Confirmatory Factor Analysis (CFA) Explained: How to Run, Read and Report It
Confirmatory factor analysis (CFA) asks a sharper question than its exploratory cousin: does a specific, theory-based measurement model fit the data? You state in advance which items measure which factors, and the analysis tests whether that structure can reproduce the observed correlations well enough. CFA is the standard tool for validating a questionnaire, and it is the measurement part of structural equation modelling. This guide explains the method, a worked example with checkable fit statistics, how to read the output and how to report it.
What is confirmatory factor analysis?
CFA is a latent-variable method in which each observed item is allowed to load only on the factors you specify, with all other loadings fixed at zero. Because the model is constrained, it makes testable predictions about the covariance matrix, and its fit can be judged. By contrast, exploratory factor analysis lets every item load on every factor and reveals the structure; CFA then tests a structure on new data. Brown (2015) and Kline (2023) are standard references.
When should you use CFA?
- Testing whether a scale has the dimensional structure its authors propose.
- Following up an EFA on an independent sample.
- Checking whether an instrument translated or adapted for a new population keeps its structure.
- Comparing competing measurement models, such as one factor versus two correlated factors.
- Preparing the measurement model for a full structural equation model.
CFA versus EFA. Use EFA when you do not know the structure and CFA when you do. Never explore and confirm on the same data, because a structure suggested by a sample will tend to fit that sample well. The SEM tool extends CFA by adding structural paths among factors.
The mathematical foundation
For observed items x and latent factors ξ, the measurement model is:
x = Λξ + δ
and the model-implied covariance matrix is:
Σ(θ) = ΛΦΛ′ + Θ
Λ holds the free loadings (zeros where an item does not load), Φ the factor variances and covariances, and Θ the residual variances. Estimation chooses the parameters θ so that Σ(θ) is as close as possible to the sample covariance matrix S, usually by maximum likelihood.
Identification and degrees of freedom. With p items there are p(p + 1)/2 unique variances and covariances. The model's degrees of freedom are that number minus the number of freely estimated parameters. A model needs df ≥ 0, and each factor needs a scale, set either by fixing one loading to 1 or by fixing the factor variance to 1.
Fit indices. With model chi-square χ²M (dfM) and the baseline (independence) model chi-square χ²B (dfB):
RMSEA = √[max(χ²M − dfM, 0) / (dfM(N − 1))]
CFI = 1 − max(χ²M − dfM, 0) / max(χ²B − dfB, χ²M − dfM, 0)
TLI = [(χ²B/dfB) − (χ²M/dfM)] / [(χ²B/dfB) − 1]
SRMR is the root of the average squared difference between observed and model-implied correlations. Hu and Bentler (1999) proposed guideline values (roughly CFI and TLI ≥ .95, RMSEA ≤ .06, SRMR ≤ .08), and these are best treated as rules of thumb, not pass-fail thresholds.
Assumptions and data requirements
- A theory-based model specified in advance. Which items load on which factors, and whether factors may correlate.
- Continuous or approximately continuous items. For ordinal items with few categories or marked skew, estimators designed for ordinal data (such as weighted least squares variants) are preferable.
- Multivariate normality for standard maximum likelihood. Check skew and kurtosis; consider robust estimators if the assumption is doubtful.
- Adequate sample size. Fixed rules are weak. Models with strong loadings and few factors can work with samples of a few hundred or fewer; weaker loadings and more complex models need more.
- At least three items per factor for a well-identified factor (two can work if the factor correlates with others).
- Independent observations. Nested data need multilevel approaches.
- Complete data or a stated missing-data strategy.
Original research example (illustrative)
Background. Following an exploratory analysis in a pilot sample, a lecturer wants to confirm the two-factor structure of a six-item student engagement scale in a new, independent sample. Items 1–3 measure behavioural engagement (F1) and items 4–6 emotional engagement (F2).
Objective and question. Does the specified two-factor model, with no cross-loadings, fit the data adequately?
Data. N = 200 new students, items rated 1–5. All values below are invented for teaching.
Step 1: Specify the model and count the degrees of freedom
Six items give 6 × 7 / 2 = 21 unique variances and covariances. The model has six loadings, six residual variances and one factor correlation, with both factor variances fixed to 1 for scaling: 13 free parameters. Degrees of freedom = 21 − 13 = 8, so the model is over-identified and testable. The independence (baseline) model estimates only the six variances, so dfB = 21 − 6 = 15.
