9.2.

2 Pivotal Quantities
We noted in the example about the uniform(0,  ) that the
coverage probability of the interval {aY,bY} did not depend
on  , while that of {Y+c,Y+d} did. This happened since the
former probability could be expressed in terms of Y/  ,
whose distribution does not depend on θ, a quantity
known as a pivotal quantity.

Definition 9.2.6 A r.v. Q(X,  ) = Q(X1,…,Xn,  ) is a

pivotal quantity if the distribution of Q(X,  ) is
independent of all parameters. That is, if
X∼F(x|  ), then Q(X,  ) has the same
distribution for all values of  .

Comment:
Note that even if the pdf of the random variable X depends
both on x and the parameter  , the density function of the
pivotal quantity Q(X,  ) does not depend on the parameter
.

1
Example 9.2.7 (Location-scale pivots)
Form of pdf Type of pdf Pivotal quantity
f(x – μ) Location X -μ
(1/σ)f(x/σ) Scale X /σ
(1/σ)f((x-μ)/σ) Location-scale ( X - μ)/S

Comments:
- There are lots of pivots in location and scale cases.
- If X1,…,Xn iid n(μ,σ2) then the t statistic ( X -μ)/(S/ n )
is a pivot because the t distribution does not depend μ
and σ2.
- How do we find pivots? Differences are pivotal for
location problems and ratios (or products) are pivotal
for scale problems.

2
Example 9.2.8 (Gamma pivot)
X1,…,Xn iid exponential(λ) ∼ (1/λ)e-(x/λ), λ > 0.
T = ΣXi ∼ gamma(n, λ) and is sufficient for λ.

The gamma pdf:

n 1
1 t 1 ( t /  )
f(t|n, λ) =   e
( n )    

Note that t and λ appear together as t/λ.

This pdf belongs to the scale family.
Let Q(T,λ) = 2T/λ. We use the transformation:
t = λq/2 and dt = λdq/2
Inserting into the pdf:
n 1
1 q 1 ( q / 2 )
g(q|n) =   e = gamma(n,2)
(n )  2  2

Q(T, λ) ∼ gamma(n, 2)
which does not depend on λ and is a pivot and ∼ χ2(2n) .

(Remember: gamma(p/2, 2) is χ2(p).)

3
If the pdf of a statistic T, f(t|  ), can be expressed in the
form

f(t|  ) = g(Q(t,  )) Q( t , )
t
for some function g and some monotone function Q
(monotone in t for each  ). Then it can be shown that
Q(t,  ) is a pivot.

Once we have a pivotal quantity, Q(X,  ), how can we

construct a confidence interval?

For a specified α we can determine constants a and b

which do not depend on θ to satisfy
Pθ(a ≤ Q(X,  ) ≤ b) ≥ 1 – α.
Then for each  0 ∈Θ,
A(  0) = {x: a≤ Q(X,  0)≤b}
is the acceptance region for a level α test of H0:  =  0.
We use the test inversion method to construct a 1 - α
confidence set:
C(x) = {  0: a≤ Q(X,  0)≤b}

4
Example 9.2.9 (Continuation of Example 9.2.8)

We have shown that Q(T,λ) = 2T/λ ∼  2n

2
.

If we choose two constants a and b to satisfy

P(𝑎≤  2n
2
≤b) = 1 – α,

then

Pλ  a  
2T
 b   Pλ(𝑎≤ Q(T,λ) ≤b) = P(𝑎≤  2n ≤b) = 1 – α
2
  
Inverting the set (given λ we accept H0 for t in the interval)
A(λ) = {t: 𝑎≤2t/λ ≤b}
gives (given t we believe that λ is in the interval)
C(t) = {λ: 2t/b ≤ λ≤2t/𝑎}

5
Example 9.2.10 (Normal Pivotal interval)
If X1,…,Xn are iid n(μ, σ2) it follows that ( X - μ)/(σ/ n ) is
a pivot.
If σ2 is known we can use this pivot to calculate a
confidence interval for μ.
For any constant a,

 X  
P  a   a  = P(-𝑎 ≤ Z ≤ 𝑎)
 / n 
Inverting this we have

   
 : x  a    xa 
 n n
Similarly, if σ is unknown we can use S and the Student t
distribution.

