TA: Tonja Bowen Bishop

14.
30 PROBLEM SET 9 SUGGESTED ANSWERS
TA: Tonja Bowen Bishop
Problem 1
For parts (a)-(g), please refer to the relevant sections in the handouts/textbooks.
(h) False. With the knowledge of the distribution of a statistic under
the null hypothesis, we can calculate only , but not :
(i) False. In particular, there are hypothesis tests for which one can
form a GLRT where no optimal test exists, so GLRT could not be optimal
in that case.
(j) Yes. Hypothesis tests are typically constructed by using a test
statistic T . The null hypothesis is rejected if T lies in some interval or if T
lies outside of some interval. The interval is chosen to make the test have
a desired signicance level. This procedure is in essence the same as the
construction of the condence interval.
(k) Recall that and are calculated under dierent distributions
(under the null and under the alternative hypothesis). There is no reason
that it always must be + = 1; although it is possible under some special
circumstances. Also, in general, it is not possible = = 0; again in some
cases we might have = 0 or = 0, but in many cases neither is possible.
NOTE: We might have some cases that = = 0 under very special
and unrealistic situation (can you think of an example of such case ?), and
in that case, the hypothesis test becomes trivial and uninteresting. (why ?)
Problem 2
(a) The probability of committing a Type I error is calculated as
follows:
P (Type I error) = P (reject H0 jH0 is true)
= P (Y
3:20j = 1)
Z 1
=
e y dy
3:20
= 0:04
(b) The probability of committing a Type II error when

calculated as follows:
1
= 4=3 is
14.30 PROBLEM SET 9 SUGGESTED ANSWERS
P (Type II error) = P (dont reject H0 jH1 is true)

4
= P (Y
3:20j = )
3
Z 3:20
3 3y=4
=
e
dy
4
0
Z 2:4
3
=
e u du (change in variable: u = y)
4
0
= 0:91
Problem 3
(a)
i) We rst nd formulas for
and
in terms of k:
= P (Type I error) = P (reject H0 jH0 is true)

Z k
Z k
=
f0 (x)dx =
2xdx
= k
0
2
= P (Type II error) = P (dont reject H0 jHA is true)

Z 1
Z 1
=
fA (x)dx =
(2 2x)dx
k
= k2
2k + 1
Note that we were able to write down expressions for and because
with only 1 observation x; we knew that x had to be our statistic, and
furthermore our test would be of the form reject H0 for x < kbecause we
are more likely to have observed a small x if HA is true.
min( + ) = min(2k 2
k
@(2k 2
So
2k + 1)
= 4k
@k
is minimized at k = 1=2:
ii)
1
4
1
4
2k + 1)
2=0
= 12 :
iii) With k = 1=2, the testing procedure is then to reject the null for
x < k = 1=2: So we do not reject the null for x = 0:6
(b)
Z
f0 (x)dx =
i) Find k such that

Z
2xdx = k 2 = 0:1 [why 0.1, when the restriction is
this implies k =
0:1:
0:10 ?]
ii) Then
(2
2x)dx
0:1
x2 ]1p0:1
p
2 0:1
= [2x
= 1:1
iii) For x = 0:4, we do not reject because

p
x = 0:4 > k = 0:1
(c) With ten observations, you could use, for instance

T (X) = X
and reject the if X < k (for appropriate k).
Problem 4
We have two simple hypotheses, so we should be able to derive an optimal test statistic using the Neyman-Pearson lemma. The lemma tells
us that the best test can be achieved by constructing the test statistic
1 ;:::;xn j =1)
T = ff01 (x
(x1 ;:::;xn j =0) and then rejecting the null ( = 0) whenever T > k,
where k is the critical value such that the probability of type I error is equal
to = 0:005. We know that the likelihood function is
n
f (x1 ; :::; xn j ) =
i=1
f (xi j )
1
p e
i=1 2
= (2 )
n
2
)2
(xi
2
1
2
P
(xi
)2
Then we have
T
(2 )
n
2
(2 )
1
2
= e
= enX
e
n
2
1
2
P
(xi 1)2
1 P 2
xi
2
P
( 2xi +1)
n
2
Thus we should reject the null hypothesis if enX 2 > k, or equivalently,

reject the null if X > 2 ln(k)+n
. Thus we have a test of the form "reject if
2n
X > c" for some c (n), and we know how to determine the proper value of c
for our desired condence level. We want Pr X > c j = 0 = = 0:005,
so we use
0:005 = Pr X > c j = 0
0
X 0
c 0
= Pr @ q
> q j
1
n
= Pr Z >
1
n
nc
From the Z-table, I nd that we must have
= 0A
nc = 1:65, so c =
the nal form of our test is that we reject the null if X >
Problem 5
(a) H0 :
X
if
= 7 versus H1 :
1:65
p .
n
Thus
1:65
p .
n
= 6.
(b) The
And under the null
p test statistics is sample mean: T = X: p
N (7; 1= 10);and under the alternative X N (7; 1 10): We reject H0
1
1
p
(0:05)
10
1
= 7 p (1:65)
10
= 7 (0:32)(1:65)
= 6:47
X < 7
So we reject H0 because X = 6:2 < 6:47:

