Exploratory Factor Analysis

Exploratory Factor Analysis: Concept Architecture

Orientation and learning roadmap

Exploratory factor analysis investigates the latent structure that may account for correlations among measured variables. This page moves from the common-factor model through data preparation, factor retention, extraction, rotation, interpretation, scoring, validation, and reporting. It emphasizes that EFA is a sequence of evidence-informed judgments rather than a single software command.

EFA asks how many common dimensions the data support

EFA is useful when the number of latent factors and the pattern linking items to factors are not yet adequately specified. It estimates a parsimonious structure in which several observed variables may reflect shared underlying dimensions. The method explores a candidate measurement structure; it does not prove that the factors are real entities or causes.

The common-factor model separates shared and unique variation

For a centered vector of observed variables (\mathbf{x}), the common-factor model expresses scores as a linear combination of latent common factors plus variable-specific residuals. The loading matrix (\boldsymbol{\Lambda}) connects variables to factors, and (\boldsymbol{\varepsilon}) contains unique variance and measurement error. The model is:

[ \mathbf{x}=\boldsymbol{\Lambda}\mathbf{f}+\boldsymbol{\varepsilon}. ]

The reproduced covariance has common and unique parts

If the factor covariance matrix is (\boldsymbol{\Phi}) and the unique variances form diagonal matrix (\boldsymbol{\Psi}), the model-implied covariance matrix combines common-factor covariance and uniqueness. Orthogonal factors have (\boldsymbol{\Phi}=\mathbf{I}), while oblique factors allow off-diagonal correlations. The covariance structure is:

[ \boldsymbol{\Sigma}=\boldsymbol{\Lambda}\boldsymbol{\Phi}\boldsymbol{\Lambda}^{\mathsf T}+\boldsymbol{\Psi}. ]

Communality shows how much variance the factors explain

An item’s communality is the proportion of its variance represented by the retained common factors when variables are standardized. Under orthogonal factors, it is the sum of squared loadings across those factors; uniqueness is one minus communality. With correlated factors, the factor covariance matrix must be included rather than blindly summing squared pattern coefficients.

[ h_i^2=\boldsymbol{\lambda}_i^{\mathsf T}\boldsymbol{\Phi}\boldsymbol{\lambda}_i, \qquad \psi_i=1-h_i^2. ]

Worked communality example

Suppose a standardized item loads 0.80 on one orthogonal factor and 0.30 on another. Its communality is (0.80^2+0.30^2=0.73), so the retained factors explain 73% of its variance and its estimated uniqueness is 0.27. This calculation does not show which factor is substantively correct or whether the two-factor solution fits the full data.

EFA is not principal component analysis

Principal component analysis transforms observed variables into components that reproduce total variance, whereas EFA models common variance and includes unique residual variance. PCA can be useful for data compression, but its components do not answer exactly the same measurement question as common factors. Calling a rotated PCA an EFA obscures this difference.

EFA and confirmatory factor analysis have different roles

EFA estimates the loading pattern with relatively few prior restrictions, while confirmatory factor analysis tests a prespecified structure through explicit constraints. A structure discovered and evaluated in the same data is vulnerable to chance features of that sample. Confirmation should therefore use new data, a defensible holdout sample, or appropriately cautious resampling evidence.

Content validity comes before statistical structure

A clean loading pattern cannot compensate for items that omit important aspects of the intended construct or use language the target population interprets differently. Qualitative development, patient or stakeholder input, cognitive interviewing, and expert review establish what the items are intended to cover. EFA contributes evidence about internal structure, not the entire validity argument.

The analysis population and item set define the result

Factor solutions can change across clinical groups, languages, cultures, settings, time points, and response modes. The sample should match the population in which the instrument will be interpreted, and item exclusions should not be made only to manufacture a simple structure. Later comparisons across groups require measurement-invariance analysis rather than an assumption that one EFA generalizes everywhere.

