326

Chapter 16: Introduction to Bayesian Methods of Inference

= = ) 1 ( ) | ( ) | ( , where y = 0, 1, , n and 0 < p < 1. So,

= p p p p
y
n
p y f
y n y

so that
) (
) ( ) (
) ( ) (
) (
) 1 (
) ( ) (
) (
) (
1
0
1 1
+ +
+ +

+
=

+

+ +
n
y n y
dp p p
y
n
y m
y n y
.
The posterior density of p is then
1 1 *
) 1 (
) ( ) (
) (
) | (
+ +

+ +
+ +
=
y n y
p p
y n y
n
y p g , 0 < p < 1.
This is the identical beta density as in Example 16.1 (recall that the sum of n i.i.d.
Bernoulli random variables is binomial with n trials and success probability p).

16.7 a. The Bayes estimator is the mean of the posterior distribution, so with a beta posterior
with = y + 1 and = n y + 3 in the prior, the posterior mean is
4
1
4 4
1

+
+
+
=
+
+
=
n n
Y
n
Y
p
B
.
b. p
n
np
n
Y E
p E
B

+
+
=
+
+
=
4
1
4
1 ) (
) ( ,
2 2
) 4 (
) 1 (
) 4 (
) (
) (
+

=
+
=
n
p np
n
Y V
p V

16.8 a. From Ex. 16.6, the Bayes estimator for p is
2
1
) | (
+
+
= =
n
Y
Y p E p
B
.
b. This is the uniform distribution in the interval (0, 1).

c. We know that n Y p / = is an unbiased estimator for p. However, for the Bayes
estimator,
Chapter 16: Introduction to Bayesian Methods of Inference 327
Instructors Solutions Manual

2
1
2
1 ) (
) (
+
+
=
+
+
=
n
np
n
Y E
p E
B
and
2 2
) 2 (
) 1 (
) 2 (
) (
) (
+

=
+
=
n
p np
n
Y V
p V
B
.
Thus,
2
2
2
2
2
) 2 (
) 2 1 ( ) 1 (
2
1
) 2 (
) 1 (
)] ( [ ) ( ) (
+
+
=

+
+
+
+

= + =
n
p p np
p
n
np
n
p np
p B p V p MSE
B B B
.
d. For the unbiased estimator p , MSE( p ) = V( p ) = p(1 p)/n. So, holding n fixed, we
must determine the values of p such that
n
p p
n
p p np ) 1 (
) 2 (
) 2 1 ( ) 1 (
2
2

<
+
+
.
The range of values of p where this is satisfied is solved in Ex. 8.17(c).


16.9 a. Here, p p p y p p y L
y 1
) 1 ( ) | ( ) | (

= = , where y = 1, 2, and 0 < p < 1. So,
1 1 1
) 1 (
) ( ) (
) (
) 1 ( ) , (

+
= p p p p p y f
y

so that
) (
) 1 ( ) 1 (
) ( ) (
) (
) 1 (
) ( ) (
) (
) (
1
0
2
+ +
+ +

+
=

+
=

+
y
y
dp p p y m
y
.
The posterior density of p is then
* 2
( )
( | ) (1 )
( 1) ( 1)
y
y
g p y p p
y



+
+ +
=
+ +
, 0 < p < 1.
This is a beta density with shape parameters
*
= + 1 and
*
= + y 1.


b. The Bayes estimators are
,
1
) | ( ) 1 (
Y
Y p E p
B
+ +
+
= =
,
) )( 1 (
) 1 )( 1 (
) )( 1 (
) 1 )( 2 ( 1
) | ( ) | ( )] 1 ( [ ) 2 (
2
Y Y
Y
Y Y Y
Y p E Y p E p p
B
+ + + + +
+ +
=
+ + + + +
+ +

+ +
+
= =


where the second expectation was solved using the result from Ex. 4.200. (Alternately,
the answer could be found by solving

=
1
0
*
) | ( ) 1 ( ] | ) 1 ( [ dp Y p g p p Y p p E .


