Statistics 426: Final Exam, Fall 2023

Moulinath Banerjee
December 15, 2023

Announcement: The exam carries 80 points but the maximum number of points you can
score is 65.
1. Let X1 , X2 , . . . , Xn be i.i.d. distributed as

3x2
f (x, ✓) = 1(0 < x  ✓) .
✓3
Let ✓0 denote the true parameter. Find the MLE of ✓ and show that it converges in
probability to ✓0 . [Hint: Calculate the probability that ✓ˆM LE > ✓0 ✏ where ✏ > 0.]
(6 + 6 = 12)

2. Let (Xi , Yi )ni=1 be i.i.d. random vectors where each of them is distributed like (X, Y )
where X ⇠ Exp(✓) and Y |X ⇠ Poisson(✓X). Let ✓0 be the true data generating
parameter.
(a) Show that the density of any (Xi , Yi ) pair is given by:
2✓x
✓e (✓x)y
f ((x, y), ✓) = .
y!
(b) Show that the MLE of ✓0 is:

Y +1
✓ˆM L = .
2X
Calculate E(Y ), E(X) and use these to show that ✓ˆM L converges in probability to ✓0 .
.
(c) Find I(✓) and an asymptotic level 1 ↵ confidence interval for ✓0 (4 + (7 + 7) +
(5 + 5) = 28 points)

3. Consider a hypothesis testing problem for an observation X from a distribution. Under

H0 , X ⇠ f0 (x) where f0 (x) = Cx2 1( 1  x  1) whilst under H1 , X ⇠ f1 (x) where

1
 f1 (x) = C 0 (1 x4 ) 1( 1  x  1). Here C, C 0 are readily calculable constants so that
f0 and f1 integrate to 1 and all answers can be written in terms of these constants.
(a) Show that '(X), the most powerful test based on the Neyman-Pearson lemma,
rejects when |X|  , for some number  > 0.
(b) Find  such that the test '(X) has level ↵.
(c) Calculate the power of this level ↵ test. (8+ 6 + 6 = 20 points)

4. Consider a Bayesian setting where the parameter space is {✓0 , ✓1 }, with 0 < ✓0 , ✓1 < 1.
Let ⇥ be a random variable such that P (⇥ = ✓0 ) = p0 and P (⇥ = ✓1 ) = p1 with
p0 + p1 = 1.
We observe Bernoulli random variables (X1 , X2 , . . . , Xn ) which are i.i.d. Bernoulli(✓0 )
conditionally, given ⇥ = ✓0 , whilst given ⇥ = ✓1 , (X1 , X2 , . . . , Xn ) are conditionally
Bernoulli (✓1 ).

(a) Write down the joint probability mass function of (X1 , X2 , . . . , Xn , ⇥) i.e.
p(⇥ = ✓i , X1 = x1 , . . . , Xn = xn ) for i = 0, 1; and compute the posterior p.m.f of ⇥
given (X1 , . . . , Xn ), i.e. P (⇥ = ✓i |X1 = x1 , . . . , Xn = xn ).
(b) A statistician will pick ✓1 as the true data generating parameter if the ratio of the
posterior probability of ✓1 to that of ✓0 is adequately large, say greater than T . Show
that the criterion can be written as: pick ✓1 if:
1
X log(✓1 /✓0 ) + (1 X) log((1 ✓1 )/(1 ✓0 )) > [log(p0 /p1 ) + log T ] .
n
If ✓1 is indeed the true value that generates the data, show that the probability of
picking ✓1 goes to 1 as n ! 1 when p0 = p1 = 1/2 and ✓1 = 1/2 and ✓0 = 1/4. (10 +
10 = 20 points)

2
L(0 , 11x21 ... m)
=
e(b) =
(n)(b) = (3x) = (tw(3x) -

32n(t)
=
(en (3x4) -

3mln(t)

(10) = (4-3) = 0
E
E 0 => = No internal critical
point
=> EMLe =
max(Xn ,
Xa , ..., Xn)
Let Mr = max (X1 ,
X2 , . .

., Xn)
*
For EMLE 0
,
we need
P(/EMLE-tol >
3) - > 0 as w + & >O
 For
large n. CDF Mr :

P(Mm [
y) =
(P(X(y)) (43) you y =
-t

((to-2))v (1 32t-3 S
e

=> P(Mm-to 2) -

= =
-

=
1 -

P(Mm = to 2) -
-

> 1 asn +

=> P(IEME-fol > 2) >

- 0 as n + 2

=> FMLE =
max(X1 , X2, ..., Xn) is a consistent estimator
you
to
 Fe-zi (E)
a) Prove
+ ((x y) , b)
,
=

