You are on page 1of 5

Ch.

Reject H0 if |t| is large is equivalent to reject H0 if its square


( X −μ0 )2 −1
=n ( X−μ 0 ) ( s ) ( X−μ0 ) is large.
2 2
t = 2
s /n

The variable t2 is the squared distance from the sample mean X

to the test value μ0.

The units of distance are expressed in terms of s/ √ n or estimated

standard deviation of X .

Once X ∧¿S2 are observed, we reject H0 at significant level α if

()

2 −1 α
n ( x−μ 0 ) ( s ) ( x −μ 0 ) >t 2
n−1 ,
2

where t 2n−1 ( α2 ) denotes the upper (100 α2 ¿ th


percentile of the t-dist.

with (n-1) degrees of freedom.

The corresponding 100(1-α )% C.I. for the parameter μ from the

statistics is
α α
( X − S t (n−1), X − S t (n−1)
( ) ( )
2 2
)
√n √n
Before the sample is selected, the 100(1-α ¿ % C . I .is a random

interval, because the endpoints depend upon the random

variables X and S. The probability that the interval contains μ is

1-α means that among large number of such independent

intervals 100(1-α ¿ % of them will contain μ .


For multivariate analysis, the test of H0: μ=u0 v.s. H1: μ ≠ u0 is a

natural generalization of the squared distance in its multivariate

analog

T 2=( X−μ 0) ' (S/n)−1 ¿


)

¿ n( X−μ 0)'( S)−1 ¿


)
n n
where X = 1 ∑ X i, , S= 1
∑ ( X −X ¿ )(X j− X)T ¿ and
n i=1 n−1 j=1 j

μ0
=( μ , … , μp0 ¿ '
.
10

The statistics T2 is called Hotelling’s T2 in honor of Harold

Hotelling who first obtained its sampling distribution.

Above T2 statistics can be proved to have the following

distribution
2 ( n−1 ) p
T F
n− p ( p , n−p )

where F( p ,n− p ) denotes a random variable with an F-distribution

with p and n-p degrees of freedom.

(see the p.d.f. in the book)

Summary : Let X 1 , X 2 , …, X n be a random vectors from N p (μ , Σ). We

have
(
α =Pr ⁡¿ α ¿ ¿

( n−1 ) p
¿ Pr ⁡¿)¿ F ( p ,n− p )(α ¿ ¿
n− p

whatever the true μ and Σ.

We reject H0: μ=μ 0 v.s. H1: μ ≠ μ 0 at level α if

T2=n( X−μ 0) ' ( S )−1 ¿)¿ ( n−1 ) p F ( p ,n− p )(α ¿


n− p

We can rewrite

( )
n −1
1
T 2
=√ n( X−μ0 )' n−1 ∑ ( X j−X ) ( X j−X ) ' √n ¿)
j =1

which is of the form

¿
r.v.)¿)-1¿ r.v.)

This is analogous to t2=√ n( X−μ0 )(s2 )−1 √ n ¿) or

¿
r.v.) )-1
¿ ¿
r.v.) for the single variate case.

Since the multivariate normal and Wishart random variables are

independent, their joint dist. is the product of the marginal

normal and Wishart distribution. The dist. of T2 can be derived

from the joint dist. by using calculus.

Ex.: Let X p*n= 6 (9 10 8


6 3 ) for p=2 and n=3. Evaluate the value of
T2 for μ0 =(9,5)T.
p.f. : After computations, we have x= 8 and (6 )
S= (−34
). Thus, S =(11/9/3 41/27
−3
9 )
/9 and −1

T =3(8-9,6-5) (S )(8−9 )= 7 .
2 T -1
6−5 9

See another example in the book.

One feature of the T2 statistics is that it is invariant under

changes in the units of measurements for X of the form:

Y p∗1 =C p∗ p X p∗1+ d p∗1


, C is nonsingular

It can be proved that T 2X =T 2Y .

5.3 T2 and L.R.T.(Likelihood Ratio Test)

Another way to test H0: μ=μ 0 is to apply the L.R.T. method. It is


max L(μ , Σ∨H 0) max L( μ0 , Σ)
defined to be ( μ ,Σ )
= (Σ)
.
max L( μ , Σ) max L( μ , Σ)
(μ , Σ) (μ , Σ )

From Ch. 4, we know that for a random sample comes from

N p ( μ , Σ ) , the
maximum likelihood estimators of μ∧Σ are

n−1
X and S respectively and the corresponding
n
−n
1
max L( μ , Σ)= np/ 2
¿ ^Σ ∨¿ 2 exp ⁡¿ ¿ ).
(μ , Σ) (2 π )

Under the hypothesis H 0 : μ=μ 0, the normal likelihood is


−n
1
L ( μ0 , Σ ) = np /2
¿ Σ∨¿ 2 exp ⁡¿ ¿Following the steps mentioned in
(2 π )

Ch.4, we rewrite the value


n n
−1 −1
∑ ( X j −μ0 ¿ )' Σ ( X j−μ 0)¿= tr {Σ ∑ ( X j −μ0 ¿ )( X j−μ0 )' }¿Apply the
−1 −1
2 j=1 2 j=1
n
result 4.10 with B=∑ (X j−μ0 ¿ )( X j−μ 0)' ¿ and b= n2 ,
j=1
−n
1
we have max L( μ0 , Σ)= np/ 2
¿ ^Σ 0∨¿ 2 exp ⁡¿ ¿ ) with
(Σ) (2 π )
n
^Σ 0= 1 ∑ (X j−μ 0¿ )( X j−μ 0) ' ¿.
n j=1
From above discussion, we have
max L(μ0 , Σ)
L.R.T. = ( Σ)
=¿ ^Σ∨ ¿ ¿ or 2
=¿ ^Σ∨ ¿ ¿ where 2
is
max L(μ , Σ) ^
¿ Σ 0∨¿ ¿ Λ n ^
¿ Σ 0∨¿ , ¿ Λ n

(μ , Σ)

called Wilk’s lambda.

We will reject H 0 : μ=μ 0when the likelihood ratio is too small.

The relationship between T2 and 2


is stated in the following
Λn

theory.

You might also like