The Bifactor Model Fits Better Than the Higher-Order Model in More Than 90% of Comparisons for Mental Abilities Test Batteries

Abstract

Abstract:The factor structure of mental abilities has most often been depicted using a higher-order model. Under this model, general mental ability (g) is placed at the top of a pyramid, with "loading" arrows going from it to the other factors of intelligence, which in turn go to subtest scores. In contrast, under the bifactor model (also known as the nested factors/direct hierarchical model), each subtest score has its own direct loading on g; the non-g factors (e.g., the broad abilities) do not mediate the relationships of the subtest scores with g. Here we summarized past research that compared the fit of higher-order and bifactor models using confirmatory factor analysis (CFA). We also analyzed additional archival datasets to compare the fit of the two models. Using a total database consisting of 31 test batteries, 58 datasets, and 1,712,509 test takers, we found stronger support for a bifactor model of g than for the traditional higher-order model. Across 166 comparisons, the bifactor model had median increases of 0.076 for the Comparative Fit Index (CFI), 0.083 for the Tucker-Lewis Index (TLI), and 0.078 for the Normed Fit Index (NFI) and decreases of 0.028 for the root mean square error of approximation (RMSEA) and 1343 for the Akaike Information Criterion (AIC). Consequently, researchers should consider using bifactor models when conducting CFAs. The bifactor model also makes the unique contributions of g and the broad abilities to subtest scores more salient to test users.

Construct Validity of the WISC-V in Clinical Cases: Exploratory and Confirmatory Factor Analyses of the 10 Primary Subtests

“…Furthermore, similar results have been reported with the Differential Ability Scales (DAS; Cucina & Howardson, 2017); DAS-II (Canivez & McGill, 2016; Dombrowski, Golay, McGill, & Canivez, 2018; Dombrowski, McGill, Canivez, & Peterson, 2019), Kaufman Adolescent and Adult Intelligence Test (Cucina & Howardson, 2017), Kaufman Assessment Battery for Children (KABC; Cucina & Howardson, 2017), KABC-2 (McGill & Dombrowski, 2018b), Stanford–Binet–Fifth Edition (SB-5; Canivez, 2008; DiStefano & Dombrowski, 2006), Wechsler Abbreviated Scale of Intelligence and Wide Range Intelligence Test (Canivez, Konold, Collins, & Wilson, 2009), Reynolds Intellectual Assessment Scales (Dombrowski, Watkins, & Brogan, 2009; Nelson & Canivez, 2012; Nelson, Canivez, Lindstrom, & Hatt, 2007), Cognitive Assessment System (Canivez, 2011), Woodcock-Johnson III (Cucina & Howardson, 2017; Dombrowski, 2013, 2014a, 2014b; Dombrowski & Watkins, 2013; Strickland, Watkins, & Caterino, 2015), and the Woodcock-Johnson IV Cognitive and full battery (Dombrowski, McGill, & Canivez, 2017a, 2017b), so results of domination of general intelligence and limited unique measurement of group factors are not unique to Wechsler scales. These results and the advantages of bifactor modeling for understanding test structure (Canivez, 2016; Cucina & Byle, 2017; Gignac, 2008; Reise, 2012) indicate that comparisons of bifactor models to the higher-order models are needed.…”

“…In bifactor models, g is conceptualized as a breadth factor (Gignac, 2008) because both the general ( g ) and the group factors directly influence the subtests and are at the same level of inference. Both g and first-order group factors are simultaneous abstractions derived from the observed subtest indicators and therefore should be considered a more parsimonious and less complicated conceptual model (Canivez, 2016; Cucina & Byle, 2017; Gignac, 2008). In bifactor models, the general factor direct subtest indicator influences are easy to interpret, both general and specific subtest influences can be simultaneously examined, and the psychometric properties necessary for determining scoring and interpretation of subscales can be directly examined (Canivez, 2016; Reise, 2012).…”

Genome-Wide Homozygosity Mapping Reveals Genes Associated With Cognitive Ability in Children From Saudi Arabia

“…The models differed in the number of estimated latent trait variances and covariance, and corresponding pattern and error loadings. Overall, several models displayed good fit to the data, and the bifactor model (see Figure 1 , also Table S3 ) was chosen as the final model based on theoretical considerations as well as recent data documenting the superior fit of the bifactor model to the data in most cases (Cucina and Byle, 2017). Briefly, the bifactor models postulates the existence of orthogonal specific ability factors that load on their respective items [i.e., in this case—domain-specific factors of verbal (Aurora-V), numerical (Aurora-N), and spatial/figural ability (Aurora-F)], while all items load on the common general factor (Aurora-G).…”

Validation of the Batería III Woodcock-Munoz Pruebas de Habilidades Cognitivas for assessment of cognitive ability among children living in the Peruvian Amazon

“…Under this bi-factor model, each subtest score loads directly onto g and separately onto the cognitive subdomain the subtest is hypothesized to measure. In this way the broad cognitive subdomain abilities are not considered to mediate the relationships between the subtest scores and g. Recent analysis of 31 test batteries in 58 datasets with over 1.7 million test takers has found consistently stronger support for the bi-factor model over the higher-order model (Cucina & Byle, 2017). In this study we examine whether the factor structure of scores from the Batería III COG among children living in the Peruvian Amazon may load onto 1) the hypothetically-tested cognitive subdomains alone, 2) a general g-factor alone, 3) a higher-order model where the cognitive subdomains load onto the g-factor, or 4) a bi-factor model where the test scores load onto the cognitive domains and onto g, separately.…”