9783642033827-c1

25
CHAPTER 2 A Review of Some Basic Statistical Concepts 2.1 Variance and Covariance of Linear Combinations of Random Variables. a. Let Y D a C bX, then E.Y/ D E.a C bX/ D a C bE.X/. Hence, var.Y/ D EŒY E.Y2 D EŒa C bX a bE.X2 D EŒb.X E.X//Ł 2 D b 2 EŒX E.X2 D b 2 var.X/. Only the multiplicative constant b matters for the variance, not the additive constant a. b. Let Z D a C bX C cY, then E.Z/ D a C bE.X/ C cE.Y/ and var.Z/ D EŒZ E.Z2 D EŒa C bX C cY a bE.X/ cE.Y2 D EŒb.X E.X// C c.Y E.Y//Ł 2 D b 2 EŒXE.X2 Cc 2 EŒYE.Y2 C2bc EŒXE.X/ŁŒYE.YD b 2 var.X/ C c 2 var.Y/ C 2bc cov.X, Y/. c. Let Z D aCbXCcY, and W D dCeXCfY, then E.Z/ D aCbE.X/CcE.Y/ E.W/ D d C eE.X/ C fE.Y/ and cov.Z, W/ D EŒZ E.Z/ŁŒW E.WD EŒb.XE.X//Cc.YE.Y//ŁŒe.XE.X//Cf.YE.Y//Ł D be var.X/ C cf var.Y/ C .bf C ce/ cov.X, Y/. 2.2 Independence and Simple Correlation. a. Assume that X and Y are continuous random variables. The proof is similar if X and Y are discrete random variables and is left to the reader. If X and Y are independent, then f.x, y/ D f 1 .x/f 2 .y/ where f 1 .x/ is the marginal probability density function (p.d.f.) of X and f 2 .y/ is the marginal p.d.f. of Y. In this case, E.XY/ D xyf.x, y/dxdy D xyf 1 .x/f 2 .y/dxdy D . R xf 1 .x/dx/. R yf 2 .y/dy/ D E.X/E.Y/

Upload: anonymous-apiw0uc9

Post on 02-Dec-2015

212 views

Category:

Documents


0 download

DESCRIPTION

tiulo

TRANSCRIPT

Page 1: 9783642033827-c1

CHAPTER 2

A Review of Some Basic Statistical Concepts

2.1 Variance and Covariance of Linear Combinations of Random Variables.

a. Let Y D a C bX, then E.Y/ D E.a C bX/ D a C bE.X/. Hence,

var.Y/ D EŒY � E.Y/�2 D EŒa C bX � a � bE.X/�2 D EŒb.X � E.X//�2

D b2EŒX � E.X/�2 D b2 var.X/.

Only the multiplicative constant b matters for the variance, not the additive

constant a.

b. Let Z D a C bX C cY, then E.Z/ D a C bE.X/ C cE.Y/ and

var.Z/ D EŒZ � E.Z/�2 D EŒa C bX C cY � a � bE.X/ � cE.Y/�2

D EŒb.X � E.X// C c.Y � E.Y//�2

D b2EŒX�E.X/�2Cc2EŒY�E.Y/�2C2bc EŒX�E.X/�ŒY�E.Y/�

D b2var.X/ C c2var.Y/ C 2bc cov.X, Y/.

c. Let Z D aCbXCcY, and W D dCeXCfY, then E.Z/ D aCbE.X/CcE.Y/

E.W/ D d C eE.X/ C fE.Y/

and

cov.Z, W/ D EŒZ � E.Z/�ŒW � E.W/�

D EŒb.X�E.X//Cc.Y�E.Y//�Œe.X�E.X//Cf.Y�E.Y//�

D be var.X/ C cf var.Y/ C .bf C ce/ cov.X, Y/.

2.2 Independence and Simple Correlation.

a. Assume that X and Y are continuous random variables. The proof is similar

if X and Y are discrete random variables and is left to the reader. If X and

Y are independent, then f.x, y/ D f1.x/f2.y/ where f1.x/ is the marginal

probability density function (p.d.f.) of X and f2.y/ is the marginal p.d.f. of

Y. In this case,

E.XY/ D ’xyf.x, y/dxdy D ’

xyf1.x/f2.y/dxdyD .

Rxf1.x/dx/.

