Solutions to Selected Exercises from Chapter 11

Bain & Engelhardt - Second Edition

Andreas Alfons and Hanno Reuvers
Erasmus School of Economics, Erasmus Universiteit Rotterdam

Exercise 1  
i.i.d. 2 X̄−µ
If X1 , . . . , Xn ∼ N(µ, σ 2 ), then X̄ ∼ N µ, σn . This implies that √
σ/ n
∼ N(0, 1) is a pivotal
quantity. This pivotal quantity is used in parts (a)-(c).
(a) We have !
X̄ − µ
P −z1− α2 < < z1− α2 = 1 − α,
√σ
n

and thus also  

σ σ
P X̄ − z1− α2 √ < µ < X̄ + z1− α2 √ = 1 − α.
n n
With z1− α2 = z0.95 = 1.645 (see Table 3), a 90% confidence interval for µ is
   
σ σ 3 3
x̄ − z1− α2 √ , x̄ + z1− α2 √ = 19.3 − 1.645 √ , 19.3 + 1.645 √ = (18.067, 20.534).
n n 16 16

(b) By similar steps as in part (a), we have

! !
X̄ − µ X̄ − µ
P σ < z1−α = 1 − α, P −z1−α < = 1 − α,
√ √σ
n n
   
σ σ
P X̄ − z1−α √ < µ = 1 − α, P µ < X̄ + z1−α √ = 1 − α.
n n

With z1−α = z0.90 = 1.282 (see Table 3), one-sided 90% confidence limits for µ are
σ 3
l(x1 , . . . , xn ) = x̄ − z1−α √ = 19.3 − 1.282 √ = 18.339,
n 16
σ 3
u(x1 , . . . , xn ) = x̄ + z1−α √ = 19.3 + 1.282 √ = 20.262.
n 16

(c) The length of the confidence interval is 2z1− α2 √σn . We need

2
2z1− α2 σ

σ 1 λ
2z 1− α √ ≤λ ⇒ √ ≤ ⇒ n≥ .
2
n n 2z1− α2 σ λ

For the given numerical values, this evaluates to a required sample size of n ≥ ( 2·1.645·3
2 )2 =
24.354. We round to n = 25.

1
 X̄−µ
(d) The pivotal quantity √
s/ n
∼ t(n − 1) yields
 
X̄ − µ
P −t1− α2 < √ < t1− α2 = 1 − α,
s/ n
and  
s s
P X̄ − t1− α2 √ < µ < X̄ + t1− α2 √ = 1 − α.
n n
With t1− α2 (n − 1) = t0.95 (15) = 1.753 (see Table 6), a 90% confidence interval for µ is
  r r !
s s 10.24 10.24
x̄ − t1− α2 √ , x̄ + t1− α2 √ = 19.3 − 1.753 , 19.3 + 1.753
n n 16 16
= (17.898, 20.702).

(n−1)S 2
(e) The pivotal quantity σ2 ∼ χ2 (n − 1) yields
(n − 1)S 2
 
2
P χ α2 < < χ21− α2 = 1 − α,
σ2
and !
(n − 1)S 2 (n − 1)S 2
P 2 < σ2 < = 1 − α.
χ1− α χ2α
2 2

With χ2α (n − 1) = χ20.005 (15) = 4.60 and χ21− α (n − 1) = χ20.995 (15) = 32.80 (see Table 4),
2 2
a 99% confidence interval for σ 2 is obtained as
! 
(n − 1)s2 (n − 1)s2

15 · 10.24 15 · 10.24
, = , = (4.683, 33.391)
χ21− α χ2α 32.80 4.60
2 2

Exercise 3
   
2nX̄ 2nX̄
(a) The pivotal quantity θ ∼ χ2 (2n) yields P θ < χ2γ = γ and P 2n
χ
X̄
2 < θ = γ. With
γ

χ2γ (2n) = χ20.95 (100) = 124.34 (see Table 4), a one-sided lower 95% confidence limit for θ
is obtained as
2nx̄ 2 · 50 · 17.9
l(x1 , . . . , xn ) = 2 = = 14.396.
χγ 124.34

