Southern University of Science and Technology

Department of Statistics and Data Science

MA204: Mathematical Statistics Final Examination (Paper A)

Date: 9 June 2022 Time: 7:00 p.m. – 10:00 p.m.

Answer ALL 5 questions. Marks are shown in square brackets.

1. Give your answers to the following questions:

1.1 Let the discrete random variable (r.v.) X ∼ Bernoulli(p) with p ∈ (0, 1)
and r be a positive integer, what is the distribution of X r ? [2 ms]
1.2 Let the probability density function (pdf) of the continuous r.v. X ∼
Exponential(β) be f (x; β) = βe−βx , where x > 0 and β > 0, then the
median of X, med(X) = . [2 ms]
iid
1.3 Let X1 , . . . , Xn ∼ U (0, 1), what is the distribution of the smallest order
statistic X(1) = min(X1 , . . . , Xn )? [2 ms]
1.4 The moment generating function (mgf) of the Poisson r.v. X ∼ Poisson(λ)
is . [Hint: The probability mass function (pmf) of X ∼ Poisson(λ)
x −λ
is λ e /x!, x = 0, 1, . . . , ∞] [2 ms]
1.5 Let E(X) = µ and Var(X) = σ 2 , then the correlation coefficient of X
and −X is . [2 ms]
iid
1.6 Let X1 , . . . , Xn ∼ Bernoulli(θ) with θ > 0, and the prior distribution of θ
be Beta(α, β) with α, β > 0, then the posterior distribution of θ is
and the Bayesian estimator of θ is . [4 ms]
1.7 Let X ∼ F (ν1 , ν2 ) (i.e., F distribution with ν1 and ν2 degrees of freedom),
then the distribution of X −1 is . [2 ms]
1.8 An efficient estimator of the parameter θ is the uniformly minimum vari-
ance unbiased estimator (UMVUE) of θ, is it true? [2 ms]
2

iid
1.9 Let X1 , . . . , Xn ∼ N (µ, σ02 ) with known σ02 > 0, then the Fisher informa-
tion In (µ) = and the UMVUE of µ is . [4 ms]
1.10 State the definition of a sufficient statistic of the parameter θ. [2 ms]
iid n
ˆ (X1 , . . . , Xn )> and T (x) =
P
1.11 Let X1 , . . . , Xn ∼ Poisson(θ), x = ˆ i=1 Xi ,
then the conditional distribution

Pr{X1 = x1 , . . . , Xn = xn |T (x) = t} = . [2 ms]

1.12 State the definition of a pivotal quantity. [2 ms]

iid
1.13 Let X1 , . . . , Xn ∼ N (µ0 , σ 2 ), where µ0 is known. A sufficient statistic of
σ 2 is and a pivotal quantity of σ 2 is . [4 ms]
1.14 The definition of the Type I error function α(θ) is . [2 ms]
1.15 How to compare two given tests T1 and T2 ? In other words, under what
kind of conditions, we say that T1 is better than T2 . [2 ms]
1.16 How to compare two given estimators θ̂1 and θ̂2 of the parameter θ? In
other words, under what kind of conditions, we say that θ̂1 is better than
θ̂2 . [2 ms]
1.17 Given the null hypothesis H0 , the test statistic T and its observed value
tobs , what is the definition of the p-value for testing H0 ? [2 ms]

[Total: 40 ms]

2. Let X be a discrete r.v. with pmf pi = Pr(X = xi ) for i = 1, 2, 3 and Y be a

discrete r.v. with pmf qj = Pr(Y = yj ) for j = 1, 2, 3, 4. Given two conditional
distribution matrices
 
1/7 1/4 3/7 1/7
A =  2/7 1/2 1/7 2/7 
4/7 1/4 3/7 4/7
and  
1/6 1/6 1/2 1/6
B =  2/7 2/7 1/7 2/7  ,
1/3 1/12 1/4 1/3
 3

where the (i, j) element of A is aij = Pr(X = xi |Y = yj ) and the (i, j) element
of B is bij = Pr(Y = yj |X = xi ).

2.1 Find the marginal distributions of X and Y . [6 ms]

2.2 Find the joint distribution of (X, Y ). [3 ms]

[Total: 9 ms]
iid
3. Let X1 , . . . , Xn ∼ Beta(θ, 1) with pdf f (x; θ) = θxθ−1 I(0 < x < 1), where
θ > 0.