Ryf2.y/dy/ D E.X/E.Y/

Page 2: 9783642033827-c1

6 Badi Baltagi

Hence,

cov.X, Y/ D EŒX � E.X/�ŒY � E.Y/� D E.XY/ � E.X/E.Y/

D E.X/E.Y/ � E.X/E.Y/ D 0.

b. If Y D a C bX, then E.Y/ D a C bE.X/ and cov.X, Y/ D EŒX � E.X/�ŒY �E.Y/� D EŒX � E.X/�Œa C bX � a � bE.X/� D b var.X/ which takes the

sign of b since var(X) is always positive. Hence,

correl.X, y/ D ¡xy D cov.X, Y/p

var.X/var.Y/D b var.X/p

var.X/var.Y/

but var.Y/ D b2 var.X/ from problem 2.1a. Hence, ¡XY D b var.X/pb2.var.X//2

D˙1 depending on the sign of b.

2.3 Zero Covariance Does Not Necessarily Imply Independence.X P(X)�2 1/5�1 1/5

0 1/51 1/52 1/5

E.X/ D2X

XD�2

XP.X/ D 15

Œ.�2/ C .�1/ C 0 C 1 C 2� D 0

E.X2/ D2X

XD�2

X2P.X/ D 15

Œ4 C 1 C 0 C 1 C 4� D 2

and var.X/ D 2. For Y D X2, E.Y/ D E.X2/ D 2 and

E.X3/ D2X

XD�2

X3P.X/ D 15

Œ.�2/3 C .�1/3 C 0 C 13 C 23� D 0

In fact, any odd moment of X is zero. Therefore,

E.YX/ D E.X2.X/ D E.X3/ D 0

Page 3: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 7

and

cov.Y, X/ D E.X � E.X//.Y � E.Y// D E.X � 0/.Y � 2/

D E.XY/ � 2E.X/ D E.XY/ D E.X3/ D 0

Hence, ¡XY D cov.X,Y/pvar.X/var.Y/

D 0.

2.4 The Binomial Distribution.

a. PrŒX D 5 or 6� D PrŒX D 5� C PrŒX D 6�

D b.n D 20, X D 5, ™ D 0.1/ C b.n D 20, X D 6, ™ D 0.1/

D

205

!

.0.1/5.0.9/15 C

206

!

.0.1/6.0.9/14

D 0.0319 C 0.0089 D 0.0408.

This can be easily done with a calculator, on the computer or using the

Binomial tables, see Freund (1992).

b.

n

n � X

!

D nŠ.n�X/Š.n�nCX/Š

D nŠ.n�X/ŠXŠ

D

nX

!

Hence,

b.n, n � X, 1 � ™/ D

nn � X

!

.1 � ™/n�X.1 � 1 C ™/n�nCX

D nX

.1 � ™/n�X™X D b.n, X, ™/.

c. Using the MGF for the Binomial distribution given in problem 2.14a, we get

MX.t/ D Œ.1 � ™/ C ™et�n.

Differentiating with respect to t yields M0X.t/ D nŒ.1 � ™/ C ™et�n�1™et.

Therefore, M0X.0/ D n™ D E.X/.

Differentiating M0X.t/ again with respect to t yields

M00X.t/ D n.n � 1/Œ.1 � ™/ C ™et�n�2.™et/2 C nŒ.1 � ™/ C ™et�n�1™et.

Therefore M00X.0/ D n.n � 1/™2 C n™ D E.X2/.

Page 4: 9783642033827-c1

8 Badi Baltagi

Hence var.X/ D E.X2/ � .E.X//2 D n™ C n2™2 � n™2 � n2™2 D n™.1 � ™/.

An alternative proof for E.X/ DnP

XD0Xb.n, X, ™/. This entails factorial

moments and the reader is referred to Freund (1992).

d. The likelihood function is given by

L.™/ D f .X1, .., Xn; ™/ D ™

nP

iD1Xi

.1 � ™/n�

nP

iD1Xi

so that log L.™/ D� nP

iD1Xi

log ™ C�

n �nP

iD1Xi

log.1 � ™/

@ log L.™/

@™D

nP

iD1Xi

™�

n �nP

iD1Xi

.1 � ™/D 0.

Solving for ™ one getsnX

iD1

Xi � ™

nX

iD1

Xi � ™n C ™

nX

iD1

Xi D 0

so that O™mle DnP

iD1Xi=n D NX.

e. E. NX/ DnP

iD1E.Xi/=n D n™=n D ™.

