Canonical Correlations
Canonical Correlations
I
Like Principal Components Analysis, Canonical Correlation Analysis
looks for interesting linear combinations of multivariate observations.
I
In Canonical Correlation Analysis, a multivariate response X is
partitioned into two groups:


X(1)
 p×1 
X =  (2) 
(p+q)×1
X
q×1
I
We look for linear combinations U = a0 X(1) and V = b0 X(2) that are
highly correlated.
NC STATE UNIVERSITY
1 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
I
The covariance matrix of X is correspondingly partitioned:
Σ1,1 Σ1,2
Cov(X) = Σ =
Σ2,1 Σ2,2
I
Then
I
Without loss of generality, Var(U) = a0 Σ1,1 a and
Var(V ) = b0 Σ2,2 b can both be constrained to equal 1.
I
The first pair of canonical variables are the U1 and V1 that maximize
the correlation, subject to this constraint.
a0 Σ1,2 b
p
Corr(U, V ) = p 0
a Σ1,1 a b0 Σ2,2 b
NC STATE UNIVERSITY
2 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
I
Differentiation shows that the maximizing a and b satisfy
Σ1,2 b − αΣ1,1 a = 0
Σ2,1 a − βΣ2,2 b = 0
for some α and β.
I
Then
Corr(U, V ) = a0 Σ1,2 b = αa0 Σ1,1 a = α
= βb0 Σ2,2 b = β
so α = β =
√
λ, say.
NC STATE UNIVERSITY
3 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
I
We assume that Σ1,1 and Σ2,2 are invertible, and then
√
λa
Σ−1
1,1 Σ1,2 b =
√
−1
Σ2,2 Σ2,1 a = λb
I
Eliminate b:
−1
Σ−1
1,1 Σ1,2 Σ2,2 Σ2,1 a = λa.
I
−1
That is, a is an eigenvector of Σ−1
1,1 Σ1,2 Σ2,2 Σ2,1 and b is a multiple
of Σ−1
2,2 Σ2,1 a.
I
−1
Consequently, b is an eigenvector of Σ−1
2,2 Σ2,1 Σ1,1 Σ1,2 with the same
eigenvalue as a.
NC STATE UNIVERSITY
4 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
I
√
Since Corr(U, V ) = λ, the first canonical pair, U1 and V1 , is
determined by the largest eigenvalue λ1 = Corr(U1 , V1 )2 = ρ∗1 2 .
I
We then look for the pair U2 and V2 with the greatest correlation,
subject to Corr(U1 , U2 ) = Corr(V1 , V2 ) = 0.
I
As a bonus, Corr(U1 , V2 ) = Corr(U2 , V1 ) = 0.
I
Not surprisingly, this second canonical pair, U2 and V2 , is determined
by the next largest eigenvalue λ2 = Corr(U2 , V2 )2 = ρ∗2 2 .
I
Continuing the process, we find r = min(p, q) canonical pairs.
NC STATE UNIVERSITY
5 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
Example: head and leg measurements of chicken bones
I
Head group:
I
I
I
Leg group:
I
I
I
skull length;
skull breadth.
femur length;
tibia length.
SAS proc cancorr program and output.
NC STATE UNIVERSITY
6 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
The number of canonical pairs
I
In sample data, all r = min(p, q) canonical correlations will be
positive, even if the data are drawn from a population with only
m < r positively correlated pairs.
I
proc cancorr tests the hypothesis
H0 : ρ∗m = ρ∗m+1 = · · · = ρ∗r = 0
for each value of m, m = 1, 2, . . . , r .
I
The first null hypothesis is that the two groups are entirely
uncorrelated. If that is rejected, the second is that the correlation
between them is carried by a single linear combination of each group,
etc.
NC STATE UNIVERSITY
7 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
Standardization
I
The coefficient vectors ai and bi may be difficult to interpret,
especially when the variables are in different physical units or have
very different variances.
I
Consequently, the data are often standardized, or equivalently, the
correlation matrix is used instead of the covariance matrix.
I
Unlike in Principal Components Analysis, the canonical variables are
unchanged by any such scaling.
I
proc cancorr reports both raw and standardized forms.
NC STATE UNIVERSITY
8 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
Correlations
I
Sometimes the canonical variables can be interpreted as representing
characteristics of the items being measured.
I
The interpretation may be made based on the (raw or standardized)
coefficients.
I
The correlations between the original variables and the canonical
variables can also be informative, but “must be interpreted with
caution.”
I
proc cancorr reports the correlations of each original variable with
both sets of canonical variables.
NC STATE UNIVERSITY
9 / 10
Statistics 784
Multivariate Analysis
Canonical Correlations
Special Cases
p = q = 1: the “canonical” correlation between single variables X1 and
X2 is just the absolute value of their simple Pearson
correlation ρ.
p = 1, q > 1: the canonical correlation of a single variable X1 and a group
of variables X(2) is the multiple correlation R in the
regression of X1 on X(2) , so ρ∗1 2 = R 2 .
I
Also, in general, ρ∗k 2 is the R 2 in the regression of Uk on X(2) , and
also in the regression of Vk on X(1) : canonical correlations are closely
related to R 2 .
NC STATE UNIVERSITY
10 / 10
Statistics 784
Multivariate Analysis