(b) Note that e−t/θ is a monotone increasing transformation of θ. This implies that a lower
confidence limit for θ can be transformed into a lower confidence limit for e−t/θ . The details
are as follows:
 

− θt
 t
0.95 = P l(X1 , . . . , Xn ) < e = P ln l(X1 , . . . , Xn ) < −
θ
 
t
=P − <θ .
ln l(X1 , . . . , Xn )
In part (a) we have found the lower bound of 14.396, hence − ln l(x1t,...,xn ) = 14.396. We
t t
find l(x1 , . . . , xn ) = e− 14.396 as a one-sided lower 95% confidence limit for P(X > t) = e− θ .

Note: The exercise states ‘where t is an arbitrary known value’. Note however that the
choice t = 0 should be excluded.

2
Exercise 5
(a) The pdf of theREXP(1, η) distribution is f (x; η) = e−(x−η) and we can integrate to find the
x
cdf F (x; η) = η e−(t−η) dt = 1 − e−(x−η) for x > η (and zero otherwise). The pdf for the
minimum is thus
h in−1
fX1:n (x; η) = n[1 − F (x; η)]n−1 f (x; η) = n e−(x−η) e−(x−η) = ne−n(x−η) , x > η.

It is clear from the form of this pdf that η is a location parameter. The transformation
Q = X1:n − η (with inverse transformation X1:n = Q + η) yields

fQ (x) = fX1:n −η (x; η) = ne−nx , x > 0.

We see that Q ∼ EXP(1/n). This distribution does not depend on η and Q is thus a pivotal
quantity.
(b) An 100γ% equal tailed confidence interval is given by P (q1 < Q < q2 ) = γ where the
quantiles q1 and q2 should satify P (Q ≤ q1 ) = FQ (q1 ) = 1−γ and P (Q ≥ q2 ) = 1−FQ (q2 ) =
1−γ
R x −nt2
2 . An explicit calculation of the cdf, FQ (x) = 0
ne dt = 1 − e−nx , leads to

1−γ 1−γ
1 − e−nq1 = e−nq2 =
2   2  
1 1+γ 1 1−γ
q1 = − ln q2 = − ln
n 2 n 2

Finally,
   !
1 1+γ 1 1−γ
P − ln < X1:n − η < − ln =γ
n 2 n 2
   !
1 1−γ 1 1+γ
P X1:n + ln < η < X1:n + ln =γ
n 2 n 2
such that    !
1 1−γ 1 1+γ
x1:n + ln , x1:n + ln
n 2 n 2
is a 100γ% equal tailed confidence interval for η.
(c) It should be understood from the exercise (although this is not very clear) that the mileages
are EXP(θ, η) distributed. If X ∼ EXP(θ, η), then Y = X/θ (θ > 0) has the pdf
1 − (θy−η) η η
fY (y) = fX (θy; θ, η) |θ| = e θ θ = e−(y− θ ) , y> .
θ θ
η
This is the pdf of an EXP(1, η ∗ ) distribution where η ∗ = θ. We can use the result from
part (b) to derive
!
1 1 − γ  η 1 1 + γ 
P Y1:n + ln < < Y1:n + ln = γ,
n 2 θ n 2
!
θ 1 − γ  θ 1 + γ 
P X1:n + ln < η < X1:n + ln = γ.
n 2 n 2

3
 A 90% confidence interval for η is obtained as
   
θ 1 − γ  θ 1 + γ  850 850
x1:n + ln , x1:n + ln = 162 + ln(0.05), 162 + ln(0.95)
n 2 n 2 19 19
= (27.980, 159.705).

