Literature search strategy

This systematic review adhered to the 2020 Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines [12] and was initiated on October 26, 2021, with no time span specified regarding the initial date of publication. An updated search was conducted on June 25th 2024, and revealed three more meta-analyses, two that did not fit inclusion criteria and one that was added. The four following databases were systematically searched: PsyNet (PsycINFO and PsycARTICLES), Medline and Web of Science. The following keywords were used: (schizophren* or psychotic or psychosis or schizotyp* or “psychotic like” or “psychosis like” or “psychosis risk” or “psychotic risk” or paranoia or hallucination* or delusion*) and (bias* or distortion or “dysfunctional attitude” or “dysfunctional thought*” or “dysfunctional thinking” or distortion* or “distorted attitude*” or “distorted thinking” or “distorted thought*”) and (“meta-analysis”).

Inclusion and exclusion criteria

The decision to include or exclude reviews were made by consensus between authors.

Inclusion criteria

• We included meta-analyses that investigate the association between reasoning or interpretative cognitive biases and psychotic features. Specifically, studies were selected if they:
  1. Compare cognitive bias scores between individuals with psychotic disorders, psychotic symptoms, psychotic-like experiences, or psychosis risk and a control group.
  2. Examine correlations between reasoning or interpretative cognitive bias and symptom levels.
  3. Assess the impact of psychological interventions on cognitive biases.
• To further define the scope, included meta-analyses met these additional criteria:
  1. Metanalyses containing systematic reviews.
  2. Full-text peer-reviewed publications in either English or French, with books and conference abstracts excluded.
  3. Studies involving adult populations, defined as those with a mean participant age over 18 years.
• Analogue studies (i.e., experimental designs in which the procedures or participants used are similar but not identical to the situation of interest) in which participants did not have psychotic symptoms but could have other psychotic-like experiences, were at risk for psychosis or were rated on a psychotic-like experiences scale were included.

Exclusion criteria

• Meta-analyses about emotion recognition were excluded as these are largely considered a social cognitive deficit rather than a cognitive bias [13–16].
• Aggression bias was excluded because it refers more to a score of behavioral intention of aggressiveness than to a cognitive bias.
• Need for closure was also excluded, since it refers to the motivation to achieve finality and absoluteness in decisions, judgments, and choices, often prematurely and is typically considered an underlying motivation that can explain the jumping to conclusion bias [17], or a personality trait [17, 18].

Data extraction

Information on nine factors were retrieved when available, namely: 1) study design, 2) outcomes (e.g., type of cognitive biases, psychotic characteristics studied), 3) effect size, 4) confidence interval, 5) consistency (homogeneity, I squared, Cochran’s Q statistic), 6) number of studies and number of participants, 8) publication bias and, 9) consideration of confounding factors (e.g., age, sex, IQ).

Conclusions about effect sizes

The magnitude of the relation between a cognitive bias and a psychotic feature or symptom, as well as the effects of psychological interventions on cognitive biases, were determined based on the effect size estimated by the investigators [19]. In the current umbrella-review, effect size estimates were calculated using Cohen’s d, Hedge’s g, Pearson’s correlation coefficients or odds ratio. By convention [20], d/g estimates of 0.8, 0.55 and 0.2 were considered as large, moderate and small. Specifications about conclusions on effect sizes, R-values estimates included, are presented in supplementary material (S1 Table).

Quality assessment using the GRADE system

The GRADE system was used to assess the quality of the evidence following the data extraction [21]. Exact criteria used for the Quality assessment using the GRADE system are presented in supplementary material (S2 Table). According to this assessment system, the quality of the evidence of the meta-analyses’ results can be judged based on various factors, namely: 1) the size of the sample (the larger the better, ideally over 1000); 2) the precision of effects (i.e., the confidence interval is not too wide—as in Matheson and colleagues [22], 25% higher or lower than the effect size was considered ideal); 3) homogeneity of effects across studies (i.e., consistency of results from one study to the next), and 4) publication bias (if analyzed and if present). We checked for follow-up data (if any and the length of time) for meta-analyses regarding the effects of psychological treatments on cognitive biases, but not for the studies on the relation between biases and psychotic characteristics.