6
Suppose we want a confidence interval for σ.

Since (n-1)S2/σ2 ∼  n21 , (n-1)S2/σ2 is a pivotal quantity.

Consider constants 𝑎 and b such that

 ( n  1) S 2 

P a  b  = P(𝑎≤  n21 ≤ b) = 1 – α
 2 
 
Inverting this gives the 1 – α confidence interval
 2 (n  1) s 2 (n  1) s 2 
 :  
2

 b a 
or, equivalently,
 (n  1) s 2 (n  1) s 2 
 :   
 b a 

Comments:
- Usually one selects the constants 𝑎 and b such that
one has equal probabilities in the tails. This is,
however, not obvious to do, since the distribution is
not symmetric. (See Section 9.3).

- We have constructed intervals for μ and σ separately.

An alternative would be to find a confidence region
for (μ, σ) instead (Exercise 9.14).

Exercises: 9.12, 9.13

7
 p 3 2 1 0
0.235 0.0130 0.1267 0.4126 0.4477
0.236 0.0131 0.1277 0.4133 0.4459
0.237 0.0133 0.1286 0.4139 0.4442
0.238 0.0135 0.1295 0.4146 0.4425
0.239 0.0137 0.1304 0.4152 0.4407
0.240 0.0138 0.1313 0.4159 0.4390
0.241 0.0140 0.1323 0.4165 0.4372
0.242 0.0142 0.1332 0.4171 0.4355
0.243 0.0143 0.1341 0.4178 0.4338
0.244 0.0145 0.1350 0.4184 0.4321
0.245 0.0147 0.1360 0.4190 0.4304
0.246 0.0149 0.1369 0.4196 0.4287
0.247 0.0151 0.1378 0.4202 0.4270
0.248 0.0153 0.1388 0.4207 0.4253
0.249 0.0154 0.1397 0.4213 0.4236
0.250 0.0156 0.1406 0.4219 0.4219
0.251 0.0158 0.1416 0.4224 0.4202
0.252 0.0160 0.1425 0.4230 0.4185
0.253 0.0162 0.1434 0.4235 0.4168
0.254 0.0164 0.1444 0.4241 0.4152
0.255 0.0166 0.1453 0.4246 0.4135
----- ----- ----- ----- -----
0.273 0.0203 0.1625 0.4329 0.3842
0.274 0.0206 0.1635 0.4333 0.3827
0.275 0.0208 0.1645 0.4336 0.3811
0.276 0.0210 0.1655 0.4340 0.3795
0.277 0.0213 0.1664 0.4344 0.3779
0.278 0.0215 0.1674 0.4348 0.3764
0.279 0.0217 0.1684 0.4351 0.3748
0.280 0.0220 0.1693 0.4355 0.3732
0.281 0.0222 0.1703 0.4358 0.3717
0.282 0.0224 0.1713 0.4361 0.3701
0.283 0.0227 0.1723 0.4365 0.3686
0.284 0.0229 0.1732 0.4368 0.3671
0.285 0.0231 0.1742 0.4371 0.3655
0.286 0.0234 0.1752 0.4374 0.3640
0.287 0.0236 0.1762 0.4377 0.3625
0.288 0.0239 0.1772 0.4380 0.3609
0.289 0.0241 0.1782 0.4383 0.3594
0.290 0.0244 0.1791 0.4386 0.3579
0.291 0.0246 0.1801 0.4388 0.3564
0.292 0.0249 0.1811 0.4391 0.3549
0.293 0.0252 0.1821 0.4394 0.3534
0.294 0.0254 0.1831 0.4396 0.3519

