Study selection

Articles for consideration were identified using a three-step procedure. First, a literature search was conducted in the electronic databases PubMed and Web of Science in the period between 1980 and April 2004, using the keywords INSIGHT or UNAWARENESS combined with PSYCHOSIS or SCHIZOPHRENIA combined with COGNITI^* or NEUROPSYCHOLOG^* or INTELLIGENCE, MEMORY or WCST. This yielded 215 papers, of which 33 fulfilled our criteria (see below). Studies were included after consensus was reached between two authors (A.A. and N.A.) that the paper met all inclusion criteria. Second, we screened the reference list of a recent review by Morgan & David (Reference Morgan, David, Amador and David2004), which yielded two additional papers for inclusion in the meta-analysis (Reference Takai, Uermatsu and UekiTakai et al, 1992; Reference Goldberg, Green-Paden and LehmanGoldberg et al, 2001). Finally, in order not to miss new articles that might not yet have been electronically indexed, a hand search of key journals was conducted for the period April 2003 to April 2004; this included the American Journal of Psychiatry, Archives of General Psychiatry, British Journal of Psychiatry, Journal of Nervous and Mental Disease, Psychiatry Research, Psychological Medicine, Schizophrenia Bulletin and Schizophrenia Research. This did not yield any additional paper suitable for inclusion.

To be included in the meta-analysis, studies had to report correlations between insight scales and cognitive performance measures. In addition, they had to satisfy the following criteria:

1. (a) the sample comprised patient groups with a psychotic disorder, whether affective or non-affective;

2. (b) sufficient information was reported to compute effect sizes, or to determine the direction and significance of the effect;

3. (c) the article had been published in a peer-reviewed English-language journal.

Studies had to include a valid measure of insight, such as the insight item from the Positive and Negative Syndrome Scale (PANSS; Reference Kay, Fiszbein and OplerKay et al, 1987) or Present State Examination (PSE; Reference Wing, Cooper and SartoriusWing et al, 1974); the 11-item Schedule for the Assessment of Insight, original and Expanded version (SAI-E; Reference Kemp, David and BlackwellKemp & David, 1997; Reference Sanz, Constable and Lopez-IborSanz et al, 1998); the 74-item Scale to Assess Unawareness of Mental Disorder (SUMD; Reference Amador, Strauss and YaleAmador et al, 1993); or the 11-item Insight and Treatment Attitudes Questionnaire (ITAQ; Reference McEvoy, Apperson and AppelbaumMcEvoy et al, 1989).

With regard to the cognitive domains, we defined five domains. The first was total cognition, in which all cognitive test results published in the same paper were pooled. Examples of such tests are the Wechsler Adult Intelligence Scale (WAIS), or subtests of the WAIS such as Digit Span; the Mini-Mental State Examination (MMSE); the National Adult Reading Test (NART); the Trail Making Test; the Wisconsin Card Sorting Test (WCST); and measures of attention, memory tests and verbal fluency tests (see Reference LezakLezak, 1995 for a description of these measures). The second domain was IQ only, which in most cases pertained to the WAIS. Other estimates of intelligence with an established validity were also included. In cases where only a number of WAIS sub-tests were included, these subtests were pooled. The third domain, memory, included established measures of verbal and visual memory performance (see Reference Aleman, Hijman and De HaanAleman et al, 1999). Fourth was frontal executive function, which included the Trail Making Test B, verbal fluency and the WCST. Finally, a separate analysis limited to the WCST was included, because several authors have postulated a specific relationship between perseveration as assessed by the WCST and poor insight (cf. Reference Morgan, David, Amador and DavidMorgan & David, 2004). Furthermore, in a construct validation factor analysis (involving data from 473 clinical cases) WCST scores loaded independently of other neuropsychological variables, indicating that the WCST contributes uniquely to neuropsychological evaluation (Reference Greve, Ingram and BianchiniGreve et al, 1998). For the WCST analysis we included categories completed as well as perseverative responses, which were pooled when both were reported in a single paper (cf. Reference Nieuwenstein, Aleman and De HaanNieuwenstein et al, 2001). These parameters of the WCST have been shown to load on the same factor, termed ‘perseveration’ (Reference Cuesta, Peralta and CaroCuesta et al, 1995).

