Exploratory Factor Analysis: Discovering Measurement Structure

Exploratory factor analysis is a statistical method that examines the correlations among a set of observed variables to discover how many underlying latent factors they represent and which items belong to each. You use it when you do not yet have a theory about your measurement structure and you want the data to suggest one. It is the discovery stage of scale development, the step that comes before you ever try to confirm a structure.

The contrast with confirmatory factor analysis defines both methods. Confirmatory analysis tests a structure you specified in advance. Exploratory factor analysis imposes nothing: it lets the pattern of correlations reveal how the items cluster, how many dimensions exist, and where items load weakly or on more than one factor. When you have a brand-new pool of questionnaire items and no prior evidence about their structure, exploration comes first.

When to use exploratory factor analysis

Reach for exploratory factor analysis when you are developing a new scale, when you are adapting an existing instrument so heavily that its original structure may no longer hold, or when you have a large set of items and suspect they reduce to a smaller number of underlying dimensions. It answers questions such as how many factors my items represent, which items hang together, and which items are redundant or do not belong.

It is the wrong tool once you already have a hypothesized structure. If theory or a previous study tells you the factors and their items, you should be confirming that structure, not rediscovering it. Running exploration when you already know the answer wastes the opportunity to test it formally.

The key decisions

Several choices shape the result, and reviewers scrutinize each one.

The first is how many factors to retain. The old habit of keeping every factor with an eigenvalue above one is now considered unreliable. Better approaches include parallel analysis, which compares your eigenvalues to those from random data, and the scree plot, read alongside interpretability and theory. Retaining too many factors splinters the structure; retaining too few merges distinct constructs.

The second is the rotation. Rotation makes the loadings easier to interpret. An oblique rotation, which allows factors to correlate, is usually more realistic for psychological and social constructs than an orthogonal rotation that forces them to be independent, because real constructs are rarely uncorrelated.

The third is the extraction method, such as principal axis factoring or maximum likelihood, chosen partly on whether your data meet distributional assumptions. These decisions interact, and documenting them transparently is part of a defensible analysis our survey data analysis team handles routinely.

Reading the output

The heart of the output is the factor loading matrix, showing how strongly each item relates to each retained factor. You look for items that load cleanly, with a strong loading on one factor of roughly 0.40 or higher and weak loadings elsewhere. An item that cross-loads, loading moderately on two factors, is ambiguous and may need to be revised or dropped. An item that loads weakly everywhere is not measuring any of the factors and is a candidate for removal.

Before trusting any of this, check that the data are suitable for factoring at all. The Kaiser-Meyer-Olkin measure of sampling adequacy should be acceptable, generally above 0.60, and Bartlett's test of sphericity should be significant, indicating the items are correlated enough to factor. Skipping these checks risks extracting structure from data that has none.

Sample size

Exploratory factor analysis needs a reasonable number of cases per item for stable results, and stronger, cleaner factor structures require fewer cases than weak, muddy ones. Fixed ratios are rough guidance only. The reliable path is to plan the sample with the same care you would for any analysis, using the reasoning in our power analysis and sample size guide and our sample size calculator to sanity-check feasibility before collecting data.

What comes next

Exploratory analysis suggests a structure; it does not confirm one. The standard workflow is to explore on one sample and then confirm the resulting structure on an independent sample with a confirmatory measurement model, and ultimately to build that validated measurement model into a structural equation model if your study tests relationships among the constructs. Treating an exploratory result as final, without confirming it on fresh data, is a common shortcut that weakens a scale-development paper.

Common mistakes

1. Using the eigenvalue-greater-than-one rule alone. It over-extracts. Use parallel analysis and a scree plot together.
2. Forcing orthogonal rotation on correlated constructs. Allow factors to correlate when theory says they should.
3. Keeping ambiguous items. Cross-loading and weakly loading items muddy the structure and should be revised or removed.
4. Skipping the suitability checks. Run the sampling adequacy measure and the sphericity test before interpreting factors.
5. Treating exploration as confirmation. A discovered structure must be confirmed on an independent sample.

