantified the overall fit.
Finally, to estimate the effects of continuous variables on verbal learning, age, and percentage of women were included as separate moderators in the immediate recall model, which had the largest number of studies. Due to missing data, the delayed and total recall models had too few studies to allow for moderation analyses. A significance threshold of p < 0.05 was applied to test if the confounding variables were significant moderators in the immediate model. If so, the significant moderator was added as a covariate in the model.
Results
Studies included
The initial search yielded 14,726 articles; most of the duplicates were then removed using EndNote (version 20.5), and the rest were removed manually (see Supplementary Figure S1). In all 9,486 articles were exposed to the screening process. Based on the abstract, the records were rejected if found irrelevant. The remaining 1,149 articles were assessed for eligibility based on the full text. Of these, 923 were excluded for different reasons: (i) 324 articles had no healthy group, (ii) 98 articles had mix groups, (iii) 172 articles assessed cognitive tasks from different domains, (iv) 83 articles assessed a different verbal episodic memory task, (v) 212 had missing cognitive data, (vi) 15 overlapped with data from another study, and (vii) 19 articles did not receive an answer from the authors regarding the request to access full-text and/or supplement materials. Therefore, 226 studies were included in the meta-analysis: 29 in the FHR group, 16 in CHR, 14 in ST, 32 in UHR, 67 in FEP, and 68 in the chronic schizophrenia group (see Supplementary Tables S2–S7 for studies’ characteristics).
Univariate meta-analyses
Immediate recall model
Effect size estimates increased progressively across groups, with FHR (d = 0.45) having the least deficit, followed by CHR (d = 0.46), ST (d = 0.52), UHR (d = 0.57), FEP (d = 1.00), and then chronic schizophrenia (d = 1.23) (see Table 1). Residual heterogeneity, as measured by tau2, was estimated at 0.0791 (SE = 0.011), corresponding to a tau value of 0.281. The I 2 statistic indicated that 69.11% of the total variance remained unexplained, suggesting significant heterogeneity beyond sampling variability (H 2 = 3.24). No publication bias was observed, as indicated by the symmetric funnel plot and a nonsignificant Egger’s test (p > 0.05). See Supplementary Figure S2 for the forest plot.
Representation of the nonlinear relationship between the psychosis spectrum and cognitive performances in immediate recall. Note. Each point in this figure represents the effect size of impairment in verbal learning during immediate recall from an individual study, plotted by group in ordinal order. According to Cohen’s convention, an effect size of 0.2 represents a small deficit, 0.5 a medium deficit, and 0.8 a large deficit in verbal learning. A quadratic curve is included to illustrate the nonlinear trend that best fits the distribution of cognitive performance across groups. A steeper decline is observed after the UHR group. A steeper incline in impairment is noted after the UHR group. FHR = familial high risk; CHR = clinical high risk; ST = schizotypy/schizotypal risk; UHR = ultra-high risk; FEP = first episode of psychosis; Chronic = chronic schizophrenia.
Results from the model, including all 282 outcomes of immediate memory in a univariate model by group
Abbreviations: CHR, clinical high risk; CI, confidence interval; Df, degree of freedom; FEP, first episode of psychosis; FHR, familial high risk; SD, standard deviation; SE, standard error; ST, schizotypy/schizotypal risk; UHR, ultra-high risk.
Delayed recall model
Similarly, effect size estimates increased progressively across groups, with ST (d = 0.07) having the least deficit, followed by FHR (d = 0.38), CHR (d = 0.40), UHR (d = 0.50), FEP (d = 1.00), and then chronic schizophrenia (d = 1.17) (see Table 2). Residual heterogeneity, as measured by tau2, was estimated at 0.0836 (SE = 0.016), corresponding to a tau value of 0.289. The I 2 statistic indicated that 65.52% of the total variance remained unexplained, suggesting significant heterogeneity beyond sampling variability (H 2 = 2.90). No publication bias was observed, as indicated by the symmetric funnel plot and a nonsignificant Egger’s test (p > 0.05). See Supplementary Figure S3 for the forest plot.