Correlation choice must match the response scale

Pearson correlations may be reasonable for approximately continuous variables with suitable relationships and distributions. Ordered categorical items often call for polychoric correlations and an estimator designed for ordinal data, while binary items may use tetrachoric correlations under their assumptions. Sparse response categories, extreme thresholds, and small samples can make these matrices unstable.

Missing data need an explicit strategy

Complete-case analysis can discard information and bias the correlation matrix when missingness is related to observed or unobserved values. Pairwise correlations can be based on different subsets and may produce a matrix that is not positive definite. Full-information, multiple-imputation, or other model-compatible approaches should follow a stated missing-data assumption.

Factorability is assessed with several pieces of evidence

The correlation matrix should contain enough common association to support a factor model without being dominated by duplicates or near-linear dependencies. Item correlations, partial correlations, anti-image information, the Kaiser–Meyer–Olkin statistic, Bartlett’s test, determinant behavior, and sample adequacy can all contribute. No single cutoff establishes that EFA is appropriate.

Sample size has no universal item-to-person rule

Required sample size depends on communalities, number of factors, factors per item, loading strength, overdetermination, response distributions, estimator, missingness, and model complexity. Fixed rules such as a single minimum total or respondent-to-item ratio can be misleading. Simulation or evidence tailored to the proposed structure is stronger when the analysis is consequential.

Extraction estimates the common-factor solution

Principal-axis factoring, minimum residual methods, maximum likelihood, and weighted least-squares approaches use different objectives and assumptions. Maximum likelihood can support likelihood-based inference under distributional assumptions, while robust or residual-based methods may be preferable when those assumptions are not credible. Extraction choice should be reported and sensitivity checked when it affects the solution.

The number of factors should not come from one rule

Retaining too few factors can merge distinct constructs, while retaining too many can create unstable or uninterpretable dimensions. Parallel analysis, scree inspection, minimum average partial criteria, fit or residual evidence, theoretical coherence, and factor replicability can contribute to the decision. The eigenvalue-greater-than-one rule alone is often too crude.

Parallel analysis supplies a noise benchmark

Parallel analysis compares observed eigenvalues with eigenvalues generated from data containing no common-factor signal under a chosen reference procedure. Factors are retained when the observed evidence exceeds the appropriate simulated benchmark. Results depend on whether the procedure is configured for principal components or common-factor analysis and whether it respects ordinal data.

Rotation improves interpretability without changing dimensionality

An unrotated solution is often difficult to interpret because factor axes are not uniquely determined. Rotation reorients the factor solution to seek a simpler loading pattern while preserving the chosen number of factors and the reproduced common covariance within the rotation family. Rotation does not discover a new dataset or guarantee a substantively correct structure.

Oblique rotation is usually realistic for health constructs

Oblique rotations such as oblimin or promax allow factors to correlate, while orthogonal rotations such as varimax constrain correlations to zero. Symptoms, functioning, wellbeing, access, and experience dimensions are often plausibly related, making an oblique solution a defensible starting point. Orthogonality should be a substantive assumption, not a convenience for easier tables.

Pattern and structure matrices answer different questions

With oblique rotation, the pattern matrix contains regression-like coefficients for each factor controlling for the others, while the structure matrix contains item–factor correlations. These matrices differ when factors correlate. Interpretation and reporting should label them correctly rather than mixing coefficients from both.

Loadings require magnitude, sign, and content together

A factor loading describes the association between an item and a factor under the selected model and rotation. Large loadings can help identify a factor, but factor naming requires the shared meaning of the item set, not the highest coefficient alone. The sign of an entire factor can reverse without changing model fit, so direction must be anchored to item wording.

Cross-loadings may reveal overlap or a poor specification

An item with meaningful loadings on more than one factor may measure overlapping constructs, contain ambiguous wording, or reflect a genuinely complex structure. Automatically deleting every cross-loading item can narrow content validity and produce an artificially tidy measure. Decisions should consider loading size, conceptual role, residuals, stability, and consequences for scoring.