(c)
P ower = 1
= 1
=
=
=
=
(d) With unknown
P (dont reject H0 jHA is true)

6:47 6
p
1
1
1= 10
6:47 6
p
1= 10
(1:49)
0:93
2,
the test statistic under the null is
T =
X
p
S= n
t(10
1)
we reject H0 if
X
p
S= n
6:2 7
p
p
1:5= 10
<
t9 ( ), where
is a signicance level
,
=
2:5 <
1:83 (from the table)
So we reject H0 :
H1 :
(e) In this case, the test then becomes two-sided [H0 : = 7 versus
6= 7], so all the procedures used before should be modied accordingly.
Problem 6
Suppose Xi
N ( X ; 1); i = 1; :::; nX and Zj
N ( Z ; 1); j = 1; :::; nZ
and all the observations are independent. You want to test the hypothesis
that the means are equal against the alternative that they are not. Use the
statistic
T =p
(X
Z)
1=nX + 1=nZ
(a) The test statistic T can take any real number, and if the absolute
value of T is su cientlydierent from 0; we will reject the null hypothesis
that the means are equal.
(b) Note that T is a linear transformation of (nX + nZ ) normal

random variables - hence, we can guess that T
N (E[T ]; V ar[T ]):The
mean and variance of T are:
"
#
(X Z)
E[T ] = E p
1=nX + 1=nZ
1
= p
E[(X Z)]
1=nX + 1=nZ
1
= p
E[X] E[Z]
1=nX + 1=nZ
1 X
1
1 X
= p
E
Xi
E
Zj
nX
nZ
1=nX + 1=nZ
1
= p
( X
Z)
1=nX + 1=nZ
= 0 (under the null X = Z )
"
(X
V ar[T ] = V ar p
Z)
1=nX + 1=nZ
1
V ar[X Z]
1=nX + 1=nZ
1
V ar[X] + V ar[Z] (independence)
1=nX + 1=nZ
1
1 X
1 X
V ar
Zj
Xi + V ar
1=nX + 1=nZ
nX
nZ
1
1
1
nX (1) + 2 nZ (1)
1=nX + 1=nZ
n2X
nZ
!
1
1
1
+
nX +nZ
nX
nZ
n n
=
=
=
=
nX nZ
nX + nZ
nX + nZ
nX nZ
= 1
So T
N (0; 1):
(c) Since the distribution of the test statistic T is the standard normal under the null, we reject the null when jtj > z =2 :
Problem 7
(a) Consider the test statistic T =
and
(nY
SY2 is
1)SY2
dened analogously.
2
Y
2
(nY
1) ,
2 =
where SX
We know that
n
Y
P
i=1
nX 1
2
(nX 1)SX
2
X
2
X
2,
Y
so we can rewrite T as
n
X
P
Xi
i=1
2
(nX 1)
n
X
P
Un-
2
(nX 1)SX
2 (n
1
)
X X
2
(nY 1)SY
2 (n
Y 1)
Y
Xi
1; nY
=
i=1
Yi
= 8:75. So T =
118=5
8:75=3
and
1) : So we reject for T > c, where c is dened by Pr (F (nX
(b) We have nX = 6, nY = 4, X = 12, Y = 2:75,

118, and
and that these two statistics are independent.
der the null hypothesis,

F (nX 1; nY
= 0:10.
2
SX
,
SY2
= 8:09. From the F-table,
our critical value, c, is equal to 5.31, so we reject the null hypothesis.
1) > c) =

TA: Tonja Bowen Bishop

Uploaded by

Copyright:

Available Formats

TA: Tonja Bowen Bishop

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

TA: Tonja Bowen Bishop

Uploaded by

Copyright:

Available Formats

14.

30 PROBLEM SET 9 SUGGESTED ANSWERS

TA: Tonja Bowen Bishop

(b) The probability of committing a Type II error when

14.30 PROBLEM SET 9 SUGGESTED ANSWERS

P (Type II error) = P (dont reject H0 jH1 is true)

= P (Type I error) = P (reject H0 jH0 is true)

= P (Type II error) = P (dont reject H0 jHA is true)

i) Find k such that

2xdx = k 2 = 0:1 [why 0.1, when the restriction is

14.30 PROBLEM SET 9 SUGGESTED ANSWERS

iii) For x = 0:4, we do not reject because

(c) With ten observations, you could use, for instance

Thus we should reject the null hypothesis if enX 2 > k, or equivalently,

14.30 PROBLEM SET 9 SUGGESTED ANSWERS

From the Z-table, I nd that we must have

So we reject H0 because X = 6:2 < 6:47:

(d) With unknown

P (dont reject H0 jHA is true)

the test statistic under the null is

14.30 PROBLEM SET 9 SUGGESTED ANSWERS

1:83 (from the table)

(b) Note that T is a linear transformation of (nX + nZ ) normal

14.30 PROBLEM SET 9 SUGGESTED ANSWERS

1) : So we reject for T > c, where c is dened by Pr (F (nX

(b) We have nX = 6, nY = 4, X = 12, Y = 2:75,

and that these two statistics are independent.

der the null hypothesis,

= 8:09. From the F-table,

our critical value, c, is equal to 5.31, so we reject the null hypothesis.

You might also like