Confirmatory Factor Analysis (CFA): Testing Hypothesized Models

Defining the Hypothesized Model

A fundamental part of CFA involves specifying a hypothesized model prior to data analysis. This model, based on prior research or theory, shows how observed variables load onto latent factors. Factors are unobservable variables influencing observed variables.

In psychological research, for example, theorists may hypothesize that observed behaviors reflect underlying traits like anxiety or depression. CFA helps determine if the observed variables load onto specific factors, either unidimensional (one factor) or multidimensional (multiple factors).

Testing the Model Fit

CFA tests how well data fits the hypothesized structure, focusing on factor loadings, error variances, and potential covariances between latent factors. Several fit indices are used to evaluate this model, including:

• Chi-Square Test: A non-significant result indicates good fit, but it is sensitive to large samples.
• Comparative Fit Index (CFI): A value above 0.90 suggests acceptable fit.
• Tucker-Lewis Index (TLI): Similar to CFI but penalizes model complexity.
• Root Mean Square Error of Approximation (RMSEA): Values below 0.06 reflect good fit.
• Standardized Root Mean Square Residual (SRMR): A value below 0.08 suggests good fit.

Model Modifications

If the model does not fit well, researchers may modify it by adjusting paths between variables and factors. Modifications should be theoretically justified to avoid overfitting, which reduces generalizability. Modification indices help guide this process but must be approached cautiously.

Evaluating Competing Models

Sometimes researchers compare multiple models to identify the best fit. For example, a unidimensional model is compared to a multidimensional one. Criteria like the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) are used, with lower values indicating a better balance between fit and complexity.

Parameter Estimation

CFA estimates factor loadings, error variances, and factor correlations, usually using Maximum Likelihood (ML) estimation. When assumptions like multivariate normality aren't met, alternatives like robust maximum likelihood (MLR) or weighted least squares mean and variance adjusted (WLSMV) are applied.

Factor loadings closer to 1 indicate stronger relationships, while values near 0 suggest weak or non-significant relationships.

Measurement Invariance

Establishing measurement invariance ensures that factor structures are comparable across groups (e.g., genders or cultures). Invariance testing occurs in stages, from configural invariance (same factor structure) to metric invariance (equal loadings) and scalar invariance (equal intercepts). This ensures valid comparisons between groups.