Introduction

Anorexia nervosa (AN) is a severe mental illness characterized predominantly by a preoccupation and distorted experience of weight and/or shape, an intense fear of gaining weight and behaviours that interfere with weight gain. Two subtypes of AN are described in the Diagnostic and Statistical Manual of Mental Disorders (DSM-5, American Psychiatric Association, 2013): (a) the restricting subtype, recognized by restricted energy intake and weight loss which is primarily achieved through dieting, fasting and/or excessive exercise, and (b) the binge-eating/purging subtype which is also characterized by restricted energy intake, but in combination with episodes of binge eating and/or purging behaviours (e.g. self-induced vomiting or laxative abuse), and/or excessive exercise. The prevalence rates of AN have been proposed to range between 0% and 0.9% (Hoek, 2006), and it has been suggested that the age of onset has decreased during the last two decades (Galmiche, Déchelotte, Lambert, & Tavolacci, 2019). In Denmark, the peak age of onset has gone down from 16–19 years to 12–15 years from 1995 to 2010 (Steinhausen & Jensen, 2015). In England, the prevalence of young patients (10- to 14-year olds) with AN has increased from 2.5/100,000 to 7.5/100,000 (Holland, Hall, Yeates, & Goldacre, 2016) – and cases as young as eight have been detected (Nicholls, Lynn, & Viner, 2011). A vast body of knowledge concerning genetic and neurobiological mechanisms in AN has accumulated in recent years, and there is increasing support for genetic, temperamental, neurobiological as well as psychosocial processes as influential factors for the onset of AN (Zipfel, Giel, Bulik, Hay, & Schmidt, 2015). Neuroimaging studies have shown brain atrophy in both young and older patients with AN, and it is unclear whether these reductions in brain tissue fully restore after recovery (Griffiths et al., 2020).

Neuropsychological tests are increasingly used for assessing cognitive functioning in psychiatric populations (Keefe, 1995), and it has been proposed that the cognitive inflexibility and compulsive behaviours commonly observed in AN (e.g. ritualized eating, body checking and exercise procedures) are reflections of underlying neuropsychological deficits in executive functioning and visuospatial processing (Lang, Stahl, Espie, Treasure, & Tchanturia, 2014; Lang & Tchanturia, 2014). In fact, both visuospatial processing and set shifting/cognitive flexibility have been proposed as endophenotypes, and a vulnerability trait that is characteristic of patients with AN (Kanakam, Raoult, Collier, & Treasure, 2013). However, this hypothesis has primarily been based on studies comprising of adult participants with recent findings being less consistent in younger populations (Brown et al., 2018; Weinbach, Bohon, & Lock, 2019). Furthermore, this research in younger patients has yet to be systematically reviewed. In a recent umbrella review by Smith, Mason, Johnson, Lavender, and Wonderlich (2018), only 28% of the quantitative reviews included young patients, and only two had participants under the age of 18 as their primary focus (Lang, Stahl, et al., 2014; Lang & Tchanturia, 2014).

Further, it is unclear whether potential neuropsychological inefficiencies could be associated with the state of malnutrition and/or duration of illness, which could leave a ‘scarring’ effect that persists after recovery. These potential confounding variables have been emphasized as a key area to address in future research (Smith et al., 2018). Smith et al. (2018) note that ‘it is also possible that deficits (e.g. set shifting, decision-making) are associated with duration of illness and/or age, though little research has addressed these issues’ (p. 816). By determining whether or not young patients with AN show inefficient neuropsychological functioning we could be closer to finding a putative distinct neuropsychological profile associated with AN.

The primary aim of the current study was to provide a comprehensive meta-analysis of neuropsychological functioning in young patients with AN, based on traditional domain classifications (Lezak, Howieson, Loring, Hannay, & Fischer, 2004; Strauss, Shermann, & Spreen, 2006), using standardized neuropsychological tests. Our second aim was to investigate potential important moderators for neuropsychological test performance in young patients with AN.