3.1 Find a sufficient statistic of θ. [3 ms]

3.2 Find the MLE of the mean parameter µ = θ/(θ + 1). [3 ms]
3.3 Prove that the cdf of the population r.v. X ∼ f (x; θ) is F (x; θ) = xθ for
x ∈ (0, 1). [3 ms]
Pn
ˆ −
3.4 Define T = log Xi . Prove that T /n is an unbiased estimator of
i=1
τ (θ) = 1/θ. [Hint: −2 ni=1 log F (Xi ; θ) ∼ χ2 (2n)]
P
[3 ms]
3.5 Find the Fisher information In (θ). [3 ms]
3.6 Prove that T /n is an efficient estimator of τ (θ); hence T /n is the unique
UMVUE for τ (θ). [3 ms]
3.7 Find the 100(1 − α)% equal-tail confidence interval (CI) of θ. [3 ms]

[Total: 21 ms]
iid
4. Let X1 , . . . , Xn ∼ N (µ0 , σ 2 ), where µ0 is known and σ 2 is unknown.

4.1 Using the Neyman–Pearson lemma, to find a most powerful test (MPT)
of size α for testing the simple null hypothesis H0 : σ 2 = σ02 against the
simple alternative hypothesis H1 : σ 2 = σ12 , where σ12 > σ02 . [10 ms]
4.2 Find a uniformly most powerful test (UMPT) of size α for testing the null
hypothesis H0 : σ 2 6 σ02 against H1 : σ 2 > σ02 . [10 ms]
4

[Total: 20 ms]

5. In the 98 year period from 1900 to 1997, there were 159 U.S. land falling
hurricanes. The numbers of hurricanes per year are summarized as follows:

Times of hurricanes per year (i) 0 1 2 3 4 5 6 Total

[2mm] Frequency of years (Ni ) 18 34 24 16 3 1 2 98

Does the number of land falling hurricanes per year follow a Poisson distribution
when the approximate significance level is taken to be 0.05? [χ2 (0.05, 3) = 7.81,
χ2 (0.05, 4) = 9.49, χ2 (0.05, 5) = 11.07, where χ2 (α, ν) denotes the upper α-th
quantile of χ2 (ν)]
[Total: 10 ms]

********** END OF THE PAPER A **********

 5

1. Suggested Solutions.

d
1.1 X r = X ∼ Bernoulli(p).

1.2 med(X) = (1/β) log 2.

Solution: Let F (x) denote the cdf of X, then F (x) = 1 − e−βx . From 0.5 =
F (med(X)), we have med(X) = (1/β) log 2. k

1.3 X(1) ∼ Beta(1, n) with pdf n(1 − x)n−1 · I(0 < x < 1).
Solution: Let X ∼ U (0, 1), then the pdf and cdf of X are given by
f (x) = 1 · I(0 < x < 1)
and
F (x) = 0 · I(x 6 0) + x · I(0 < x < 1) + 1 · I(x > 1).
So the pdf of X(1) is
g1 (x) = nf (x){1 − F (x)}n−1 = n(1 − x)n−1 · I(0 < x < 1). k

1.4 The mgf of the Poisson random variable is exp{λ(et − 1)}.

1.5 Corr(X, −X) = −1.

Solution: Note that E(−X) = −µ, Var(−X) = σ 2 and
E(−X 2 ) = −E(X 2 ) = −{Var(X) + [E(X)]2 } = −σ 2 − µ2 .
According to the definition of correlation coefficient
Cov(X, −X)
Corr(X, −X) = p
Var(X) · Var(−X)
E(−X 2 ) − E(X)E(−X) (−σ 2 − µ2 ) + µ2
= = = −1. k
σ2 σ2
6
Pn
1.6 θ|x ∼ Beta(α+x+ , β +n−x+ ) with x+ = i=1 xi . The Bayesian estimator
of θ is (α + nX̄)/(α + β + n).
Solution: Let x = (x1 , . . . , xn )> be the realizations of x = (X1 , . . . , Xn )>. See
Example 3.11 in page 123. k

1.7 1/X ∼ F (ν2 , ν1 ).

1.8 Yes, it is true.

1.9 In (µ) = n/σ02 and X̄ is the UMVUE of µ.

Solution: See Example 3.18 in pages 135–136. k

1.10 A statistic T (x) is said to be a sufficient statistic of θ if the conditional

distribution of x, given T (x) = t, does not depend on θ for any value of t. In
discrete case, this means that

Pr{X1 = x1 , . . . , Xn = xn ; θ|T (x) = t} = h(x)

does not depend on θ. k

1.11 See Example 3.21 in page 140:

t! 1
Pn−1 · t. k
x1 ! · · · xn−1 !(t − i=1 xi )! n

iid
1.12 Assume that X1 , . . . , Xn ∼ f (x; θ) and T = T (X1 , . . . , Xn ) is a sufficient
statistic of θ. Let P = P (T, θ) be a function of T and θ. If the distribution of
P does not depend on θ, then P is called a pivotal quantity or pivot. k
 7
Pn
1.13 T (x) =
ˆ i=1 (Xi − µ0 )2 is a sufficient statistic of σ 2 . Then
n  2
X Xi − µ0
P = ∼ χ2 (n)
i=1
σ

is a pivotal quantity of σ 2 . k

1.14 The Type I error function is defined as

α(θ) = Pr(Type I error) = Pr(rejecting H0 |H0 is true)