Hence, NX is unbiased for � . Also, var. NX/ D var.Xi/ = n D �.1 � �/ = n

which goes to zero as n ! 1. Hence, the sufficient condition for NX to be

consistent for ™ is satisfied.

f. The joint probability function in part d can be written as

f.X1, .., Xn; ™/ D ™n NX.1 � ™/n�n NX D h. NX, ™/g.X1, : : : , Xn/

where h. NX, ™/ D ™n NX.1�™/n�n NX and g.X1, .., Xn/ D 1 for all Xi0s. The latter

function is independent of ™ in form and domain. Hence, by the factorization

theorem, NX is sufficient for ™.

g. NX was shown to be MVU for ™ for the Bernoulli case in Example 2 in

the text.

h. From part (d), L.0.2/ D .0.2/

nP

iD1Xi

.0.8/n�

nP

iD1Xi

while L.0.6/ D .0.6/

nP

iD1Xi

.0.4/n�

nP

iD1Xi

with the likelihood ratio L.0.2/L.0.6/

D � 13

�nP

iD1Xi

2n�

nP

iD1Xi

Page 5: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 9

The uniformly most powerful critical region C of size ’ � 0.05 is given by� 1

3

�nP

iD1Xi

2n�

nP

iD1Xi � k inside C. Taking logarithms of both sides

�nX

iD1

Xi.log 3/ C

n �nX

iD1

Xi

!

log 2 < log k

solving � nX

iD1

Xi

!

log 6 � K0 ornX

iD1

Xi � K

where K is determined by making the size of C D ’ � 0.05. In this case,nP

iD1Xi � b.n, ™/ and under Ho ; ™ D 0.2. Therefore,

nP

iD1Xi � b.n D 20, ™ D

0.2/. Hence, ’ D PrŒb.n D 20, ™ D 0.2/ � K� � 0.05.

From the Binomial tables for n D 20 and ™ D 0.2, K D 7 gives PrŒb.n D 20,

™ D 0.2/ � 7� D 0.0322. Hence,nP

iD1Xi � 7 is our required critical region.

i. The likelihood ratio test is

L.0.2/

L.O™mle/D .0.2/

nP

iD1Xi

.0.8/n�

nP

iD1Xi

� NX�nP

iD1Xi �

1 � NX�n�nP

iD1Xi

so that LR D �2 log L.0.2/ C 2 log L.O™mle/

D �2

" nX

iD1

Xi.log 0.2 � log NX/

#

� 2

"

n �nX

iD1

Xi

!

.log 0.8 � log.1 � NX//

#

.

This is given in Example 5 in the text for a general ™o. The Wald statistic is

given by W D . NX � 0.2/2

NX.1 � NX/=n

and the LM statistic is given by LM D . NX � 0.2/2

.0.2/.0.8/=nAlthough, the three statistics, LR, LM and W look different, they are all

based on j NX � 2j � k and for a finite n, the same exact critical value could

be obtained from the binomial distribution.

2.5 d. The Wald, LR, and LM Inequality. This is based on Baltagi (1995). The

likelihood is given by equation (2.1) in the text.

L��, ¢2� D .1=2 ¢2/n=2e

�.1=2¢2/nP

iD1.Xi��/2

(1)

Page 6: 9783642033827-c1

10 Badi Baltagi

It is easy to show that the score is given by

S��, ¢2� D

0

@

n. NX��/

¢2nP

iD1.Xi��/2�n¢2

2¢4

1

A , (2)

and setting S.�, ¢2/ D 0 yields O� D NX and O¢2 DnP

iD1.Xi � NX/2=n. Under

Ho, Q� D �0 and

Q¢2 DnX

iD1

.Xi � �0/2 =n.

Therefore,

log L� Q�, Q¢2� D �n

2log Q¢2 � n

2log 2  � n

2(3)

and

log L� O�, O¢2� D �n

2log O¢2 � n

2log 2  � n

2. (4)

Hence,

LR D n log

2

664

nP

iD1.Xi � �0/2

nP

iD1

�Xi � NX�2

3

775 . (5)

It is also known that the information matrix is given by

I

¢2

!

