Perhaps the oldest and most well-known example of the trend of converging aptitudes within a domain of individual differences was the development of the general intelligence factor (g), as initially proposed by Spearman at the fin de siècle (turn of the 20th century), based on evidence that children’s abilities on seemingly unrelated subjects were significantly correlated (Spearman, 1904). Although the existence of g was historically contentious, it is presently largely undisputed, based on evidence derived from psychometric factor analysis and genetic research (Plomin and Spinath, 2002; Jensen, 1987).

On the contrary, general cognitive intelligence has also been split into different components, prominently by Cattell who divided it into fluid and crystallized intelligence (Cattell, 1963). This model was subsequently developed into what is presently known as the Cattell-Horn-Carroll theory of intelligence which includes narrow, broad, and general ability as the three respective components of intelligence (Flanagan et al., 2013). In this sense, the most accepted modern theory of cognitive intelligence is hierarchical and compatible with g.

Nevertheless, other researchers, including Gardner, have suggested that there are a much larger number of different intelligences, such as musical, bodily-kinesthetic, and interpersonal and intrapersonal intelligence (Visser et al., 2006; Klitgaard and Gardner, 1984). Notably, Gardner’s multiple intelligence theory remains disputed in the scientific community due to scant substantive empirical evidence that would connect it to real-life outcomes, especially relative to g (Gottfredson, 2003; Ferrero et al., 2021; Tuncdogan et al., 2024). Additionally, research has indicated that g directly and significantly loads onto many of Gardner’s multiple intelligences (Visser et al., 2006).

Significantly, however, two of Gardner’s multiple intelligences (i.e., interpersonal and intrapersonal intelligence) seem to be closely tied to the construct that thereafter became known as emotional intelligence (EI) (Salovey and Mayer, 1990). EI subsequently developed into what can be regarded as the second major subsection of differential psychology and particularly gained cachet following the publication of Goleman’s eponymous book (Goleman, 1995). EI has bloomed into one of the most popular elements of contemporary psychology, enjoying particular prominence in the field of business psychology (Schmitt, 2006). As EI became increasingly predominant, the multitude of various measures that claimed to assess it generally coalesced into two major categories: ability and trait emotional intelligence (Petrides, 2011). Put succinctly, the difference between the two definitions is that whereas the former posits that emotional intelligence is an aptitude best assessed through performance tests, the latter regards EI as a set of personality-like dispositions that are best measured through self-evaluation (Petrides, 2011). Typically, correlations between the two have been reported to be modest, e.g., a corrected meta-analytic correlation of ρ = 0.26 (Joseph and Newman, 2010).

The third major subsection of differential psychology is perhaps the widest and most disparate and can be broadly described as personality (Furnham and Kanazawa, 2020). One of the earliest researchers that suggested a multifactorial system of personality was the eminent psychologist, Hans Eysenck, who proposed a bifactorial model that was later expanded to three dimensions, viz., extraversion, neuroticism, and psychoticism (van Kampen, 2009). Eysenck’s three-factor model was in significant contention with a competing five-factor model that was advocated by Costa and McCrae (1995). Interestingly, both camps of researchers argued that one another’s model could be reduced to the respective number of factors of which their preferred model was composed—a fact that could be construed as calling into question the factorial nature of both models. However, one of the models of personality that gained the most popularity was the Big Five model, which is composed of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism (Vassend and Skrondal, 2011; Anglim et al., 2022), although a significant number of competing models had also been developed.

Intriguingly, and perhaps predictably, a debate similar to the one that unfolded surrounding cognitive intelligence erupted in the domain of personality. While models of personality were composed of varying numbers of dimensions as initially proposed, evidence emerged that suggested that these dimensions were not fundamental, but in fact showed consistent intercorrelations, as demonstrated by theoretical reviews (Digman, 1997; Hofstee, 2003; Rushton and Irwing, 2009), single empirical studies (e.g., Musek, 2007), and meta-analytic evidence (van der Linden et al., 2010). This led to the positing of higher-order factors of personality, such as the two-factor model consisting of stability and plasticity (van der Linden et al., 2017). Furthermore, van der Linden et al. (2017) presented a model of personality wherein additionally to the Big Five traits loading onto stability and plasticity, the latter two factors further load onto an apex factor called the General Factor of Personality (GFP). Interestingly, the GFP had been meta-analytically found to be very strongly correlated (r = 0.86) with trait emotional intelligence (trait EI) (van der Linden et al., 2017), which led the authors to suggest that the two constructs may in fact be synonymous.

