Written and maintained by CASRAI Editorial Board
Last updated
SPSS does not have a separate menu for principal component analysis (PCA). It lives inside the same dialog as exploratory factor analysis (EFA) — Analyze → Dimension Reduction → Factor — with Principal Components as the default extraction method. That shared dialog is the single biggest source of confusion between the two techniques: PCA is a data-reduction transform of total variance, while EFA models a latent common factor. For the conceptual distinction, see Exploratory Factor Analysis: Extraction, Rotation, and How Many Factors to Retain. This guide covers the SPSS mechanics specific to running PCA: the dialog options, retention rules, and how to save component scores as new variables.
Before extracting: check whether the data is even suitable
PCA reduces a correlation matrix to a smaller number of components, which only works if the variables are correlated enough to begin with. Run this check before extraction, not after:
- Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy — a value below roughly 0.60 signals the correlations are too weak for PCA to meaningfully reduce.
- Bartlett’s test of sphericity — tests whether the correlation matrix differs from an identity matrix (no correlation at all); a non-significant result means PCA has nothing to work with.
Both appear as checkboxes on the Descriptives tab of the Factor dialog (KMO and Bartlett’s Test of Sphericity). Full walkthrough of the output and interpretation thresholds: KMO and Bartlett’s Test.
Running the extraction: the Extraction tab
On the Extraction tab:
- Method: leave it on Principal components (the default). Changing it to Principal Axis Factoring, Maximum Likelihood, or another extraction method turns the analysis into common factor analysis, not PCA — that is the exact switch that produces an EFA solution from the same dialog.
- Analyze: Correlation matrix is the standard choice when variables are on different scales (it standardizes each variable first); Covariance matrix is only appropriate when all variables share the same unit and scale is itself meaningful.
- Display: check Unrotated factor solution and Scree plot.
- Extract: this is where the retention rule is set (next section).
Retention rules: eigenvalues, the scree plot, and their real limitations
SPSS’s Extract options default to Eigenvalues over 1 — the Kaiser criterion, from Kaiser, H. F. (1960), “The application of electronic computers to factor analysis,” Educational and Psychological Measurement, 20(1), 141–151. The rule keeps every component whose eigenvalue exceeds 1, on the reasoning that a component should explain at least as much variance as a single original variable. It is also known to over-extract in practice: Kaiser’s own derivation is a lower bound on the number of components in the population correlation matrix, not a precise stopping rule for a sample.
The alternative built into the same tab, scree plot inspection (Cattell’s elbow method), asks a human to identify the point where the plotted eigenvalues stop dropping sharply and flatten out, keeping only the components before the bend. It is judgment-based and two analysts can reasonably disagree on ambiguous plots.
The methodologically preferred alternative to both — parallel analysis, which compares your eigenvalues against eigenvalues generated from random data of the same size — is not built into SPSS’s Factor dialog at all. O’Connor, B. P. (2000), “SPSS and SAS programs for determining the number of components using parallel analysis and Velicer’s MAP test,” Behavior Research Methods, Instruments, & Computers, 32(3), 396–402, documents SPSS syntax add-ins written specifically to close this gap. That tooling gap, not a methodological argument for the Kaiser rule, is a large part of why eigenvalues-over-1 still dominates published PCA results run in SPSS.
If you already know how many components you want (a fixed a priori structure), the Extract options also let you specify a Fixed number of factors directly instead of relying on either rule.
Reading the output: Total Variance Explained and the Component Matrix
The Total Variance Explained table lists, per component, its eigenvalue, percent of variance, and cumulative percent — both for the full initial solution and for the retained components after extraction. Because PCA analyzes total variance and places 1.00 on the diagonal of the correlation matrix for every variable, the initial communalities are always exactly 1 in a PCA run; that is a defining property of the technique, not a computed estimate the way common-factor communalities are.
The Component Matrix (unrotated) shows each variable’s loading on each retained component — for PCA these loadings are correlations between the observed variable and the component, since a component is defined as a weighted linear combination of the observed variables themselves.