D"

n¢2 00 n

2¢4

#

. (6)

Therefore,

W D . O� � �0/2 OI11 D n2 � NX � �0�2

nP

iD1

�Xi � NX�2

, (7)

where OI11 denotes the (1,1) element of the information matrix evaluated

at the unrestricted maximum likelihood estimates. It is easy to show from

(1) that

log L� Q�, O¢2� D �n

2log O¢2 � n

2log 2  �

nP

iD1.Xi � �0/2

2 O¢2 . (8)

Page 7: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 11

Hence, using (4) and (8), one gets

�2 log�L� Q�, O¢2� =L. O�, O¢2/

� D

nP

iD1.Xi � �0/2 � n O¢2

O¢2 D W, (9)

and the last equality follows from (7). Similarly,

LM D S2 � Q�, Q¢2� QI11 D n2. NX � �0/2

nP

iD1.Xi � �0/

2, (10)

where QI11 denotes the (1,1) element of the inverse of the information matrix

evaluated at the restricted maximum likelihood estimates. From (1), we

also get

log L� O�, Q¢2� D �n

2log Q¢2 � n

2log 2  �

nP

iD1.Xi � NX/2

2 Q¢2 (11)

Hence, using (3) and (11), one gets

�2 log�L� Q�, Q¢2� =L

� O�, Q¢2�� D n �

nP

iD1

�Xi � NX�2

Q¢2 D LM, (12)

where the last equality follows from (10). L. Q�, Q¢2/ is the restricted max-

imum; therefore, log L. Q�, O¢2/ � log L. Q�, Q¢2/, from which we deduce

that W � LR. Also, L. O�, O¢2/ is the unrestricted maximum; therefore

log L. O�, O¢2/ � log L. O�, Q¢2/, from which we deduce that LR � LM.

An alternative derivation of this inequality shows first that

LMn

D W=n1 C .W=n/

andLRn

D log

1 C Wn

.

Then one uses the fact that y � log.1 C y/ � y=.1 C y/ for y D W=n.

2.6 Poisson Distribution.

a. Using the MGF for the Poisson derived in problem 2.14c one gets

Mx.t/ D eœ.et�1/.

Page 8: 9783642033827-c1

12 Badi Baltagi

Differentiating with respect to t yields

M0x.t/ D eœ.et�1/œet.

Evaluating M0X.t/ at t D 0, we get

M0x.0/ D E.X/ D œ.

Similarly, differentiating M0X.t/ once more with respect to t, we get

M00x .t/ D e�.et�1/

��et�2 C e�.et�1/�et

evaluating it at t D 0 gives

M00x .0/ D �2 C œ D E.X2/

so that

var.X/ D E.X2/ � .E.X/2/ D �2 C � � �2 D �.

Hence, the mean and variance of the Poisson are both equal to œ.

b. The likelihood function is

L .�/ D e�n��

nP

iD1Xi

X1ŠX2Š..XnŠ

so that

log L.�/ D �n� C nX

iD1

Xi

!

log � �nX

iD1

log XiŠ

@ log L.�/

@�D �n C

nP

iD1Xi

�D 0.

Solving for œ, yields Oœmle D NX.

c. The method of moments equates E(X) to NX and since E.X/ D œ the solution

is Oœ D NX, same as the ML method.

d. E. NX/ DnP

iD1E.Xi/=n D nœ=n D œ. Therefore NX is unbiased for œ. Also,

var. NX/ D var.Xi/n D �

n which tends to zero as n ! 1. Therefore, the

sufficient condition for NX to be consistent for œ is satisfied.

Page 9: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 13

e. The joint probability function can be written as

f.X1, .., Xn; �/ D e�n��n NX 1X1Š..XnŠ

D h� NX, �

�g .X1, .., Xn/

where h. NX, œ/ D e�nœœn NX and g.X1, .., Xn/ D 1X1Š..XnŠ

. The latter is inde-

pendent of œ in form and domain. Therefore, NX is a sufficient statistic

for œ.

f. log f.X; œ/ D �œ C X log œ � log X!

and

@ log f.X; �/

@�D �1 C X

@2 log f.X; �/

@�2 D �X�2 .

The Cramer-Rao lower bound for any unbiased estimator Oœ of œ is given by

var� O�

� � 1

nE�

@2 log f.X;�/

@�2

D �2

n E.X/D �

n.

But var. NX/ D œ=n, see part (d). Hence, NX attains the Cramer-Rao

lower bound.

g. The likelihood ratio is given by

L.2/

L.4/D e�2n2

nP

iD1Xi

e�4n4nP

iD1Xi

The uniformly most powerful critical region C of size ’ � 0.05 is given by

e2n�

12

� nP

iD1Xi

� k inside C

Taking logarithms of both sides and rearranging terms, we get

� nX

iD1

Xi

!

log 2 � K0

ornX

iD1

Xi � K

Page 10: 9783642033827-c1

14 Badi Baltagi

where K is determined by making the size of C D ’ � 0.05. In this case,nP

iD1Xi � Poisson.nœ/ and under Ho; œ D 2. Therefore,

nP

iD1Xi � Poisson

(œ D 18). Hence ’ D PrŒPoisson.18/ � K� � 0.05.

