Exploratory Factor Analysis (EFA) Explained: How to Run, Read and Report It
Exploratory factor analysis (EFA) answers one question: can the correlations among many observed variables be explained by a smaller number of unobserved factors? If you have built a questionnaire and want to know how many constructs it actually measures, which items belong together, and which items misbehave, EFA is usually the first analysis to run. This guide explains the method, walks through a fully worked example, shows how to read each part of the output, and gives an APA 7 reporting template.
What is exploratory factor analysis?
EFA is a family of latent-variable methods that models each observed variable as a weighted combination of a few common factors plus a unique part. It is "exploratory" because you do not specify in advance which item loads on which factor; the data suggest the structure. When you do have a specified structure to test, you need confirmatory factor analysis (CFA) instead.
When should you use EFA?
- Developing or shortening a scale and checking its dimensionality.
- Investigating the structure of a translated or adapted instrument in a new population.
- Reducing correlated indicators to interpretable constructs before regression or group comparison.
- Screening items prior to a CFA on an independent sample (never confirm on the same data you explored).
EFA versus PCA. They are often confused. Principal component analysis (PCA) creates weighted composites that summarise total variance. EFA models only the shared variance and treats the rest as unique variance plus error. If your aim is to represent latent constructs, EFA is conceptually appropriate (Fabrigar et al., 1999). If your aim is only data compression, see the PCA calculator.
The mathematical foundation
For p standardised items and m common factors, the common factor model is:
xi = λi1F1 + λi2F2 + … + λimFm + εi
Here λij is the loading of item i on factor j, Fj is a common factor, and εi is the item's unique part. In matrix form the model-implied correlation matrix is:
R = ΛΦΛ′ + Ψ
Λ holds the loadings, Φ the factor correlations (an identity matrix under orthogonal rotation) and Ψ the diagonal matrix of unique variances. Two quantities follow directly. With orthogonal factors, the communality of item i is hi² = Σj λij², and its uniqueness is ψi = 1 − hi². Communality is the proportion of an item's variance explained by the retained factors.
Assumptions and data requirements
- Measurement level. Continuous or approximately continuous items. Likert items with five or more categories and reasonable distributions are commonly analysed with Pearson correlations; short or highly skewed ordinal items are better served by polychoric correlations.
- Linear relationships among items and no extreme multivariate outliers.
- Factorability. Items must correlate enough to share variance. Check the correlation matrix, the Kaiser–Meyer–Olkin (KMO) measure (Kaiser, 1974) and Bartlett's test of sphericity (Bartlett, 1954).
- No perfect multicollinearity. A determinant near zero or correlations above about .90 signal redundant items.
- Adequate sample size. Rules such as "10 cases per item" are weak guides. MacCallum et al. (1999) showed that adequacy depends on communalities and factor overdetermination: with communalities around .60 or higher and several items per factor, modest samples can recover structure well; with low communalities you need far more.
Bartlett's test statistic is:
χ² = −[(N − 1) − (2p + 5)/6] · ln|R|, df = p(p − 1)/2
A significant result only shows that R is not an identity matrix; with large N it is almost always significant, so rely on KMO and the correlation pattern as well. Overall KMO values above roughly .80 are usually described as good, and values below .60 as poor.
Original research example (illustrative)
Background. A university lecturer develops a six-item scale of student engagement in an online course. Items 1–3 were written to capture behavioural engagement (logging in, completing tasks, attending live sessions); items 4–6 to capture emotional engagement (interest, enjoyment, sense of belonging).
Objective and question. Does the scale show a two-factor structure, and do items load on the intended factors?
Data. N = 200 students, each item rated 1–5. The values below are invented for teaching.
Step 1: Check factorability
Illustrative results: KMO = .81; Bartlett's χ²(15) = 412.60, p < .001. With six items, df = 6 × 5 / 2 = 15. Working backward, 412.60 corresponds to ln|R| ≈ −2.103, i.e. |R| ≈ .122 (since (200 − 1) − 17/6 ≈ 196.17). The determinant is well above zero, so multicollinearity is not a concern.
Step 2: Decide how many factors
| Factor | Eigenvalue | Parallel-analysis threshold (95th percentile) | Retain? |
|---|---|---|---|
| 1 | 2.61 | 1.31 | Yes |
| 2 | 1.54 | 1.19 | Yes |
| 3 | 0.62 | 1.09 | No |
| 4 | 0.49 | 1.00 | No |
| 5 | 0.43 | 0.90 | No |
| 6 | 0.31 | 0.75 | No |
The eigenvalues sum to 6.00, the number of items, as they must for a correlation matrix. The Kaiser rule (eigenvalue > 1; Kaiser, 1960) and parallel analysis (Horn, 1965) both support two factors here. In real data they often disagree. Kaiser's rule tends to over-retain factors, so treat parallel analysis, Velicer's MAP (Velicer, 1976) and the scree plot as the more defensible guides, and always check interpretability.