We also added a specific section for confounding factors considered, which could include controlling for study biases, trial quality, as well as other variables that could influence the results. Following the recommended guidelines, no points were given for the effect size. As such, a point is given for each element of the GRADE system measured (e.g., sample size, precision, consistency, publication bias and confounding factors; and follow-up data in the case of trials), with meta-analyses being rated from poor, poor to moderate, moderate, moderate to high, high–quality evidence, to very high-quality evidence, for a maximum of 6 points. If information was missing in the article, the authors of the meta-analysis were contacted and if they did not respond within two months, we considered the information as missing and no point was given for this factor. For each meta-analysis, four of the authors rated the different components with the GRADE system. Differences were discussed and a consensus was made for a final decision. Given the stringent criteria involved, consensus was easily reached (over 95% initial agreement). The authors of this article already used the GRADE system in previous umbrella and meta-reviews [23, 24].

Included and excluded studies

Overall, our search generated 1753 potential articles, from which 748 duplicates were removed, 848 articles were excluded based on titles and 69 were excluded based on abstract. Based on full text-read, 62 articles were excluded because they were not linked to interpretation or reasoning cognitive biases or were not reporting statistics specific to cognitive biases (e.g., some studies reported a general score of social cognition), six articles were excluded because they were not meta-analyses, one was removed because participants did not have any psychoticfeatures, three were removed because of substantial overlap in studies with another selected meta-analysis. Overall, 128 results were extracted and graded from the 16 meta-analyses remaining (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. PRISMA 2020 Flow diagram of the articles’ selection process.

https://doi.org/10.1371/journal.pone.0314965.g001

Nomenclature.

In order to extract the results, a careful analysis of definitions, terms and measures used to measure the cognitive biases was conducted. This led to Table 1, which synthesizes the concepts measured in the meta-analyses under a single nomenclature, enabling us to aggregate together studies measuring the same concepts but that might have used different terminologies (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Cognitive bias classification and measures.

https://doi.org/10.1371/journal.pone.0314965.t001

Evidence for cognitive biases studied in relation to psychotic characteristics

The full results of the GRADE analysis are presented in supplementary material (S3 Table). Reasoning and interpretation biases can be broken down into smaller concepts, with specific biases studied. In Fig 2, we have attempted to synthesize the cognitive biases studied in meta-analyses on these two categories (reasoning and interpretation biases). As can be seen, attributional biases are a subtype of interpretation biases, and can be broken down into externalisation of cognitive events, or source monitoring, (with their valence–positive, neutral or negative), a mix of specific attributional biases, self-serving bias, personalizing bias, hostility attribution bias, and aberrant salience. Some authors also studied specific attributional biases regrouped together. Reasoning biases hold two larger categories, jumping to conclusions (JTC) and/or data gathering bias, and belief inflexibility being broken down into bias against disconfirmatory evidence (BADE), bias against confirmatory evidence (BACE), and liberal acceptance (LA) (Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Classification of the cognitive biases studied.

https://doi.org/10.1371/journal.pone.0314965.g002

Evidence for the effects of interventions on cognitive biases

In terms of the quality of the evidence for the effects of interventions on cognitive biases, psychological interventions in general [25] and Meta-Cognitive Training (MCT) in relation to the data-gathering bias more specifically [41] have been documented in meta-analyses. When analyzed together, the effects of psychological interventions MCT [9], and MCT adaptations, Maudsley review training programs (MRTP [85], MRTP in combination with the Interpretative Bias Modification (CBM-I; [86] or its adaptation, Thinking Well (TW) program [87], Reasoning Training (RT; [88], Cognitive Bias Correction (CBC; [89] and Cognitive Bias Modification (CBM; [90]) had small effect sizes on cognitive biases, with moderate-high overall quality evidence [25].

