Using pivots to construct confidence intervals

In Example 41 we used the fact that

X
Q(X, ) = N (0, 1) for all .
/ n

We then said |Q(X, )| z/2 with probabil-

ity 1 , and converted this into a statement
about .

Definition 21 Given a data vector X , a ran-

dom variable Q(X , ) is a pivotal quantity if
the distribution of Q(X , ) is independent of
all unknown parameters.

Suppose we want a (1 )100% confidence

interval for . Try to find a function of the data
that also depends on but whose probability
distribution does not depend on .

191
Ideally, the random variable Q(X , ) will de-
pend on X only through a sufficient statistic.

Let 1 and 2 be positive and such that 1 +

2 = . Let q1 and q2 be the 1 and 1 2
quantiles of the distribution of Q(X , ). Then

P (q1 < Q(X , ) q2) = 1 .

Now, try to find C(X ) such that

C(X ) iff q1 < Q(X , ) q2.

It then follows that C(X ) is a (1 )100%
confidence set for .

Example 42 Let X1, . . . , Xn be a random sam-

ple from the U (0, ) distribution. Would like
to construct a (1 )100% confidence interval
for .

192
What kind of transformation of the data would
make the distribution of the transformed data
free of ?
X1 Xn
,..., are i.i.d. U (0, 1).

It follows that Q(X , ) = max(X1/, . . . , Xn/)

has the same distribution as the maximum of
a random sample from the U (0, 1) distribution.
Therefore, Q(X , ) is a pivotal quantity. Note
that
X(n)
Q(X , ) = ,

and thus depends on the data through a suffi-
cient statistic.

Choose a and b so that

!
X(n)
P a< < b = 1 .

193
 ! Z b
X(n)
P a< <b = ntn1 dt = bn an.
a
Since
! !
X(n) X(n) X(n)
P a< <b =P << ,
b a

[X(n)/b, X(n)/a] is a (1 )100% confidence

interval for so long as 0 < a, b < 1 and bn
an = 1 .

It would be desirable to find an interval of

the above form that has the shortest possible
length. The length is

1 1
X(n) .
a b

We have no control over X(n), but we could

choose a and b to minimize a1 b1 subject
to the constraint bn an = 1 .

194
It is straightforward to show that the solution
to the previous problem is

b=1 and a = 1/n.

Therefore, among (1 )100% confidence in-

tervals of the form [X(n)/b, X(n)/a], the short-
est one is [X(n), X(n)/1/n].

Example 43 Let X1, . . . , Xn be a random sam-

ple from a density of the form

1 x
f ,

where f is known. The parameter space is

= {(, ) : < < , > 0}.

The family of densities

1 x
f : (, )

is called a location-scale family.
195
Define
n
1 X
X = Xi
n i=1
and
n
X
1
S2 = (Xi X)2.
n 1 i=1
The random variable (X )/S is a pivotal
quantity, a fact which leads to a confidence
interval for .

Proof: Consider

Xi = (Xi ) +
= (Xi )/ +
= Zi +
We then have X = Z, where
n
X
Z = n1 Zi.
i=1

196
 n
1 X
S2 = (Zi Z)2
n 1 i=1
= 2SZ
2.

It follows that
(X ) Z
= .
S SZ
It is easy to verify that Zi has density f , which
is free of any unknown parameters, and the
result follows.

The N (, 2) family of distributions is an ex-

ample of a location-scale family, with
1 y2/2
f (y) = e .
2
In this case
(X )
T =
S/ n
is a pivotal quantity having the Students t-
distribution with n 1 degrees of freedom.
197
Let tp,n1 be the (1 p) quantile of Students
t-distribution with n 1 degrees of freedom. A
(1 )100% confidence interval for is
" #
S S
X t2,n1 , X t11,n1 ,
n n
where 1 + 2 = .

For any location-scale or scale family, S 2/ 2

is a pivotal quantity. This fact leads to confi-
dence intervals for 2 (or ).

When X1, . . . , Xn are i.i.d. N (, 2),

(n 1)S 2 2
2
n1 .

Let a and b be such that P (a < 2

n1 < b) =
1 . Then
" #
(n 1)S 2 (n 1)S 2
,
b a
is a (1 )100% confidence interval for 2.
198
Shortest length confidence intervals

Subject to the confidence coefficient being 1

, we would like our confidence interval to be
of shortest length.

Example 43 (continued) X1, . . . , Xn i.i.d. N (,

2). The length of the previously derived con-
fidence interval for is
S
(t11,n1 + t2,n1).
n
Subject to the constraint 1 + 2 = , choose
1 and 2 to minimize

t2,n1 t11,n1.

199
Theorem 12 Let f be a unimodal pdf. If the
interval [a, b] satisfies

Rb
(i) a f (x) dx = 1

(ii) f (a) = f (b) > 0

(iii) a x b where x is the mode of f ,

then [a, b] is the shortest of all intervals that

satisfy (i).

Proof: See Casella and Berger, p. 442.

Continuing Example 43, since the t-distribution

is unimodal, we may apply Theorem 13. Now,
the t-distribution is symmetric, so t2,n1
t11,n1 is minimized subject to 1 + 2 =
when 1 = 2 = /2.
200
The shortest interval has the form
" #
S S
X t/2,n1 , X + t/2,n1 .
n n

In some situations we can minimize length di-

rectly.

Example 44 In Example 42 we had X1, . . . , Xn

i.i.d. U (0, ), and found that [X(n)/b, X(n)/a] is
a (1)100% confidence interval for for each
(a, b) such that bn an = 1 . Here, we cant
apply Theorem 12, but we could minimize the
length directly.

201