Exploratory Factor Analysis Step by Step: KMO, Bartlett and Loadings Explained

Exploratory factor analysis (EFA) finds the small number of hidden themes, called factors, that explain why groups of questionnaire items are answered in similar ways. You check your data are suitable with the KMO measure and Bartlett's test, then read the factor loadings to see which items belong together. Most people reach EFA with a 20-item survey and a supervisor asking "what does your scale actually measure?" This guide walks through each step, from suitability checks to an APA 7 write-up.

Key Takeaways

  • Exploratory factor analysis groups correlated items into underlying factors. It is used to develop or check questionnaires and scales.
  • The Kaiser–Meyer–Olkin (KMO) measure should be at least .60 before running EFA. Values of .80 or higher are considered meritorious.
  • Bartlett's test of sphericity must be significant (p < .05). A significant result shows your items are correlated enough to factor.
  • A common rule treats loadings of .40 or higher as meaningful. Items that load on two factors (cross-loadings) are candidates for removal.
  • StatRyx runs KMO, Bartlett, extraction, rotation and loadings in one pass and writes the APA 7 results paragraph for you.

What is exploratory factor analysis, in plain terms?

Exploratory factor analysis looks for patterns in how people answer your items. Suppose respondents who agree with "I worry about deadlines" also tend to agree with "I feel tense at work." Both items are probably tapping one underlying trait, such as work anxiety. EFA finds these clusters statistically, without you telling it which items belong together.

That last point is what makes it exploratory. If you already have a theory about which items form which factor, confirmatory factor analysis (CFA) is the better tool. EFA is the right choice when you are building a new scale, adapting one to a new population, or have no firm structure in mind.

How do I know my data are suitable for factor analysis?

EFA needs items that are correlated, but not perfectly. Two tests check this before you extract anything. You also need enough people. A widely used guideline is at least 5–10 participants per item, with around 200 or more for a stable solution.

What does the KMO test tell you?

The Kaiser–Meyer–Olkin measure of sampling adequacy estimates how much of the variation among items is shared, meaning how much of it could come from common factors. It ranges from 0 to 1. Kaiser's widely cited benchmarks are:

  • .90 and above: marvelous
  • .80s: meritorious
  • .70s: middling
  • .60s: mediocre
  • below .60: unsuitable for EFA

Also check the KMO value for each individual item, shown on the diagonal of the anti-image correlation matrix. Items below .50 are usually dropped.

What does Bartlett's test of sphericity tell you?

Bartlett's test asks whether your correlation matrix differs from an identity matrix, which is a matrix where every item is completely unrelated to every other item. A significant result (p < .05) means the items share enough correlation to be worth factoring. With large samples, Bartlett is almost always significant, so treat it as a minimum hurdle rather than strong evidence.

KMO measure Bartlett's test of sphericity
Question it answers Is there enough shared variance? Are the items correlated at all?
Output Value from 0 to 1 χ², df, p
Pass criterion ≥ .60 (ideally ≥ .80) p < .05
Sensitivity to sample size Low High (almost always significant when N > 200)
Item-level diagnostic? Yes (anti-image diagonal) No

Exploratory factor analysis step by step: extraction and rotation

Step 1: Choose an extraction method. Principal axis factoring and maximum likelihood are true factor analysis methods. Principal component analysis (PCA) is technically a different technique, although many software packages present it alongside EFA. For identifying latent constructs, prefer principal axis factoring.

Step 2: Decide how many factors to keep. The "eigenvalue greater than 1" rule is the default in SPSS, but it tends to overestimate the number of factors. Parallel analysis is more accurate. It keeps factors whose eigenvalues exceed those produced by random data of the same size. Check the result against the scree plot and against theory.

Step 3: Rotate. Rotation makes the factor structure easier to interpret. Use an oblique rotation such as oblimin or promax when factors may correlate, which is almost always the case with psychological constructs. Use varimax only if you have good reason to expect the factors to be unrelated.

How do I interpret factor loadings?