8
0.298 0.0265 0.1870 0.4406 0.3459
0.299 0.0267 0.1880 0.4408 0.3445
0.300 0.0270 0.1890 0.4410 0.3430
0.301 0.0273 0.1900 0.4412 0.3415
0.302 0.0275 0.1910 0.4414 0.3401
0.303 0.0278 0.1920 0.4416 0.3386
0.304 0.0281 0.1930 0.4418 0.3372
0.305 0.0284 0.1940 0.4420 0.3357
0.306 0.0287 0.1950 0.4421 0.3343
0.307 0.0289 0.1959 0.4423 0.3328
0.308 0.0292 0.1969 0.4425 0.3314
0.309 0.0295 0.1979 0.4426 0.3299
----- ----- --- ----- -----
0.361 0.0470 0.2498 0.4422 0.2609
0.362 0.0474 0.2508 0.4420 0.2597
0.363 0.0478 0.2518 0.4419 0.2585
0.364 0.0482 0.2528 0.4417 0.2573
0.365 0.0486 0.2538 0.4415 0.2560
0.366 0.0490 0.2548 0.4413 0.2548
0.367 0.0494 0.2558 0.4412 0.2536
0.368 0.0498 0.2568 0.4410 0.2524
0.369 0.0502 0.2578 0.4408 0.2512
0.370 0.0507 0.2587 0.4406 0.2500
----- ----- ----- ----- -----
0.630 0.2500 0.4406 0.2587 0.0507
0.631 0.2512 0.4408 0.2578 0.0502
0.632 0.2524 0.4410 0.2568 0.0498
0.633 0.2536 0.4412 0.2558 0.0494
0.634 0.2548 0.4413 0.2548 0.0490
0.635 0.2560 0.4415 0.2538 0.0486
0.636 0.2573 0.4417 0.2528 0.0482
0.637 0.2585 0.4419 0.2518 0.0478
0.638 0.2597 0.4420 0.2508 0.0474
0.639 0.2609 0.4422 0.2498 0.0470
0.640 0.2621 0.4424 0.2488 0.0467
----- ----- ----- ----- -----
0.693 0.3328 0.4423 0.1959 0.0289
0.694 0.3343 0.4421 0.1950 0.0287
0.695 0.3357 0.4420 0.1940 0.0284
0.696 0.3372 0.4418 0.1930 0.0281
0.697 0.3386 0.4416 0.1920 0.0278

0.760 0.4390 0.4159 0.1313 0.0138

0.761 0.4407 0.4152 0.1304 0.0137
0.762 0.4425 0.4146 0.1295 0.0135
0.763 0.4442 0.4139 0.1286 0.0133

9
Table 9.2.2 Acceptance region and confidence set for
Sterne´s construction, X ∼ binomial(3,p) and 1 - α = .442

p Accept. region x Confidence set

[.000,.238] {0}
(.238,.305) {0,1} 0 [.000,.305)⋃(.362,.366)
[.305,.362] {1}
(.362,.366) {0,1} 1 (.238,.634)
[.366,.634] {1,2}
(.634,.638) {2,3} 2 [.366,.762)
[.638,.695] {2}
(.695,.762) {2,3} 3 (.634,.638) ⋃(.695,1.00)
[.762,1.00] {3}

10
9.2.3 Pivoting the CDF
We have seen that a pivot, Q, leads to a confidence set of
the form
C(x) = {  0 : a ≤ Q(x,  0) ≤ b}.

If, for every x, Q is a monotone function of  , then the

confidence set C(x) is guaranteed to be an interval.

The examples so far, using location and scale

transformations, resulted in monotone Q functions, and
hence, confidence intervals.

11
Now assume our CI construction for a parameter  is
based on a statistic T with cfd FT(t|  ).

Assume T to be a continuous variable.

FT(T|  ) is a random variable that is uniform(0,1) and is a

pivotal quantity.

If α1 + α2 = α, then an α-level acceptance region of the

hypothesis H0:  =  0 is

A(  0) = {t: α1 ≤ FT(t|  0) ≤ 1 - α2},

with associated confidence set

C(t) = {  : α1 ≤ FT(t|  ) ≤ 1 - α2}.

To guarantee that the confidence set is an interval, we

need to have FT(t|  ) to be monotone in  , either
increasing or decreasing.