Outside of general personality assessment, instruments have emerged that focus on explicitly assessing the aversive aspects of personality, prominently including narcissism, psychopathy, and Machiavellianism, which are presently combined in models such as the Dark Triad and, in combination, with sadism, the dark tetrad (Dinić et al., 2021; Bonfá-Araujo et al., 2022). Narcissism is characterized by a disposition toward self-absorption, grandiosity, and is often associated with low empathy (di Giacomo et al., 2023). Psychopathy refers to an assortment of antisocial tendencies including boldness, social dominance, meanness, disinhibition, and is likewise associated with low empathy (Patrick, 2022; De Brito et al., 2021). Unlike the prior two traits, Machiavellianism is associated positively with empathy, but Machiavellians harness their empathic and interpersonal abilities cynically and exploitatively to achieve selfish aims (Jones and Mueller, 2021). Finally, sadism is defined as the tendency to behave cruelly toward others and especially by the desire to intentionally inflict pain (Foulkes, 2019).

Significant positive correlations have been found amongst the dark traits and emotional intelligence, despite the fact that EI is understood to be a positive rather than a maladaptive aspect of personality. In particular, recent meta-analyses have repeatedly found EI to be positively associated with narcissism (Michels and Schulze, 2021; Nguyen et al., 2022). A proposed theory explains this relationship through reference to a purported dark side of trait emotional intelligence (Davis and Nichols, 2016).

Finally, the fourth subsection of differential psychology consists of an assortment of miscellaneous integrative and contextual factors that do not neatly fit into the aforementioned categories of cognitive intelligence, emotional intelligence, and personality, but that do have significant associations with constructs belonging to said categories. These are primarily performance, outcome, or indicator measures. Amongst them are academic achievement; career-related factors: job performance, job satisfaction, leadership, and cultural intelligence (CQ); creative and aesthetic measures: creative achievement, creativity, and general aesthetic taste (T); wellbeing and related measures: self-concept, self-esteem, general self-efficacy, and satisfaction with life; as well as spirituality/religiosity. Of these different categories of measures, the data obtained throughout the meta-meta-analytic process provided linkages with the primary three categories only in the case of creativity.

Creativity is broadly defined as the ability to come up with novel and original solutions to a particular task (Zare and Flinchbaugh, 2018). Creativity has been found to be significantly correlated with cognitive ability, with one second-order meta-analytic estimate suggesting a relationship of r = 0.17 (Da Costa et al., 2015). However, correlations amongst these two constructs have been found to vary widely dependent on how creativity and intelligence are defined respective to one another and the type of creativity assessed (e.g., creative potential vs. outcome; Plucker et al., 2015, Runco, 2022).

Creativity and emotional intelligence are also reported to be positively related—though meta-analytic evidence has indicated that the relationship of creativity with trait EI is significantly stronger than the relationship between creativity and ability EI (r = 0.35 vs. r = 0.08; Xu et al., 2019).

Finally, with respect to the relationship between measures of creativity and the Big Five, meta-analytic data has indicated that creativity is most strongly linked to openness (r = 0.47), followed by extraversion (r = 0.26) and conscientiousness (r = 0.13; Karwowski and Lebuda, 2015).

In summary, while a number of miscellaneous performance, outcome, and contextual measures have been linked to the major domains of individual differences, amongst them creativity emerges as a linking construct. Namely, creativity has been found to be positively correlated with both cognitive ability and trait emotional intelligence as well as with personality measures. This therefore indicates that creativity may serve as a bridge across the aforementioned domains of individual differences and thereby is the miscellaneous measure that most merits special attention.

The general aim of this study was to identify whether a single latent factor can account for the shared variance amongst a broad assortment of differential psychology measures (including cognitive intelligence, emotional intelligence, personality, and miscellaneous constructs) by means of meta-analytic correlational data.

Conducting data analysis on meta-analytic data poses certain unique challenges that necessitate discussion. The process of fitting Structural Equation Models (SEM) to meta-analytic data is referred to as MASEM, or meta-analytic structural equation modeling (Jak et al., 2021). One of the advantages of the MASEM approach is that it empowers researchers to extend beyond the scope of the standard meta-analysis which is generally restricted to investigation of a single bivariate relationship (Oh, 2020). Another benefit is that it enables researchers to synthesize a complete correlational matrix, even if no single included meta-analysis includes relationships amongst all variables of interest (Oh, 2020).

