Homework 13 - Solution

(p.554: 10.105)
Let Y1, Y2, ..., Yn denote a random sample from a normal distribution with mean μ (unknown)
and σ2. For testing H0: σ2 = σ02 against Ha: σ2 > σ02, show that the likelihood ration test is
equivalent to the χ2 test given in Section 10.9.

Solution:

The hypotheses are H0: σ2 = σ02 vs. Ha: σ2 > σ02. The null hypothesis specifies Ώ0 = {σ2 : σ2 = σ02},
so in this restricted space the MLE for σ2 is σ02. Now we calculate the MLE for μ. The likelihood
function is
( yi   )2

L(  ,  )  i 1  (2 )  n / 2 ( 2 )  n / 2 exp(i 1 ( y2i 2 ) ) .
2 n n 2
1
2 
e 2 2

Then the log-likelihood function is

l (  ,  2 )  (n / 2) log(2 )  (n / 2) log  2  i 1 ( y2i 2 ) .

n 2

Taking the derivative with respect to μ, we have

l (  ,  2 )  i 1 ( yi 2 ) .
 n


Setting it to zero and solving for μ, we get the MLE for μ is Y .

For the unrestricted space Ώ = {σ2 : σ2 ≥ σ02}, the MLE for μ is still Y . Now we calculate the MLE
of σ2.
The likelihood function is
( yi   )2

L(  ,  )  i 1  (2 )  n / 2 ( 2 )  n / 2 exp(i 1 ( y2i 2 ) ) .
2 n n 2
1
2 
e 2

Replacing μ with its MLE, the likelihood function becomes

( yi  y )2

L( y ,  2 )  i 1  (2 )  n / 2 ( 2 )  n / 2 exp(i 1 ( y2i  y2 ) ) .
n n 2
1
2 
e 2

Then the log-likelihood function is

l ( y ,  2 )  (n / 2) log(2 )  (n / 2) log  2  i 1 ( y2i  y2 ) .

n 2

Taking the derivative with respect to σ2, we have

l ( y ,  2 )   2n 2  21 4 i 1 ( yi  y ) 2 .
 n
 2


n
Setting it to zero and solving for σ2, we get the solution is 1n i 1
( yi  y ) 2 .
Since the unrestricted space is Ώ = {σ2 : σ2 ≥ σ02}, the MLE for σ2 is

ˆ 2  max( 1n i 1 ( yi  y ) 2 ,  0 2 ) .
n

The likelihood ratio statistic is

 
( yi  y )2
i1
n
1 exp(  )
 ( yi  y ) 2 
exp  i12 2  i12ˆ 2  .
n n
ˆ )
L( 2  02  02 n/2 ( yi  y ) 2
 L(
0
ˆ)  ( yi  y )2
 ˆ 2
 02
i1
n
1
2 ˆ 2
exp( 
ˆ 2
)  0

If ˆ   0 , λ = 1. If ˆ   0 ,
2 2 2 2

( yi  y )2
i1
n
1 exp(  ) n/2
 ( yi  y ) 2   
  i1n 2  exp  i12 2  n2  ,
n n
ˆ )
L( 2  02  02 ( yi  y ) 2
 L(
0
ˆ)  ( yi  y )2
i1
n
1
2 ˆ 2
exp( 
ˆ 2
)  0
  0

and H0 is rejected when λ ≤ k. Let V = (n - 1)S2/σ02. By thm 7.3, we know that V is a chi-square
distributed random variable with n - 1 degrees of freedom. λ can be rewritten as

  Vn  exp V2  n2 
n/2

which is a function of V.
Now we calculate the derivative of λ with respect to V,
n n 1  V2  n2
d
dV  12 n 2 V 2 e (n  V ) .

Under the assumption that ˆ   0 , we have

2 2

n  V  n  n21 S 2  n  i1 2  n(1  i1n 2

n n
( yi  y ) 2 ( yi  y ) 2
)  n(1  ˆ 2 )  0 .
2

0 0 0 0

Thus, the function is monotonically decreasing function of V, the test λ ≤ k is equivalent to V ≥ c,

where c is chosen so that the test is of size α.

(p.555: 10.107)
Let S12 and S22 denote, respectively, the variances of independent random samples of sizes n
and m selected from normal distributions with mean μ1 and μ2 and common variance σ2. If
μ1 and μ2 are unknown, construct a likelihood ratio test of H0: σ2 = σ02 against Ha: σ2 = σa2,
assuming that σa2 > σ02.

Solution:

Let X1, X2, ..., Xn and Y1, Y2, ..., Ym denote the two samples. Under H0, the quantity

V  i1
( X i  X ) 2   ( Yi Y ) 2
n n
( n 1) S12  ( m 1) S 22
2
i 1
  02
0

has a chi-square distribution with n + m - 2 degrees of freedom.

The likelihood function is
( xi   )2 ( yi   )2
 
L(  ,  )  i 1  exp(i 1 ( xi221 )  i 1 ( yi222 ) )
n m  m2 n  m2 n n 2 m 2
2 1
2 
e 2 2
i 1
1
2 
e 2 2
 (2 ) ( 2 )
If Ha is true, then both S12 and S22 will tend to be larger than σ02. Under restriction, the MLEs for μ1
and μ2 are X and Y respectively. The MLE of σ2 is σ02. Thus, the maximized likelihood is
ˆ )
L(  1 1
exp( V2 ) .
0 ( 2 ) ( m n ) / 2  0m n

In the unrestricted space, the MLE of σ2 is either σ02 or σa2. For the former, the likelihood ratio will
ˆ )
L(
be equal to 1. But, for k < 1, the rejection region is   L(
0
ˆ)  k only if ˆ   a . In this case,


ˆ )
L(
L(
0
ˆ)     a mn
0 
exp  V2  V2     
 02
 a2
 a mn
0  
exp  V2 1   02 ,
a
2

which is a decreasing function of V. Thus, we reject H0 if V is too large, and the rejection region is
{V ≥ c}.