Why Do Bi-Factor Models Outperform Higher-Order g Factor Models? A Network Perspective | JoVE Visualize

Area of Science:

Psychometrics
Cognitive Psychology
Statistical Modeling

Background:

Bi-factor models of intelligence often show superior statistical fit compared to higher-order general intelligence (g) factor models.
Existing explanations for this discrepancy include the bi-factor model reflecting the true data-generating mechanism, bias in fit indices, or an underlying network structure.
Investigating these competing explanations is crucial for accurate intelligence modeling.

Purpose of the Study:

To investigate the validity and plausibility of three competing explanations for the statistical superiority of bi-factor models over higher-order g factor models.
To determine whether fit indices are biased against higher-order models or if a network structure better explains intelligence data.
To provide guidance for future intelligence model selection procedures.

Main Methods:

A Monte Carlo simulation was employed, generating 3000 data sets based on bi-factor, higher-order factor, and network models.
Parameter values were derived from confirmatory analyses of the Wechsler Scale of Intelligence IV.
Three models were refitted to each simulated data set, fit statistics were obtained, and model selection procedures were performed.

Main Results:

No evidence of bias in fit measures themselves was found; however, biased inferences can occur when using approximate or incremental fit indices inappropriately.
The network explanation for intelligence structure was validated and found plausible, consistent with prior empirical findings.
Empirical findings contradicted the hypothesis that a bi-factor model represents the true underlying structure of intelligence.

Conclusions:

Network models of intelligence are a plausible alternative and should be considered in future model selection, particularly when higher-order g factor models are rejected in favor of bi-factor models.
The study highlights potential pitfalls in interpreting fit indices, emphasizing the need for careful application.
The findings challenge the assumption that bi-factor models necessarily represent the true structure of intelligence.