The general procedure of conducting a MASEM consists of two steps wherein 1) a correlational matrix is constructed based on meta-analytic correlations derived from prior meta-analyses and 2) said input matrix is then utilized to carry out subsequent analyses (Oh, 2020; Cheung, 2014). Furthermore, there are two approaches to how the analyses are conducted: fixed- and random-effects models (Cheung, 2014; Sheng et al., 2016). The former family of approaches only deals with point estimates of effect sizes while ignoring the variances around them, leading to SEM results that unpredictably both under- or overestimate parameters (Oh, 2020; Jak and Cheung, 2022). The alternative option is the use of the two-stage SEM (TSSEM) method (Sheng et al., 2016) that allows for the unambiguous determination of sample size, while simultaneously accounting for sampling variance between meta-analyses by weighing the primary studies appropriately, i.e., by giving less weight to studies with greater amounts of sampling variance (Cheung and Chan, 2005).

However, TSSEM requires researchers to have access to the data of all primary studies that are included within each meta-analysis incorporated in the investigation (Oh, 2020). In instances where this is considered impossible or impractical, the full-information MASEM (FIMASEM) or the traditional SEM approaches are suggested (Oh, 2020). In the absence of standard deviation data, an option is to populate the input correlational matrix with previously derived meta-analytic estimates along with some corresponding sample size (Oh, 2020). However, such an approach may not sufficiently account for effect size heterogeneity (Jak and Cheung, 2022).

Considering the large number of meta-analyses that were incorporated in the present study, it was determined to proceed with SEM analyses using a correlational matrix composed of meta-analytic estimates as input.

Given extant meta-analytic evidence that indicates significant relationships amongst pairs of constructs belonging to the four enumerated primary aspects of differential psychology, the present meta-meta-analysis was designed to holistically examine the pattern of correlations amongst measures belonging to the aforementioned categories. Meta-analyses were included if they reported on the correlations between at least two constructs under investigation. In instances where correlational data for construct pairs was not obtained, the affected constructs were dropped from analysis.

The primary goal of the current study was to establish whether the relationships amongst differential psychology constructs form a latent construct that is strongly related to trait EI. It was expected that a latent factor would be obtained with observed uneven construct loadings. The identification of such a factor, however, even if it accounted for a small percentage of variance, would support the presence of a common thread underlying diverse constructs in differential psychology. This, in turn, could represent a psychological tendency that influences behavior across all domains. The breadth of constructs investigated in this research will allow us to conduct a preliminary examination of the hypothesis that a single background variable underpins mental activity in its entirety (Petrides, 2021).

Statistical methods, primarily designed to study variation and difference, are often ill-suited for the discovery of a quality that is omnipresent and equally present in mental activity. Nevertheless, any evidence of commonality across the diverse constructs examined herein would support this hypothesis. Such analyses would also contribute to the broader discussion of the role and nature of general factors in differential psychology.

In accordance with meta-analytic best practices, a protocol was prepared according to PRISMA-P guidelines (Shamseer et al., 2015) and registered on Open Science Framework (OSF).¹ There were deviations from the registered protocol, each of which had a minimal impact on the conclusions for the following reasons: a) Mini meta-analyses were not performed to avoid introducing bias from primary studies, and instead, only published meta-analyses were used to maximize data quality and homogeneity; b) Quality assessments were omitted because all included studies were published meta-analyses, for which standard primary-study quality ratings are not appropriate, while the meta-analytic process itself screens out low-quality research; c) When multiple meta-analyses existed for a construct pair, recency was prioritized over sample size to best reflect the current state of the literature; d) Fisher’s z-transformed matrices could not be analyzed due to non-positive definite matrices, so Pearson correlations were used instead, which is a standard approach in meta-analytic SEM; e) FE/RE TSSEM and OSMASEM methods were not used due to insufficient overlap amongst studies, so a pooled fixed-effects meta-analytic correlation matrix was adopted as a recognized alternative; f) Models including subfactors did not converge, which itself is informative and does not bias the main finding of a latent factor; g) Though sensitivity analyses were not performed, robustness was checked by means of alternative model specifications. In all cases, analytic choices were transparently data-driven and aimed at maximizing the validity and interpretability of the results.

