In the first post on factor analysis, we examined computing the estimated covariance matrix \(S\) of the rootstock data and proceeded to find two factors that fit most of the variance of the data. However, the variables in the data are not on the same scale of measurement, which can cause variables with comparatively large variances to dominate the diagonal of the covariance matrix and the resulting factors. The correlation matrix, therefore, makes more intuitive sense to employ in factor analysis. In fact, as we saw previously, most packages available in R default to using the correlation matrix when performing factor analysis. There are several benefits to using \(R\) over \(S\), not only that it scales non-commensurate variables, but it is also easier to calculate the factors as the matrix does not need to be decomposed and estimated like \(S\).
Factor Analysis with the Correlation Matrix
Similar to factor analysis with the covariance matrix, we estimate \(\Lambda\) which is \(p \times m\) where \(D\) is a diagonal matrix of the \(m\) largest eigenvalues of \(R\), and \(C\) is a matrix of the corresponding eigenvectors as columns.
Where \(\theta_1, \theta_2, \cdots, \theta_m\) are the largest eigenvalues of \(R\).
Thus the correlation matrix \(R\) does not require decomposition, and we can proceed directly to finding the eigenvalues and eigenvectors of \(R\).
Load the rootstock data and name the columns. From the previous post:
The rootstock data contains growth measurements of six different apple tree rootstocks from 1918 to 1934 (Andrews and Herzberg 1985, pp. 357-360) and were obtained from the companion FTP site of the book Methods of Multivariate Analysis by Alvin Rencher. The data contains four dependent variables as follows:
- trunk girth at four years (mm \(\times\) 100)
- extension growth at four years (m)
- trunk girth at 15 years (mm \(\times\) 100)
- weight of tree above ground at 15 years (lb \(\times\) 1000)
root <- read.table('ROOT.DAT', col.names = c('Tree.Number', 'Trunk.Girth.4.Years', 'Ext.Growth.4.Years', 'Trunk.Girth.15.Years', 'Weight.Above.Ground.15.Years'))
Compute the correlation matrix of the data.
R <- cor(root[,2:5])
round(R, 2)
## Trunk.Girth.4.Years Ext.Growth.4.Years
## Trunk.Girth.4.Years 1.00 0.88
## Ext.Growth.4.Years 0.88 1.00
## Trunk.Girth.15.Years 0.44 0.52
## Weight.Above.Ground.15.Years 0.33 0.45
## Trunk.Girth.15.Years
## Trunk.Girth.4.Years 0.44
## Ext.Growth.4.Years 0.52
## Trunk.Girth.15.Years 1.00
## Weight.Above.Ground.15.Years 0.95
## Weight.Above.Ground.15.Years
## Trunk.Girth.4.Years 0.33
## Ext.Growth.4.Years 0.45
## Trunk.Girth.15.Years 0.95
## Weight.Above.Ground.15.Years 1.00
Then find the eigenvalues and eigenvectors of \(R\).
r.eigen <- eigen(R)
r.eigen
## eigen() decomposition
## $values
## [1] 2.78462702 1.05412174 0.11733950 0.04391174
##
## $vectors
## [,1] [,2] [,3] [,4]
## [1,] 0.4713465 0.5600120 0.6431731 0.2248274
## [2,] 0.5089667 0.4544775 -0.7142114 -0.1559013
## [3,] 0.5243109 -0.4431448 0.2413716 -0.6859012
## [4,] 0.4938456 -0.5324091 -0.1340527 0.6743048
We can check the proportion of each eigenvalue respective to the total sum of the eigenvalues.
Where \(p\) is the number of variables. The quick and dirty loop below finds the proportion of the total for each eigenvalue and the cumulative proportion.
cumulative.proportion <- 0
prop <- c()
cumulative <- c()
for (i in r.eigen$values) {
proportion <- i / dim(root[,2:5])[2]
cumulative.proportion <- cumulative.proportion + proportion
prop <- append(prop, proportion)
cumulative <- append(cumulative, cumulative.proportion)
}
data.frame(cbind(prop, cumulative))
## prop cumulative
## 1 0.69615676 0.6961568
## 2 0.26353043 0.9596872
## 3 0.02933488 0.9890221
## 4 0.01097793 1.0000000
As in the case of the covariance matrix, the first two factors account for nearly all of the sample variance and thus can proceed with \(m = 2\) factors.
The eigenvectors corresponding to the two largest eigenvalues are multiplied by the square roots of their respective eigenvalues as seen earlier to obtain the factor loadings.
factors <- t(t(r.eigen$vectors[,1:2]) * sqrt(r.eigen$values[1:2]))
round(factors, 2)
## [,1] [,2]
## [1,] 0.79 0.57
## [2,] 0.85 0.47
## [3,] 0.87 -0.45
## [4,] 0.82 -0.55
Computing the communality remains the same as in the covariance setting.
h2 <- rowSums(factors^2)
The specific variance when factoring \(R\) is \(1 - \hat{h}^2_i\).
u2 <- 1 - h2
According to the documentation of the principal() function (called by
`?principal), there is another statistic called complexity, which is
the number of factors on which a variable has moderate or high loadings
(Rencher, 2002 pp. 431), that is found by:
In the most simple structure, the complexity of all the variables is \(1\). The complexity of the variables is reduced by performing rotation which will be seen later.
com <- rowSums(factors^2)^2 / rowSums(factors^4)
com
## [1] 1.831343 1.553265 1.503984 1.737242
mean(com)
## [1] 1.656459
As seen in the previous post, the principal() function from the psych
package performs factor
analysis with the principal component method.
library(psych)
Since we are using \(R\) instead of \(S\), the covar argument remains
FALSE by default. No rotation is done for now, so the rotate
argument is set to none.
