CODEWITHSUNDEEP
r
Fanta
Fanta

Reputation: 3089

How can I compute the principal component scores with the same result as returned by prcomp() in $x?

When I try to compute the principal components (PC) scores (i.e., the dataset projected in the PC space), I obtain a different result from what prcomp() reports in the $x attribute of its result. I center and scale the data.

I run prcomp() as

pca.result = prcomp(USArrests, scale=TRUE)

Then I can access the the wanted PC scores as

pca.result$x  # 1

From the doc., I understand the above centers and scales the data before multiplying them by the rotation matrix (a.k.a. variable loadings, a.k.a. eigenvectors). I am trying to reproduce the result above by projecting (rotating) the dataset myself, as in:

((as.matrix(USArrests)-pca.result$center)/pca.result$scale) %*% pca.result$rotation  # 2

But the result I get from # 2 is different from the result I get from # 1.

How should I correct # 2 so that it gives the same result as # 1?

Upvotes: 0

Views: 514

Answers (1)

dcarlson
dcarlson

Reputation: 11056

This is a bit simpler. Just use the score() function to compute the standardized data:

pcscores <- scale(USArrests) %*% pca.result$rotation

If you want to use the statistics in pca.result:

means <- pca.result$center
sdevs <- pca.result$scale
center <- sweep(USArrests, 2, means)
zscores <- sweep(center, 2, sdevs, "/")
pcscores2 <- as.matrix(zscores) %*% pca.result$rotation

Your method subtracts means down each column instead of across columns.

Upvotes: 1

Related Questions