Results from the model, including all 146 outcomes of delayed memory in a univariate model by group
Abbreviations: CHR, clinical high risk; CI, confidence interval; Df, degree of freedom; FEP, first episode of psychosis; FHR, familial high risk; SD, standard deviation; SE, standard error; ST, schizotypy/schizotypal risk; UHR, ultra-high risk.
Total recall model
Again, effect sizes estimates increased progressively across groups, with ST (d = 0.14) having the least deficit, followed by CHR (d = 0.36), FHR (d = 0.42), UHR (d = 0.55), FEP (d = 1.06), and then chronic schizophrenia (d = 1.23) (see Table 3). Residual heterogeneity, as measured by tau2, was estimated at 0.0695 (SE = 0.0143), corresponding to a tau value of 0.2637. The I 2 statistic indicated that 69.11% of the total variance remained unexplained, suggesting significant heterogeneity beyond sampling variability (H 2 = 3.24). No publication bias was observed, as indicated by the symmetric funnel plot and a nonsignificant Egger’s test (p > 0.05). See Supplementary Figure S4 for the forest plot.
Results from the model, including all 148 outcomes of the total recall in a univariate model by group
Abbreviations: CHR, clinical high risk; CI, confidence interval; Df, degree of freedom; FEP, first episode of psychosis; FHR, familial high risk; SD, standard deviation; SE, standard error; ST, schizotypy/schizotypal risk; UHR, ultra-high risk.
Comparison of models
Immediate recall models
For immediate recall, the three models (linear, nonlinear, and categorical) compared to the intercept (no predictors) were significant (p < 0.05). However, the added complexity of the quadratic term is justified, as it had the lowest AIC and BIC values (see Table 4) and provided a significantly better fit to the data (see Supplementary Table S8), followed by the categorical model, and the linear model displaying the lowest fit. LRTs confirmed the quadratic model’s superiority over both alternative models (p < 0.001). It also explained the most variance (R 2 = 56.83%), with the categorical model next, and the linear model last. Notably, the predictor in the categorial model indicates a significant (p < 0.001) moderate-to-large difference in verbal learning impairments between the at-risk and clinical groups. This difference is further illustrated in Figure 1, which shows a steeper increase in effect size starting after the UHR group. These findings show that the relationship between psychosis and cognitive performance in immediate recall follows a nonlinear trajectory, with a more pronounced decline in clinical populations.
Linear, categorial, and nonlinear models compared to the intercept
Note: In bold is the selected model for immediate recall.
Abbreviations: AIC, Akaike Information Criterion; BIC, Bayesian Information Criterion; DF, degree of freedom; LRT, likelihood ratio test.
Finally, to examine the effects of continuous variables, age and percentage of women were tested as moderators in separate models. Age was not significant and was excluded from further analyses. The percentage of women was retained because it was a significant moderator (p < 0.05), though about 10% of studies were excluded due to missing data. Still, when sex ratio was included as a predictor, the categorical and quadratic models no longer differed significantly (p > 0.05), but both still outperformed the linear model in terms of AIC and BIC.
Delayed recall models
For delayed recall, all three models showed significant improvements over the intercept-only model (p < 0.001). Although the quadratic model explained slightly more variance (R 2 = 57.66%) than the categorical model (R 2 = 56.64%), the latter had lower AIC and BIC values, suggesting a trend toward a categorical distinction in performance (see Table 4). Moreover, LRTs provided no strong evidence favoring a quadratic model over the categorical (p = 0.49) or linear (p = 0.06) models, suggesting that the simpler categorical model is sufficient. Additionally, the categorial model indicates again a significant (p < 0.001) moderate-to-large difference in verbal learning impairments between the at-risk and clinical groups. Unlike immediate recall, where a quadratic model provided the best fit, the pattern in delayed recall is less clear, with categorical distinctions offering a slightly better, though not statistically superior, fit.
Total recall models
For total recall, all three models (linear, quadratic, and categorical) showed significant improvements over the intercept-only model (p < 0.05). Although AIC and BIC differences were modest, the quadratic model had the lowest AIC, and the categorical model had the lowest BIC. The quadratic model also explained the highest variance (R 2 = 61.88%); however, LRTs revealed that adding the quadratic term did not significantly improve model fit compared to the categorical (p = 0.13) or linear (p = 0.09) models. This indicates that, despite explaining more variance, the improvement was insufficient to definitively favor it over the categorical model. Also, the predictor in this model indicates a significant (p < 0.001) moderate-to-large difference in verbal learning impairments between the at-risk and clinical groups. Taken together, these findings suggest that the quadratic model may not offer enough benefit to justify its added complexity over the simpler, yet robust categorical model.