Step 2: Evaluate the fit
| Statistic | Value | Guideline | Verdict |
|---|---|---|---|
| χ²(8) | 14.20, p = .077 | Non-significant preferred | Acceptable |
| CFI | .984 | ≥ .95 | Good |
| TLI | .971 | ≥ .95 | Good |
| RMSEA | .062 | ≤ .06 (good), ≤ .08 (acceptable) | Acceptable |
| SRMR | .041 | ≤ .08 | Good |
Check the arithmetic. Baseline model: χ²(15) = 412.60. RMSEA = √[(14.20 − 8) / (8 × 199)] = √(6.20 / 1592) = √.003894 = .062. CFI = 1 − (14.20 − 8) / (412.60 − 15) = 1 − 6.20 / 397.60 = 1 − .0156 = .984. TLI = (412.60/15 − 14.20/8) / (412.60/15 − 1) = (27.507 − 1.775) / 26.507 = 25.732 / 26.507 = .971. For χ²(8) = 14.20, the 5% critical value is 15.51, so the model is not rejected (p ≈ .077).
Step 3: Examine the loadings and reliability
| Item | Factor | Loading | Loading² | Residual |
|---|---|---|---|---|
| 1. Completes weekly tasks | F1 | .78 | .61 | .39 |
| 2. Logs in regularly | F1 | .81 | .66 | .34 |
| 3. Attends live sessions | F1 | .74 | .55 | .45 |
| 4. Finds content interesting | F2 | .83 | .69 | .31 |
| 5. Enjoys the course | F2 | .77 | .59 | .41 |
| 6. Feels part of the class | F2 | .72 | .52 | .48 |
Factor correlation (F1 with F2): r = .38 (illustrative). Composite reliability = (Σλ)² / [(Σλ)² + Σ residuals]. For F1, Σλ = 2.33, so (2.33)² = 5.43 and residuals sum to 1.19: CR = 5.43 / (5.43 + 1.19) = .82. For F2, Σλ = 2.32, (2.32)² = 5.38, residuals 1.20: CR = 5.38 / 6.58 = .82. Average variance extracted (AVE) = mean of squared loadings: F1 = (.6084 + .6561 + .5476)/3 = .60; F2 = (.6889 + .5929 + .5184)/3 = .60 (Fornell & Larcker, 1981).
Step 4: Interpret
The two-factor model fits well by conventional guidelines, all standardised loadings are .72 or higher, both factors have composite reliability of .82 and AVE of .60, and the factors correlate moderately (.38), which supports treating them as related but distinct. Conclusion: the illustrative data support the specified structure. This is evidence of structural validity in this sample, not proof: it should be complemented by evidence of reliability, convergent and discriminant validity, and, where relevant, measurement invariance across groups. A good fit also does not rule out other models that fit equally well.
How to run CFA with the StatFAS CFA Calculator
Open statfas.com/confirmatory-factor-analysis-online.html.
- Prepare your data. One row per participant, one column per item, a header row and numeric codes. Reverse-score negatively worded items. Use a sample independent of any prior EFA.
- Specify the model. Assign each item to its factor, decide whether factors correlate, and set the scaling method.
- Set options. Estimator, handling of missing data and which fit indices to display.
- Run and review. Check that the model converged without warnings, then fit indices, then loadings and residuals.
- Download results. Verify every number and sentence before it enters a manuscript.
statfas-cfa-calculator-interface.pngAlt text: "StatFAS confirmatory factor analysis calculator interface before data entry."
Caption: The CFA Calculator on StatFAS.com.
statfas-cfa-example-dataset.png · Alt: "Six-item engagement dataset loaded in the StatFAS CFA calculator."Screenshot 3:
statfas-cfa-model-specification.png · Alt: "Two-factor model specification in StatFAS CFA."Screenshot 4:
statfas-cfa-fit-indices.png · Alt: "Model fit indices table from StatFAS CFA."Screenshot 5:
statfas-cfa-loadings-or-diagram.png · Alt: "Standardised loadings table or path diagram from StatFAS CFA." Use only genuine captures of the tool.Reading the output
- Convergence and warnings. Negative residual variances (Heywood cases), standardised loadings above 1 or non-convergence signal a problem. Mistake: interpreting fit from a solution that has warnings.