12
Theorem 9.2.12 (Pivoting a continuous cdf)
Let T be a statistic with continuous cdf FT(t|  ).
Let α1 + α2 = α with 0 < α < 1 be fixed values. Suppose
that for each t ∊ 𝓣, the functions  L(t) and  U(t) can be
defined as follows.

i. If FT(t|  ) is a decreasing function of  for each t,

define  L(t) and  U(t) by

FT(t|  U(t)) = α1, FT(t|  L(t)) = 1 - α2.

ii. If FT(t|  ) is an increasing function of  for each t,

define  L(t) and  U(t) by
FT(t|  U(t)) = 1 - α2, FT(t|  L(t)) = α1.

Then the random interval [  L(T),  U(T)] is a 1 – α

confidence interval for  .

Proof: Part i. Assume we have an acceptance region

{t: α1 ≤ FT(t|  0) ≤ 1 - α2}.
Since FT(t|  0) is a decreasing function of  for each t, and
1-α2 > α1,  L(t) <  U(t) and the values  L(t) and  U(t) are
unique. Also, (show a picture!)
FT(t|  )< α1 ⟺  >  U(t)
FT(t|  )> 1 - α2 ⟺  <  L(t)
and hence
{  : α1 ≤ FT(t|  ) ≤ 1 - α2} = {  : c≤  ≤  U(t)}.

13
Example 9.2.13 (Location exponential interval)
Let X1,…,Xn be iid from f(x|μ) = e-(x-μ)I[μ,∞)(x).

Then Y = min{ X1,…,Xn} is sufficient for μ with pdf

fY(y| μ) = ne-n(y-μ) I[μ,∞)(y).

Here we can find the cdf to be

FY(y| μ) = 1 – e-n(y-μ).
Since this is a decreasing function of μ for each y, we use
part i. of the theorem:
FY(y| μU(y) ) = 1 - e- n(y - u ( y )) = α/2
FY(y| μL(y) ) = 1 - e-n(y - L ( y)) = 1 - α/2
We obtain
μU(y) = y + (1/n)log(1 – α/2)
μL(y) = y + (1/n)log(α/2)

Hence the random interval

 1  1   
C(Y) =  : Y  log     Y  log1  
 n 2 n  2 

is a 1 – α confidence interval for μ.

14
Theorem 9.2.14 (Pivoting a discrete cdf)
Let T be a discrete statistic with cdf FT(t|  ) = P(T≤t|  ).
Let α1 + α2 = α with 0 < α < 1 be fixed values. Suppose
that for each t∊𝓣,  L(t) and  U(t) can be defined as follows.

i. If FT(t|  ) is a decreasing function of  for each t,

define  L(t) and  U(t) by

P(T≤t|  U(t)) = α1, P(T≥t|  L(t)) = α2.

ii. If FT(t|  ) is an increasing function of  for each t,

define  L(t) and  U(t) by
P(T≥t|  U(t)) = α1, P(T≤t|  L(t)) = α2.

Then the random interval [  L(T),  U(T)] is a 1 – α

confidence interval for  .

15
Example 9.2.15 (Poisson interval estimator)

Let X1,…,Xn be iid from Poisson(λ).

Let Y = ΣXi.
Y is sufficient for λ and Y ∼ Poisson(nλ).
Let α1 = α2 = α/2.
If Y = y0 is observed we should solve for λ in the equations:

y0 ( n ) k  ( n ) k
 n  n
e = α/2 and  e = α/2.
k 0 k! k  y0 k!

Link between Poisson and gamma distributions gives

  n ( n  )
 
y0 k
= e = P(Y≤y0|λ) = P  22( y0 1)  2n
2 k 0 k!
Thus we take
1 2
λ=  2( y0 1), / 2
2n
Similarly we have
 
 n ( n )
 
k
=  e = P(Y≥y0|λ) = P  22 y0  2n
2 k  y0 k!
Thus
 1 2 1 2 
 :  2 y0 ,1 / 2     2( y0 1), / 2 
 2n 2n 

16