advanced 11 min read
Social sciences · Topic
Confirmatory Factor Analysis
linear algebra · optimization · probability theory · eigenvalues
Confirmatory Factor Analysis (CFA) tests a researcher-specified factor structure rather than discovering it empirically. Fixed-zero constraints on cross-loadings encode theoretical assumptions, and overall model fit is evaluated with chi-squared tests and approximate fit indices, providing a formal hypothesis-testing framework for measurement models.

Fixed and Free Parameters

In CFA, each indicator $y_j$ loads on a pre-specified subset of factors. The implied covariance matrix is:

\[\boldsymbol{\Sigma}(\boldsymbol{\theta}) = \boldsymbol{\Lambda}\boldsymbol{\Phi}\boldsymbol{\Lambda}^\top + \boldsymbol{\Theta}\]

where $\boldsymbol{\Phi}$ is the factor correlation (or covariance) matrix and $\boldsymbol{\Theta} = \text{diag}(\psi_1, \ldots, \psi_p)$ contains unique variances. Free parameters are estimated; fixed parameters (usually 0 for cross-loadings) are held at their specified values.

Identification

Each factor requires at least one scale-setting constraint: either fix one loading per factor to 1 (marker indicator) or fix each factor variance to 1. A sufficient condition for local identification is the two-indicator rule (each factor has $\ge 2$ indicators with no cross-loadings) combined with a positive-definite $\boldsymbol{\Phi}$.

Model Fit and Modification

The chi-squared test statistic $\chi^2 = (N-1)F_{\min}$ has $df = p^* - q$ degrees of freedom. Because $\chi^2$ is sensitive to sample size, approximate indices are preferred:

Index Acceptable Good
CFI $\ge 0.90$ $\ge 0.95$
RMSEA $\le 0.08$ $\le 0.06$
SRMR $\le 0.10$ $\le 0.08$

Modification indices (MI) estimate the $\chi^2$ drop if a fixed parameter is freed. Freeing parameters post-hoc requires cross-validation to avoid capitalizing on chance.