Imagine you have developed a 24-item questionnaire to measure employees' attitudes toward digital transformation. You believe the items capture several different aspects, perhaps perceived usefulness, anxiety, organisational support and skills. But how do you know whether the items really group into these dimensions? Maybe some items do not belong where you expected, or maybe there are fewer or more dimensions than you thought.
Exploratory factor analysis (EFA) is a statistical technique designed to answer exactly this kind of question. It is widely used in questionnaire development, scale adaptation and data reduction. This post explains what EFA is, when you need it, the key decisions involved, how to interpret the output, and how to report it in your thesis.
What is exploratory factor analysis?
EFA is a multivariate statistical method that identifies the underlying structure among a set of observed variables, usually questionnaire items. It assumes that correlations among items can be explained by a smaller number of unobserved (latent) variables called factors.
For example, if items about “feeling nervous using new software” and “worrying about making mistakes with technology” are highly correlated with each other, but less correlated with items about “management provides training”, EFA may identify them as belonging to a common factor, which you might interpret as “technology anxiety”.
EFA is called exploratory because it does not require you to specify in advance which items belong to which factor. The analysis suggests a structure based on the data, which you then interpret using theory.
When do you need EFA?
EFA is useful in several situations:
1. Developing a new scale
When you create new items, EFA helps you discover how many dimensions they measure and which items work well. It is a standard step in scale development.
2. Using a scale in a new context
If you adapt or translate an existing scale, or use it in a very different population, the original structure may not hold. EFA can explore the structure in your data, although if a clear structure is already well established, confirmatory factor analysis (CFA) is often more appropriate.
3. When theory about the structure is weak
If previous research does not agree on the number of dimensions, or if no validation evidence exists, EFA provides an initial exploration.
4. Data reduction
EFA can reduce a large number of items into a smaller set of factors for further analysis, such as regression.
When EFA may not be needed
If you use a well-established scale with a strong, replicated factor structure in similar populations, CFA is usually more suitable, because it tests whether your data fit the expected structure. Some examiners question the use of EFA on well-established scales unless there is a clear reason.
EFA vs. principal component analysis (PCA)
SPSS places both EFA and PCA under Analyze → Dimension Reduction → Factor, and the default extraction method is principal components. This causes confusion.
- PCA is a data reduction technique that creates components as weighted combinations of the observed variables, accounting for total variance. It does not assume underlying latent constructs.
- EFA (using extraction methods such as principal axis factoring or maximum likelihood) models only the shared (common) variance among items and assumes latent factors cause the item responses.
Many methodologists recommend EFA methods over PCA when the goal is to identify latent constructs, such as in scale development. If you use PCA, justify it and describe it accurately as PCA, not factor analysis.
Is your data suitable for EFA?
Sample size
There is no single rule. Older guidelines suggest ratios such as 5 or 10 participants per item, or absolute minimums such as 200 or 300. Research on EFA has shown that the required sample size depends strongly on data characteristics: when communalities are high and each factor has several strongly loading items, smaller samples may be adequate; when communalities are low, larger samples are needed. Aim for as large a sample as feasible and cite guidance from your field.
Correlations among items
EFA needs items that are correlated. Check the correlation matrix: many correlations should be above about .30. Very high correlations (such as above .90) may indicate redundancy or multicollinearity.
Kaiser–Meyer–Olkin (KMO) measure
KMO assesses sampling adequacy, with values from 0 to 1. Kaiser suggested descriptive labels: values in the .90s “marvellous”, .80s “meritorious”, .70s “middling”, .60s “mediocre”, .50s “miserable”, and below .50 “unacceptable”. A value of at least .60 is commonly treated as a minimum.
Bartlett's test of sphericity
This tests whether the correlation matrix differs significantly from an identity matrix (a matrix with no correlations). A significant result (p < .05) indicates that there are correlations to analyse. With large samples, it is almost always significant, so it is a minimal check.
Key decisions in EFA
1. Extraction method
Common choices are principal axis factoring (robust when data are not multivariate normal) and maximum likelihood (allows some fit statistics but assumes multivariate normality).
2. Number of factors to retain
This is one of the most important decisions. Methods include:
- Kaiser's criterion (eigenvalues greater than 1): the SPSS default. It is widely criticised for often retaining too many factors and should not be used alone.
- Scree plot: look for the point where the curve flattens (the “elbow”). Useful but subjective.
- Parallel analysis: compares your eigenvalues with those from random data. It is widely recommended as one of the more accurate methods. SPSS does not provide it through the menus, but syntax and free tools are available.
- Theory and interpretability: the final solution should make conceptual sense.
