Personality and Individual Differences
Factor Rotation Playground
Twelve markers in a plane. Turn the axes and every loading changes while not one point moves.
Simulated — fictional adjective markers with illustrative coordinates
Learning objective
By the end you should be able to say exactly what rotation changes and what it leaves untouched, and defend simple structure without claiming it finds the true solution.
About 20 minutes. Nothing you do here is saved or sent anywhere.
How to use this
- Say what you think rotation does to model fit.
- Rotate the first axis and watch the loading table.
- Watch the invariants underneath it — total variance and every communality — refuse to move.
- Let the axes come off 90° and see the factors start to correlate.
- Try the marker set with a genuine cross-loading.
First, a prediction
A researcher extracts two factors, then rotates them.
The playground
Key terms
- Loading
- How strongly one marker relates to one factor.
- Communality
- The share of a marker's variance accounted for by the factors together.
- Simple structure
- A solution in which each marker loads clearly on one factor and near zero on the others.
- Orthogonal and oblique
- Whether the axes are held at 90°, keeping the factors uncorrelated, or allowed to tilt so that the factors correlate.
- Cross-loading
- A marker relating substantially to both factors at once.
Two factors, twelve markers
Fictional adjective markers with illustrative coordinates. The positions came out of the extraction and never change; only the axes move.
What has not changed
Tentative factor labels
- Factor 1 is anchored by
- Factor 2 is anchored by
Labels are an interpretive act, not a result. These are the highest-loading markers, nothing more.
Loadings
| Marker | F1 | F2 | Communality | Salient on |
|---|
Challenge — what actually changed?
Rotate until you have found the simplest structure you can, then answer.
What this demonstrates
Rotation changes the description, not the data
Every loading changes; the configuration of points does not. The rotated model reproduces the correlations exactly as well as the unrotated one, each variable's communality is untouched, and the total variance explained is identical. Rotation is a change of coordinate system, and coordinate systems are not discoveries.
Which is why it is neither a trick nor a revelation
Two opposite overreactions are common. The first is that rotation manufactures results — it cannot, because fit is invariant. The second is that the rotated solution is the true structure — it is not, because the mathematics does not privilege it. Simple structure is a criterion people adopted, for good reasons: it yields solutions that are interpretable and comparable between studies. Those are real virtues, and they are not the same as truth.
Orthogonal or oblique is a substantive choice
Forcing the axes to 90° asserts that the constructs are uncorrelated. For most psychological constructs that is implausible, and the third marker set shows what it costs. When the clusters genuinely sit closer to 36° apart than to 90°, no orthogonal rotation puts an axis through both, and the solution looks worse than it is. Allowing the axes to move fixes the picture and produces correlated factors — which is a claim about the world, to be defended rather than avoided.
Some items will not settle, and that is information
In the second marker set, one pair of markers sits between the clusters and no rotation makes it simple. That is what a genuine cross-loading is: not a nuisance to be rotated away but a fact about a variable that belongs partly to both factors. Deleting such items to improve simple structure is common and quietly changes what the factors mean.
For teaching elsewhere: take this activity as one self-contained block of HTML, on the clipboard or as a file. Either way it is styled so that it will not disturb the page you put it into.