Methods Literature search strategy and information sources

The research was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines (Liberati et al., 2009) and was part of a larger systematic review on neuropsychological functioning in AN (Stedal, Broomfield, Hay, Touyz, & Scherer, 2021). Six major neuropsychological domains and 10 subdomains were the focus of our study. The domains, subdomains and neuropsychological tests were classified and chosen based on traditional neuropsychological classifications as determined by experts in the field (Lezak et al., 2004; Strauss et al., 2006) – as well as previous studies of comparable patient populations (Abramovitch et al., 2015; Geller et al., 2018).

The literature search was conducted in July 2019, updated in May 2020 and again in September 2020. A librarian at the Medical Library, Oslo University Hospital, conducted the search using the computerized databases of MEDLINE, PsycINFO, ISI Web of Science and Epistemonikos. Full search strategy for all databases, including limits used, can be found in Appendix S1.

Screening procedures Inclusion and exclusion criteria

Studies were considered a priori for inclusion if they: (a) evaluated a sample of young patients (mean age < 18 years) diagnosed with AN using DSM or ICD criteria; (b) included a comparison of a young (mean age < 18 years) healthy comparison (HC) group; (c) compared the groups using standardized neuropsychological tests to assess one or more of the following neuropsychological domains: attention, executive functions, memory, processing speed, visuospatial abilities, working memory; and (d) were written in English or had an available English-translated publication. Studies were excluded if they lacked a healthy control group, if they did not use validated, traditional and standardized neuropsychological tests (Lezak et al., 2004; Strauss et al., 2006), or if they employed outcome measures not considered a part of the original tests. Studies using modified versions of the original tests (e.g. emotional colour word interference test) were excluded. Only the original article, or the one with the most extensive report of the target information, was included when the same dataset was reported in more than one study. Reviews, theses, abstracts, books and book chapters were excluded.

Title, abstract and full text screening

After removal of duplicates, the search resulted in 3,313 titles. Studies which included ‘adults’ in the title or which were clearly ineligible were excluded. Equivocal and plausible titles were retained. After title screening, 2,888 out of the 3,313 publications were omitted and 425 titles were retained for abstract review. The abstracts were screened for eligibility by two authors (KS and CB) and labelled ‘include’ or ‘exclude’. Any discrepancies were resolved through discussion to reach consensus. Inter-rater agreement for extraction has previously been demonstrated to be high between the two reviewers for both title (κ = 0.80, p < .01) and abstract (κ = 0.87, p < .001) extraction (Stedal et al., 2021).

After reviewing 425 abstracts, 109 studies were retained for full text review and were screened by the first author. Seventy-three of these studies were conducted on an adult population and consequently excluded. Of the remaining 36 studies, four were excluded due to not providing a DSM diagnosis, five studies did not include sufficient results for calculation of effect sizes, three studies lacked a healthy control group, four studies combined patient groups (e.g. AN was combined with bulimia nervosa) or investigated patients with a lifetime diagnosis of AN. Finally, four studies employed experimental or modified tasks or the test was performed in the scanner. In total, we included 16 primary studies and 59 effect sizes. References for all included studies are provided in Appendix S2, and a detailed overview of the screening process is presented in Figure S1. Table S1 presents the domains, subdomains and tests coded from the studies and a description of included variables and coding procedures can be found in Appendix S3.

We computed Hedges’ g as an effect size of the standardized mean difference, utilizing the means and standard deviations for the AN and the HC group, and the corresponding sampling variances (Borenstein, Hedges, Higgins, & Rothstein, 2009). Negative effect sizes indicated lower performance of participants in the AN group on the neuropsychological tests compared to the HC group.