MCT had a small-medium effect size regarding the effect on jumping to conclusions (data-gathering bias) with poor overall quality evidence, since the results were not precise, inconsistent, the sample size was small, and publication bias was not considered [41]. MCT’s effect on cognitive biases as a large category, based on 19 studies (and 931 participants) reports a small effect size (g = 0.16) with moderate quality of evidence, based on good consistency, fairly large sample size and control for confounding variables, although when only RCTs were included (14 studies, 658 participants), the effect size became non-significant and the quality of the evidence also decreased to poor [26]. Furthermore, Penney et al. compared follow-up scores (less than 1 year, 10 studies, 658 participants) to both post-intervention and pre-intervention scores, revealing either small or very small effect sizes. Similarly, a very small effect size was observed between follow-up scores (more than 1 year, 3 studies, 328 participants) and post-intervention scores, but the quality of the evidence in these instances was poor or poor to moderate [26].

Cognitive biases and psychotic characteristics

By grouping cognitive biases studied in meta-analyses under similar concepts, we can conclude that there is good evidence regarding the presence of the following alongside psychotic characteristics.

Interpretations biases

Individuals with paranoia (clinical and subclinical regrouped) tend to show large interpretation biases when compared to controls. We also see small to small-medium correlations between the severity of paranoia and interpretation biases [29]. This concurs with existing theoretical models (e.g., Freeman).

Attributional biases. Two meta-analyses that studied multiple attributional biases [30, 31] grouped together found no effect to medium-large effect sizes, with insufficient quality evidence. However, Murphy et al. [33] showed that individuals with persecutory delusions had a greater tendency for a mix of specific attributional biases (external-personal attribution, personalizing bias, internal attribution for negative events, externalizing bias) (effect size was medium), with moderate quality of evidence. These contradictory results could be an artifact of studies looking at several biases together, given the amalgam of different concepts measures, and often small sample sizes.

Externalisation of cognitive events. Individuals with hallucinations and hallucination-prone individuals (clinical and non-clinical grouped together) have a greater tendency (effect size was medium-large) to misattribute internally generated cognitive events to external sources [32]. Although sub-analyses on stimuli valence were inconclusive and more quality studies are warranted, the current results seem congruent with the literature, whereby negative situations are often attributed to external causes to protect self-esteem [34].

Self-serving bias/Externalizing bias. Individuals with schizophrenia-spectrum disorders with persecutory delusions demonstrate a stronger self-serving bias than non-clinical controls, and then individuals with schizophrenia-spectrum disorders with remitted persecutory delusions (with small to small-medium effect sizes). There is no effect or a very-small effect size when comparing individuals with schizophrenia (with and without persecutory delusions) and controls, meaning the self-serving bias would be specific to persecutory delusions.

However, these results should be interpreted with caution, since sub-analyses reveal an absence of effect when using the ASQ [50] or IPSAQ [66], both widely used questionnaires for this bias. It is unclear at this moment if the other measures used are flawed, or measure the concept differently, or if the IPSAQ and ASQ have methodological issues, but studies using other measures report large effect sizes. Results also show that individuals with schizophrenia spectrum disorders have a greater tendency for this type of attribution compared to individuals with major depression, with a very large effect size. This is consistent with the literature on major depression, as depressed individuals lack the self-serving bias often found with controls [91].

Personalizing bias. The evidence was too poor to conclude on the relation between the personalizing bias and schizophrenia (with no effect or very small effect size) [35] and positive psychotic-like experiences in healthy samples and ultra-high risk samples (with no effect or very small effect sizes to large effect sizes) [36]. This bias was only measured with the IPSAQ. None of the meta-analyses in our review addressed the specific relation between the personalizing bias and paranoia or the severity of paranoia within clinical populations.

