Fit indices (CFI/RMSEA/SRMR)

Description

Quantitative indices that summarise how well a factor or structural equation model reproduces the observed data, used to accept, revise or reject a measurement model.

Because the chi-squared test of exact fit rejects almost any model in large samples, SEM uses approximate fit indices. CFI compares the model with a null model of uncorrelated variables (values near 1 are better). RMSEA estimates misfit per degree of freedom with a confidence interval (values near 0 are better). SRMR is the average standardised residual between observed and implied correlations. Hu and Bentler's widely used cut-offs (CFI at or above 0.95, RMSEA at or below 0.06, SRMR at or below 0.08) were derived from specific simulation conditions and should not be treated as universal thresholds. Rigour means reporting several indices together, examining local misfit, and avoiding chasing fit through unjustified modifications.

Attributes

Loss type
Siloed
Stage
Definition
Effort
Days

Steps

  1. Fit the model and report chi-squared with degrees of freedom and p-value even if expected to be significant.
  2. Report CFI and TLI (target at or above 0.95; 0.90 as marginal), RMSEA with 90% CI (at or below 0.06 good, above 0.10 poor) and SRMR (at or below 0.08).
  3. Inspect standardised residuals and modification indices for local misfit, investigating residual correlations above 0.10.
  4. Only modify the model where a change is theoretically justified, and cross-validate any modification on a holdout sample.
  5. Compare non-nested alternative models using AIC or BIC.

Sources

  • Principles and Practice of Structural Equation Modeling (Rex Kline)
  • A Primer on Partial Least Squares Structural Equation Modeling (Joseph Hair et al.)
  • DAGs+SEM Playbook (Vivek Gopinath)
  • Hu & Bentler (1999) Cutoff criteria for fit indexes in covariance structure analysis
  • Principles & Practice of SEM (Kline)
  • Primer on PLS-SEM (Hair et al.)
  • DAGs+SEM Playbook (Gopinath)
  • The Inquiry Series, Ch 49 — Structural Equation Modeling — From Quotes to Coefficients

Also known as

  • Model fit indices
  • Goodness-of-fit statistics
  • CFA model fit
  • Goodness-of-fit evaluation
  • Confirmatory factor analysis fit indices (CFI/RMSEA)

Categories

  • Analytics & Behaviour
  • Survey & Psychometrics