Use several criteria together and report them.
3. Rotation
Rotation makes factors easier to interpret.
- Orthogonal rotation (such as Varimax) assumes factors are uncorrelated.
- Oblique rotation (such as Promax or Direct Oblimin) allows factors to correlate.
In social and behavioural research, factors are usually expected to correlate, so oblique rotation is often recommended. If factors turn out to be nearly uncorrelated, oblique and orthogonal solutions will be similar.
Interpreting the output
Factor loadings
A loading shows how strongly an item relates to a factor. With oblique rotation, SPSS gives a pattern matrix (unique contribution of each factor to the item, usually interpreted) and a structure matrix (simple correlations between items and factors).
Common guidelines treat loadings of about .40 or higher as meaningful, although thresholds of .30 to .50 are used depending on sample size and field. Cite the guideline you use.
Cross-loadings
An item that loads substantially on two or more factors is a cross-loading item. It may be ambiguous and a candidate for revision or removal, especially if the difference between loadings is small.
Communalities
The communality is the proportion of an item's variance explained by all the factors. Very low communalities suggest the item does not fit well with the others.
Variance explained
Report the percentage of variance explained by the factors. There is no universal minimum, but very low values may suggest the factors capture limited information.
Refining the solution
EFA is often iterative. You may remove items with low loadings, high cross-loadings or low communalities, then re-run the analysis. However:
- Remove items one at a time and re-examine the solution.
- Consider the content of each item: removing items can narrow the meaning of a construct.
- Keep a record of every decision and the reason.
- Ensure each factor retains enough items, often at least three, to be measured reliably.
Naming factors
Examine the items that load on each factor and identify the common theme. Name the factor to reflect this meaning, drawing on theory. Avoid names that claim more than the items measure.
Running EFA in SPSS: the basic steps
- Go to Analyze → Dimension Reduction → Factor and move your items into the Variables box.
- Click Descriptives and tick KMO and Bartlett's test of sphericity. You may also request the correlation matrix.
- Click Extraction. Change the method from Principal components to Principal axis factoring (or Maximum likelihood). Request the Scree plot. If you have already decided the number of factors using parallel analysis, choose “Fixed number of factors”.
- Click Rotation and choose Promax or Direct Oblimin for an oblique rotation.
- Click Options and tick Sorted by size and Suppress small coefficients (for example, below .30) to make the loading table easier to read. Keep the unsuppressed output for your records.
- Click Paste to save the syntax, then run it.
In the output, look first at KMO and Bartlett's test, then at the eigenvalues and scree plot, then at the pattern matrix, communalities and factor correlation matrix.
A short illustration
Suppose your 24 items on digital transformation produce a KMO of .87 and a significant Bartlett's test. Parallel analysis suggests four factors, the scree plot shows an elbow after the fourth factor, and eigenvalues greater than 1 suggest five. You examine both four- and five-factor solutions. In the five-factor solution, the fifth factor contains only two items with weak loadings and no clear meaning, so you choose four factors. Two items cross-load strongly on two factors; after checking their wording, you remove them one at a time. The final solution has 22 items on four interpretable factors: perceived usefulness, technology anxiety, organisational support and digital skills. Factor correlations range from .21 to .48, supporting the use of oblique rotation. (This example is illustrative.)
Notice how the decision combines several statistical criteria with interpretability, and how each change is documented. This is exactly what examiners want to see.
Reporting EFA in your thesis
- Justification for using EFA.
- Sample size and data suitability (KMO, Bartlett's test).
- Extraction method and rotation, with justification.
- Criteria for the number of factors (eigenvalues, scree plot, parallel analysis).
- Loading thresholds and item removal decisions.
- A table of the final pattern matrix with loadings, communalities, eigenvalues and variance explained.
- Factor correlations (for oblique rotation).
- Reliability of each factor.
Common mistakes to avoid
- Using PCA but calling it factor analysis.
- Relying only on eigenvalues greater than 1.
- Using Varimax without considering whether factors are correlated.
- Deleting many items purely to improve statistics.
- Running EFA and CFA on the same sample and presenting CFA as independent confirmation.
- Not reporting the decisions that shaped the final solution.
Final thoughts
Exploratory factor analysis helps you discover how questionnaire items group together and whether they measure the dimensions you intended. It involves several judgement calls, including extraction method, number of factors, rotation and item retention, so transparency is essential. Combine statistical criteria with theory, document each decision, and, where possible, confirm your structure with CFA in an independent sample. Done well, EFA gives you strong evidence that your measures are meaningful.
Are you running EFA for your questionnaire? Share your questions about the output in the comments.