Data analysis Random-effects meta-analyses

We performed random-effects meta-analyses to obtain weighted average effect sizes () and indicators of the heterogeneity in the data, such as the variance component τ2 and the heterogeneity index I2 (Borenstein et al., 2009). However, some primary studies contributed multiple effect sizes due to the administration of multiple neuropsychological tests or subtests – hence the sampling errors within such studies are dependent and represent correlated effects (Pustejovsky & Tipton, 2021). As a consequence, we estimated random-effects models with robust variance estimation (RVE), assuming a constant within-study correlation between sampling errors (ρ = 0.30) that was informed by influential primary studies, systematic reviews and meta-analyses (e.g. Friedman & Miyake, 2017; Jewsbury, Bowden, & Strauss, 2016). We specified the meta-analytic models in the R packages ‘robumeta’ (Fisher, Tipton, & Zhipeng, 2017), ‘metafor’ (Viechtbauer, 2010) and ‘clubSandwich’ (Pustejovsky, 2021). Please find the detailed R code and output in the Data S1–S5.

Moderator analyses

The extracted moderators represented ‘participant characteristics’ (e.g. education, age, body mass index [BMI] and intelligence quotient [IQ]), ‘patient participant characteristics’ (e.g. duration of illness, illness onset and clinical severity), or ‘study characteristics’ (e.g. study quality score, study country, publication year and control for depression/anxiety). Table S2 presents the details of these characteristics. To examine whether these variables moderated the effect sizes, we extended the random-effects models to mixed-effects meta-regression models with RVE and inspected the regression coefficients (Hedges, Tipton, & Johnson, 2010). Categorical moderators were transformed into dummy-coded variables, and continuous variables were scaled. Categorical moderators with more than 10 categories were introduced as an additional level of analysis and the respective variance components were tested against zero via model comparisons (Cheung, 2013). To further examine the importance of moderators, we performed a multilevel random-effects, machine-learning meta-analysis, utilizing random forests in the R package ‘MetaForest’ (Van Lissa, 2020a). This algorithm was based on a 100-fold replicated feature selection and retained moderators that showed a positive variable importance in at least 50% of the replications in the preselection step (Van Lissa, 2020b). The final model was based on 10,000 regression trees Please find the detailed model specifications in Data S2.

Publication bias, influential effect sizes and sensitivity analyses

We evaluated possible publication bias via trim-and-fill analyses, the test for funnel plot asymmetry (Duval & Tweedie, 2000), Begg’s rank correlation test (Begg & Mazumdar, 1994) and the precision-effect estimate with standard error (PEESE; Fernández-Castilla et al., 2021). The trim-and-fill analyses were based on multilevel meta-analyses and the two estimators and (Fernández-Castilla et al., 2021). We further identified influential effect sizes using Viechtbauer and Cheung's (2010) diagnostics. Finally, we examined the sensitivity of the meta-analytic model parameters to (a) the choice of the within-study sampling correlation (ρ); (b) the exclusion of influential effect sizes; (c) the choice of a meta-analytic model and (d) the treatment of missing moderators.

Results Description of the primary studies

The total sample consisted of k = 59 effect sizes from m = 16 primary studies, and a total of n = 1,333 participants, of which n = 665 were participants diagnosed with AN, and n = 668 were healthy controls. Only one study specified subtype of AN (n = 24, diagnosed with the restrictive form of AN). Table S2 presents the participant characteristics and a detailed overview of the included studies can be found in Table S3.

Overall effect size and moderator effects Baseline model

Combining all effect sizes resulted in a small negative pooled effect size ( = −0.144, 95% CI [−0.328, 0.041], p = .12; see Figure 1), which did not differ significantly from zero. This effect size exhibited moderate heterogeneity (I2 = 65.8%, τ2 = 0.132) and was not sensitive to the choice of ρ (i.e.  = −0.144 and τ2 = 0.132–0.134 for ρ = 0.0–1.0). An alternative, three-level random-effects model showed a slightly higher effect ( = −0.181, 95% CI [−0.306, −0.057], p < .01;  = 0.102), yet did not support hierarchical effects with studies as clusters (=0.000). Hence, we used the random-effects model with RVE and correlated effects as the baseline model for the subsequent moderator analyses (for details, see Data S1).

