Exploratory Factor Analysis (EFA) in SPSS: How to Run and Interpret It

Exploratory factor analysis, usually shortened to EFA, answers a question that comes up in almost every questionnaire-based thesis: do all these survey items really measure a smaller number of underlying ideas? If you have twenty questions about study motivation, for example, EFA helps you find out whether they group into two or three factors, which items belong together, and which items do not belong anywhere. This guide walks you through the whole process in SPSS, from checking your data to reporting the result in APA style.

What exploratory factor analysis does, and when to use it

Factor analysis is a data-reduction technique. It looks at the correlations among your measured variables and tries to explain them with fewer unobserved factors. In a questionnaire context, each factor is a latent construct, such as anxiety, engagement, or satisfaction, that you cannot measure directly but infer from several items.

Use EFA when you are exploring structure rather than testing a structure you already fixed in advance. That distinction matters. If you already have a firm theory that says exactly which items load on which factors, and you want to test that model, you need confirmatory factor analysis instead. If you are developing a scale, cleaning up a questionnaire, or trying to understand what dimensions are hiding in your items, EFA is the right first step.

EFA is also useful before other analyses. Once you have a defensible factor structure, you can compute factor scores or composite scores and use them in correlations, t-tests, or regression, instead of running dozens of separate tests on individual items.

Before you start: the checks that save you later

EFA is only as good as the data you feed it. Work through this checklist before you open the factor analysis dialog.

Sample size and cases per item

There is no single magic number, but a widely used rule of thumb is at least five to ten participants per item, with larger samples giving more stable solutions. A 20-item questionnaire therefore calls for roughly 100 to 200 cases at minimum, and more is better. Very small samples tend to produce factors that do not replicate.

Level of measurement and correlations

Your items should be measured on an interval or Likert-type scale, and they should actually correlate with each other. If items barely correlate, there is no shared structure for factors to capture. SPSS will test this formally with the Kaiser-Meyer-Olkin measure and Bartlett’s test, described below, but a quick scan of the correlation matrix first is good practice.

Missing data and outliers

Decide how to handle missing responses before you run the analysis, and check for extreme cases that could distort correlations. Keep your decisions consistent with the rest of your thesis, and describe them in your methods section.

Factorability

In plain terms, factorability means your data are suitable for factoring at all. The Kaiser-Meyer-Olkin (KMO) measure and Bartlett’s test of sphericity, which SPSS reports by default when you ask for them, are the formal checks. A KMO value close to 1 indicates that patterns of correlations are compact enough for factor analysis to yield distinct, reliable factors. Values below about 0.50 are generally considered unacceptable, values in the 0.60s are mediocre, and values of 0.80 and above are good. Bartlett’s test should be significant, which tells you the correlation matrix is significantly different from an identity matrix, in other words, that your variables are related enough to factor.

How to run EFA in SPSS, click by click

Assume your questionnaire items are in separate columns, one item per variable, with one row per respondent.

Step 1. Go to Analyze, then Dimension Reduction, then Factor.

Step 2. In the Factor Analysis dialog, move all the items you want to factor into the Variables box. Do not include demographic variables or total scores, only the items whose structure you are exploring.

Step 3. Click Descriptives. Tick Initial solution, KMO and Bartlett’s test of sphericity, and Anti-image if you want the individual measures of sampling adequacy. Click Continue.

Step 4. Click Extraction. Under Method, choose Principal components if you want a straightforward data-reduction solution, which is the most common starting point for questionnaire work. Under Extract, keep Eigenvalues greater than 1 to start with, and also ask for a scree plot. Click Continue.

Step 5. Click Rotation. Choose Varimax for an orthogonal rotation, which keeps factors uncorrelated and usually gives the cleanest, easiest-to-read loadings for a first analysis. If you have good reason to expect your factors to correlate, for example subscales of the same broader construct, choose Direct Oblimin instead and report that choice. Click Continue.