= Pr(x ∈ C|θ ∈ Θ0 ),

which is a function of θ defined in Θ0 . k

1.15 If αT1 (θ), αT2 (θ) 6 α∗ and βT1 (θ) 6 βT2 (θ), then T1 is better than T2 ,
where α∗ (0 < α∗ < 1) is a preassigned (small) level.
Alternatively, if pT1 (θ) > pT2 (θ), then T1 is better than T2 . k

1.16 If MSE(θ̂1 ) 6 MSE(θ̂2 ), then θ̂1 is better than θ̂2 . k

1.17 The p-value (or probability value) is defined as the probability, under
the null hypothesis H0 , the test statistic T is equal to or more extreme than
the observed value tobs ; i.e.,

Pr(T is equal to or more extreme than the observed value tobs |H0 ). k
8

2. [L/E] Solution. Q2 is the same as Q1.4 in Exercise 1.

Note that SX = {x1 , x2 , x3 } and SY = {y1 , . . . , y4 }. By using point-wise IBF,
the marginal distribution of X is given by

X x1 x2 x3
pi = Pr{X = xi } 0.24 0.28 0.48

Similarly, the marginal distribution of Y is given by

Y y1 y2 y3 y4
qj = Pr{Y = yj } 0.28 0.16 0.28 0.28

The joint distribution of (X, Y ) is given by

 
0.04 0.04 0.12 0.04
P =  0.08 0.08 0.04 0.08  .
0.16 0.04 0.12 0.16
 9

3. Solution.

3.1 The joint pdf of X1 , . . . , Xn is

n n
!θ n
!−1
Y Y Y
f (x1 , . . . , xn ; θ) = θxθ−1
i = θn xi × xi .
i=1 i=1 i=1

ˆ ni=1 Xi is a sufficient s-
Q
From the factorization theorem, we know that S =
ˆ − log(S) = − ni=1 log Xi is also a sufficient statistic
P
tatistic of θ. Thus, T =
of θ.

3.2 [see Example 3.29 on the page 152 of the Textbook] The likelihood
function of θ is L(θ) = ni=1 θxiθ−1 , so that the log-likelihood function is
Q

n
X
`(θ) = n log θ + (θ − 1) log xi .
i=1
Pn
Let 0 = `0 (θ) = n/θ + i=1 log xi , the MLE of θ is given by
n n
θ̂ = Pn = .
− i=1 log Xi T

Thus, the MLE of µ is

θ̂ n
µ̂ = = .
θ̂ + 1 n+T

3.3 The cdf of X is

Z x Z x Z x x
θ−1
F (x; θ) = f (t; θ)dt = θt dt = dtθ = tθ = xθ , x ∈ (0, 1). (1)
0 0 0 0

3.4 From (4.3) on page 165 of the Textbook, we have

n
X
−2 log F (Xi ; θ) ∼ χ2 (2n)
i=1
10

for any continuous cdf; hence

n n
X (1) X
−2 log F (Xi ; θ) = −2θ log Xi = 2θT ∼ χ2 (2n). (2)
i=1 i=1

We have E(2θT ) = 2n, i.e.,

 
T 1
E = = τ (θ), (3)
n θ

indicating that T /n is an unbiased estimator of τ (θ) = 1/θ.

On the other hand, from (2), Var(2θT ) = 2 × 2n, i.e.,
 
T 1
Var = 2. (4)
n nθ

In particular, in (3) and (4) let n = 1, we obtain

1 1
E(− log X1 ) = = τ (θ) and Var(− log X1 ) = 2 . (5)
θ θ

3.5 Let X ∼ f (x; θ) = θxθ−1 , x ∈ (0, 1). Then, from (3.24) in the Textbook,
we have
 2  2
d log f (X; θ) 1
I(θ) = E =E + log X = E [− log X1 − τ (θ)]2
dθ θ
(5) (5) 1
= Var(− log X1 ) =
θ2
while the method on page 152 of the Textbook is much easier; and hence
n
In (θ) = nI(θ) = .
θ2

3.6 Now, T /n is an unbiased estimator of τ (θ), and

{τ 0 (θ)}2
 
T (4) 1
Var = = ,
n nθ2 In (θ)
 11

i.e., the variance attains the CR lower bound. Then T /n is an efficient estimator
of τ (θ), and hence T /n is the unique UMVUE for τ (θ).