The common factor model, and why it is not principal components

The most consequential decision is made before any rotation: factor analysis and principal components analysis are different models, and the software offers both under confusingly similar menus. The common factor model partitions each item's variance into a part shared with the other items (the communality) and a part unique to that item (measurement error plus item-specific variance), and it models only the shared part:

x_i = l_i1*F1 + l_i2*F2 + ... + l_im*Fm + e_i

Here the latent factors F explain the correlations among items, the loadings l are the regression weights of items on factors, and e is the unique factor. Principal components analysis makes no such split: components are weighted sums of the items that reproduce total variance, error included, so they are summaries, not latent causes. For scale development you almost always want the common factor model (principal axis factoring or maximum likelihood extraction), because the question is what unobserved construct generates the responses, not how to compress the items. Reporting "factor analysis" while having run principal components is one of the quiet errors a measurement reviewer catches immediately.

Deciding the number of factors without the eigenvalue rule

Parallel analysis is the right default, but it has its own choices. Horn's parallel analysis compares each observed eigenvalue to the distribution of eigenvalues from many random datasets of the same size; retain a factor only while the observed eigenvalue exceeds the 95th percentile (not merely the mean) of the random eigenvalues, which is the more conservative and reproducible threshold. Run it on the factor model, not on components, so the random baseline matches what you are estimating. Pair it with Velicer's minimum average partial test, which stops adding factors when the average squared partial correlation among items starts to rise, and treat strong disagreement between the two as a signal to examine the borderline factor for interpretability rather than to trust a single rule. When you extract by maximum likelihood you additionally get a formal model-fit chi-square and approximate fit indices, which give an independent check on whether a given number of factors reproduces the correlation matrix.

Likert items need polychoric correlations

Most questionnaires use ordered categories (strongly disagree to strongly agree), yet the default factor analysis runs on Pearson correlations, which assume continuous, normally distributed variables. With five or fewer categories, or visibly skewed items, Pearson correlations are attenuated and can manufacture spurious "difficulty factors" that merely group items by how extreme their response distributions are. The correct treatment is to factor the matrix of polychoric correlations, which estimate the association between the continuous traits assumed to underlie each pair of ordinal items, and to extract with a robust estimator such as weighted least squares or unweighted least squares. Treating ordinal data as interval is defensible only when items have many categories and roughly symmetric distributions.

Factor scores are indeterminate, and the rotation output has two matrices

Two technical points routinely trip people up when they move from extraction to use. First, under an oblique rotation the software prints both a pattern matrix and a structure matrix: the pattern matrix holds the unique loading of each item on each factor controlling for the other factors (interpret the structure from this one), while the structure matrix holds item-factor correlations inflated by the factor intercorrelations. Always report the factor correlation matrix alongside, because correlations near or above 0.7 suggest the factors are not really distinct. Second, factor scores are indeterminate: because the common factor model has more unknowns than equations, infinitely many sets of scores are consistent with the same loadings, so a factor score is an estimate, not the latent value. If you must carry scores forward, the Bartlett or Anderson-Rubin methods have better properties than simple regression scores, but the cleaner path is to model the latent variable directly in a structural equation model rather than scoring and regressing in two error-prone steps.

A worked exploration in R

library(psych)

# Suitability: sampling adequacy and sphericity before anything else
KMO(items); cortest.bartlett(items)

# How many factors: parallel analysis on the FACTOR model, plus Velicer's MAP
fa.parallel(items, fm = 'ml', fa = 'fa', n.iter = 500)   # 95th-percentile rule
vss(items, n = 6)                                         # MAP and very-simple-structure

# Extraction by maximum likelihood with an oblique (oblimin) rotation
f <- fa(items, nfactors = 3, fm = 'ml', rotate = 'oblimin')
print(f$loadings, cutoff = 0.30)   # pattern matrix
f$Phi                              # factor correlation matrix

# Ordinal items: factor the polychoric correlations instead
fa(items, nfactors = 3, fm = 'wls', rotate = 'oblimin', cor = 'poly')

Bringing it together

Exploratory factor analysis is the disciplined discovery of measurement structure: let the data suggest how many factors exist and which items belong, document every decision, and then confirm the result on fresh data. It is the first step in building a scale that will eventually survive peer review.

If you are developing or refining a questionnaire and want the factor structure done rigorously, our survey data analysis team runs the analysis and documents every choice. Request a quote and tell us about your item pool.