0% found this document useful (0 votes)
17 views5 pages

Practice Session 1 With Answers

Uploaded by

Mds Dms
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
17 views5 pages

Practice Session 1 With Answers

Uploaded by

Mds Dms
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

LSTAT2050 - Analyse statistique II - Practice

session 1

Exercise 1 - Exponential family


Let X ∼ Beta(a, b), i.e. X has density function
Γ(a + b) a−1
pa,b (x) = x (1 − x)b−1 1[0,1] (x),
Γ(a)Γ(b)
where a, b > 0. Show that p belongs to the exponential family in the following scenarios
a) a is known and b is unknown;
a,b

b) Both a and b are unknown.


Solution

a) a is known and b is unknown


We want to be able to write :
Pb (x) = C(b)h(x)exp[Q(b)T (x)]
We have
Γ(a + b) xa−1
Pb (x) = 1[0,1] (x) b log(1 − x)]
exp[|{z}
Γ(a) + Γ(b) | 1 − x} | {z }
| {z } {z Q(b) T (x)
C(b) h(x)

b) a and b are unknown


We want to be able to write :
Pa,b (x) = C(a, b)h(x)exp[Q1 (a, b)T1 (x) + Q2 (a, b)T2 (x)]
We have
Γ(a + b) 1
Pa,b (x) = 1[0,1] (x) exp[ |{z}
a log(x) + |{z}b log(1 − x)]
Γ(a) + Γ(b) x(1 − x) | {z } | {z }
| {z }| {z } Q1 (a,b) T (x) Q2 (a,b) T2 (x)
1
C(a,b) h(x)

Exercise 2 - Suciency
Find a sucient statistic for the given parameter in the following scenarios :
a) Let X , . . . , X ∼ P ar(θ , θ ) the density of which is given by p (x) = θ θ x
iid θ1 −θ1 −1
1[θ2 ,∞) (x)
and where θ = (θ , θ ) with θ , θ > 0.
1 n 1 2 θ 1 2
T
1 2 1 2

b) Let X , . . . , X ∼ p (x), where p (x) = (2x/θ) exp[−x /θ]1 (x), with θ > 0.
1 n
iid
θ θ
2
[0,∞)

c) Let X , . . . , X ∼ U [0, θ], with θ > 0.


1 n
iid

LSTAT2050 - Analyse statistique II - Practice session 1 1


Solution

Subquestion a)
Because of the presence of the parameter in the support, we are obviously not dealing with
the EF. A good move consists in using the factorisation theorem (theorem 2.1). Given this, we
need to nd T (x) such that p (x) = g [T (x)]h(x). Let us do this.
θ θ,n
n n
1[θ2 ,∞) (Xi )
Y Y
pθ (x) = θ1n θ2nθ1 Xi−θ1 −1
i=1 i=1
| {z }| {z }
(I) (II)

Let us rewrite (I) and (II). " #


n
Y n
X
Xi−θ1 −1 = exp −(θ1 − 1) log(Xi )
i=1 i=1
n
1[θ2 ,∞) (Xi ) = 1[Xi ∈ [θ2 ; ∞)∀i = 1, . . . , n]
Y

i=1
= 1[X(1) ≥ θ2 ]1[X(n) < ∞]
Using this, we can reformulate p"(x) as θ
#
n
1[X(1) ≥ θ2 ] 1[X(n) < ∞]
X
pθ (x) = θ1n θ2nθ1 exp −(θ1 − 1) log(Xi )
i=1
| {z }
| {z } h(x)
gθ,n [T (x)]

In conclusion, T (x) = Pn


log(Xi ), X(1)

is sucient for θ. Note for completeness that (I)
didn't need to be rewritten. i=1

Subquestion b)
It is easy to see that p θ ∈ EF (1) . Indeed,  
1 1 2
pθ (x) = 2x1[θ2 ,∞) (x) exp − x
θ θ
Hence, the privileged statistic is . In order to conclude that it is sucient for
Pn
T (x) = i=1 Xi2
θ, we still need to check the conditions of theorem 2.6. Noting that the natural parameter is
λ = −1/θ it is trivial to determine that p is of minimal dimension and has non-empty interior.
Hence, is indeed sucient for θ.
λ
T (x)
Note : in order to understand the avour of this last condition let us provide an example of
a situation where Λ has empty interior. Consider a N (θ, θ ). We have λ = g(θ) and Λ = {g(θ) :
2