Weighted average effect sizes and their 95% confidence intervals for the neuropsychological (sub)domains

Publication bias and influential effect sizes

Egger’s regression test (B = 0.66, SE = 2.16, p = .77), the PEESE (B = −0.38, SE = 1.79, p = .83) and the funnel plot test (B = 0.00, SE = 0.01, p = .81) did not indicate any selection bias and pointed to the symmetry of the funnel plot (see Data S1). Begg’s regression test supported this conclusion (r = −.01, p = .90). The trim-and-fill analyses confirmed that no additional effect sizes were missing ( =  = 0). No influential effect sizes or possible outliers could be detected.

Mixed-effects meta-regression models

The moderator analyses revealed an age effect, with older adolescent samples exhibiting a stronger negative effect (B = −0.23, SE = 0.05, p < .01). Moreover, samples of participants with a higher BMI also showed a stronger negative effect (B = −0.20, SE = 0.05, p = .01), as did samples of participants with more years of education (B = −0.32, SE = 0.10, p = .02). These effects remained even after treating missing data with multiple imputation (see Data S2). Better test quality was associated with less negative effect sizes (B = 0.33, SE = 0.15, p = .05). Compared to the effect size of executive functions, we observed a tendency towards lower and more negative effects for memory (B = −0.30, SE = 0.14, p = .08) and processing speed (B = −0.52, SE = 0.23, p = .06), and evidence for significantly lower and more negative effects for visuospatial abilities (B = −0.41, SE = 0.15, p = .03). The comparison between a three-level random-effects model with RVE and the baseline model suggested significant variation between countries (χ2[1] = 6.06, p = .01;  = 0.069). A similar comparison did not exhibit variation between neuropsychological outcomes and tests. Figure 2 provides an overview of the test-specific effect sizes ( = −0.72–0.12), and Data S5 contains the detailed parameters.

Weighted average effect sizes and their 95% confidence intervals for the neuropsychological tests and outcomes. bd, block design test; ce, commission errors; con, congruent trial; dr, delayed recall; dsy, digit-symbol test; int, interference trial; ir, immediate recall; lm, logical memory; oe, omission errors; org, organization score; pe, perseverative errors; pr, perseverative responses

Machine learning meta-analysis

The final random-forests model provided positive estimates of the out-of-bag and cross-validation variance explanation for a new data set ( = 0.524,  = 0.162) and identified seven moderators with positive variable importance that could be retained for at least 50% of the replications (see Figures 3a,b): Weighted average age of participants, country, diagnostic system, study quality, test quality, associations with eating disorders, and associations with depressions. Convergence was achieved with the 10,000 trees, and the cumulative MSE stabilized already after 5,000 trees (see Figure 3c). Figure 3b shows the variable importance and indicates that the sample age, country, and the diagnostic system were the most important predictors. Figure 3d shows the shape of the marginal effects for each of the seven moderators. These findings support the relevance of the previously identified moderators and provide additional information about their relative importance.

Results of the multilevel meta-analysis based on random forests. Gbr, United Kingdom; Can, Canada; Esp, Spain; Pol, Poland; Usa, United States of America; Den, Denmark; Deu, Germany. D3R, DSM-III-R; D3, DSM-III; D4TR, DSM-IV-TR; D4, DSM-IV; D5, DSM-5; I10, ICD-10; Y1, Yes (BMI); Y2, Yes (BMI and ED); Y3, Yes (BMI, illness duration and onset). ED, eating disorder; MSE, mean squared error; w_age, weighted average age in the sample (in years)

Domain-specific effect sizes and moderators

As a final step, we estimated the weighted average effect sizes for each domain and subdomain separately. Figure 1 exhibits the resultant effects and Table S4 displays the detailed model parameters. Overall, the domain-specific effect sizes ranged between  = −0.43 (central coherence) and  = 0.02 (attention). Only the effects for memory, nonverbal memory, visuospatial abilities, central coherence and working memory were statistically significant. Given that only few effect sizes were available for some domains and subdomains, we tested for moderator effects only for executive functions, memory and visuospatial abilities (see Data S3).