16.10 a. The joint density of the random sample and is given by the product of the marginal
densities multiplied by the gamma prior:
328 Chapter 16: Introduction to Bayesian Methods of Inference
Instructors Solutions Manual

[ ]
( )

=


=

+
=
+

=
1
exp
) (
/ exp
) (
) / exp(
) (
1
) exp( ) , , , (
1
1
1
1
1
1
1
n
i
i
n
n
i
i
n
n
i
i n
y
y
y y y f

b.

=
+

+



=
0
1
1
1
1
exp
) (
1
) , , ( d
y
y y m
n
i
i
n
n
, but this integral resembles
that of a gamma density with shape parameter n + and scale parameter
1
1
+

=
n
i
i
y
.
Thus, the solution is
+
=

+

+

=

n
n
i
i
n
y
n y y m
1
) (
) (
1
) , , (
1
1
.

c. The solution follows from parts (a) and (b) above.


d. Using the result in Ex. 4.111,

( )
( ) ( ) 1
1
1 1
1
1
1
) 1 (
1
) | / 1 ( ) | (
1 1
1
1
* *
+
+
+
=
+
+
=

+
+

=

= = =


= =

=
n n
Y
n
Y
n
Y
E E
n
i
i
n
i
i
n
i
i
B
Y Y

e. The prior mean for 1/ is
) 1 (
1
) / 1 (

= E (again by Ex. 4.111). Thus,
B
can be
written as
( )

+


+

+
=
1
1
1
1
1

n n
n
Y
B
,
which is a weighted average of the MLE and the prior mean.


f. We know that Y is unbiased; thus E(Y ) = = 1/. Therefore,
( ) ( )

+


+

+
=

+


+

+
=
1
1
1
1
1
1
1
1
1
1
1
) ( ) (
n n
n
n n
n
Y E E
B
.
Therefore,
B
is biased. However, it is asymptotically unbiased since
0 / 1 ) (
B
E .
Also,
Chapter 16: Introduction to Bayesian Methods of Inference 329
Instructors Solutions Manual

( )
0
1
1
1
1
1
) ( ) (
2 2
2
2
2

+
=

+
=

+
=
n
n
n
n
n n
n
Y V V
B
.
So, / 1
p
B
and thus it is consistent.




16.11 a. The joint density of U and is

=


=

= =
+

1
exp
) ( !
) / exp(
) ( !
) / exp(
) (
1
!
) ( exp ) (
) ( ) | ( ) , (
1
1
1
n u
n
n
u
n
u
n n
g u p u f
u
u
u
u
u

b.

d
n u
n
u m
u
u
0
1
1
exp
) ( !
) ( , but this integral resembles that of a
gamma density with shape parameter u + and scale parameter
1 +

n
. Thus, the
solution is
+

+

=
u
u
n
u
u
n
u m
1
) (
) ( !
) ( .

c. The result follows from parts (a) and (b) above.

d.

+ = = =
1
) ( ) | (

* *
n
U U E
B
.

e. The prior mean for is E() = . From the above,
( )

+
+

+ =

=
1
1
1 1

1
n n
n
Y
n
Y
n
i
i B
,
which is a weighted average of the MLE and the prior mean.


f. We know that Y is unbiased; thus E(Y ) = Therefore,

+
+

+
+

=
1
1
1 1
1
1
) ( )

(
n n
n
n n
n
Y E E
B
.
So,
B

is biased but it is asymptotically unbiased since

(
B
E 0.
Also,
330 Chapter 16: Introduction to Bayesian Methods of Inference
Instructors Solutions Manual

( )
0
1 1 1
) ( )

(
2
2 2

+

=

+

=

=
n
n
n
n
n n
n
Y V V
B
.
So,
p
B

and thus it is consistent.

