Confirmatory factor analysis model fit indices (CFI, TLI, and RMSEA) are numbers that tell you how well your proposed factor structure matches your actual data. A CFI and TLI of .95 or higher and an RMSEA of .06 or lower are the most widely used signs of good fit. If you have run a CFA for your thesis and are staring at a wall of output, wondering which numbers matter and whether yours are "good enough", this guide walks you through it.
Key Takeaways
- CFI and TLI compare your model to a "nothing is related" baseline model. Higher is better, and values of .95 or above indicate good fit.
- RMSEA measures how badly your model misfits per degree of freedom. Lower is better: .06 or below is good, and above .10 is poor.
- The most cited cutoffs come from Hu and Bentler (1999): CFI/TLI ≥ .95, RMSEA ≤ .06, and SRMR ≤ .08.
- The chi-square test is almost always significant in large samples, so researchers rely on fit indices instead of chi-square alone.
- APA 7 expects you to report χ², df, p, CFI, TLI, RMSEA with its 90% CI, and usually SRMR.
What are model fit indices in confirmatory factor analysis?
In confirmatory factor analysis, you specify in advance which questionnaire items belong to which latent factor. For example, you might propose that items 1–4 measure anxiety, items 5–8 measure depression, and items 9–12 measure stress. The software then estimates what the correlations between items should look like if your model were true. It compares those predicted correlations to the ones you actually observed.
Model fit indices summarize the size of that gap. A small gap means your theory describes the data well. A large gap means the structure you proposed doesn't hold up. If you're unsure whether you need CFA at all, see our guide on exploratory vs confirmatory factor analysis.
Why not just use the chi-square test?
The chi-square (χ²) test checks whether your model fits the data perfectly. With the 200–500 participants typical of CFA studies, it flags even trivial misfit as significant. This is why most CFA papers report a significant χ² alongside fit indices that still indicate good fit. Report χ², but judge fit using the indices below.
What do CFI and TLI mean?
The Comparative Fit Index (CFI) asks: how much better is my model than a baseline model in which no items are correlated? It ranges from 0 to 1. A CFI of .97 means your model reduces the baseline's misfit by roughly 97%.
The Tucker-Lewis Index (TLI), also called the Non-Normed Fit Index, asks the same question but adds a penalty for complexity. Adding parameters that barely improve fit lowers TLI. It can occasionally exceed 1, which simply means the model fits very well.
Both indices follow the same rule of thumb: ≥ .95 is good, .90–.95 is acceptable but marginal, and below .90 is poor.
What does RMSEA mean?
The Root Mean Square Error of Approximation (RMSEA) estimates how much misfit remains per degree of freedom, adjusted for sample size. Because it measures error, lower is better. Common benchmarks are ≤ .05 close fit, .05–.08 reasonable fit, .08–.10 mediocre fit, and > .10 poor fit.
RMSEA comes with a 90% confidence interval, and you should report it. A narrow interval sitting below .08 is reassuring. A wide interval that crosses .10, which is common with small samples or few degrees of freedom, signals uncertainty.
CFI vs TLI vs RMSEA: key differences
| Index | What it measures | Direction | Good fit | Acceptable | Penalizes complexity? |
|---|---|---|---|---|---|
| CFI | Improvement over baseline model | Higher = better | ≥ .95 | ≥ .90 | No |
| TLI | Improvement over baseline, per df | Higher = better | ≥ .95 | ≥ .90 | Yes |
| RMSEA | Misfit per df, sample-adjusted | Lower = better | ≤ .06 | ≤ .08 | Yes |
| SRMR | Average residual correlation | Lower = better | ≤ .08 | ≤ .10 | No |
Worked example: interpreting real CFA output
Suppose 300 participants complete a 12-item scale, and you test a three-factor model. The output shows:
- Your model: χ² = 98.40, df = 51
- Baseline model: χ² = 2150.00, df = 66
Step 1: Chi-square. χ²(51) = 98.40, p < .001. The result is significant, but that is expected with N = 300, so you move on to the indices.
Step 2: CFI. CFI = 1 − (98.40 − 51) / (2150 − 66) = 1 − 47.40/2084 = .977. This is above .95, so fit is good.
Step 3: TLI. Baseline χ²/df = 32.58, and model χ²/df = 1.93. TLI = (32.58 − 1.93) / (32.58 − 1) = .971. This is also good, which tells you the model isn't overparameterized.
Step 4: RMSEA. RMSEA = √[(98.40 − 51) / (51 × 299)] = √.00311 = .056, 90% CI [.040, .073]. The point estimate meets the .06 cutoff, and the upper bound stays below .08.
Step 5: SRMR. The software reports SRMR = .041, well under .08.
Every index agrees, so you can conclude the three-factor structure fits well. If the indices had disagreed, for example CFI = .96 but RMSEA = .09, you would check modification indices and factor loadings rather than declaring success.
How do I report CFA fit indices in APA 7?
Report the chi-square test first, then the fit indices, and give RMSEA with its confidence interval:
A confirmatory factor analysis indicated that the three-factor model fit the data well, χ²(51, N = 300) = 98.40, p < .001, CFI = .977, TLI = .971, RMSEA = .056, 90% CI [.040, .073], SRMR = .041.
Note that p has no leading zero, and that CFI, TLI, RMSEA, and SRMR are also reported without a leading zero because none of them can exceed 1 in practice. Name your estimator (e.g., maximum likelihood) in the Method section, and cite your cutoff source, typically Hu and Bentler (1999).
Which software can run a CFA?
Base SPSS Statistics does not include CFA. You need IBM's separate AMOS add-on, which adds cost. R's lavaan package is the research standard but requires coding. JASP and jamovi offer free point-and-click CFA modules, though you still interpret and write up the output yourself. Stata handles CFA through its sem command.
StatRyx runs the CFA, flags which indices meet or miss the cutoffs, and drafts the APA 7 results paragraph for you. That last step is where most students lose hours. If your fit is poor, StatRyx also points to the weakest loadings, which gives you a concrete place to start instead of guessing.
Stop calculating this by hand — run it free in StatRyx → Try StatRyx
Frequently Asked Questions
What is a good CFI value in confirmatory factor analysis?
A CFI of .95 or higher indicates good fit, based on Hu and Bentler's (1999) widely cited guidelines. Values between .90 and .95 are often described as acceptable, especially in older literature. Below .90, the model usually needs revision.
What if my CFI is good but my RMSEA is bad?
Disagreement between indices is common, especially with small samples or models with few degrees of freedom, where RMSEA tends to be inflated. Report all indices honestly and check the RMSEA confidence interval. Then inspect modification indices and low loadings to find the source of misfit.
Do I have to meet every cutoff for my model to be acceptable?
No. Cutoffs are guidelines, not strict rules, and their authors warned against applying them mechanically. Most reviewers look for a consistent pattern across indices. Justify your conclusion with your theory and the factor loadings, not one number.
Should I report SRMR as well as CFI, TLI, and RMSEA?
Yes. SRMR is commonly reported alongside these indices, and Hu and Bentler recommended pairing it with CFI or RMSEA. SRMR at or below .08 indicates good fit. Reporting it gives reviewers a measure based on residual correlations as well as the comparative indices.