3.7 From (2), we know that 2θT ∼ χ2 (2n) is a pivotal quantity. Thus, using
the equal-probability (or equal-tail) method, we have
n o
1 − α = Pr χ2 (1 − α/2, 2n) 6 2θT 6 χ2 (α/2, 2n)
 2
χ2 (α/2, 2n)

χ (1 − α/2, 2n)
= Pr 6θ6 ;
2T 2T

that is,
χ2 (1 − α/2, 2n) χ2 (α/2, 2n)
 
[Lp , Up ] = ,
2T 2T
is a 100(1 − α)% CI for θ. k
12

4. Solution.
4.1 First we consider to test

H0 : σ 2 = σ02 against H1 : σ 2 = σ12 (> σ02 ). (6)

The likelihood function is given by

n
(xi − µ0 )2
 
2
Y 1
L(σ ) = √ exp −
i=1 2πσ 2 2σ 2
 Pn 2

2 −n/2 i=1 (xi − µ0 )
= (2πσ ) exp − .
2σ 2
Then ( )
n/2 n
L(σ02 ) σ12

1 1 1 X
= exp − (xi − µ0 )2 6k
L(σ12 ) σ02 2 σ12 σ02 i=1
is equivalent to
n
X 2 log k − n log(σ12 /σ02 )
(xi − µ0 )2 > 1 =
ˆ c.
i=1 σ2
− σ12
1 0
Pn Pn
Note that i=1 (Xi − µ0 )2 /σ 2 ∼ χ2 (n). When H0 is true, we have i=1 (Xi −
2
µ0 ) /σ02 2
∼ χ (n). Thus, the c can be determined by the size
( n )
X
α = Pr (Xi − µ0 )2 > c σ 2 = σ02
i=1
n
µ0 )2
P 
i=1 (Xi −c
= Pr > 2 σ 2 = σ02
σ2 σ
 Pn 2

i=1 (Xi − µ0 ) c
= Pr > 2
σ02 σ0
2 2
= Pr{χ (n) > c/σ0 }
= Pr{χ2 (n) > χ2 (α, n)},

i.e., c = σ02 χ2 (α, n). Therefore, by the Neyman–Pearson lemma, a test ϕ with
critical region
Pn
C = {x : i=1 (xi − µ0 )2 > σ02 χ2 (α, n)}
 13

is a most powerful test of size α for testing (6).

4.2 Since the critical region C depends only on n, σ02 , α and the fact σ12 > σ02 ,
but not on the value of σ12 , the test ϕ is also a UMPT of size α for testing

H0 : σ 2 = σ02 against H1 : σ 2 > σ02 . (7)

Next, the supremum of the power function is given by

sup pϕ (σ 2 ) = sup Pr{ ni=1 (Xi − µ0 )2 > c|σ 2 }

P
σ 2 6σ02 σ 2 6σ02
 Pn
µ0 ) 2

i=1 (Xi − c
= sup Pr > 2
σ 2 6σ02 σ2 σ
= sup Pr{χ2 (n) > c/σ 2 }
σ 2 6σ02

= max2 1 − Pr{χ2 (n) < c/σ 2 }

σ 2 6σ0

= 1 − min2 Pr{χ2 (n) < c/σ 2 }

σ 2 6σ0

= 1 − Pr{χ2 (n) < c/σ02 }

= 1 − Pr{χ2 (n) < χ2 (α, n)}
= Pr{χ2 (n) > χ2 (α, n)}
= α = pϕ (σ02 ),

where the fact that Pr{χ2 (n) < c/σ 2 } is a decreasing function of σ 2 is utilized.
Then, the test ϕ is also a UMPT of size α for testing H0 : σ 2 6 σ02 against
H1 : σ 2 > σ02 . k
14

5. Solution. We wish to test

H0 : The distribution is Poisson against

H1 : The distribution is not Poisson.

Under H0 , the maximum likelihood estimate of λ is

159
λ̂ = x = ≈ 1.622.
98
Now
5
λ̂j −λ̂ X
p̂j0 = pj0 (λ̂) = e , j = 0, 1, . . . , 5, p̂6,0 = 1 − p̂j0 ,
j! j=0

and n = 98, we obtain

j 0 1 2 3 4 5 6(≥ 6)
Nj 18 34 24 16 3 1 2
p̂j0 0.1974 0.3203 0.2598 0.1405 0.0570 0.0185 0.0064
np̂j0 19.3466 31.3889 25.4635 13.7711 5.5857 1.8125 0.6317

Those classes with expected frequencies less than 5 should be combined with
the adjacent class. Therefore, we combine the last 3 classes, and the revised
table is

j 0 1 2 3 4(≥ 4)
Nj 18 34 24 16 6
p̂j0 0.1974 0.3203 0.2598 0.1405 0.0819
np̂j0 19.3466 31.3889 25.4635 13.7711 8.0299

So we have
4
X (Nj − np̂j0 )2
Q̂98 = = 1.2690 < χ2 (0.05, 5 − 1 − 1) = 7.81.
j=0
np̂ j0

Thus, we cannot reject H0 when the approximate significance level is taken to

be 0.05.