θ → θ }. Hence, all points in Λ lie on a parabola and, hence, it has empty interior.
2

Subquestion c)
Let us use the factorisation theorem (theorem 2.1). In order to use it, we rewrite p as θ
n
1
1[0 ≤ Xi ≤ θ]
Y
pθ (x) =
θn i=1
1
= 1[X(n) ≤ θ]1[X(1) ≥ 0]
θn

LSTAT2050 - Analyse statistique II - Practice session 1 2


Hence, X is sucient for θ.
(n)

Exercise 3 - Ancillarity
In the so-called location-scale family of distribution, nding an ancillary statistic becomes a
trivial task. This exercise illustrates this point. Let us dene the following families.
 Location family : P = {p (x) : p (x) = g(x − µ)}. This means that there exists some
Z such that Z = X − µ and its density is given by p (z) = g(z), free of µ.
1 µ µ

 Scale family : P = {p (x) : p (x) = g( )}. This means that there exists some Z such
Z
1 x

that Z = and its density is given by p (z) = g(z), free of σ.


2 σ σ σ σ
X

 Location-scale family : P = {p (x) : p = g( )}. This means that there


σ Z
1 x−µ

exists some Z such that Z = and its density is given by p (z) = g(z), free of σ and
3 µ,σ µ,σ(x) σ σ
X−µ

µ.
σ Z

Use the properties of these families to nd ancillary statistics in the following scenarios :
a) X , . . . , X ∼ U [θ, θ + 1]
1 n
iid

b) X , . . . , X ∼ U [0, θ]
1 n
iid

c) X , . . . , X ∼ U [θ , θ ]
1 n
iid
1 2

Subquestion a)
We have that U [θ, θ + 1] ∈ F . Indeed 1

pθ (x) = 1[θ ≤ x ≤ θ + 1]
= 1[0 ≤ x − θ ≤ 1] = g(x − θ)
Consequently, Z = X − θ has a density p free of θ. Hence, S(x) = X − X(1) = Z(n) − Z(1)
is ancillary.
Z (n)

Subquestion b)
We have that U [0, θ] ∈ F . Indeed 2

1
pθ (x) = 1[0 ≤ x ≤ θ]
θ
1 h x i 1 x
= 1 0≤ ≤1 = g
θ θ θ θ
Consequently, Z = has a density p free of θ. Hence, S(x) =
X
θ Z
X(1)
X(n)
=
Z(1)
Z(n)
is ancillary.
Subquestion c)
We have that U [θ , θ ] ∈ F . Indeed
1 2 3

1
pθ1 ,θ2 (x) = 1[θ1 ≤ x ≤ θ2 ]
θ2 − θ1
1
= 1 [0 ≤ x − θ1 ≤ θ2 − θ1 ]
θ2 − θ1
   
1 x − θ1 1 x − θ1
= 1 0≤ ≤1 = g
θ2 − θ1 θ2 − θ1 θ2 − θ1 θ2 − θ1

LSTAT2050 - Analyse statistique II - Practice session 1 3


Consequently, Z = X−θ1
has a density p free of (θ , θ ). Hence, S(x) = X(n) −X(1)
=
Z(n) −Z(1)

is ancillary.
θ2 −θ1 Z 1 2 X(n−1) −X(1) Z(n−1) −Z(1)

Exercise 4 - Completeness and minimal suciency


Let X , . . . , X iid
∼ U [θ, θ + 1] . Prove that T (x) = [X (1) , X(n) ] is minimal sucient for θ but not
complete.
1 n

Solution

Minimal suciency
We are going to use theorem 2.2, which states that provided that pθ (x)
is free of θ if and only
if T (x) = T (y), then T (x) is minimal sucient for θ. We have
pθ (y)