Step 6. Click Scores if you want SPSS to save factor scores as new variables, for example using the Regression method. This step is optional and only needed if you plan to use the factors in later analyses.

Step 7. Click Options. Under Missing Values, choose Exclude cases pairwise or listwise, consistent with your earlier decision. Under Coefficient Display Format, tick Sorted by size and Suppress small coefficients, and set the suppression value to about 0.30 so that tiny loadings do not clutter the rotated matrix. Click Continue, then OK.

SPSS will now produce the factor analysis output. The next section explains how to read it in order.

Paste-ready SPSS syntax

If you prefer syntax, or you need a reproducible record for your thesis appendix, the following commands reproduce the dialog steps above. Replace Item1 to Item20 with your own variable names.

FACTOR
 /VARIABLES Item1 Item2 Item3 Item4 Item5 Item6 Item7 Item8 Item9 Item10 Item11 Item12 Item13 Item14 Item15 Item16 Item17 Item18 Item19 Item20
 /MISSING PAIRWISE
 /ANALYSIS Item1 Item2 Item3 Item4 Item5 Item6 Item7 Item8 Item9 Item10 Item11 Item12 Item13 Item14 Item15 Item16 Item17 Item18 Item19 Item20
 /PRINT INITIAL KMO EXTRACTION ROTATION
 /PLOT EIGEN
 /CRITERIA MINEIGEN(1) ITERATE(25)
 /EXTRACTION PC
 /CRITERIA ITERATE(25)
 /ROTATION VARIMAX
 /METHOD=CORRELATION.

Run this from a syntax window with Run, then All, and SPSS will produce the same output as the dialog version.

How to interpret the SPSS output, table by table

Read the output in this order. Skipping ahead to the rotated matrix before checking factorability is the most common beginner mistake.

KMO and Bartlett’s test

Find the KMO and Bartlett’s Test table first. Check the KMO value against the guide above: below 0.50, stop and reconsider, because your data may not factor well. Check Bartlett’s test: a significant result supports proceeding. If KMO is low, look at the anti-image matrix diagonals, the individual measures of sampling adequacy. Items with very low values are candidates for removal, one at a time, re-running the analysis after each removal.

Communalities

The Communalities table shows, for each item, how much of its variance is explained by the extracted factors. The Extraction column is the one to read. Low communalities, commonly below about 0.30, mean an item shares little with the other items. Such items are candidates for removal, but do not delete items on this number alone. Check their loadings first.

Total Variance Explained

This table shows how much variance each factor accounts for, and it is where the eigenvalues-greater-than-one rule shows up. Factors with eigenvalues above 1 are retained by default. Also look at the cumulative percentage of variance explained by the retained factors. There is no universal pass mark, but for questionnaire research a solution that explains roughly half or more of the total variance is often considered reasonable, provided the factors are interpretable.

Scree plot

The scree plot graphs eigenvalues against factor number. Look for the elbow, the point where the steep drop levels off into a flatter tail. Factors before the elbow are usually worth keeping; factors after it often add little. When the eigenvalue rule and the scree plot disagree, which happens often, favour interpretability: a solution with an extra factor that makes no conceptual sense is rarely an improvement, and a solution that merges two clearly different ideas into one factor is rarely defensible either.

Rotated Component Matrix

This is the table you will spend the most time with, and the one your readers will care about. Each row is an item, each column is a factor, and each cell is a factor loading, essentially the correlation between that item and that factor.

Read it column by column and look for a simple structure: each item should load strongly on one factor and weakly on the others. As a working guide, loadings of about 0.40 and above are often treated as meaningful, and loadings of 0.70 and above are strong. For example, an item that loads 0.75 on Factor 1 and 0.12 on Factor 2 clearly belongs to Factor 1. An item that loads around 0.45 on two factors is cross-loading, and you need to decide whether to keep it, rewrite it, or drop it, and to justify that decision in your write-up.

Once the pattern is clear, name each factor from the items that load on it. A factor made up of items about worry, tension, and restlessness might be named Anxiety. A factor made up of items about enjoyment, interest, and persistence might be named Engagement. Good factor names describe the shared content of their items, not just the first item in the list.