As this was an exploratory research project, it was necessary for the authors to first prepare a list of keywords to target for inclusion prior to initiating the meta-meta-analysis. Therefore, 24 articles that were previously known to the authors were examined in order to generate a list of keywords that would be utilized to search for a general factor within the domain of differential psychology (see Supplementary Table S1). Thereafter, this initial list was expanded by including logical extensions and the resultant 120 keywords were sorted into five categories on the basis of their general categorical themes: namely, cognitive intelligence, emotional intelligence, personality, general factors, and miscellaneous.

Following the construction of this initial keyword list, test searches were conducted to assess the viability of the search strategy, with select meta-analyses retrieved and inspected to identify the constructs and measures they included. These constructs were subsequently combined and categorized, for a total of 24 construct categories that included 143 base constructs. These categories were: Overall Emotional Intelligence; Academic Achievement; Motivation; Empathy; General Factor of Personality; Cognitive Intelligence; Antisocial Personality Disorder; General Dark Triad; Cultural Intelligence; Generic Performance; General Creative Achievement; Defensiveness; Divergent Thinking; Career-Related Factors; General Taste; Leadership; Self-Concept; Self-Efficacy; Values; Interpersonal Factors; Response Time; WellBeing; and Religiosity. Following a literature review, this initial list was subsequently pared down to 19 categories.

Using this as foundation, an updated list of search keywords was created, with new keywords added while some older keywords were truncated. Simultaneously, a list of 34 target constructs was developed by means of author consensus on the basis of the aforementioned searches in order to precisely enumerate the constructs between which correlations would be subsequently sought and extracted. The list of targeted constructs was as follows: Ability Emotional Intelligence; Trait Emotional Intelligence; Emotional Manipulation (EMS); Academic Achievement (including Academic Performance and Academic Success); Affective Empathy; Cognitive Empathy; Openness; Conscientiousness; Extraversion; Agreeableness; Neuroticism / Emotional Stability; Honesty-Humility; Eysenckian Lie scale; Eysenckian Psychoticism; Full-Scale IQ / General Mental Ability; Crystallized Intelligence; Fluid Intelligence; Antisocial Personality Disorder (including Antisocial Behavior); Machiavellianism; Narcissism; Psychopathy; Cultural Intelligence (CQ); Job Performance; Job Satisfaction; Creative Achievement; Creativity; General Aesthetic Taste (T); Leadership; Self-Concept / Self-Esteem (including General Self-Concept, Self-Perceptions, and Self-Esteem); Self-Efficacy; General Values; General WellBeing (including WellBeing (various types) and SWB (Subjective WellBeing)); Satisfaction with Life; and Spirituality/Religiosity.

Following the development of the updated keyword lists (containing a total of 138 keywords; see Supplementary Table S2), a series of pilot searches were conducted to validate the search algorithm. This was done by checking whether certain meta-analyses known to the authors as relevant and related to the research question were appearing in the results as expected. Additionally, a small sample of the papers (n = 5) was examined to determine whether the obtained studies contained the required data. Given the fact that certain meta-analyses previously known to the authors failed to appear in the results, the search algorithm was consequently amended. Specifically, rather than conducting searches within the author-specified keywords specifically, searches were conducted within all available fields instead, including paper abstracts.

As the present study was exploratory in nature, the preparatory stage was conducted in a systematic and iterative manner so as to identify and categorize the relevant constructs for inclusion in the meta-meta-analysis. First, preliminary keywords were collated from an initial collection of 24 articles. This group of keywords was then expanded to 120 terms across thematic categories. Following test searches and a literature review, 19 categories of interest were developed, on the basis of which 34 constructs were targeted. Once the keyword lists were compiled, pilot searches were conducted to validate the search strategy. The algorithm was subsequently adjusted to search all fields to ensure all relevant articles were retrieved. In the end, a comprehensive and validated set of constructs and keywords was developed, which lay the foundation for the proceeding phases of the study.

The present meta-meta-analysis sought to identify all English-language studies that reported at least one meta-analytic correlation between two or more targeted constructs. The primary set of searches was conducted in January 2024 utilizing a search algorithm on Web of Science that included all studies published through to December 31st, 2022. A publication date cutoff was incorporated to enable interrater replicability. Due to the large number of articles that were expected to be retrieved, no search expansion approach was utilized. Corresponding authors were not contacted as the research was designed to not incorporate any primary or non-publicly-accessible data to ensure replicability and extension. The searches were repeated in June 2024 to update the scope of the meta-meta-analysis.

