I haven't run this by my stats colleagues, but my intuitive feeling is the same as yours. I don't think the subset of PCA loadings is the same as the PCA loadings of a subset.

I picture the PCA procedure as drawing a vector through a cloud of data points, in the direction of maximal variation. It seems to me that the variation of the entire dataset must differ from that of any subsets so at the very least, you'll get different values by running PCA on a subset.

I guess it would be easy enough to confirm this using R; just generate a small matrix/data frame with some toy data, run prcomp() on it, subset the matrix of rotations and plot PC1 v PC2. Then run prcomp() on the subsets, do the same plot and compare.

Thanks Neil. I did exactly what you recommended using the iris data and the results are strikingly different. Looks like it's what both of us expected - you likely can't simply use the PCs computed on all the data for an analysis with a subset.

data(iris)

# Fit prcomp to all 150 observations
pca_all <- prcomp(~Sepal.Length + Sepal.Width + Petal.Length + Petal.Width, 
                  data=iris)

# Extract data rotated multiplied by rotation matrix and merge with orig data.
iris_pca_all <- data.frame(iris,pca_all$x)

# Plot PC2 by PC1 for only the subset of 'setosa' species, add linear trendline
with(subset(iris_pca_all, Species=='setosa'), plot(PC1,PC2, pch=20))
abline(lm(PC2~PC1, data=subset(iris_pca_all, Species=='setosa')))

# Fit prcomp to only the 50 setosa species individuals
pca_setosa <- prcomp(~Sepal.Length + Sepal.Width + Petal.Length + Petal.Width, 
                     data=subset(iris_pca_all, Species=='setosa'))

# Plot PC2 by PC1 for PC loadings computed from the subset only.
with(as.data.frame(pca_setosa$x), plot(PC1,PC2,pch=20))
abline(lm(PC2~PC1, data=as.data.frame(pca_setosa$x)))