For executive functions, no difference between the effects of planning and response inhibition existed, and countries did not exhibit substantial variation in the effects. Similar to the entire sample, we found negative moderation effects of age (B = −0.26, SE = 0.05, p < .01), BMI (B = −0.24, SE = 0.06, p = .02), and years of education (B = −0.35, SE = 0.11, p = .03). Higher comorbidity in the AN group was associated with a less negative effect size (B = 0.19, SE = 0.06, p = .09), and so was higher test quality (B = 0.35, SE = 0.14, p = .04). Studies controlling for depression or anxiety exhibited a similar tendency.

For memory, we found a less negative effect size for verbal memory than for nonverbal memory (B = 0.23, SE = 0.06, p = .02). Besides, studies with AN early onset exhibited larger negative effects (B = −0.23, SE = 0.06, p = .02).

For visuospatial abilities, we could not identify any moderator among the sample, study and measurement characteristics.

Discussion

The primary aim of this study was to provide a broad assessment of neuropsychological functioning in young patients with AN, based on traditional domain classifications. A corresponding study on adults was recently published (Stedal et al., 2021) and the current results enable comparisons between these two populations. Our overall analyses revealed a small, non-significant, difference between patients and controls ( = −0.14). This is inconsistent with our findings in adults of a moderate, negative and statistically significant difference between patients and controls ( = −0.43, 95% CI [−0.50, −0.36]) (Stedal et al., 2021). Thus, the results support the position that underperformance is more prominent in older patients, and this may be related to their longer durations of illness (Grau, Magallón-Neri, Faus, & Feixas, 2019).

In the current study there were few significant effects for the majority of the included domains and subdomains. Exceptions were memory and working memory, as well as visuospatial abilities (including central coherence). It is important to note that there was an overlap of assessment tools across some domains. In particular, nonverbal memory and visuospatial abilities, including central coherence, which were all primarily assessed using the Rey Complex Figure Test (RCFT, Osterrieth, 1944). The RCFT is a complex task which involves elements of visuospatial processing, attention, visual-motor functioning, planning, memory and executive skills, including organization abilities (Shin, Park, Park, Seol, & Kwon, 2006). In fact, research adopting the RCFT (copy and organization condition) were the only studies which discriminated between individuals with AN and healthy controls with a medium, significant effect size. Scores on the copy and organization condition of the RCFT have traditionally been interpreted within the visuospatial memory and processing domain (Lezak et al., 2004; Strauss et al., 2006). In the eating disorder field, difficulties with global processing – often referred to as weak central coherence (Happé & Frith, 2006) – is frequently reported (Lang, Lopez, Stahl, Tchanturia, & Treasure, 2014). The current study confirms the utility of employing the RCFT when assessing individuals with AN. It also accentuates previous research highlighting the usefulness of a process oriented approach when examining this patient population (Stedal et al., 2019), together with underscoring the need for assessment tools which reflect everyday functioning, using an ecologically valid approach (Stedal & Dahlgren, 2015).

Executive functioning was found to be the most frequently assessed domain (k = 25), which was not unexpected given the widespread focus on this domain (particularly set shifting/cognitive flexibility) in the field of eating disorder research. The current study confirms more recent findings which demonstrate a comparable performance between young individuals with AN and healthy control participants on tasks of set shifting and other executive functions (van Noort, Pfeiffer, Ehrlich, Lehmkuhl, & Kappel, 2016). Interestingly, this is in contrast to our meta-analysis of adults, where we found not just an overall lower performance in the patient group, but that the patient group performed worse on all domains, including executive functions (Stedal et al., 2021). Thus, the current study confirms previous findings questioning the hereditability of set shifting/cognitive flexibility difficulties in individuals with AN (Weinbach et al., 2019), and suggest that executive functioning difficulties may develop over the course of illness – possibly as a consequence of prolonged starvation and/or malnutrition.