Step 3: Extract and rotate
Principal axis factoring (or maximum likelihood) with two factors, followed by rotation. Because engagement dimensions are likely to be related, an oblique rotation (oblimin or promax) would be the natural first choice in practice. To keep the arithmetic transparent, the table below shows an orthogonal (varimax) solution, where communalities are simple row sums of squared loadings.
| Item | Factor 1 | Factor 2 | h² | Uniqueness |
|---|---|---|---|---|
| 1. Completes weekly tasks | .78 | .12 | .62 | .38 |
| 2. Logs in regularly | .81 | .08 | .66 | .34 |
| 3. Attends live sessions | .74 | .19 | .58 | .42 |
| 4. Finds content interesting | .10 | .83 | .70 | .30 |
| 5. Enjoys the course | .15 | .77 | .62 | .38 |
| 6. Feels part of the class | .21 | .72 | .56 | .44 |
| Sum of squared loadings | 1.89 | 1.86 | 3.75 | |
| % of total variance | 31.5 | 30.9 | 62.4 |
Check the arithmetic. Item 1: .78² + .12² = .6084 + .0144 = .6228 ≈ .62. Factor 1 sum of squares: .6084 + .6561 + .5476 + .0100 + .0225 + .0441 = 1.8887 ≈ 1.89. The total 3.75 divided by 6 items = 62.4%. The extracted variance (3.75) is smaller than the sum of the first two eigenvalues (4.15) because common-factor extraction analyses only shared variance.
Step 4: Interpret
Each item loads strongly (≥ .72) on its intended factor and weakly (≤ .21) on the other, so there are no cross-loadings. Factor 1 is interpretable as behavioural engagement and Factor 2 as emotional engagement. All communalities exceed .55, which supports the adequacy of a sample of 200. Conclusion: the illustrative data support the intended two-dimensional structure. This is evidence of structure, not proof of validity; a CFA on new data and reliability estimates should follow.
How to run EFA with the StatFAS EFA Calculator
Open statfas.com/exploratory-factor-analysis-online.html. According to the page, you can run the analysis without an account, and the tool is described as checking assumptions, reporting effect sizes and providing APA7-ready results to download.
The workflow below follows the general logic of any EFA; the specific controls must be confirmed against the live interface before publication.
- Prepare your data. One row per participant, one column per item, a header row, numeric codes only. Reverse-score negatively worded items first.
- Load the data and select the items. Choose only the variables that belong in the analysis; do not include IDs or demographic columns.
- Set analysis options. Typical choices are extraction method, number-of-factors rule, rotation and loading display cut-off.
- Run and review. Check the factorability diagnostics first, then the factor-retention evidence, then the loadings.
- Download results. The page states that APA7-ready results can be downloaded. Verify every number and sentence before it enters a manuscript.
statfas-efa-calculator-interface.pngAlt text: "StatFAS exploratory factor analysis calculator interface before data entry."
Caption: The EFA Calculator on StatFAS.com.
statfas-efa-example-dataset.png · Alt: "Six-item engagement dataset loaded in the StatFAS EFA calculator."Screenshot 3:
statfas-efa-analysis-settings.png · Alt: "Extraction, rotation and factor-number settings for EFA in StatFAS."Screenshot 4:
statfas-efa-loadings-table.png · Alt: "Rotated factor loadings and communalities table from StatFAS EFA."Screenshot 5:
statfas-efa-scree-plot-or-report.png · Alt: "Scree plot or APA7 report download for EFA." Use only genuine captures of the tool.Reading the output
- KMO and Bartlett's test. They tell you whether the matrix is worth factoring. Mistake: treating a significant Bartlett test as evidence of a good factor structure.
- Eigenvalues, scree plot, parallel analysis. They guide the number of factors. Mistake: retaining factors solely by "eigenvalue > 1".
- Loadings. Standardised weights linking item and factor. In an oblique solution, report the pattern matrix and the factor correlation matrix (and, if useful, the structure matrix). A common working convention is |λ| ≥ .30 to .40 for meaningful loadings, but this is a convention and not a statistical test.
- Communalities. Low values (e.g. below .30) indicate items that share little variance with the others; consider removing them if theory allows.