Data analysis

For each study an effect size was calculated, which was the mean r weighted for sample size (Reference Hunter and SchmidtHunter & Schmidt, 1990). When the precise r values were not given, an estimate was computed from exact P, t or F values (Reference Lipsey and WilsonLipsey & Wilson, 2001). A problem in meta-analysis concerns the handling of missing values owing to studies reporting that a result was non-significant without providing the exact statistics. The most common method is to exclude such studies completely; this is unfortunate, and will bias the results considerably when there is a substantial number of studies with this problem (as in the present case). Another method is to give such studies values of r=0, which is a highly conservative approach that may obscure the existence of small effects. A third method is to assign such studies the mean effect size of the studies that report an insignificant effect size. A problem of this method is that it may yield an overestimation of the non-significant effect sizes (as authors will more readily report non-significant effect sizes of some magnitude than very small ones). A reasonable compromise, in our view, is to include the lowest value of the 95% confidence interval of the mean effect size for non-significant studies that do report precise statistics. This is a conservative approach, but not as extremely conservative as the zero substitution approach. We adopted this strategy for the overall analysis, to be able to include as many informative papers as possible. After computing effect sizes for each study we applied meta-analytic methods in order to obtain a combined effect size, which indicated the magnitude of the association across all studies (cf. Reference Nieuwenstein, Aleman and De HaanNieuwenstein et al, 2001). Effect sizes were weighted for sample size, in order to correct for upwardly biased estimation of the effect in small sample sizes (Reference RosenthalRosenthal, 1991). The corresponding z-value and significance level provide an indication of the statistical significance of the association. We also calculated a homogeneity statistic, Q, to test whether the studies can be taken to share a common population effect size. A significant Q statistic indicates heterogeneity of the individual study effect sizes which poses a threat to a reliable interpretation of the results. However, when a small number of studies are included in the meta-analysis, the Q statistic can underreport heterogeneity, and an exploration of heterogeneity is warranted through subgroup analyses. All analyses were carried out in the random-effects model, using the Comprehensive Meta-Analysis package (see http://www.meta-analysis.com). In contrast to the fixed-effects model, the random-effects model permits generalisation to studies not yet in the sample, and is to be preferred when studying psychiatric phenomena which may be not only empirically but also conceptually heterogeneous (Reference Rosenthal and DiMatteoRosenthal & DiMatteo, 2001).

Publication bias

A concern in the interpretation of meta-analytic findings is the possibility of an upward bias of the mean effect due to the omission of unpublished studies with null effects. Small studies with significant results tend to be published, whereas small studies without statistically significant findings tend to remain in the file drawer. Figure 4 is a funnel plot of the studies included in the primary analysis: this is a scatterplot of the effect size by sample size of each study, which should take the shape of a funnel, as there should be greater variability among the effect sizes based on small samples than those based on large samples (Reference Wang and BushmanWang & Bushman, 1998). The underrepresentation of small studies with small effects could be indicative of publication bias. Figure 4 shows that such studies are indeed somewhat underrepresented, but the asymmetry is not strong. To assess whether sampling bias would be large enough to render the mean effect size insignificant, we computed an additional statistic, the fail-safe n (Reference OrwinOrwin, 1983; Reference RosenthalRosenthal, 1991). Specifically, we determined the number of studies with an effect size of zero needed to reduce the mean effect size to a negligible effect (r=0.05). The fail-safe n was 84 for the overall analysis (35 studies) and 40 for the analysis confined to the diagnosis of schizophrenia (11 studies). It is unlikely that such a large number of unpublished studies with null effects reside in file drawers.

Fig. 4

Funnel plot of the studies included in the primary analysis: the effect size r (x-axis) is plotted against the sample size (y-axis) of each study (the vertical line indicates the mean weighted effect size).

Neuropsychological dysfunction

It has been a matter of debate in recent years whether poor insight in psychotic disorders can, in part, be explained by neuropsychological dysfunction. The association of insight with general cognitive function observed in our meta-analysis may imply that the interpretation and attribution of anomalous mental experiences is indeed hampered by poor general cognition. Notably, although the effect sizes were in the small to medium range, a recent meta-analysis meta-analysis of insight studies reported comparable mean correlations for the association with positive and negative symptoms; Mintz et al (Reference Mintz, Dobson and Romney2003) observed a correlation of 0.25 for positive symptoms (based on 22 studies) and a correlation of 0.23 for negative symptoms (based on 20 studies). Hence, it could be argued that poor insight in psychosis is part and parcel of psychopathology and cognitive dysfunction in roughly equal measure. However, although psychotic symptoms have been suggested to be largely independent of cognitive functioning (Reference Nieuwenstein, Aleman and De HaanNieuwenstein et al, 2001), any conclusion regarding independent contributions of psychopathological disorder and cognitive deficits to impaired insight should be confirmed by a regression model in the same sample. Furthermore, interpretations of association should not assume causality. It is possible that the assessment of insight is confounded by cognition, since a clinician's tendency to infer insight may be coloured by the patient's ability to express complex attitudes and experiences. Independent rating of insight and cognition would minimise this effect.