If the solution looks messy

Do not panic, and do not keep re-running the analysis until you get a tidy picture by chance. Instead, work methodically. Remove one problematic item at a time, starting with the lowest communality or the worst cross-loader, and re-run. Check whether a different number of factors, forced in the Extraction dialog, gives a more interpretable solution. Check whether an oblique rotation fits better if your factors are plausibly correlated. And always ask whether the problem is statistical or conceptual: sometimes items that refuse to group together simply do not measure the same construct, and that is a finding, not a failure.

Reporting in APA style (7th ed.)

Report EFA transparently enough that a reader could reproduce it. Name the extraction method, the rotation, the criteria for retaining factors, and the sample size. Then report the factorability checks, the number of factors retained, the variance explained, and the loading pattern, usually in a table.

Here is a template paragraph you can adapt. Fill in your own numbers, and do not copy the illustrative values as if they were your results:

“An exploratory factor analysis was conducted on the 20 questionnaire items with principal components extraction and varimax rotation. The Kaiser-Meyer-Olkin measure verified sampling adequacy, and Bartlett’s test of sphericity indicated that correlations between items were sufficiently large for factor analysis. Three factors had eigenvalues greater than 1 and were retained, together explaining a substantial proportion of the total variance. The scree plot supported the three-factor solution. Items loading below the chosen threshold and items with substantial cross-loadings were examined and removed one at a time, and the final solution showed a clear simple structure. Factor 1, labelled [name], comprised [number] items; Factor 2, labelled [name], comprised [number] items; and Factor 3, labelled [name], comprised [number] items.”

When you report numbers, give the KMO value, the Bartlett chi-square with degrees of freedom and p value, the eigenvalues or variance explained for each retained factor, and the range of loadings for each factor. Put the full rotated matrix in a table so readers can judge your interpretation for themselves.

Key Takeaways

  • EFA explores structure; it does not test a pre-specified model. For that, use confirmatory factor analysis.
  • Check factorability first: KMO and Bartlett’s test tell you whether factoring your data is defensible.
  • Use the eigenvalue rule and the scree plot together, and let interpretability break the tie.
  • Read the rotated matrix for simple structure: strong loadings on one factor, weak loadings elsewhere.
  • Remove problematic items one at a time and justify each removal; do not delete a batch of items at once.
  • Report extraction, rotation, retention criteria, KMO, Bartlett’s test, variance explained, and the full loading matrix.

Frequently asked questions

How is EFA different from principal component analysis?

In SPSS, principal components is offered as an extraction method inside the factor analysis dialog, which causes confusion. Strictly, principal component analysis is a data-reduction method that uses all variance, while factor analysis models shared variance only. In practice, many questionnaire studies use principal components extraction as a first step and still call the procedure exploratory factor analysis. What matters most is that you state your extraction method clearly so readers know what you did.

How many factors should I keep?

Start with eigenvalues greater than 1 and the scree plot, then check interpretability. If a factor has only one or two items loading on it, or its items make no conceptual sense together, it is probably not a real factor. If two factors look like halves of the same idea, consider a solution with one fewer factor.

What should I do with cross-loading items?

First, check whether the item is worded ambiguously or double-barrelled. If it genuinely taps two constructs, you may need to rewrite it for future data collection, or drop it from the current analysis and say so. Keeping a cross-loading item without comment weakens your factor definitions.

Can I run EFA on a small sample?

You can run it, but the solution may be unstable, meaning a different sample could give a different factor pattern. With a small sample, be extra cautious about over-interpreting small differences in loadings, report your sample size prominently, and consider collecting more data before finalising a scale.

Do I need to report every item I dropped?

Yes. Readers need to know which items were removed, at which stage, and why. A scale that looks clean because problematic items were silently deleted is misleading. Report the original item count, the final item count, and the reason for each removal.

Need this done on your data? Get a free quote