- Variance explained. Descriptive only. A high percentage does not prove that the factors are meaningful.
- Factor correlations. If an oblique rotation yields correlations above about .70 to .80, the factors may not be empirically distinct.
Remember that EFA output has no p-value for individual loadings in the usual implementation, and the "effect size" analogue is the loading magnitude itself. Report what the tool provides rather than imposing inferential language on descriptive quantities.
APA 7 reporting example (based on the illustrative data)
An exploratory factor analysis was conducted on six engagement items (N = 200) using principal axis factoring with varimax rotation. The Kaiser–Meyer–Olkin measure indicated adequate sampling (KMO = .81), and Bartlett's test of sphericity was significant, χ²(15) = 412.60, p < .001. Parallel analysis supported retaining two factors, which together accounted for 62.4% of the variance. Items 1–3 loaded on a behavioural engagement factor (loadings = .74–.81), and items 4–6 loaded on an emotional engagement factor (loadings = .72–.83). No item cross-loaded above .21. Communalities ranged from .56 to .70. Table 1 presents the rotated loadings.
Essential elements: sample size, extraction method, rotation (with justification), factorability statistics, the factor-retention criterion, loadings in a table, variance explained and how you handled problem items. Report the retention rule you actually used, whatever the software printed by default.
Common mistakes and troubleshooting
- Using PCA but calling it EFA. Match the method to the aim and state which you used.
- Defaulting to varimax. If the factors are plausibly correlated, forcing them to be uncorrelated distorts the solution. Run oblique first and inspect the factor correlations.
- Keeping factors with Kaiser's rule alone. Triangulate with parallel analysis and interpretability.
- Heywood cases (communality ≥ 1). Usually caused by too many factors, too few items per factor or a small sample. Reduce factors or collect more data.
- Factors with only one or two items. They are unstable. Aim for at least three strongly loading items per factor.
- Repeated deletion of items until the output "looks clean". Delete one item at a time, with a stated rationale, and re-run.
- Exploring and confirming on the same sample. Split the sample or collect new data for CFA.
- Missing data. Listwise deletion can bias results and shrink N. Examine missingness patterns before running the analysis.
Limitations and alternatives
EFA solutions depend on the extraction method, rotation and retention rule, and they can vary across samples. Results cannot establish that a construct exists; they describe correlation structure. Alternatives include CFA within SEM when you hold a specified model, PCA for pure data reduction, and item response theory or network psychometrics for other measurement questions. Before an EFA, inspect item distributions with descriptive statistics and the item correlation matrix.
Frequently asked questions
What sample size do I need for EFA?
There is no fixed rule. Higher communalities and more items per factor reduce the sample needed (MacCallum et al., 1999). Treat fixed ratios such as 10:1 as rough heuristics only.
Should I use PCA or EFA?
Use EFA to model latent constructs, and PCA to summarise variance in composites.
Which rotation should I choose?
Choose oblique rotation (oblimin or promax) when factors may correlate, which is typical in behavioural research. Use varimax only with justification.
Can I run EFA on Likert items?
Often yes, especially with five or more response options and mild skew. For few categories or strong skew, polychoric correlations are preferable.
Conclusion
Good EFA is less about pressing "run" and more about a chain of defensible decisions: checking factorability, choosing the number of factors with several criteria, selecting an appropriate extraction and rotation, and interpreting loadings against theory. When you are ready to apply it to your data, the StatFAS EFA Calculator provides a browser-based route, and its downloadable results should be checked by you before publication.
References
- Bartlett, M. S. (1954). A note on the multiplying factors for various χ² approximations. Journal of the Royal Statistical Society: Series B, 16(2), 296–298.
- Fabrigar, L. R., Wegener, D. T., MacCallum, R. C., & Strahan, E. J. (1999). Evaluating the use of exploratory factor analysis in psychological research. Psychological Methods, 4(3), 272–299.
- Horn, J. L. (1965). A rationale and test for the number of factors in factor analysis. Psychometrika, 30(2), 179–185.
- Kaiser, H. F. (1960). The application of electronic computers to factor analysis. Educational and Psychological Measurement, 20(1), 141–151.
- Kaiser, H. F. (1974). An index of factorial simplicity. Psychometrika, 39(1), 31–36.
- MacCallum, R. C., Widaman, K. F., Zhang, S., & Hong, S. (1999). Sample size in factor analysis. Psychological Methods, 4(1), 84–99.
- Velicer, W. F. (1976). Determining the number of components from the matrix of partial correlations. Psychometrika, 41(3), 321–327.