- Chi-square. A significant value indicates exact-fit rejection, but with large samples it is sensitive to trivial misfit. Report it and rely on approximate-fit indices too.
- CFI, TLI, RMSEA, SRMR. Use together, not singly, and as guidelines (Hu & Bentler, 1999). Report the RMSEA confidence interval.
- Standardised loadings. Values of about .50–.70 or higher indicate items that represent their factor well. Low loadings flag weak items.
- Factor correlations. Values near .85 or above question whether the factors are empirically distinct.
- Modification indices and residuals. They point to local misfit, such as correlated residuals or cross-loadings. Mistake: freeing parameters purely to improve fit without a theoretical reason, which turns confirmation into exploration.
APA 7 reporting example (based on the illustrative data)
A confirmatory factor analysis was conducted on six engagement items (N = 200) using maximum likelihood estimation, testing a two-factor model with items 1–3 on behavioural engagement and items 4–6 on emotional engagement. The model showed good fit to the data, χ²(8) = 14.20, p = .077, CFI = .984, TLI = .971, RMSEA = .062, SRMR = .041. Standardised loadings ranged from .72 to .83, and the factors were moderately correlated (r = .38). Composite reliability was .82 for both factors. Table 1 presents the standardised parameter estimates.
Essential elements: the model specified and its theoretical basis, sample size and its independence from any exploratory sample, estimator, handling of missing data, χ² with df and p, CFI, TLI, RMSEA (with interval) and SRMR, standardised loadings in a table, factor correlations, reliability, and any model modifications with justification.
Common mistakes and troubleshooting
- Confirming on the exploratory sample. Use a new sample or a split.
- Judging fit by one index or by fixed cut-offs. Use several indices and consider the context.
- Specification searching. Repeatedly freeing parameters from modification indices capitalises on chance; cross-validate any changes.
- Under-identified models. Check the degrees of freedom and that each factor has a scale.
- Heywood cases. Usually from small samples, too few items per factor or misspecification.
- Ignoring non-normality or ordinal data. Use robust or ordinal estimators when needed.
- Treating good fit as proof of validity. Equivalent models can fit equally well, and fit says nothing about whether the factors mean what you think.
- Comparing non-nested models by chi-square difference. Use information criteria or other appropriate comparisons.
Limitations and alternatives
CFA is demanding about sample size and model specification, and its assumption of exact zero cross-loadings is strict. Alternatives and extensions include exploratory structural equation modelling, which allows small cross-loadings; Bayesian CFA with informative priors on small cross-loadings; bifactor and higher-order models for hierarchical structures; and item response theory. To go from measurement to structural paths, use structural equation modelling. To explore structure first, use EFA and PCA, and inspect items with the descriptive statistics calculator and the correlation calculator. A heatmap of the item correlation matrix is a useful first look.
Frequently asked questions
What is the difference between EFA and CFA?
EFA discovers a structure; CFA tests a structure you specified in advance by fixing some loadings to zero.
Which fit indices should I report?
The chi-square with df and p, CFI, TLI, RMSEA with its confidence interval and SRMR. Treat the usual benchmarks as guidelines.
Can I run CFA on the same sample as EFA?
It is not advisable. Confirm the structure in an independent sample or a held-out split.
What sample size do I need for CFA?
There is no single rule. Strong loadings, few factors and several items per factor need less; weak loadings and complex models need more. Consider a simulation-based power analysis for the planned model.
Conclusion
Good CFA is a chain of defensible decisions: a model specified from theory, an independent sample, a suitable estimator, several fit indices read with judgement, and any modification justified and cross-validated. When you are ready to apply it to your data, the StatFAS CFA Calculator provides a browser-based route, and its results should be checked by you before publication.
References
- Brown, T. A. (2015). Confirmatory factor analysis for applied research (2nd ed.). Guilford Press.
- Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18(1), 39–50.
- Hu, L., & Bentler, P. M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6(1), 1–55.
- Kline, R. B. (2023). Principles and practice of structural equation modeling (5th ed.). Guilford Press.