Hostility attribution bias. Hostility attribution bias was associated with negative-psychotic-like experiences in healthy samples (with a medium effect size), but quality evidence was poor-to-moderate [34]. Although these results are consistent with Rector and Beck’s theory [92] which suggests that negative symptoms of psychosis may be linked to social aversion and defeatist attitudes, the limited quality of evidence, the restricted sample (comprising only individuals with positive psychotic-like experiences in healthy or ultra-high-risk groups), and the lack of meta-analyses addressing this specific attributional bias in individuals with psychotic disorders collectively constrain our ability to confirm the role of hostility attribution bias in psychotic symptoms. The hostility attribution literature in schizophrenia and psychosis appears limited at this point, with a paucity of studies, mostly from the same team, suggesting links between hostility attribution bias and suspiciousness [93, 94].

Aberrant salience bias. Aberrant Salience bias was associated with positive psychotic-like experiences in healthy samples (with a very-large effect size), but with poor quality evidence. No meta-analysis in our review addressed this cognitive bias in individuals with psychotic disorders. This was particularly surprising as the aberrant salience literature has exploded, with over 200 studies in the past decade alone. For instance, a recent study [95] found a strong link between the aberrant salience bias and delusions, specifically ideas of reference. A meta-analysis on aberrant salience bias in psychosis is definitely warranted.

Reasoning bias. Belief inflexibility bias, bias against disconfirmatory evidence, bias against confirmatory evidence, and liberal acceptance bias. Belief inflexibility is associated with global delusion severity in patients with psychotic disorders and does not seem specific to delusions. It is also associated with delusion conviction, delusional preoccupation and delusional distress. Belief inflexibility biases show a medium effect size on global severity of delusion and a medium-large effect size delusional conviction, but only a small effect on delusional preoccupation and delusional distress suggesting that belief inflexibility biases bring more conviction in beliefs, but not much more distressed about them. These studies however did not look at the content of the delusions, which might further explain this result as paranoid delusions and grandiose delusions can trigger different levels of distress. Regarding studies on healthy and ultra-high-risk samples, the quality of the evidence is too poor at the moment to conclude.

As mentioned by Zhu and colleagues [37], the choice of measure can greatly influence the results here as some studies used items from the Maudsley Assessment of Delusions Schedule (MADS [81]), which overlap with measures of psychotic symptoms. and thus increase the effect size, particularly when compared to studies where belief inflexibility is measured with a task unrelated to the individual’s delusions (e.g., the BADE task [83]). As such, we found a large effect size between delusional conviction and MADS subscales but only a small-medium effect size between delusional conviction and BADE tasks, although the quality evidence was insufficient to clearly conclude about the effects of specific tasks on the results.

Regarding the Bias Against Disconfirmatory evidence [37, 38], insufficient overall quality evidence limits our ability to confirm or not the link with psychotic symptoms.

Overall quality evidence was also insufficient regarding the Bias against confirmatory evidence and the Liberal acceptance bias, both measured with scores derived from the BADE task. Methodologically sound studies using the BADE task, perhaps using emotion-triggering as well as neutral information, are warranted before we can conclude if this measure captures these reasoning biases well in individuals with psychotic characteristics.

Jumping to conclusions (JTC) and/or data gathering bias. Individuals with psychosis show a larger jumping to conclusions bias (Data-Gathering bias) than healthy individuals, both when comparing numbers of draws to decision and when considering only extreme answers (i.e. deciding after 1 or 2 beads), with medium to large effect sizes [40] and moderate to high quality evidence (Table 2). Individuals with psychosis also show a larger JTC (Data-Gathering) bias than individuals with other mental health problems, when comparing number of draws before making a decision, with a medium-large effect size [39, 40].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Effect sizes of cognitive bias concepts with moderate to high-quality evidence.

https://doi.org/10.1371/journal.pone.0314965.t002

Only extreme responding in JTC (i.e., deciding after a single or two beads/fish) is associated with the presence or severity of delusions in psychosis, with a medium-large effect size [40]. It is not clear however if extreme responding reveals more a lack of understanding of the task, a ‘need for closure’ (motivation to achieve absoluteness in judgements) [17], greater impulsivity or a stronger reasoning bias [40].