1[θ ≤ Xi ≤ θ + 1]
Qn
pθ (x)
= Qi=1
i=1 1[θ ≤ Yi ≤ θ + 1]
n
pθ (y)
1[X(1) ≥ θ]1[X(n) ≤ θ + 1]
=
1[Y(1) ≥ θ]1[Y(n) ≤ θ + 1]
which is free of θ if and only if X = Y(1) and X = Y(n) . Hence, T (X) is minimal sucient
for θ
(1) (n)

Non completeness
Several options are available.
Using the denition
We need to nd some integrable function g(.) such that E [g(X , X )] = 0, ∀θ ∈ Θ but
it is not the case that g(X , X ) = 0 a.s. We rst need to compute E (X ) and E (X ).
θ (1) (n)

First, noting p and P the density and cumulative distribution of the U [θ, θ + 1], we have
(1) (n) θ (1) θ (n)
θ θ

pX(1) (u) = npθ (u) [1 − Pθ (u)]n−1 = n [θ + 1 − u]n−1 1[θ ≤ u ≤ θ + 1]


pX(n) (t) = npθ (t) [Pθ (t)]n−1 = n [t − θ]n−1 1[θ ≤ t ≤ θ + 1]

After some computations, we obtain


n
Eθ (X(1) ) = θ + 1 −
n+1
n
Eθ (X(n) ) = +θ
n+1
Therefore
2n
Eθ [X(n) − X(1) ] = 1 +
n+1
Taking g(X , X ) = X − X − 1 − , we have E [g(X 2n
(1) , X(n) )] = 0∀θ . But it is not the
case that g(X , X ) = 0 a.s. Hence, T (X) is not complete.
(1) (n) (n) (1) n+1 θ
(1) (n)

LSTAT2050 - Analyse statistique II - Practice session 1 4


Using Basu's theorem
Let us rst recall Basu's theorem. Let S(X) be a sucient and complete statistic for θ and
U (X) be ancillary for θ. Then S(X) and U (X) are independent.
Assume per contra that T ((X)) is complete. We need to nd a statistic S(X) which is
ancillary and not independent with T (X). We would then reach a contradiction, proving that
T (X) cannot be complete.
By exercise 3, we know that S(X) = X −X is ancillary for θ. However, S(X) is obviously
not independent from T (X). Hence, T (X) is not complete
(n) (1)

Using directly an ancillary statistic


A general procedure for assessing the non completeness of a statistic T (X) lies in following
the subsequent steps.
1. Construct from T (X) an ancillary statistic S(X) = S(T ).
2. Because S(X) is ancillary, we have E [S(T )] = k, where k is a constant not depending
on θ.
θ

3. Dene g(T ) = S(T ) − k, we have E [g(T )] = 0∀θ.


θ

4. If it is not the case that g(T ) = 0 a.s. then T (X) is not complete.
Here, the ancillary statistic is again S(X) = X − X . Hence, we do not have g(T ) = 0
a.s. and T (X) is not complete.
(n) (1)

Exercise 5 - Basu theorem


This exercise illustrates the idea that Basu theorem may be used to compute moments of complex
∼ U [0, θ]. Use Basu theorem to compute E[X/X(n) ].
statistics. Let X1 , . . . , Xn iid

Solution
Notice that, E[X] = E[ X ] = E[ ]E[X ], where the last equality only applies in case
X
(n)
X
(n)

of independence. In order to show the independence between and X , we will use Basu
X(n) X(n)
X

Theorem.
X(n) (n)

We now show that X is minimal sucient for θ and X/X is ancillary for θ.
 It is easy to prove that X is minimal sucient by theorem 2.2 (as in exercise 4).
(n) (n)

 It is easy to prove that X/X is ancillary once we recognize that U [0, θ] is in the scale
(n)

family.
(n)

Then we can apply Basu theorem to conclude that X/X is independent from X .(n) (n)

Finally, we can easily show that E[X] = θ/2 and E[X ] = θ. Hence, E[X/X ] = (n)
n
n+1 (n)
n+1
E[X]/E[X(n) ] = 2n

LSTAT2050 - Analyse statistique II - Practice session 1 5

You might also like