From the Poisson tables, for œ D 18, K D 26 gives PrŒPoisson.18/ �26� D 0.0446. Hence,

nP

iD1Xi � 26 is our required critical region.

h. The likelihood ratio test is

L.2/

L. NX/D e�2n2

nP

iD1Xi

e�n NX � NX�nP

iD1Xi

D en. NX�2/

�2NX� nP

iD1Xi

so that

LR D �2 log L.2/ C 2 log L. NX/ D �2n. NX � 2/ � 2nX

iD1

Xi�log 2 � log NX� .

In this case,

C .�/ Dˇˇˇˇ@2 log L .�/

@�2

ˇˇˇˇ D

ˇˇˇˇˇˇˇˇ

nP

iD1Xi

�2

ˇˇˇˇˇˇˇˇ

and

I .�/ D �E

@2 log L.�/

@�2

D n�

�2 D n�

.

The Wald statistic is based upon

W D � NX � 2�2 I

�O�mle

D . NX � 2/2 � n

NXusing the fact that Oœmle D NX. The LM statistic is based upon

LM D S2.2/I�1.2/ D n2. NX � 2/2

4� 2

nD n. NX � 2/2

2.

Note that all three test statistics are based upon j NX � 2j � K, and for finite

n the same exact critical value could be obtained using the fact thatnP

iD1Xi

has a Poisson distribution, see part (g).

Page 11: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 15

2.7 The Geometric Distribution.

a. Using the MGF for the Geometric distribution derived in problem 2.14d,

one gets

Mx.t/ D ™et

Œ1 � .1 � ™/et�.

Differentiating it with respect to t yields

M0X.t/ D ™et �1 � .1 � ™/et�C .1 � ™/et™et

Œ1 � .1 � ™/et�2D ™et

Œ1 � .1 � ™/et�2

evaluating M0X.t/ at t D 0, we get

M0x.0/ D E.X/ D ™

™2 D 1™

.

Similarly, differentiating M0X.t/ once more with respect to t, we get

M00x.t/ D ™et �1 � .1 � ™/et�2 C 2

�1 � .1 � ™/et� .1 � ™/et™et

Œ1 � .1 � ™/et�4

evaluating M00X.t/ at t D 0, we get

M00x .0/ D E.X2/ D ™3 C 2™2.1 � ™/

™4 D 2™2 � ™3

™4 D 2 � ™

™2

so that

var.X/ D E.X2/ � .E.X//2 D 2 � ™

™2 � 1™2 D 1 � ™

™2 .

b. The likelihood function is given by

L .™/ D ™n.1 � ™/

nP

iD1Xi�n

so that

log L.™/ D n log ™ C nX

iD1

Xi � n

!

log.1 � ™/

@ log L.™/

@™D n

™�

nP

iD1Xi � n

.1 � ™/D 0

solving for ™ one gets

n.1 � ™/ � ™

nX

iD1

Xi C n™ D 0

Page 12: 9783642033827-c1

16 Badi Baltagi

or

n D ™

nX

iD1

Xi

which yields

O™mle D n=

nX

iD1

Xi D 1NX.

The method of moments estimator equates

E.X/ D NX

so that

1O™ D NX or O™ D 1

NXwhich is the same as the MLE.

2.8 The Uniform Density.

a. E.X/ D R 10 x dx D 1

2 Œx2�10 D 12

E.X2/ D R 10 x2dx D 1

3

�x3�1

0 D 13

so that var.X/ D E.X2/ � .E.X//2 D 13 � 1

4 D 4�312 D 1

12 .

b. PrŒ0.1 < X < 0.3� D R 0.30.1 dx D 0.3 � 0.1 D 0.2. It does not matter if

we include the equality signs PrŒ0.1 � X � 0.3� since this is a continuous

random variable. Note that this integral is the area of the rectangle, for X

between 0.1 and 0.3 and height equal to 1. This is just the length of this

rectangle, i.e., 0.3 � 0.1 D 0.2.

2.9 The Exponential Distribution.

a. Using the MGF for the exponential distribution derived in problem 2.14e,

we get

Mx.t/ D 1.1 � ™t/

.

