# factor loadings = eigenvectors in R output?

I'm trying to make sense of a principal component analysis using R (either princomp or prcomp, I get similar results) with a correlation matrix analysis. In particular, I'm having trouble understanding the factor loadings output.

The data are in a data frame called ds. Here is the eigenvalue/vector analysis of the correlation matrix.

> eigen(cor(ds))
$values [1] 3.9831253 0.5483985 0.4520555 0.4191041 0.3506337 0.2466829$vectors
[,1]       [,2]         [,3]       [,4]       [,5]        [,6]
[1,] -0.4292649  0.2561060 -0.298728176  0.2474691  0.3939538  0.66668007
[2,] -0.3938608  0.7037562 -0.066953299  0.1803110 -0.4130921 -0.37677712
[3,] -0.4060290  0.1537541  0.365977815 -0.7055394  0.4027306 -0.13259964
[4,] -0.4166136 -0.4038936 -0.004352731  0.4760061  0.3984587 -0.52719349
[5,] -0.3951931 -0.4041873 -0.621151418 -0.3939178 -0.3732825 -0.01071668
[6,] -0.4074327 -0.2983273  0.621683956  0.1634262 -0.4624437  0.34343296


Now, when I ask for a principal component analysis, I get the following initial output.

> pca.out <- princomp(ds, cor=TRUE)
> summary(pca.out)
Importance of components:
Comp.1     Comp.2     Comp.3     Comp.4     Comp.5     Comp.6
Standard deviation     1.9957769 0.74053931 0.67235070 0.64738253 0.59214334 0.49667185
Proportion of Variance 0.6638542 0.09139974 0.07534258 0.06985069 0.05843896 0.04111382
Cumulative Proportion  0.6638542 0.75525396 0.83059653 0.90044722 0.95888618 1.00000000


which all makes sense, given the output of the eigen() function.

However, my understanding is that loadings are computed as the product of the eigenvector and the square root of the eigenvalue. When I ask for the loadings from the pca, I get

Loadings:
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6
A -0.429  0.256 -0.299  0.247 -0.394  0.667
B -0.394  0.704         0.180  0.413 -0.377
C -0.406  0.154  0.366 -0.706 -0.403 -0.133
D -0.417 -0.404         0.476 -0.398 -0.527
E -0.395 -0.404 -0.621 -0.394  0.373
F -0.407 -0.298  0.622  0.163  0.462  0.343

Comp.1 Comp.2 Comp.3 Comp.4 Comp.5 Comp.6
SS loadings     1.000  1.000  1.000  1.000  1.000  1.000
Proportion Var  0.167  0.167  0.167  0.167  0.167  0.167
Cumulative Var  0.167  0.333  0.500  0.667  0.833  1.000


Not only do the cumulative and proportion variance not match the initial output, but the loadings simply match the eigenvectors, with no adjustment for eigenvalue. Furthermore, the claim that the first component captures 66% of the variance is impossible with these loading values, because every single variable in the data set (A-F) has a later component with a higher (absolute) loading.

Can someone please straighten out my confusion/error? For the record, I'm running R version 3.3.3

• ttnphns comment here: R documentation labels eigenvectors "loadings" ... (to say it soft) ... recklessly. This issue was asked by users here million of time. – Antoni Parellada Jul 14 '17 at 6:28
• @AntoniParellada - thanks. I did actually google search for this and also looked at the suggested links that came up when posting my question. I hope that maybe my title will make the issue easy to identify as a repeat in future. – J Taylor Jul 14 '17 at 20:41

## 1 Answer

However, my understanding is that loadings are computed as the product of the eigenvector and the square root of the eigenvalue.

I depends on definition of loading you use. In princomp loadings are simply coefficients of principal components (recall that principal components are linear combinations of original variables) that are equal to eigenvectors entries. This has one inconvenience: since variance of each PC equals corresponding eigenvaule, loadings defined this way are not correlations between PC's and original variables. Correction by square root of eigenvalue is done to standardize the variance of PC scores to 1 and therefore to allow for correlation interpretation of loadings. These standardized loadings are sometimes called loadings as well. See for example PCA function from FactoMineR package. It never uses a word loadings, it uses word coordinates for standardized loadings.

Not only do the cumulative and proportion variance not match the initial output

loadings function doesn't give you cumulative and proportion variance. It just gives you sum of squares of each PC's loadings. And this, by definition, is 1. So, you'll always see this kind of output. It sounds ridicullus but works well when you apply loadings function to Explanatory Factor Analysis. In PCA, second part of loadings output is simply useless.

the claim that the first component captures 66% of the variance is impossible with these loading values, because every single variable in the data set (A-F) has a later component with a higher (absolute) loading

Actually it is possible, since loadings here are just eigenvectors not standardized loadings.

• +1 but when you say "standardized loadings" (e.g. in the last sentence) do you mean standardized to SS=1 or scaled to variance=eigenvalue? If the latter, it's really strange to call it "standardized" because "standardized" usually means scaling to variance=1. I'm also confused by this sentence: "Correction by square root of eigenvalue is done to standardize their variance to 1 and therefore to allow for correlation interpretation." -- corr interpretation is only possible if variance=eigenvalue not variance=1! – amoeba Jul 14 '17 at 7:34
• @Łukasz Deryło - thank you for your clear explanation. In all the reading I've been doing it seems like "standardized loadings" gets abbreviated to just loadings. (Given that people write that the "loadings" provide the correlations between the components and the original variables.) It seems redundant to call the eigenvectors loadings when you can simply call them eigenvectors. However, I thank you for your explanation. – J Taylor Jul 14 '17 at 16:42
• @amoeba - I believe he means the component scores are standardized to z-scores (i.e. M=0, SD=1). It makes it possible to use the standardized loadings as the regression coefficients in a standardized regression predicting z for a measured variable from the z-scores of the components. And in that regression equation, the correlation interpretation is possible, because the components are uncorrelated (in the initial solution or in subsequent orthogonal rotations). – J Taylor Jul 14 '17 at 16:46
• @JTaylor This is exactly what I meant – Łukasz Deryło Jul 15 '17 at 6:18
• Lukasz (and @JTaylor): so to be completely clear - when you say "standardized loadings" you mean "loadings corresponding to the standardized scores" i.e. loadings having sum of squares = eigenvalue? I find it confusing terminology but okay. – amoeba Jul 16 '17 at 21:09