CHAPTER 7.

MULTIVARIATE
STATISTICAL MODELS

7.2. Covariance and correlation matri-

ces

Let Y = (Y1, · · · , Yd)T be a random

vector. Define the expectation of Y as
E(Y) = (E(Y1), · · · , E(Yd))T , covariance
of Y as
COV (Y) = (Cov(Yi, Yj ))1≤i,j≤d
= E{(Y − E(Y))(Y − E(Y))T },

1
correlation matrix of Y as

CORR(Y) = S−1COV (Y)S−1,

where S = diag(σY1 , · · · , σYd ) and σYi is

the standard deviation of Yi.

Example 7.1. CRSPday covariances and

correlations

2
7.3. Linear combinations:

i) E(wT Y) = wT E(Y);

ii) V ar(wT Y) = wT COV (Y)w;

T Y , wT Y ) = wT COV (Y )w ;
iii) Cov(w1 2 1 2

T Y , wT Y ) = wT COV (Y , Y )w
iv) Cov(w1 1 2 2 1 1 2 2

3
7.4 Scatterplot Matrices

A correlation coefficient is only a sum-

mary of the linear relationship between
variables. Interesting features such as
nonlinearity or the joint behavior of ex-
treme values, remain hidden when only
correlations are examined. A solution
to this problem is the so-called scatter-
plot of matrix.

4
7.5. Multivariate normal distribution

The random vector Y = (Y1, · · · , Yd)T

has a d-dimensional multivariate nor-
mal distribution with mean µ = (µ1, · · · , µd)T
and covariance matrix Σ if its density
function is
φd(y|µ, Σ) = 1 ×
(2π)d/2|Σ|1/2
1 (y
exp{− 2 − µ)T Σ−1(y − µ)}.

5
Fig 7.2. Contour plots of a bivariate
normal densities with N (0, 1) marginals
and correlations of 0.5 or −0.95.

(a) corr = 0.5 (b) corr = −0.95

0.02
2

2
0.06
0.1
1

1
14
0.
X2

X2
0

0
6
0.1

2
0.1
−1

−1

0.08
0.04
−2

−2

−2 −1 0 1 2 −2 −1 0 1 2

X1 X1

6
Important property: If Y has a mul-
tivariate normal distribution, then any
linear combination cT Y has a normal
distribution with mean cT µ and vari-
ance cT Σc.

7
7.6 Multivariate t-distribution

The random vector Y has a multivari-

ate tν (µ, Λ) distribution if
ν
s
Y =µ+ Z,
W
where W is chi-squared distributed with
ν degrees of freedom, Z ∼ Nd(0, Λ),
and W and Z are independent.

8
For ν > 1, E(Y) = µ; For ν > 2,
COV (Y) = ν−2ν Λ := Σ. We call Λ the

scale matrix. Any linear combination

wT Y has a univariate t-distribution with
mean wT µ and variance wT Σw.

9
7.7 Fitting Multivariate t-distribution

R codes: selm in sn package

Example 7.4. Fitting the CRSPday data.

10
7.8 Elliptically Contoured Densities

The multivariate normal and t-distributions

have elliptically contoured densities. A
d-variate multivariate density f is ellip-
tically contoured if it can be expressed
as

f (y) = |Λ|−1/2g((y − µ)T Λ−1(y − µ)),

where g is a nonnegative function such

that
Z
1= g(||y ||2) dy ,

µ is a d × 1 vector, Λ is a d × d sym-
metric, positive definite matrix.
11
7.9 Multivariate Skewed t-distribution

In the multivariate case, in addition to

the shape parameter ν determining the
tail weight, the skewed t-distribution
has a vector α = (α1, · · · , αd)T of shape
parameters determining the amounts
of skewness in the components of the
distribution.

12
R codes: mst.mple in sn package.

Example 7.5. Fitting the skewed t-

distribution to CRSPday.

13
7.10 Fisher Information Matrix Fisher
Information Matrix:
∂2
Iij (θ ) = −E{ log L(θ )}
∂θi∂θj

14
Theorem 7.6. Under suitable assump-
tions, the maximum likelihood estima-
tor is approximately normally distributed
with mean equal to the true parame-
ter vector and with covariance matrix
equal to the inverse of the Fisher in-
formation matrix.

Homework: Exercises 1, 2, 5

15