Heywood cases signal a problem to investigate

A negative uniqueness or a standardized communality above one is an improper solution often called a Heywood case. It may arise from sampling fluctuation, too many factors, model misspecification, extreme collinearity, weak identification, or estimation problems. Constraining the estimate without investigating the cause can conceal an invalid solution.

Factor scores are estimates, not observed traits

Regression, Bartlett, and other scoring methods can assign estimated factor scores to individuals, but different methods can produce different values because latent scores are indeterminate. Reliability, shrinkage, scaling, and the intended downstream use should be considered before scores enter regression, economic models, or clinical classification. Simple summed scores require separate justification and may not equal model-based factor scores.

Reliability is not established by factor analysis alone

EFA may support a dimensional structure needed for scale scoring, but it does not by itself establish score reliability, responsiveness, measurement error, criterion validity, or interpretability. Internal-consistency estimates should be calculated for defensible unidimensional scores and with assumptions suited to the measure. High reliability also does not prove that the intended construct is measured.

Health-economic use often begins with measurement instruments

EFA can help develop or refine patient-reported outcome, experience, preference, capability, and implementation measures used alongside economic evaluation. A factor solution may guide subscale formation or identify redundant items, but deleting content solely to maximize statistical neatness can weaken decision relevance. If factor scores feed utility mapping or economic models, their uncertainty and transportability should be propagated where material.

Excel is suitable for auditing, not full EFA estimation

Excel can inspect missingness, descriptive statistics, correlation matrices, loading tables, communalities, and simple residual calculations. For an orthogonal loading row in B2:D2, communality can be audited with =SUMSQ(B2:D2) and uniqueness with =1-SUMSQ(B2:D2). Extraction, polychoric correlations, rotation, fit assessment, and uncertainty should use validated statistical software.

A disciplined EFA workflow

EFA requires linked decisions about items, data, correlations, factor number, extraction, rotation, and interpretation. Each decision should be made before later results are used to justify it retroactively. A practical workflow is:

1. Define the construct. State the intended content, population, setting, and use of the scores.
2. Prepare the items. Check coding, reverse-keyed items, category frequencies, missingness, and descriptive distributions.
3. Choose the correlation matrix. Match Pearson, polychoric, tetrachoric, or mixed correlations to the data and assumptions.
4. Assess factorability. Examine common association, partial correlations, adequacy, and matrix conditioning.
5. Compare factor counts. Use parallel analysis, scree evidence, fit, residuals, stability, and theory.
6. Select extraction. Match the estimator to data properties and inferential goals.
7. Rotate transparently. Prefer oblique rotation when correlated constructs are plausible.
8. Interpret the full solution. Review pattern, structure, factor correlations, communalities, cross-loadings, and residuals.
9. Test robustness. Vary defensible choices and examine bootstrap, split-sample, or replication stability.
10. Confirm separately. Evaluate the proposed structure in independent data with an appropriate confirmatory model.

Common analytical errors

EFA is vulnerable to apparently reasonable defaults that answer a different question from the one intended. Transparent decisions and sensitivity analysis are more informative than one supposedly universal threshold. Common failures include:

• Calling principal component analysis exploratory factor analysis.
• Selecting factors only because their eigenvalues exceed one.
• Forcing orthogonal rotation when factors are plausibly correlated.
• Using Pearson correlations for sparse ordinal items without assessment.
• Treating a significant Bartlett test as proof of a good factor model.
• Applying a fixed sample-size rule without considering communalities and structure.
• Naming factors from one high-loading item while ignoring item content.
• Deleting cross-loading items without evaluating content validity.
• Treating factor scores as error-free observed variables.
• Claiming confirmation from the same sample used to discover the structure.

What a complete EFA report should disclose

A reproducible report identifies the construct, population, items, sample, missing-data handling, correlation matrix, factorability evidence, retention criteria, extraction method, rotation, software, convergence, improper solutions, loadings, factor correlations, communalities, residuals, and sensitivity analyses. It should explain how factors were named and how items or scores will be used, including decisions that were not based on statistical thresholds alone. Separate confirmation, reliability, validity, and invariance evidence should be reported without implying that EFA supplies all of them.