How to Run PCA on Mac & Windows

PCA boils many correlated variables down to a few that capture most of the variation. Here's how to run it on Mac & Windows and read what comes back.

When you have a dozen numeric variables that overlap, principal component analysis finds a handful of new axes — the principal components — that capture as much of the variation as possible. It's the workhorse for exploring and compressing wide data.

Run it

Open your dataset and choose PCA from the analysis menu.
Select the numeric variables to include (PCA standardizes them for you).
Read the report: how much variance each component explains.

A PCA scree plot in Stratum — The scree plot: variance explained by each component.

How many components to keep

Look at the scree plot — the variance each component explains. Keep components up to the “elbow” where the curve flattens, or enough to reach a cumulative variance you're comfortable with (70–90% is common). The first two or three often capture the bulk of wide datasets.

Read the biplot

The biplot draws your observations as points and your original variables as arrows on the first two components at once. Points that cluster are similar; arrows pointing the same way are correlated variables; an arrow's length shows how much that variable contributes.

A PCA biplot in Stratum — The biplot: observations as points, variables as arrows.

Use the components downstream. Stratum can add the component scores back as new columns, so you can feed two or three components into a regression or clustering instead of a dozen correlated originals.

To interpret what the components mean, see Understanding PCA Loadings & Scores, or the full PCA lesson.

Download Stratum

Frequently asked questions

What does PCA do?

It reduces many correlated numeric variables to a few uncorrelated “principal components” that capture most of the variation — useful for visualizing, compressing, and de-correlating wide data.

How many principal components should I keep?

Keep components up to the “elbow” of the scree plot, or enough to reach a cumulative variance you're comfortable with (often 70–90%). The first two or three usually dominate.

Do I need to standardize my data before PCA?

Usually yes — otherwise variables with large units dominate. Stratum standardizes the inputs for you so each variable contributes on equal footing.

What's the difference between a scree plot and a biplot?

A scree plot shows how much variance each component explains (to decide how many to keep). A biplot shows observations and original variables together on the first two components (to interpret them).