Comparisons between individuals with schizophrenia or a psychotic disorder with delusions versus those without delusions show a small effect size, with high quality evidence, which appears to support the idea that JTC might measure proneness to delusions but not necessarily delusions themselves [96].

When using the Peters et al’s Delusions Inventory [84] to assess delusional thinking, JTC biases were found in people from the general population, individuals with current delusions, individuals with previous delusions, individuals with anxiety or depressive disorders, individuals with ultra-high-risk for psychosis, individuals with obsessive-compulsive disorder and individuals in a new religious movement), with small effect sizes. There is no evidence that JTC, as measured with a self-report questionnaire, is related to positive- or negative-psychotic-like experiences in healthy and ultra-high-risk for psychosis individuals [36]. Validation studies on the Cognitive Bias Questionnaire [56] do suggest more important biases, including JTC bias, in people with psychosis when compared to the non-clinical controls but do not show convergent validity with JTC or other cognitive bias tasks. A meta-analysis that addresses JTC using self-report questionnaires in clinical populations is required.

Effects of psychological interventions on cognitive biases

Regarding the effects of psychological interventions on cognitive biases, we found moderate-high-quality evidence and a small effect size when grouping together interventions targeting cognitive biases (Table 3). MCT was the intervention used in the majority of studies in Sauve et al.’s meta-analysis [25] and all of the studies in Penney et al’s meta-analysis [26]. When looking at the effects of MCT on the data gathering bias only, the quality of evidence is weaker, given that the results were not precise, not consistent, the sample size was small and publication bias was not considered. However, these results were part of a larger meta-analysis on the effects on MCT on cognition, which explains the small sample size in the sub analyses and why publication bias was not verified in this particular analysis. Overall, we find a small effect size on cognitive biases, with moderate quality of the evidence overall, but no effect and poor or poor to moderate quality of the evidence when including follow-up data or only high-quality studies. This might appear surprising as meta-analyses show that MCT improves psychotic symptoms (small-medium effect size), but cognitive biases that are theorized and targeted in the training to improve these symptoms, only show a small effect size in terms of improvements. Are we faced with a sensitivity-measure issue, whereby changes in biases are not detected by the measures used in the trials, or are other mediating variables at play? Future studies need to explore this question as the theoretical underpinnings of MCT could otherwise be questioned.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Evidence for the effect of interventions targeting cognitive biases.

https://doi.org/10.1371/journal.pone.0314965.t003

It is also important to mention that no meta-analysis specifically addressing the effects of cognitive behaviour therapy for psychosis (CBTp) on cognitive biases was found in our review. CBTp is a cognitively oriented intervention that targets cognitive biases in order to modify beliefs underlying hallucinations and delusions [97]. CBTp is a well-recognized evidence-based treatment [98], considered effective in reducing psychotic (positive and negative) symptoms as well as depressive and social anxiety symptoms, and in improving functioning and mood [99].

Relations between cognitive biases

While this umbrella-review investigated the associations between cognitive biases and psychotic characteristics, it does not allow us to extrapolate on the relationships between cognitive biases, as was theoretically hypothesized in a recent review [1]. Several cognitive models of psychopathology suggest roles for more than one cognitive bias in the maintenance of different disorders, with bidirectional effects. This phenomenon has been named the combined effect of cognitive biases hypothesis [101].