The search keywords included all possible combinations from Supplementary Table S2, along with the term “meta-analysis,” targeting all text fields. Article titles and keywords were collected for all results. In addition to the studies retrieved using the primary approach, manual searches were conducted across other databases (Ovid PsycArticles, Wiley Online Library, Elsevier ScienceDirect, Medline/Pubmed, Taylor and Francis, Google Scholar, and ResearchGate, in that respective order) for construct pairs lacking correlational data or when the sample size was considered insufficient. In all, two additional meta-analyses were obtained and retained using this secondary methodology.

Figure 1 illustrates the flow of articles through the screening stages. The number of studies that were initially retrieved was 3,322 following deduplication. Next, the papers were screened in three main phases (keyword/title, abstract, and full text screening). It was expected based on the pilot results, that the title and keyword screening phase would exclude a significant quantity of papers and was deemed necessary due to the high number of papers that were expected to be retrieved. The remaining stages of the screening process were standard and were inspired by past meta-analyses that were conducted in this area (Bücker et al., 2018; Fong et al., 2021; Gesel et al., 2020).

The papers were first screened based on their titles and author keywords. To minimize the risk of bias, ISSN/DOI data, publication dates, and journal names were hidden in this phase. The criteria were that the papers had to be meta-analytical and not focused on children and/or adolescents (<17 in age). Additionally, studies conducted on clinical populations were excluded except for when the construct under investigation was Anti-Social Disorder (ASD). Non-human research was also excluded. Following this initial screening phase, 781 studies were retained while 2,541 studies were excluded. Of the 2,541 studies that were excluded, the top reasons for exclusion were: 1,357 papers were excluded as they did not appear to involve related concepts, 607 were removed because they were conducted on clinical populations, and 237 because they were conducted on either children or adolescents.

Next, the 781 retained papers from the prior step were screened based on their abstracts according to the same criteria, but also excluding experimental designs, studies that did not include related constructs, and excluding studies that were unretrievable, duplicates, or retracted. A further 36 studies were excluded during this stage, with 745 studies retained for full-text analysis.

The full text screening phase excluded meta-analyses that incorporated primary studies focused primarily on adolescents (<17), contained internally inconsistent data, were missing summary correlational data, did not contain relevant data, or had sample sizes that were N < 100. The majority of the 584 articles excluded at the full text screening phase were excluded for the following reasons, in descending order: 365 did not contain relevant data, 78 did not contain summary level data (they included only data broken by category, for instance, only by gender), 71 were not true meta-analyses (for instance, were internal meta-analyses), and 43 were conducted on adolescents/children. Correlations were extracted from 162 studies.

The search repetition, conducted separately at a later date, yielded a further 70 studies, of which 26 were retained on the basis of title/keyword screening, of which 10 were kept following abstract screening. Full-text analysis revealed that one of these papers contained a relevant correlation.

In the first phase, 162 studies were found to contain relevant data, of which 92 contained meta-analytic correlations that were directly extracted, 13 referred to prior meta-analyses from which data was extracted, and 57 did not contain novel meta-analytic correlational data (viz., said data was already available in other cited meta-analyses). One additional correlation was extracted from the search repetition and two studies were retrieved separately thereafter in order to avoid data gaps. In total, there were 404 samples extracted from 108 articles (see Supplementary Table S3). For each study, the following data was extracted: construct pairing, correlational value, correlation type, N, k, uncorrected mean correlations, corrected correlations (rho), data source information (original meta-analysis vs. cited meta-analysis), and metadata (author names, year of publication, article title). In addition, a list of measures that were used to operationalize all constructs was extracted from each included meta-analysis, concatenated, and is presented in Supplementary Table S4. For the narcissism construct, all data were derived from instruments that specifically assess grandiose narcissism (and not vulnerable narcissism). Therefore, the narcissism indicator in the current study should be interpreted as representing grandiose narcissism only. For the Big Five intercorrelations specifically, the decision was made to reuse correlational data specified in a prior meta-analysis, as the sample size was substantial (van der Linden et al., 2010).

The secondary researcher performed a duplication of the searches, with the goal of confirming that the same studies were retrieved. Next, they screened the articles based on the agreed-upon inclusion/exclusion criteria in order to reduce potential selection bias. In all phases, all discrepancies were discussed and resolved prior to moving on to the next stage. Researcher interrater agreement was assessed by means of Cohen’s kappa and was calculated to be 0.94, which was interpreted as nearly perfect agreement.