Our moderator analyses also showed a significant effect of age, revealing that older participants exhibited a stronger negative effect, that is, poorer neuropsychological test performance. Previous studies have suggested that prolonged illness may amplify neuropsychological difficulties (Lang, Stahl, et al., 2014) with preliminary evidence from a recently conducted systematic review suggesting this may be particularly relevant in the domains of decision-making, intellectual functioning and set shifting abilities (Broomfield, Stedal, & Touyz, 2021). Harper, Brodrick, Van Enkevort, and McAdams (2017) proposed that neuropsychological functioning is an important predictor of patients’ ability to recover. For example, difficulties with set shifting/cognitive flexibility can make it challenging to address some of the central treatment objectives of cognitive therapies – including goal setting, collaboration and thought experiments. Hence, if poor neuropsychological functioning can impair treatment response and complicate clinical progress in patients experiencing AN, it is imperative that patients are adequately assessed and treatments get tailored to address individual needs. Important to note is that we did not find an association between neuropsychological functioning and duration of illness, nor age of onset. However, only four of the included studies reported age of illness onset and less than half (37%) of the studies specified duration of illness. Therefore, we are cautious about making inferences regarding associations between illness duration and test performance. Furthermore, the incomplete data on illness duration limits conclusions that can be drawn pertaining to the association between years of education and neuropsychological test performance, since education will typically be largely confounded by both age and potential duration of illness in a child/adolescent sample. This also underscores our concern regarding the consistent lack of in-depth investigations of potential confounding variables that could influence results on neuropsychological tests.

To examine as many moderators as possible – despite low number of primary studies – we utilized machine-learning meta-analysis. This is a novel method that, to the authors’ knowledge, has not previously been used with this population. Machine learning moderator analysis can give an indication of which moderators most strongly predict effect sizes and has been shown to have sufficient power, even for small meta-analyses with few included studies (Van Lissa, 2017). Our machine learning analysis confirmed age as a strong effect modifier. This is not just of interest when discussing whether or not cognitive underperformance is a vulnerability trait for developing AN, but it also highlights some important considerations relating to test-selection in younger populations. The majority of neuropsychological tests utilized in the source studies build on the assumption that the examinee is an adult with a fully developed nervous system, and systematic observations of adults are the foundation of most neuropsychological theories (Reynolds & Mason, 2009). This is troublesome, since a child’s nervous system – which is under development – is significantly different from that of an adult (Holmes-Bernstein & Waber, 1990). Holmes-Bernstein and Waber (1990) state that the ‘direct application of adult methodology to the assessment of children is inappropriate’ (p. 313).

For example, the Wisconsin Card Sorting Test (WCST, Heaton & Staff, 1993) which was employed in six of the included studies, was originally designed for use with adults, and it has been demonstrated that adult level performance on the WCST is not reached before the age of 15 years (Lin, Chen, Yang, Hsiao, & Tien, 2000). Studies on nonclinical young populations have promoted a unity-but-diversity view, where executive function consists of components which are related, but separable (Best & Miller, 2010). Within this framework, complex tasks like the WCST – which tap multiple components of executive functions – are particularly poor at providing a nuanced picture of executive functioning in young individuals, since these components seem to mature at different ages (Best & Miller, 2010). This is further complicated by previous findings which indicate that young individuals experiencing AN might demonstrate disruptions to the expected age-related improvements in set shifting/cognitive flexibility (van Noort et al., 2016). Consequently, we encourage future studies to take into consideration the age of the participants when selecting neuropsychological assessment tools and to strive to employ tests specifically aimed at young participants undergoing cognitive development. In addition, for future primary studies – especially of younger individuals – it might be helpful to examine particular components of executive functioning by utilizing less complex tasks, which tap individual dimensions of executive functions including set shifting, inhibition and working memory (Miyake et al., 2000).