Discussion
Despite being well-documented in schizophrenia and observed in at-risk populations, verbal learning impairments had not been directly compared by way of meta-analysis. Moreover, the question of whether cognitive deficits in the psychosis spectrum are better explained by a categorical, linear, or nonlinear model remains unresolved, yet holds crucial implications for understanding the nature of the psychopathology of the psychosis spectrum. Our meta-analysis combined 226 studies across the psychosis spectrum, using three standardized list learning tests. We found that verbal learning impairments progressively increased across groups, with effect sizes around 0.50 for FHR, CHR, ST, and UHR; around 1.00 for FEP; and above 1.00 for chronic schizophrenia. Further analyses revealed significant differences between the at-risk groups and clinical populations (p < 0.001). Additionally, in immediate recall, which was the measure with the most statistical power, a nonlinear model provided the best fit to the data (p < 0.001), while the linear model showed the weakest fit. Results were not influenced by age and sex ratio. Likewise, the lowest fit was also observed with the linear model in the cases of both delayed recall and total recall.
In the FHR group, effect sizes for immediate, delayed, and total recall ranged from 0.38 to 0.45, indicating moderate deficits. Comparable results were observed in the CHR group, with effect sizes between 0.36 and 0.46. The ST group showed a wider range of effect sizes, from 0.07 (delayed recall) to 0.52 (immediate recall); however, estimates for delayed and total recall were based on a small number of studies (n < 5), raising concerns about their reliability. Despite some variation in point estimates, the confidence intervals for effect sizes in the FHR, CHR, and ST groups overlapped substantially, suggesting that these groups are not clearly distinguishable in terms of severity. In contrast, the UHR group demonstrated slightly larger impairments among at-risk populations, with effect sizes ranging from 0.50 to 0.57. These findings are consistent with previous meta-analyses on cognition, which typically reported deficits of around 0.50 in at-risk individuals [Reference Bora, Lin, Wood, Yung, McGorry and Pantelis22–Reference Catalan, Salazar de Pablo, Aymerich, Damiani, Sordi and Radua24]. While some evidence suggests that the severity of risk of symptoms may vary between schizotypal, FHR, CHR, and UHR individuals [Reference McGorry, Hickie, Yung, Pantelis and Jackson11], cognitive performance is relatively homogeneous across most of these at-risk entities. In clinical populations, our findings also aligned with prior meta-analytic evidence, where deficits around 1.00 and over 1.00 are reported in individuals with a FEP and with chronic schizophrenia, respectively [Reference Catalan, McCutcheon, Aymerich, Pedruzo, Radua and Rodríguez15, Reference Lee, Cernvall, Borg, Plavén-Sigray, Larsson and Erhardt16, Reference Fioravanti, Bianchi and Cinti18]. Indeed, effect sizes across immediate, delayed, and total recall ranged from 1.00 to 1.06 for the FEP group, and from 1.16 to 1.23 for the chronic schizophrenia group, indicating large deficits in both groups. In brief, our primary results are consistent with the existing literature and provide a comprehensive understanding of verbal learning deficits across the psychosis spectrum.
These results confirmed that verbal learning impairments are present across the psychosis spectrum relative to healthy controls, challenging a categorical perspective assuming that verbal deficits significantly deteriorate at the onset of psychosis [Reference Insel, Cuthbert, Garvey, Heinssen, Pine and Quinn5]. However, the dimensional approach assumes that differences between at-risk individuals and those with established psychosis are merely quantitative, with impairments increasing linearly along the psychosis spectrum. This perspective de-emphasizes the notion of schizophrenia as a distinct disorder and instead places it within a continuum of psychotic manifestations of linearly increasing severity [Reference Kotov, Jonas, Carpenter, Dretsch, Eaton and Forbes10]. Yet, it remains possible that a qualitative difference exists between subclinical and clinical entities, which would manifest through a nonlinear distribution of cognitive deficits across the psychosis spectrum. In support of this perspective, model comparisons of studies on immediate recall indicated that a nonlinear model provided the best fit for the data, with the linear model performing the worst. This suggests that the relationship between verbal learning impairments and the progression of psychosis may not follow a simple linear trajectory but may rather involve a qualitative change that is not adequately captured by linear approaches. Although analyses on delayed and total recall were based on a relatively small subset of studies, results also showed that the linear model provided the worst model fit.