For instance, Broyd et al.’s [102] model on the formation and maintenance of delusional beliefs suggests that some cognitive biases (and other cognitive factors) would be related to the formation of delusions, whereas others would be more related to the maintenance of delusions. Cognitive factors such as JTC (which, according to the authors, would be mediated by the Liberal acceptance bias and by Salience) would interfere with the individual’s ability to process the information encountered using a “top down” process, i.e. by considering the information already known to interpret new stimuli. A study using a Bayesian mathematical model suggested that not using top-down processes prevents new information from being appropriately integrated [103]. Other cognitive vulnerabilities such as difficulties related to metamemory [104] (i.e. metacognitive knowledge and processes that necessitate memory [105], confirmation bias and BADE would prevent the person from invalidating their delusional beliefs. It would be relevant to further clarify which cognitive biases are linked to the formation and/or maintenance of delusions, since these could allow us to adjust cognitive interventions according to the specific phases of psychosis (e.g., prodromal/at-risk phase or residual phase).

Limitations

Our umbrella-review is limited by the meta-analyses it included. Several reviews reported inconsistent results. This could be partly explained by the lack of consensus in the literature regarding the nomenclature, the classification and the measures used regarding cognitive biases. As such, cognitive bias names are used by different authors to designate different concepts. The term “externalizing bias” for instance is used by Brookwell et al. [32] to designate “misattribution of internally generated cognitive events to an external source”, whatever their valence (neutral, positive and negative), by Savla et al. [35] and Livet et al. [36] to designate “attributing external causes to negative situations and/or attributing internal causes to positive situations”, and by Murphy et al. [33] to designate a group of attributional biases (including the external-personal attribution score, the personalizing bias, the internality attribution score for negative events and the externalizing bias score). Several other small distinctions between concepts and nomenclature can be found across the meta-analyses used in our review. The plethora of different measures for the same concepts can also explain the important discrepancies in effect sizes across studies.

There are likely multiple studies on specific cognitive biases in psychosis or schizophrenia that were not described here because they were not included in a meta-analysis. For instance, we did not find meta-analyses on other specific interpretation or reasoning biases such as the catastrophizing bias, the dichotomous thinking bias or the emotional-based reasoning bias [56]. This could be explained by the paucity of individual studies on some of these biases in psychosis, as well as the few well-validated instruments available to measure them. However, there are some individual studies [106, 107] reporting results on catastrophizing bias, using the catastrophizing interview procedure [108], as well as studies using the catastrophizing scale of Peters and al.’s Cognitive Bias Questionnaire for psychosis [56]–they have yet to be included in a meta-analysis. We did not find meta-analyses regarding cognitive biases in people with substance-induced psychosis, nor meta-analyses looking at CBTp’s effects on cognitive biases. Furthermore, meta-analyses on cognitive biases and psychotic features measured longitudinally were absent from our search. Moreover, other cognitive biases (attention to threat, aberrant salience) were only reported in Livet et al. [36] within analogue studies involving individuals with psychotic experiences, and have not been reviewed in meta-analyses that include clinical populations. Finally, results specifically on the personalizing bias were also only reported in Livet et al. [36] in analogue studies, and in Savla et al. [35] as a sub analysis of a larger study on social cognition, making the conclusion we can draw from the results limited. It would be important to address the personalizing bias in a future meta-analysis, and to compare it with results linked to the self-serving bias (often called externalizing bias) and the hostility attribution bias to determine if the personalizing bias, or the hostility attribution bias is more associated with persecutory delusions than the self-serving bias.

Strengths

An important strength of this umbrella-review is the use of the GRADE system to assess the quality of meta-analyses investigating relationships between cognitive biases and psychotic characteristics, including both healthy individuals and individuals with psychotic disorders. This enabled us to clearly identify the links with psychotic symptoms and comparisons between populations on biases supported by strong evidence, as well as document gaps in the literature and the need for future studies. This umbrella-review covered several specific biases, under two larger categories of biases, namely reasoning and interpretation biases. These biases are currently the most studied cognitive biases in schizophrenia, although a growing literature is developing on other cognitive biases such as memory and attention biases [109, 110]. Our umbrella-review was also quite thorough, carefully reporting sub-analyses as well as main analyses and considering several definitions and measures.