As we discussed previously, the point behind rotation is to maximize the loading of the variables on a specific component, which helps in simplifying the interpretation by reducing/eliminating the correlation among these components. The method to conduct orthogonal rotation is known as varimax. There are other non-orthogonal rotation methods that allow correlation across factors/components. The choice of the rotation methodology that you will use in your profession should be based on the pertinent literature, which exceeds the scope of this chapter. Feel free to experiment with this dataset. I think that when in doubt, the starting point for any PCA should be an orthogonal rotation.

For this process, we will simply turn back to the principal() function, slightly changing the syntax to account for five components and orthogonal rotation, as follows:

> pca_rotate <- psych::principal(train_scale, nfactors = 5, rotate = "varimax")

Given the number of features, I just normally save this into a CSV file and examine it in a spreadsheet, in particular with a subject matter expert. Here, we save it and I'll come back with what are high-level summaries. When I worked in oncology market research, we always ended up with a component around the drug's efficacy, one around the drug's side effect profile, and then maybe one or two components regarding dosing, cost, or something of that ilk. The code here just removes the crazy loading class from the object so we can save it as a dataframe:

> pca_loading <- unclass(pca_rotate$loading)

> pca_loading <- data.frame(pca_loading)

> pca_loading$features <- row.names(pca_loading)

> readr::write_csv(pca_loading, "pca_loading.csv")

Welcome back! There is no correct answer, but my guess as to how to summarize these components would be something like this:

• PC1: A catchall component; 44 features have loading higher than 0.5
• PC2: Hips, thighs, and buttocks...with a dash of waist and chest
• PC3: Neck, shoulders, arms
• PC4: Some height measures
• PC5: Oddly enough, head and foot measures

This can be a fun exercise naming the components. I fondly recall the days of naming such components compassionate conservatives, pragmatic practitioners, and so on. Be that as it may, we need to create scores from these components so we can give supervised learning a go.