Rotation: the Rotation tab
Unrotated components are mathematically optimal for explaining variance but often difficult to interpret substantively, because early components tend to load on most variables at once. Rotation redistributes the loadings without changing the total variance explained:
- Varimax (orthogonal) keeps components uncorrelated and is the most commonly reported rotation for PCA used as a data-reduction step.
- Direct Oblimin or Promax (oblique) allow the components to correlate with each other, which is more defensible when there’s a substantive reason to expect the underlying dimensions aren’t independent.
With rotation requested, SPSS adds a Rotated Component Matrix (and, for oblique rotations, a separate Pattern and Structure Matrix) alongside the unrotated one.
Saving component scores as new variables
The Scores tab has a Save as variables checkbox that appends one new variable per retained component directly onto the working dataset, using whichever score method is selected: Regression, Bartlett, or Anderson-Rubin. These same three method names are also used for factor scores in common factor analysis run from the identical dialog, which invites a real point of confusion worth being explicit about: for PCA specifically, component scores are exact linear combinations of the observed variables, not estimates of an unobserved latent variable, so they carry no indeterminacy problem. That is a genuine, well-documented difference from factor scores proper — see Factor Scores: Methods and When Not to Use Them for the indeterminacy issue that applies to common-factor scores but not PCA component scores.
PCA vs. EFA in the same SPSS dialog: a one-line reminder
Because both live under Analyze → Dimension Reduction → Factor, the only thing separating a PCA run from an EFA run in SPSS is the Extraction Method dropdown (Principal components vs. Principal Axis Factoring/Maximum Likelihood/etc.) plus how the results get reported: components (no error term, no latent variable) versus factors (a modeled latent construct with unique variance). See Exploratory Factor Analysis for the full model-level distinction, and Confirmatory Factor Analysis (CFA) if the goal is testing a pre-specified factor structure rather than exploring one.
Related SPSS procedures
PCA is frequently run alongside reliability analysis on the same item set — see Cronbach’s Alpha in SPSS. If the reduced components are then used as predictors, Multiple Regression in SPSS covers the next step. For reproducing a PCA run exactly (syntax rather than menus, useful once the dialog choices above are finalized), see SPSS Syntax: Writing Reproducible Analysis Files. New to SPSS generally? Start with What Is SPSS?, or compare it against SPSS vs. R and SPSS vs. Stata if you’re choosing a package.
Frequently asked questions
Is PCA the same as factor analysis in SPSS?
No, even though both run from the same Analyze → Dimension Reduction → Factor dialog. Changing the Extraction Method from Principal components to any of the other options (Principal Axis Factoring, Maximum Likelihood, etc.) switches the analysis to common factor analysis, a different statistical model with a latent variable and unique/error variance that PCA does not have.
What does “Eigenvalues over 1” actually mean, and should I trust it?
It is SPSS’s default component-retention rule (the Kaiser criterion): keep every component whose eigenvalue exceeds 1. It is widely used because it is the SPSS default and requires no judgment call, but it is also well documented to over-extract components relative to more rigorous alternatives like parallel analysis, which SPSS does not offer natively.
How do I save PCA component scores as new variables in SPSS?
On the Scores tab of the Factor dialog, check “Save as variables” and choose a method (Regression, Bartlett, or Anderson-Rubin). SPSS appends one new column per retained component to the working dataset.
Does SPSS have parallel analysis for deciding how many components to keep?
Not built into the menus. It requires separately obtained syntax add-ins, documented in O’Connor (2000); the base Factor dialog only offers the eigenvalue-over-1 rule, a fixed number of components, or the scree plot.
Should I rotate the components in a PCA?
Rotation (Varimax for uncorrelated components, Direct Oblimin or Promax if they may correlate) does not change the total variance explained, only how it’s distributed across components — it’s a standard step when the goal is an interpretable, reportable component structure rather than a single unrotated solution.