Our machine-learning moderator analysis also revealed that the country in which the study originated from was the second most influential moderator. This indicates that there could be some bias in the source studies, which should be taken into consideration when interpreting the results. The risk of bias is high with any study, and can occur at all stages of the research – from design of the study to writing up the results, and the conclusions which are drawn from the findings (Pannucci & Wilkins, 2010). Compared to adults, neuropsychological research on young patients with AN is scarce. Therefore, it is not surprising that several of the included studies are conducted by the same research labs. Indeed, three out of the four studies from Germany were conducted by the same research group, as were two of the three studies from the UK. Additionally, all of the studies from both Spain and Poland had overlapping authors. It is noteworthy that all the studies included in the current meta-analysis were conducted in western countries, despite eating disorders being regarded as a growing mental health issue in developing countries, together with increasing reports of self-starvation syndrome across all socioeconomic and ethnic backgrounds (Merry & Andrés, 2001; Palmer, 2008). Multiple factors may have contributed to this bias, including the availability of funding for conducting research, accessibility to healthcare services as well as diagnostic practices. For example, it has been suggested that certain diagnostic instruments, including the Eating Disorder Inventory (Garner, 2004), might skew eating disorder pathology results depending on the ethnicity of the individual (Wildes, Emery, & Simons, 2001). Consequently, our findings highlight the need for greater diversity amongst participants recruited in research, and replications of findings across laboratories. This would improve the accuracy of outcomes and mend social and cultural factors, which may impact appropriate assessment of eating disorders in minority groups. In addition, checklists to avoid different types of bias, like the one provided by Pannucci and Wilkins (2010), could aid in the planning and execution of future primary studies.

Limitations and future directions

The current study reveals an apparent lack of studies reporting possible confounding variables. This makes the interpretation of findings challenging, and it is difficult to disentangle what is the result of neurocognitive deficits and what is caused by other variables. In addition, the precision of the effect size interval estimate depends on several factors, including sample size and design. Due to the low number of available effect sizes for some of the domains; there is large uncertainty in the weighted average effect size estimates of these domains.

Conclusion

The present meta-analysis revealed a small cognitive underperformance in children and adolescents with AN, compared to controls, especially in memory, working memory and visuospatial abilities. However, some caution should be taken when interpreting these findings as they are largely based on neuropsychological tests developed for the assessment of adults. We urge future primary studies to utilize instruments specifically targeted at younger participants. Furthermore, there is currently no agreement on which variables to include as confounders, which may explain why several of the included studies neglected to assess or report on the impact of potential confounders. In order to move the field forward and facilitate the interpretation of findings, there is a need for more widespread inclusion of potential confounding variables – both in primary and meta-analytical studies.

Acknowledgements

This work was supported by the Clinic for Mental Health and Addiction, Oslo University Hospital, Ullevål HF, Norway (K.S., Grant No. 44079). The authors (K.S., R.S., S.T., P.H., C.B.) have no conflicts of interest to declare, all work on this paper was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. P.H. and S.T. are consultants to Takeda Pharmaceuticals.

Preregistration This work was registered in the Open Science Framework under the following protocol: https://doi.org/10.17605/OSF.IO/5SCRH. Key pointsCognitive underperformance has been demonstrated in adults with anorexia nervosa, but no meta-analysis has examined overall neuropsychological functioning in children and adolescents. The random-effects meta-analyses revealed that patients with anorexia nervosa did not differ from controls in overall neuropsychological functioning, but performed significantly lower in the memory, working memory, and visuospatial abilities domains. Mixed-effects meta-regression demonstrated a stronger negative effect in older participants and machine learning meta-analyses also revealed country and assessment quality as influential moderators. Our findings emphasise the importance of including potential confounding variables – both in primary and meta-analytical studies, and to utilize neuropsychological assessment tools specifically developed for younger participants.