Factor Rotation

Factor Rotation: Concept Architecture

Orientation and learning roadmap

The page starts with an extracted factor solution and shows why analysts reorient its axes to interpret the measured variables. It then compares orthogonal and oblique choices, reads an illustrative health questionnaire pattern, and explains which mathematical properties are preserved. The last sections cover reporting, measurement validity, and common interpretation errors.

Locate rotation in the factor-analysis workflow

Exploratory factor analysis (EFA) represents shared variation among observed variables with a smaller set of latent factors. Extraction produces an initial loading matrix, but its axes may distribute meaningful associations across several columns and make the substantive dimensions hard to name. Rotation reorients those axes under a stated criterion so that the pattern can be easier to interpret; it does not itself choose the number of factors or improve the underlying data.

In a common-factor model for centred observations,

$$\mathbf{x}=\boldsymbol\Lambda\mathbf{f}+\boldsymbol\varepsilon, \qquad \operatorname{Cov}(\mathbf{x}) =\boldsymbol\Lambda\boldsymbol\Phi\boldsymbol\Lambda^{\mathsf T} +\boldsymbol\Psi,$$

where $\boldsymbol\Lambda$ is the factor pattern, $\boldsymbol\Phi$ the factor correlation matrix for standardised factors, and $\boldsymbol\Psi$ the unique-variance matrix. The precise estimation method and model assumptions need separate evaluation before interpreting rotated loadings. Principal components analysis can also rotate component axes, but its components are not automatically latent common factors.

Choose whether factors may correlate

An orthogonal rotation keeps the factor axes at right angles and imposes zero correlations between distinct factors. Varimax is a familiar orthogonal criterion that seeks a clearer loading pattern, often by encouraging relatively high and low loadings within each factor. That restriction can be inappropriate when the underlying constructs, such as physical and emotional health, plausibly covary.

An oblique rotation allows nonzero factor correlations; common methods include promax and oblimin. It returns a pattern matrix, a factor-correlation matrix, and a structure matrix of observed-variable-to-factor correlations. The pattern and structure entries can differ substantially, so the analyst should report which matrix is being interpreted.

The choice should follow the measurement question and empirical evidence, with sensitivity to alternative plausible rotations. An oblique solution does not require the factors to be strongly correlated; it permits the data to reveal their association. A clean-looking table alone is not a sufficient reason to impose independence.

Read a health questionnaire example

Suppose an EFA explores four standardised questionnaire items intended to represent physical and emotional functioning. The following oblique pattern loadings are invented for instruction and are not estimates from a validated instrument. A larger absolute loading indicates a stronger conditional association with a factor in this model, while small cross-loadings can make interpretation more straightforward.

If the two factors correlate at 0.35, the structure correlation for climbing stairs with the emotional factor is $0.10+0.80(0.35)=0.38$, even though its emotional pattern loading is only 0.10. For standardised oblique factors, the corresponding relation is $\mathbf{S}=\boldsymbol\Lambda\boldsymbol\Phi$. This illustrates why reading the structure correlation as a second direct pattern loading would overstate the item's conditional association with that factor.

Understand what a valid rotation preserves

With an orthogonal transformation $\mathbf{T}$ satisfying $\mathbf{T}^{\mathsf T}\mathbf{T}=\mathbf{I}$, the transformed loadings can be written $\boldsymbol\Lambda^{*}=\boldsymbol\Lambda\mathbf{T}$. The common-factor covariance reproduced by the retained factor space is unchanged because $\boldsymbol\Lambda^{*}(\boldsymbol\Lambda^{*})^{\mathsf T}=\boldsymbol\Lambda\boldsymbol\Lambda^{\mathsf T}$ when factor covariance is the identity. Individual loadings and the allocation of variance across rotated factors can change even while this total common covariance is preserved.

For a nonsingular oblique transformation, the pattern and factor covariance must be transformed together: $\boldsymbol\Lambda^{*}=\boldsymbol\Lambda\mathbf{T}$ and $\boldsymbol\Phi^{*}=\mathbf{T}^{-1}\boldsymbol\Phi\mathbf{T}^{-\mathsf T}$. Their product $\boldsymbol\Lambda^{*}\boldsymbol\Phi^{*}(\boldsymbol\Lambda^{*})^{\mathsf T}$ then reproduces the same common covariance for an equivalent solution. Different fit statistics or estimated communalities in software output may reflect other changes in estimation, extraction, or factor count rather than rotation alone.

Connect interpretation to health economic measurement

Factor rotation can help investigate the dimensional structure of patient-reported outcome measures, experience surveys, and other multi-item instruments used in health research. A plausible physical or emotional factor can guide further instrument evaluation and domain scoring, but it does not by itself establish content validity, cross-population invariance, responsiveness, or a preference-based utility tariff. Health economic models should use the measure and scoring system appropriate to their decision question and evidence base.

The sign and ordering of factors are arbitrary: reversing a factor's sign reverses its loadings without changing the modelled construct. Loadings need interpretation alongside item wording, cross-loadings, factor correlations, sample size, missing-data treatment, and the selected number of factors. A rotation that yields a convenient label in one sample may fail to reproduce in another setting.

Report the analysis so readers can judge it

Reporting should make it possible to distinguish a defensible measurement result from a visually tidy table. State the extraction method, factor-number decision, input correlation type, rotation criterion, and treatment of missing or ordinal responses. For an oblique solution, include factor correlations and identify whether the displayed loadings are pattern or structure values.

• Explain why the selected rotation allows or constrains factor correlations for these constructs.
• Report material cross-loadings and unexpected item behaviour instead of hiding them behind a display cutoff.
• Distinguish exploratory factor analysis from rotated principal components when describing latent constructs.
• Check stability across samples or meaningful subgroups where the instrument will be used.
• Assess the broader measurement properties before using factor scores or domains as outcomes in an economic evaluation.

Sources and technical basis

The software documentation describes orthogonal and oblique criteria and the distinct matrices returned for oblique solutions. The COSMIN methods guidance places structural validity within the broader evaluation of patient-reported measures. The questionnaire loadings and factor correlation on this page are illustrative values constructed to show the pattern–structure calculation.

• SAS, PROC FACTOR documentation.
• UCLA Statistical Methods and Data Analytics, Factor Analysis annotated output.
• UCLA Statistical Methods and Data Analytics, Exploratory Factor Analysis annotated output.
• COSMIN, Guideline for systematic reviews of patient-reported outcome measures, version 2.0.