The finding that in patients with a psychotic disorder the relationship between WCST performance and insight was significantly stronger than the association with IQ may imply a specific role of perseveration. With regard to the cognitive mechanisms of insight, WCST performance has been suggested to be of particular relevance (Reference Drake and LewisDrake & Lewis, 2003). Cognitive flexibility is important, as it refers to the capacity ‘to hold an abstract representation related to an actual situation, but different from it, at the same time as the more obvious immediate representation’ (Reference Drake and LewisDrake & Lewis, 2003). This capacity would enable people to evaluate their own perceptions, thoughts and behaviour in relation to knowledge of symptoms of mental illness (shaped by social and cultural influences). In a sample of 33 people with acute psychosis, Drake & Lewis (Reference Drake and Lewis2003) reported a specific strong correlation (r=0.59) between insight and perseverative errors, rather than more general measures of abstraction. This is consistent with the notion that a failure to change cognitive set and to monitor error responses may lead to impaired insight. It should be kept in mind, none the less, that although ‘perseveration’ is widely recognised as a key cognitive process called upon during WCST performance, the cognitive processes underlying WCST performance remain poorly understood.

Diagnosis of schizophrenia

Whereas the WCST-insight association was stronger than the IQ-insight association in samples of patients with psychotic disorders in general, there was no difference between these associations when analyses were limited to samples of patients with a diagnosis of schizophrenia. This was not owing to a smaller effect size for WCST, but could be attributed to a larger effect size for IQ in patients with schizophrenia. Schizophrenia involves more negative symptoms and more pronounced intellectual impairment in comparison with other psychotic disorders, and this could lead to a larger correlation between IQ and insight; that is, a certain IQ level might be required to enable the person to make the complex judgements involved in self-reflection and monitoring of unusual or ambiguous experiences. Another difference between the analysis confined to schizophrenia proper concerned the relationship between memory performance and insight. This association was significant in schizophrenia, but was absent when all psychoses were included in the analysis. In schizophrenia, large effect sizes of memory impairment have been documented (Reference Aleman, Hijman and De HaanAleman et al, 1999) and a certain level of memory function may be necessary for intact insight.

An alternative interpretation of the different results in the analyses including schizophrenia v. ‘all psychoses’ would be that this does not imply a specific role of perseveration in impaired insight, but rather reflects an effect of clinical diagnosis. That is, in patients with less severe psychotic disorders only executive and attentional functioning might be impaired, whereas schizophrenia is characterised by generalised cognitive impairment. Thus, a relationship with insight is found only for cognitive functions that are impaired. Direct comparison between different patient groups is needed to test this possibility. The results of a study published after our meta-analysis was completed may be consistent with this notion (Reference Keshavan, Rabinowitz and DeSmedtKeshavan et al, 2004): in a sample of 535 patients with first-episode psychosis, the authors observed significant associations of impaired insight with multiple cognitive domains, including memory, learning and executive functions. However, another recent study of a sample comprising 122 out-patients with schizophrenia failed to find significant associations between insight and neuropsychological tests sensitive to frontal lobe function (Reference Freudenreich, Deckersbach and GoffFreudenreich et al, 2004).

Different scales

With regard to the use of different insight scales, the PANSS, SAI, SUMD and ITAQ apparently yielded effect sizes in the same range. This is consistent with studies that observed high correlations between these different insight measures (Reference Sanz, Constable and Lopez-IborSanz et al, 1998; Reference Cuesta, Peralta and ZarzuelaCuesta et al, 2000; Reference Drake and LewisDrake & Lewis, 2003). Unfortunately, owing to the small number of studies reporting separate data for different components of insight scales, it was not possible to analyse differential relations in this regard. It has been suggested that reduced symptom relabelling abilities might be more closely related to deficits in cognitive functioning than other domains of insight such as illness awareness and treatment compliance (Reference Morgan, David, Amador and DavidMorgan & David, 2004). The latter components would be less independent of external factors such as social and cultural variations (Reference Saravanan, Jacob and PrinceSaravanan et al, 2004).

A possible explanation for the small and inconsistent correlations observed in a number of studies might be that the relationship between insight and cognitive deficits is non-linear. According to Startup (Reference Startup1996), both motivational and cognitive deficits affect insight, with a trade-off between the two processes, giving rise to quadratic relationships between insight and cognitive test performance. Unfortunately, the other studies included in our meta-analysis did not examine curvilinear associations.

It should be noted that the effect sizes were rather small, implying that cognitive function may explain only a limited portion of the observed variance in insight ratings. The method of meta-analysis has been criticised for mixing dissimilar studies, publication bias and inclusion of poor-quality studies. We addressed these issues by imposing strict inclusion criteria and exploring publication bias. Indeed, as Rosenthal & DiMatteo (Reference Rosenthal and DiMatteo2001) have noted; ‘criticisms of meta-analysis that are applicable are equally applicable to traditional, non-quantitative, narrative reviews of the literature’.

Future research should focus more in detail on the role of metacognitive processes, and examine separate components of the insight construct. Koren et al (Reference Koren, Seidman and Poyurovsky2004) reported in a sample of 30 patients with first-episode schizophrenia that prediction of poor insight was significantly improved by adding the new, free-choice metacognitive measures to the conventional WCST measures. These preliminary results suggest that metacognition might be an important mediator between basic cognitive deficits and poor insight, and might be even more relevant to poor insight than cognitive deficits per se.