!
y
-

X-Exp(f)(fx(x , b) =
Ge
xyox > 0
(EcYe- *
YIX ~
Poisson (8X)(8x1x (y(x ) ,
=

!
you y ES0 ,
1
,
2 ...
%
y
y f)
fx y
(x , =
=

fy(x(y(x , b) +x(x) x
,

(c)Ye-oxte
=

_
=
)yyxbe -20x

y!
b) Prove :
EMLE =
Likelihood
junction L(bi (x yi))
:
,
= De-Zi
e(b) = ((n0 lu(txi)yi 20xi)
+

!
-

Yi

= (Int +
yiln(xi) Inyi - ! -

28xi)

nufyilnYilni-ni-
=
Inyyiln-2
Nint +

a l() = + -
+y 2 n -
= 0

E) - = i = 1
Yi + n

25 xi

T+ 1
=> EMLE =
where F =
Yi , = Xi
2

E(Y) =
E(E(Y(X) = E(OX) =
GE(X) =

0xz = 1
 => Y =1 asn + 2

Sce X
E(X) = x =

- =
fasn + 2

=> Fre ==
Do

6) We have Fisher
Information :
I(f) =
-

E((t)]
l() =_ +
02

+Si) = =
E(n
-

E( -n +
Y =

=> I(b)
=
EMLE =
N(to ,) No =

1- 2CI : Ente
e zrEmleze EMlez =
a) For fo(x)
J.. Code =
1

Es
/" d = 1

El
Cx
X3
For fe(ce) :
(! C (1 -

x4)dxc = 1

(21-x ]
*
=
1

ec'(2 - E) = 1

= 5

=> fo() =

- .
1) 1 = x(1)

fe() =

=(1 xi)1(
- -
14x - 1)

Likelihood ratio X(X) :

M(X)
=)
Reject Ho
: X) i

E)1 *
4x -

hx
*
E) 17X +

Gn) X + x4
b) x =
P/reject Ho /Hotrue) =
P(X1 => k) X-fo)
Since
X-fo(x) =

21(-1*1) :

X =

J fo() do Je =

=
= k =
213
2) Power =
P(reject Holth true) = P((X) - k(X-fe)
=

Jinfe(t)dx /(1 =
-

x4)dx =

=) (1 -
x4)d

=x(2k E) -

zx(2k -) = (k E)
-

= (213 -

((25) = (213 -
a) For i = 0 1
, ,
PMF :

P(0 = Gi , Xn = 2 , ...)
Xm =
xu) =
P(0 ti)P(Xe
= = x , ...)
Xn = xn)0 = Gi)
Since (X1 , X2, ...) Xn) ale i .
id Bernoulli (ti)
given O =
Gi :

P(Xn =
(1) ..., Xm = xn)0 = (i) = ↑ P(Xj xj)0 (i)
= =

= )
= Fi (1-Fi)m-EXj
=> P(G = Ei , X = c , ...,
Xn =) =

pi1(1-0) -Enj
 Posterior
probability : P(G =
Gi(X1 =1 ...) Xn = xn)
By Bayes' Theorem :

P(G Ei , Xi
x) Xm xn)
P(G Gi(Xn xn)
= =

Xn
=
--.,
= =
x 1 ..., = =

P(X1 = > , . . .,
Xn = (n)

P(X1 =
() ...,
Xn = xu) = P(G =
Gi)P(X1 =
x) - - -

)Xn = xu)0 Gi)

=

=
pot(1-0) -E1 +
pet(1-)- E
=> P(G =
Eil Xn =
11 ,
...
>
Xn = xn)

=
Pit (1-Ei) -Enj
Po1j(1-0) -E (1 E1)n Ej 1x
-

=
+
-

P(G (1)X1
b) Will pick Fr Y :
=
=
x ,...,
Xn =
kn) >T
P(O = Go(X1 =
(1) ...

,
Xn =
kn)
Pron(1-)-E
E potr
(1-50) -E + per (1-0)-E ;

7 T
pot(1-to-Enj
poto (1-50)m- E + por (1-0) -E "
;
 PrEE(1-0) Enj
=

E) >T
Pot (1-00) -E1]

log() +

((xlog()
xj)log)(J log(t) + ( -

log() () log (log() fl (1 -

+

Let =

X
So the criterior is
log () + (1- * )
log () (log() 10g(T) +

We have :

P((xlog() + ( -

xj)log)(J) log(t) - 1asm +

We have
:[Vilog ()
+
( xj)
-

log))] >
log(T)
Es [Xjlog(1) + (1 -

Xj) log(5)] >

log(t)
 By
=X-an-
LN :

We have
: [Xjlog(2) + (1 -

X]) log(5)]
=
Flog() + (1 Y)log(z)
-
=

Elog(2) zlog(z)
+

&log()
For large =
n
n(log() >
-
0 asn +

=> P(nzlog(t)) log(T) + 1am +