A factor loading is the correlation between an item and a factor. A loading of .70 means the item is strongly driven by that factor. Most researchers use .40 as the threshold for a meaningful loading. Some use .32, following Tabachnick and Fidell's guideline.

Look for three problems:

  • Cross-loadings: an item loads above .30 on two or more factors.
  • Weak items: an item loads below .40 on every factor.
  • Single-item factors: a factor has fewer than three items loading on it.

Remove problem items one at a time, rerun the analysis, and recheck. Every removal changes the whole solution. Once you settle on a final structure, test each factor's internal consistency. Our guide on Cronbach's alpha explains that next step.

Worked example: a 12-item stress questionnaire

A researcher collects responses from 210 university students on a 12-item stress questionnaire. That gives 17.5 participants per item, which is comfortable.

Suitability. KMO = .84, which falls in the meritorious range. All individual item KMOs exceed .70. Bartlett's test gives χ²(66) = 1124.37, p < .001. The 66 degrees of freedom come from 12 × 11 / 2. The data are suitable for factoring.

Extraction. Principal axis factoring with parallel analysis supports three factors, which together explain 54.6% of the variance.

Rotation and loadings. An oblimin rotation shows factor correlations from r = .31 to .42, which confirms that an oblique rotation was the right choice. Item 7 cross-loads at .38 and .35, so it is removed. After rerunning, the remaining 11 items load between .48 and .81 on their intended factors. These factors are labelled academic pressure, financial strain and social isolation.

APA 7 write-up:

The Kaiser–Meyer–Olkin measure indicated good sampling adequacy, KMO = .84, and Bartlett's test of sphericity was significant, χ²(66) = 1124.37, p < .001. Principal axis factoring with oblimin rotation was conducted, and parallel analysis supported a three-factor solution explaining 54.6% of the variance. Item 7 was removed due to cross-loading. In the final 11-item solution, loadings ranged from .48 to .81, and factor correlations ranged from .31 to .42.

Report the full pattern matrix in a table, with loadings below .30 suppressed for readability.

How can I run EFA without wrestling with SPSS menus?

In SPSS, EFA means working through the Factor dialog's Descriptives, Extraction, Rotation and Options sub-menus. SPSS also offers no built-in parallel analysis. R (via the psych package), jamovi and JASP all handle EFA well. R requires code, while jamovi and JASP are free desktop installs.

StatRyx is an AI-powered statistical analysis tool that replaces manual SPSS workflows with automated, APA 7-formatted reporting. Upload your item data and StatRyx checks KMO and Bartlett, runs parallel analysis, applies an appropriate rotation, and flags cross-loading items. It then writes a results paragraph like the one in the worked example.

Stop calculating this by hand — run it free in StatRyx → Try StatRyx

Frequently Asked Questions

What is a good KMO value for factor analysis?

A KMO of .60 is the usual minimum, and .80 or above is considered meritorious. If your overall KMO is low, check the individual item values and remove items below .50. Then rerun the test.

What if Bartlett's test is not significant?

A non-significant Bartlett's test means your items are not correlated enough to form factors. EFA is not appropriate in that case. Check whether the items were meant to measure related constructs, and whether your sample is very small.

What is the minimum factor loading to keep an item?

The most common cutoff is .40, although .32 is also widely accepted. Items that load below the cutoff on every factor, or above .30 on two factors, are typically removed one at a time.

Should I use varimax or oblimin rotation?

Use oblimin or promax when your factors are likely to correlate, which is typical in psychology and the social sciences. If the oblique solution shows factor correlations below about .30, varimax gives a similar result and is acceptable.

What is the difference between EFA and PCA?

EFA models only the shared variance among items to identify latent constructs. PCA summarises total variance, including error variance. For scale development, EFA is generally preferred, but both often produce similar loading patterns in practice.

Stop calculating this by hand. Upload your dataset and StatRyx's AI runs the correct test and returns copy-paste-ready APA 7 output in seconds — no SPSS license, no syntax.

Run your data through StatRyx free →
← All posts