=

n
i
i
Y v
1
2
0
) ( is chisquare with n degrees of freedom.
Then, the density function for U (conditioned on v) is given by
( )
2 / 2 / 1 2 /
2 /
2 / 1 2 /
2 /
2 ) 2 / (
1
2 ) 2 / (
1
) ( ) | (
uv n n
n
uv n
n
W U
e v u
n
e uv
n
v uv f v v u f

= = .

a. The joint density of U and v is then
.
2
2
exp
2 ) ( ) 2 / (
1
) / 2 / exp(
2 ) ( ) 2 / (
1
) / exp(
) (
1
) 2 / exp(
2 ) 2 / (
1
) ( ) | ( ) , (
1 2 / 1 2 /
2 /
1 2 / 1 2 /
2 /
1 2 / 1 2 /
2 /

=


=

= =
+

u
v v u
n
v uv v u
n
v v uv v u
n
v g v u f v u f
n n
n
n n
n
n n
n
U

b. dv
u
v v u
n
u m
n n
n

=
0
1 2 / 1 2 /
2 /
2
2
exp
2 ) ( ) 2 / (
1
) ( , but this integral
resembles that of a gamma density with shape parameter n/2 + and scale parameter
2
2
+

u
. Thus, the solution is
+

+

=
2 /
2 /
1 2 /
2
2
) 2 / (
2 ) ( ) 2 / (
) (
n
n
n
u
n
n
u
u m .

c. The result follows from parts (a) and (b) above.


d. Using the result in Ex. 4.111(e),
( ) 2 2
2
2
2
1 2 /
1
) 1 (
1
) | / 1 ( ) | (
* *
2 2
+
+
=

+
+
=

= = =
n
U U
n
U v E U E
B
.

e. The prior mean for
) 1 (
1
/ 1
2

= = v . From the above,
( )

+


+

+
=
+
+
=
2 2
) 1 ( 2
) 1 (
1
2 2 2 2
2

2
n n
n
n
U
n
U
B
.


16.13 a. (.099, .710)
b. Both probabilities are .025.
Chapter 16: Introduction to Bayesian Methods of Inference 331
Instructors Solutions Manual

c. P(.099 < p < .710) = .95.
d.-g. Answers vary.
h. The credible intervals should decrease in width with larger sample sizes.


16.14 a.-b. Answers vary.

16.15 With y = 4, n = 25, and a beta(1, 3) prior, the posterior distribution for p is beta(5, 24).
Using R, the lower and upper endpoints of the 95% credible interval are given by:
> qbet a( . 025, 5, 24)
[ 1] 0. 06064291
> qbet a( . 975, 5, 24)
[ 1] 0. 3266527

16.16 With y = 4, n = 25, and a beta(1, 1) prior, the posterior distribution for p is beta(5, 22).
Using R, the lower and upper endpoints of the 95% credible interval are given by:
> qbet a( . 025, 5, 22)
[ 1] 0. 06554811
> qbet a( . 975, 5, 22)
[ 1] 0. 3486788

This is a wider interval than what was obtained in Ex. 16.15.


16.17 With y = 6 and a beta(10, 5) prior, the posterior distribution for p is beta(11, 10). Using
R, the lower and upper endpoints of the 80% credible interval for p are given by:
> qbet a( . 10, 11, 10)
[ 1] 0. 3847514
> qbet a( . 90, 11, 10)
[ 1] 0. 6618291

16.18 With n = 15,

=
n
i
i
y
1
= 30.27, and a gamma(2.3, 0.4) prior, the posterior distribution for
is gamma(17.3, .030516). Using R, the lower and upper endpoints of the 80% credible
interval for are given by
> qgamma( . 10, shape=17. 3, scal e=. 0305167)
[ 1] 0. 3731982
> qgamma( . 90, shape=17. 3, scal e=. 0305167)
[ 1] 0. 6957321

The 80% credible interval for is (.3732, .6957). To create a 80% credible interval for
1/, the end points of the previous interval can be inverted:

.3732 < < .6957
1/(.3732) > 1/ > 1/(.6957)

Since 1/(.6957) = 1.4374 and 1/(.3732) = 2.6795, the 80% credible interval for 1/ is
(1.4374, 2.6795).