Data from the included studies were manually extracted and assembled in a 34 × 34 matrix and inspected to assess what correlational data was found and what construct combinations were missing data. All extracted correlations were either corrected or uncorrected Pearson correlations. Given that the resultant matrix contained gaps, the matrix was pared down to 12 × 12, including data from the additional meta-analyses. The secondary researcher also extracted correlational data from all studies and these data extractions were cross verified to ensure no discrepancies were present. Appendix Table A5 specifies which studies were utilized for the construction of this matrix.

Statistical analyses were performed using SPSS v. 29, SPSS AMOS v. 29, and RStudio v. 2023.12.0 + 369. The general statistical procedure that was followed adhered to the recommendations in prior literature (Morin et al., 2020; Morin, 2023). The primary goal of the present study was to assess whether a general latent factor could account for the shared variance amongst a range of differential psychology constructs using meta-analytic data. In pursuit of this aim, we employed a stepwise analytic strategy, progressing from exploratory to increasingly more constrained models.

Specifically, we conducted 1) Exploratory Factor Analyses (EFA); 2) Exploratory Structure Equation Modeling (ESEM) within a Confirmatory Factor Analysis (ESEM-within-CFA) (Morin et al., 2013); 3) a series of Confirmatory Factor Analyses (CFA); 4) bifactor CFA and ESEM models; 5) and finally a unifactorial CFA solution with constraints.

EFA was conducted first in an effort to identify the underlying factor structure amongst the included constructs without any theory-derived constraints. This stage was necessary due to the breadth of the included constructs and the resultant lack of confidence in any concrete a priori model. Next, ESEM-within-CFA was used to model cross-loadings utilizing a more flexible representation that is afforded by this approach as compared to standard CFA. Following this, CFA was employed to assess whether a multifactorial or a unifactorial model better accounted for the observed pattern of correlations. In order to examine the possibility of a solution that included both a latent general factor as well as a number of domain specific factors, bifactor models were assessed. Finally, a unifactorial CFA model was used to determine if model fit could be improved by means of accounting for a number of known construct interrelations that are not captured by the latent factor in isolation. At each stage, fit indices were used to guide decisions, favoring models that were both more theoretically plausible and better fitting. By following this sequence, we were able to test alternative explanations that were driven by both data and theory and to determine whether the hypothesized latent factor could explain the observed shared variance amongst the constructs.

Prior to initiating data analysis, an input matrix was generated that included all constructs for which meta-analytic correlations were obtained. However, seeing as this matrix included gaps, the decision was made to remove the constructs for which meta-analytic estimates were not obtained. This approach was justified by the concern that with such a great number of empty cells, there was a significant risk of bias and Type I error (Sheng et al., 2016). In instances where both reliability-uncorrected and corrected correlations were provided, the corrected values were preferred. In addition, the possibility of overlapping primary studies was considered if multiple retrieved meta-analyses were allowed to contribute correlational data for the same pair of constructs (Türk et al., 2023). To avoid this issue, for each unique construct pair, only the meta-analysis with the largest sample size was retained. The resultant gap-free construct matrix is presented in Table 1.

With respect to the sample size specified for analysis, we followed the best practice that recommended using the harmonic mean as the sample size for the model specification as the most defensible choice from the standpoint of statistical rigor (Sheng et al., 2016). Finally, in accordance with the recommendations in the same article, we checked whether the input matrix was nonpositive definite prior to proceeding with analysis.

The goodness of fit of the various models assessed was carried out through the use of the Comparative Fit Index (CFI), the Tucker-Lewis Index (TLI), the Root Mean Square Error of Approximation (RMSEA) (including 90% confidence intervals), and the Standardized Root Mean Square Residual (SRMR). The fit thresholds that indicated good fit were selected from prior literature as follows: CFI > 0.90, TLI > 0.90, RMSEA < 0.08, SRMR < 0.08 (Hoyle, 1995).

Although the present meta-meta-analysis provides initial support for trait EI representing the shared variance amongst differential psychology constructs, there were limitations that ought to be discussed that were consequent of the exploratory nature of the present study.