Exercise 7
(a) We need to find the distribution of Y = X 2√when X ∼ WEI(θ, 2). The transformation
Y = X 2 has the inverse transformation X = Y such that the pdf of Y is given by
 √ 2
√ 1 2√ − y
1 1 −y
fY (y; θ) = fX ( y ; θ) √ = ye θ
√ = 2 e θ2 , y > 0.
2 y θ2 2 y θ

We conclude that Y ∼ EXP(θ2 ). Using the distributional result from Example 11.2.1, we
n
Xi2
P
2 2nȲ
have i=1
θ2 = θ2 ∼ χ2 (2n).
Pn
2 Xi2
(b) From i=1
θ2 ∼ χ2 (2n), we obtain
v v P 
Pn u Pn
2 i=1 Xi2 u 2 i=1 Xi2 u 2 ni=1 Xi2
  u
2 2
P χ 1−γ < < χ 1+γ = γ ⇒ P t <θ<t  = γ.
2 θ2 2 χ21+γ χ21−γ
2 2

!
r Pn r Pn
2 x2i 2 x2i
A 100γ% confidence interval for θ is i=1
χ21+γ
, i=1
χ21−γ
.
2 2

 
(c) Note that exp − (t/θ)2 is an increasing function in θ2 . A lower confidence
 limit2 for θ
2

can thus be manipulated into a lower confidence limit for P(X > t) = exp − (t/θ) . We
find this lower confidence limit from
 Pn Pn
2 i=1 Xi2 θ2 2 i=1 Xi2
    
2 1 2
γ=P < χγ = P n > = P θ > .
2 i=1 Xi2
P
θ2 χ2γ χ2γ

The remaining steps of the calculation are as follow

! ! !
1 χ2γ −t2 −t2 χ2γ  2
  −t2 χ2 
γ
γ=P < Pn =P > Pn =P exp −(t/θ) > exp Pn ,
θ2 2 i=1 Xi2 θ2 2 i=1 Xi2 2 i=1 Xi2

where we used t > 0 (the case t = 0 should be excluded becauseP(X > 0)= 1). A lower
  −t2 χ2
100γ% confidence limit for exp − (t/θ)2 is l(x1 , . . . , xn ) = exp 2 Pn γx2 .
i=1 i

th
(d) We first need to compute the p percentile for the given Weibull
distribution. If we denote
p
this percentile by xp , then xp satisfies the equation P X ≤ xp = 100 . With
Z x Z x  2 x
2 −( θt )2 2t 2
− θt 2 − θt 2 x2
F (x; θ) = 2
te dt = − − 2
e dt = −e = 1 − e− θ 2 , x > 0,
0 θ 0 θ 0

the pth percentile is obtained as

x2
r
− θ2p p  p 
1−e = ⇒ xp = −θ2 ln 1 − .
100 100

4
 q
p


The expression −θ2 ln 1 − 100 is again an increasing function in θ2 . An upper confi-
dence limit for θ2 will thus imply an upper confidence limit for the pth percentile of the
distribution. We have
Pn Pn !
2 2
 
2 X
i=1 i 2 X
i=1 i
γ = P χ21−γ < = P θ2 <
θ2 χ21−γ

p


and by noting that − ln 1 − 100 is a positive quantity
! s !
 −2 ln 1 − p
Pn X 2 p
−2 ln 1 − 100

Pn 2
p i=1 Xi

2 100 i=1 i
γ = P −θ ln 1 − < = P xp < .
100 χ21−γ χ21−γ
r
p Pn
th −2 ln(1− 100 ) i=1 x2i
An upper 100γ% confidence limit for the p percentile is thus χ21−γ
.

Exercise 11
The setting corresponds to a random sample from the BIN(1, p) distribution. If X ∼ BIN(1, p),
then E(X) = p and Var(X) = p(1 − p). The CLT implies that
√
n(X̄ − p) d
p −
→ Z ∼ N(0, 1).
p(1 − p)
√
d
Now note that p̂ = X̄ is a consistent estimator for p such that also √n(X̄−p) −
→ Z ∼ N(0, 1).
p̂(1−p̂)
Hence, for large n, we find
√ !
n(p̂ − p)
P −z1− α2 < p < z1− α2 ≈ 1 − α,
p̂(1 − p̂)
r r !
p̂(1 − p̂) p̂(1 − p̂)
P p̂ − z1− 2
α < p < p̂ + z1− 2
α ≈ 1 − α.
n n

5
With p̂ = 40 = 18 and z1− α2 = z0.95 = 1.645 (see Table 3), an approximate 90% confidence
interval for p is obtained as
r r !  s s 
1 7 1 7
p̂(1 − p̂) p̂(1 − p̂) 1 · 1 ·
p̂ − z1− α2 , p̂ + z1− α2 =  − 1.645 8 8 , + 1.645 8 8 
n n 8 40 8 40

= (0.039, 0.211).