332 Chapter 16: Introduction to Bayesian Methods of Inference
Instructors Solutions Manual

16.19 With n = 25,

=
n
i
i
y
1
= 174, and a gamma(2, 3) prior, the posterior distribution for is
gamma(176, .0394739). Using R, the lower and upper endpoints of the 95% credible
interval for are given by
> qgamma( . 025, shape=176, scal e=. 0394739)
[ 1] 5. 958895
> qgamma( . 975, shape=176, scal e=. 0394739)
[ 1] 8. 010663
16.20 With n = 8, u = .8579, and a gamma(5, 2) prior, the posterior distribution for v is
gamma(9, 1.0764842). Using R, the lower and upper endpoints of the 90% credible
interval for v are given by
> qgamma( . 05, shape=9, scal e=1. 0764842)
[ 1] 5. 054338
> qgamma( . 95, shape=9, scal e=1. 0764842)
[ 1] 15. 53867

The 90% credible interval for v is (5.054, 15.539). Similar to Ex. 16.18, the 90% credible
interval for
2
= 1/v is found by inverting the endpoints of the credible interval for v,
given by (.0644, .1979).

16.21 From Ex. 6.15, the posterior distribution of p is beta(5, 24). Now, we can find
) 3 . ( ) (
*
0
*
< = p P p P by (in R):
> pbet a( . 3, 5, 24)
[ 1] 0. 9525731

Therefore, ) 3 . ( ) (
* *
= p P p P
a
= 1 .9525731 = .0474269. Since the probability
associated with H
0
is much larger, our decision is to not reject H
0
.

16.22 From Ex. 6.16, the posterior distribution of p is beta(5, 22). We can find
) 3 . ( ) (
*
0
*
< = p P p P by (in R):
> pbet a( . 3, 5, 22)
[ 1] 0. 9266975

Therefore, ) 3 . ( ) (
* *
= p P p P
a
= 1 .9266975 = .0733025. Since the probability
associated with H
0
is much larger, our decision is to not reject H
0
.

16.23 From Ex. 6.17, the posterior distribution of p is beta(11, 10). Thus,
) 4 . ( ) (
*
0
*
< = p P p P is given by (in R):
> pbet a( . 4, 11, 10)
[ 1] 0. 1275212

Therefore, ) 4 . ( ) (
* *
= p P p P
a
= 1 .1275212 = .8724788. Since the probability
associated with H
a
is much larger, our decision is to reject H
0
.

16.24 From Ex. 16.18, the posterior distribution for is gamma(17.3, .0305). To test
H
0
: > .5 vs. H
a
: .5,
we calculate ) 5 . ( ) (
*
0
*
> = P P as:
Chapter 16: Introduction to Bayesian Methods of Inference 333
Instructors Solutions Manual

> 1 - pgamma( . 5, shape=17. 3, scal e=. 0305)
[ 1] 0. 5561767

Therefore, ) 5 . ( ) (
* *
= P P
a
= 1 .5561767 = .4438233. The probability
associated with H
0
is larger (but only marginally so), so our decision is to not reject H
0
.


16.25 From Ex. 16.19, the posterior distribution for is gamma(176, .0395). Thus,
) 6 ( ) (
*
0
*
> = P P is found by
> 1 - pgamma( 6, shape=176, scal e=. 0395)
[ 1] 0. 9700498

Therefore, ) 6 ( ) (
* *
= P P
a
= 1 .9700498 = .0299502. Since the probability
associated with H
0
is much larger, our decision is to not reject H
0
.

16.26 From Ex. 16.20, the posterior distribution for v is gamma(9, 1.0765). To test:
H
0
: v < 10 vs. H
a
: v 10,
we calculate ) 10 ( ) (
*
0
*
< = v P v P as
> pgamma( 10, 9, 1. 0765)
[ 1] 0. 7464786

Therefore, ) 10 ( ) (
* *
= v P P
a
= 1 .7464786 = .2535214. Since the probability
associated with H
0
is larger, our decision is to not reject H
0
.