root.fa <- principal(root[,2:5], nfactors = 2, rotate = 'none')
root.fa
## Principal Components Analysis
## Call: principal(r = root[, 2:5], nfactors = 2, rotate = "none")
## Standardized loadings (pattern matrix) based upon correlation matrix
## PC1 PC2 h2 u2 com
## Trunk.Girth.4.Years 0.79 0.57 0.95 0.051 1.8
## Ext.Growth.4.Years 0.85 0.47 0.94 0.061 1.6
## Trunk.Girth.15.Years 0.87 -0.45 0.97 0.027 1.5
## Weight.Above.Ground.15.Years 0.82 -0.55 0.98 0.022 1.7
##
## PC1 PC2
## SS loadings 2.78 1.05
## Proportion Var 0.70 0.26
## Cumulative Var 0.70 0.96
## Proportion Explained 0.73 0.27
## Cumulative Proportion 0.73 1.00
##
## Mean item complexity = 1.7
## Test of the hypothesis that 2 components are sufficient.
##
## The root mean square of the residuals (RMSR) is 0.03
## with the empirical chi square 0.39 with prob < NA
##
## Fit based upon off diagonal values = 1
The output of the principal() function agrees with our calculations.
Factor Rotation with Varimax Rotation
Rotation moves the axes of the loadings to produce a more simplified structure of the factors to improve interpretation. Therefore the goal of rotation is to find an interpretable pattern of the loadings where variables are clustered into groups corresponding to the factors. We will see that a successful rotation yields a complexity closer to \(1\), which denotes the variables load highly on only one factor.
One of the most common approaches to rotation is varimax rotation, which is a type of orthogonal rotation (axes remain perpendicular). The varimax technique seeks loadings that maximize the variance of the squared loadings in each column of the rotated matrix \(\hat{\Lambda}*\).
The varimax() function is used to find the rotated factor loadings.
For those interested, the R code for the varimax() function can be
found here.
factors.v <- varimax(factors)$loadings
round(factors.v, 2)
##
## Loadings:
## [,1] [,2]
## [1,] 0.16 0.96
## [2,] 0.28 0.93
## [3,] 0.94 0.29
## [4,] 0.97 0.19
##
## [,1] [,2]
## SS loadings 1.928 1.907
## Proportion Var 0.482 0.477
## Cumulative Var 0.482 0.959
The varimax rotation was rather successful in finding a rotation that simplified the complexity of the variables. The first two variables now load highly on the second factor while the remaining two variables load primarily on the first factor.
Since we used an orthogonal rotation technique, the communalities will not change.
h2.v <- rowSums(factors.v^2)
h2.v
## [1] 0.9492403 0.9390781 0.9725050 0.9779253
h2
## [1] 0.9492403 0.9390781 0.9725050 0.9779253
Thus the specific variances will also be unchanged.
u2.v <- 1 - h2.v
u2.v
## [1] 0.05075965 0.06092192 0.02749496 0.02207470
u2
## [1] 0.05075965 0.06092192 0.02749496 0.02207470
As stated previously, the complexity of the variables on the rotated factors should be closer to \(1\) compared to the non-rotated complexity.
com.v <- rowSums(factors.v^2)^2 / rowSums(factors.v^4)
com.v
## [1] 1.054355 1.179631 1.185165 1.074226
mean(com.v)
## [1] 1.123344
The complexity is rather close to \(1\) which provides us further acknowledgment the factors are now in a more simplified structure.
Setting the rotation argument to varimax in the principal()
function outputs the rotated factors and corresponding statistics.
root.fa2 <- principal(root[,2:5], nfactors = 2, rotate = 'varimax')
root.fa2
## Principal Components Analysis
## Call: principal(r = root[, 2:5], nfactors = 2, rotate = "varimax")
## Standardized loadings (pattern matrix) based upon correlation matrix
## RC1 RC2 h2 u2 com
## Trunk.Girth.4.Years 0.16 0.96 0.95 0.051 1.1
## Ext.Growth.4.Years 0.28 0.93 0.94 0.061 1.2
## Trunk.Girth.15.Years 0.94 0.29 0.97 0.027 1.2
## Weight.Above.Ground.15.Years 0.97 0.19 0.98 0.022 1.1
##
## RC1 RC2
## SS loadings 1.94 1.90
## Proportion Var 0.48 0.48
## Cumulative Var 0.48 0.96
## Proportion Explained 0.50 0.50
## Cumulative Proportion 0.50 1.00
##
## Mean item complexity = 1.1
## Test of the hypothesis that 2 components are sufficient.
##
## The root mean square of the residuals (RMSR) is 0.03
## with the empirical chi square 0.39 with prob < NA
##
## Fit based upon off diagonal values = 1
Interpretation of Factors
The factor analysis performed on the rootstock data yielded two latent variables that fit and explain the variance of the data quite sufficiently. We see both variables relating to measurements at four years load heavily on factor 2 while the 15-year measurements load mainly on the first factor. Thus we could designate names for the factors, or latent variables, such as '15 years growth' and '4 years growth', respectively. There isn't any standard way of 'naming' factors as the interpretation can vary widely between each case. In this example, the factors make intuitive sense based on how they load on the variables; however, factors resulting from a factor analysis may not always make logic sense to the original data. If the resulting factors do not seem logical, changes to the approach such as adjusting the number of factors or the threshold of the loadings deemed important, or even a different method of rotation can be done to improve interpretation.
References
Rencher, A. C. (2002). Methods of multivariate analysis. New York: J. Wiley.
https://en.wikipedia.org/wiki/Talk:Varimax_rotation
https://en.wikipedia.org/wiki/Varimax_rotation
http://web.stanford.edu/class/psych253/tutorials/FactorAnalysis.html