Exercise 12
 
X̄−µ
√
(a) Equation (11.3.20) implies that P ≈ γ for large n. We need to manipulate
µ < zγ
n
√
the inequality inside the probability. Having this in mind, we define θ = µ, such that
! !
X̄ − θ2
 
X̄ − µ 2 zγ
γ≈P p µ < zγ = P < zγ = P θ + √ θ − X̄ > 0 .
√θ n
n n

5
 q 2
z z z
The solutions of the quadratic equation θ2 + √γn θ − X̄ = 0 are θ1 = − 2√γn − 4nγ + X̄ and
q 2
z z
θ2 = − 2√γn + 4nγ + X̄. Since θ > 0 and θ1 < 0, it always holds that θ − θ1 > 0, and the
above can be written as
 
γ ≈ P (θ − θ1 )(θ − θ2 ) > 0 = P(θ − θ2 > 0) = P(θ2 < θ) = P(θ22 < µ)
 r !2 
zγ zγ2
= P − √ + + X̄ < µ .
2 n 4n

With zγ = z0.90 = 1.282 (see Table 3), an approximate one-sided lower 90% confidence
limit for µ is obtained as
r !2 r !2
zγ zγ2 1.282 1.2822
l(x1 , . . . , xn ) = − √ + + x̄ = − √ + + 1.7 = 1.468.
2 n 4n 2 45 4 · 45

(b) For large n, Equation (11.3.21) yields

  r !
X̄ − µ X̄
γ≈P  q < zγ = P X̄ − zγ
 <µ .
X̄ n
n

With zγ = z0.90 = 1.282 (see Table 3), an approximate one-sided lower 90% confidence
limit for µ is obtained as
r r
x̄ 1.7
l(x1 , . . . , xn ) = x̄ − zγ = 1.7 − 1.282 = 1.451
n 45

Exercise 19  2

For the pdf of the N(µ1 , σ12 ) distribution with µ1 known, we have f (x; σ12 ) = (2πσ12 )−1/2 exp − 12 (x−µ σ12
1)
.
2
Pn1 2
This pdf is a member of the RECPwith t(x) = (x − µ1 ) . S1 = i=1 (Xi − µ1 ) is thus a sufficient
n2
statistic for σ12 . Similarly, S2 = j=1 (Yj − µ2 )2 is a sufficient statistic for σ22 .
The mean and standard deviation of the normal distribution are location-scale parameters.
X −µ Y −µ
It is therefore easily shown that X1σ−µ 1
1
, . . . , nσ1 1 1 ∼ N(0, 1) and Y1σ−µ 2
2
, . . . , n2σ2 2 ∼ N(0, 1)
n1 Xi −µ1 2 n2 Y −µ
and this implies both Sσ21 = i=1 ( σ1 ) ∼ χ2 (n1 ) and Sσ22 = j=1 ( jσ2 2 )2 ∼ χ2 (n2 ). By
P P
1 2
taking ratios and rescaling we can find the pivotal quantity:
 
S1
n2 σ22 S1 σ12
/n1 d χ2 (n )/n
1 1
= = 2 ∼ F (n1 , n2 ),
n1 σ12 S2
 
S2
/n χ (n 2 )/n 2
2σ2 2

d
where = is used to denote equivalence in distribution. Denoting the α quantile of the F (n1 , n2 )
distribution by fα , we obtain
n2 S1 σ22 σ22
   
n1 S2 n1 S2
P f α2 < < f1− α = 1 − α ⇒ P f α < < f1− α = 1 − α.
n1 S2 σ12 2 2
n2 S1 σ12 2
n2 S1
σ2
 
A 100(1 − α)% confidence interval for σ22 is f α2 nn12 ss21 , f1− α2 nn12 ss21 .
1