These results carry important theoretical implications. If the categorical model were to be replaced by a dimensional one, our results suggest, based on a substantial number of studies, that the alternative dimensional approach should probably not adopt a linear conceptualization. Noteworthy, this aligns with critiques of a minority of authors who question the application of a fully (linear) dimensional perspective regarding severe mental disorders [Reference Ross and Margolis9]. Supporting this view, an image-based meta-analysis, using whole-brain T-maps from 17 neuroimaging studies on emotion processing, found no brain activation differences in response to emotional stimuli between at-risk individuals and healthy controls [Reference Fiorito, Aleman, Blasi, Bourque, Cao and Chan40]. Together, these results are inconsistent with a (linear) dimensional perspective whereby brain activity alterations are expected to be observed in at-risk individuals, albeit to a lesser extent than in psychotic patients.
Nonetheless, these findings should be interpreted with caution, as results from delayed and total recall were less definitive. The nonlinear model did not significantly outperform the categorical model (p = 0.49; p = 0.13 for delayed and total recall, respectively), nor the linear model (p = 0.06; p = 0.09). The lack of superiority of the quadratic model in delayed and total recall makes it unclear whether the nonlinear (dimensional) approach should be considered as the best model, as suggested by the immediate recall analyses. These inconclusive results may be partly explained by the smaller number of studies for these measures, leaving some clinical entities (especially ST) with few studies. That is, analyses on delayed and total recall may have lacked statistical power. Another limitation is the reliance on cross-sectional data, which do not capture within-individual changes over time, and limit conclusions about the progression of cognitive deficits, especially in at-risk individuals. Longitudinal data would be better suited to assess developmental trajectories and determine whether the observed nonlinear pattern reflects true changes. From a cognitive perspective, it must also be acknowledged that analyses had to be restricted to free (immediate and delayed) recall, due to a low number of studies reporting results on cued recall and recognition, meaning that key information on strategic retrieval and memory consolidation could not be analyzed. Finally, given the low conversion rates to psychosis in currently defined at-risk populations, a more linear trajectory might emerge if analyses were restricted to individuals who eventually transitioned to psychosis.
To conclude, our results advocate for a more nuanced understanding of verbal learning/memory deficits in the psychosis spectrum, where the transition from subclinical to clinical status might involve nonlinear shifts in cognitive abilities, rather than a simple, linear, continuous decline. In the future, refining our ability to identify individuals who are truly at high risk of transitioning to psychosis remains crucial before moving away from the categorical framework. Without this precise information, dimensional models may hide key distinctions between those who transition to a psychotic disorder and those who do not. Moreover, it will be essential to examine whether this nonlinear pattern extends beyond verbal learning to other important cognitive domains in schizophrenia to determine whether the discontinuity observed here is a key feature of broader cognitive dysfunction in psychosis. Finally, to better characterize verbal memory deficits across the psychosis spectrum, future studies should report scores for all verbal memory components. This would also allow the investigation of potential reverse associations between these components [Reference Dunn and Kirsner41], which could help clarify the functional dependence or independence of memory processes.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1192/j.eurpsy.2025.10094.
Data availability statement
In accordance with our commitment to transparency and reproducibility in research, all data utilized in our meta-analysis will be made available upon publication. This includes aggregated data from individual studies, statistical and analytic code, and supporting information. We encourage further analyses and welcome collaboration within the scope of our original study objectives. For inquiries, please contact the corresponding author.
Acknowledgements
SP is a holder of the Eli Lilly Canada Chair on schizophrenia research and FB is a holder of the Merit scholarship from the Faculty of Medicine of the University of Montreal.
Competing interests
The authors declare none.
Open access
Comments
No Comments have been published for this article.