It is important to note that the strong loading of trait EI (0.96) on the general factor is potentially partially ascribable to the specific methodology used in the present study, including the modeling approach and the indicators used. In particular, the final model incorporated a number of correlated residuals, which are justified theoretically and empirically but also shape the interpretation of the factor. It is also notable that the majority of the included indicators were self-report personality measures, while cognitive ability (a performance measure) was found to load only weakly onto the general factor. Therefore, future investigations are required to provide further supporting evidence for the centrality of trait EI in the domain of individual differences, as well as its link to the general factor of individual differences.

A potential limitation is that meta-analyses incorporate, by their nature, an assortment of primary studies, each of which potentially operationalizes constructs using different measures, some of which are further classified in divergent ways across different meta-analyses. In order to resolve this issue, one would need to perform meta-analyses strictly on primary studies, while incorporating the particular instrument used to operationalize a construct as a moderator. This is potentially extremely difficult to carry out in practice in cases when a great number of constructs are incorporated within one study, each of which may be measured by one of dozens of different instruments (see Supplementary Table S4).

This issue—wherein meta-analyses (and consequently, the meta-meta-analyses that rely upon them) treat potentially disparate measures as equivalent or even identical—is an enduring and fundamental concern. As Eysenck (1994) emphasized, meta-analyses can mislead by indicating that which primary studies do not show, obscuring that which they do, and generally giving rise to confusion and disorder. If, indeed, “apples and oranges” are aggregated indiscriminately, such investigations may obscure more than they illuminate. In the present study, we attempted to mitigate some of these concerns by explicitly avoiding primary study data, in order to reduce the risk of including low quality or idiosyncratic primary findings. Conversely, however, this means that we are necessarily at the mercy of the choices made by the prior meta-analysts with respect to their choices regarding construct operationalization and inclusion criteria. As such, our findings ought to be interpreted with caution, in acknowledgement of the persistent challenge of construct heterogeneity and conceptual ambiguity—not to mention the field’s fundamental flaw of relying on operationalism as a means of defining constructs in the first place (Petrides, 2019; Bringmann et al., 2022).

Most importantly, the present study was limited by its reliance on meta-analytic (as opposed to primary study) data, although this is also its strength. As noted in prior literature, populating an input correlation matrix with previously derived meta-analytic estimates along with some corresponding sample size is a potential source of bias (Oh, 2020). This is because such an approach wholly ignores effect size heterogeneity, which is likely to produce biased standard errors, thereby potentially giving rise to uninterpretable results (Jak and Cheung, 2022). In other words, conducting analyses on an input correlational matrix that consists of meta-analytic estimates (without accounting for the variance within each meta-analytic estimate) implies a loss of information that prevents one from assessing and accounting for heterogeneity.

This is a limitation that ought to be addressed in future studies, through either the extraction of primary study data for each included meta-analysis or the incorporation of true standard deviations for each effect size. Such an approach would also enable the use of modern MASEM analytical approaches including TSSEM and OSMASEM in the case of primary data and FIMASEM in the case of the latter approach (Oh, 2020; Jak et al., 2021; Cheung and Chan, 2005). As an additional benefit, this would also enable analysis on correlational matrices that contain gaps (that is, matrices that are missing correlations for certain construct pairings), which would allow for a far greater number of constructs to be incorporated in analysis (Jak and Cheung, 2017). On the other hand, the present analytical approach required a gap-free input correlational matrix which severely limited the number of constructs that could be incorporated within the analysis.

Additionally, reliance on prior meta-analyses implies accepting the limitations of the statistical approaches that were employed in prior studies. Specifically, the prevalence of basic statistical methods in extant psychology literature, especially including the utilization of simple correlations which assume linearity of relations between pairs of variables, implies a transference of that limitation to the present study. Therefore, in order to delve deeper into the true nature of relations between constructs, primary studies applying novel statistical methodologies are necessary.

It is theorized that the inclusion of additional constructs into a unifactorial model similar to the one presented herein (including constructs that were initially proposed for inclusion and for which meta-analytic data is included in Table 3) would yield evidence that would lend stronger and broader support for the theoretical latent factor proposed within this study.

Future studies may also validate the existence of the latent factor through the administration of a comprehensive questionnaire that includes measures of all described constructs to study participants (see Supplementary Table S4). Even greater validity and support could be indicated through the use of a longitudinal paradigm wherein participants would be retested using the same questionnaire at regular intervals. Additionally, the longitudinal paradigm combined with the neuroscientific approach could provide evidence that corroborates the activating role of trait EI in association with the achievement of various outcomes.