Page 13: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 17

Differentiating with respect to t yields

M0x.t/ D ™

.1 � ™t/2 .

Therefore

MX0.0/ D ™ D E.X/.

Differentiating M0X.t/ with respect to t yields

M00X.t/ D 2™2 .1 � ™t/

.1 � ™t/4 D 2™2

.1 � ™t/3 .

Therefore

M00X.0/ D 2™2 D E.X2/.

Hence

var.X/ D E.X2/ � .E.X//2 D 2™2 � ™2 D ™2.

b. The likelihood function is given by

L .™/ D�

1™

�n

e�

nP

iD1Xi=™

so that

log L.™/ D �nlog™ �

nP

iD1Xi

@ log L .™/

@™D �n

™C

nP

iD1Xi

™2 D 0

solving for ™ one getsnX

iD1

Xi � n™ D 0

so that

O™mle D NX.

c. The method of moments equates E.X/ D NX. In this case, E.X/ D ™, henceO™ D NX is the same as MLE.

Page 14: 9783642033827-c1

18 Badi Baltagi

d. E. NX/ DnP

iD1E.Xi/=n D n™=n D ™. Hence, NX is unbiased for ™. Also,

var. NX/ D var.Xi/=n D �2=n which goes to zero as n ! 1. Hence, the

sufficient condition for NX to be consistent for ™ is satisfied.

e. The joint p.d.f. is given by

f .X1, .., Xn; ™/ D�

1™

�n

e�

nP

iD1Xi=™ D e�n NX=�

�1™

�n

D h. NX; �/g.X1, : : : , Xn/

where h. NX; �/ D e�n NX=�� 1

�n and g.X1, : : : , Xn/ D 1 independent of ™ in

form and domain. Hence, by the factorization theorem, NX is a sufficient

statistic for ™.

f. log f.X; ™/ D � log ™ � X™

and

@ log f.X; ™/

@™D �1

™C X

™2 D X � ™

™2

@2 log f.X; ™/

@™2 D 1™2 � 2X™

™4 D ™ � 2X™3

The Cramer-Rao lower bound for any unbiased estimator O™ of ™ is given by

var�O™

� �1

nE�

@2 log f.X;™/

@™2

D �™3

nE .™ � 2X/D ™2

n.

But var. NX/ D ™2=n, see part (d). Hence, NX attains the Cramer-Rao lower

bound.

g. The likelihood ratio is given by

L.1/

L.2/D e

�nP

iD1Xi

2�ne�

nP

iD1Xi=2

D 2ne�

nP

iD1xi=2

.

The uniformly most powerful critical region C of size ’ � 0.05 is given by

2n e�

nP

iD1Xi=2 � k inside C.

Taking logarithms of both sides and rearranging terms, we get

n log2 � nX

iD1

X=2

!

� K0

Page 15: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 19

ornX

iD1

Xi � K

where K is determined by making the size of C D ’ � 0.05. In this case,nP

iD1Xi is distributed as a Gamma p.d.f. with “ D ™ and ’ D n. Under

Ho; ™ D 1. Therefore,nX

iD1

Xi � Gamma.˛ D n, ˇ D 1/.

Hence, PrŒGamma.’ D n, “ D 1/ � K� � 0.05

K should be determined from the integralR1

K1

�.n/xn�1e�x dx D 0.05 for

n D 20.

h. The likelihood ratio test is

L.1/

L. NX/D e

�nP

iD1Xi

�1NXn

e�n

so that

LR D �2 log L.1/ C 2 log L. NX/ D 2nX

iD1

Xi � 2n log NX � 2n.

In this case,

C .™/ Dˇˇˇˇ@2 log L.™/

@™2

ˇˇˇˇ D

ˇˇˇˇˇ

n™2 �

2nP

i�1Xi

™3

ˇˇˇˇˇ

Dˇˇˇˇˇ

n™ � 2nP

iD1Xi

™3

ˇˇˇˇˇ

and

I.™/ D �E

@2`nL .™/

@™2

D n™2 .

The Wald statistic is based upon

W D � NX � 1�2 I

�O™mle

D � NX � 1

�2 nNX2

using the fact that O™mle D NX. The LM statistic is based upon

LM D S2.1/I�1.1/ D nX

iD1

Xi � n

!21n

D n� NX � 1

�2 .

Page 16: 9783642033827-c1

20 Badi Baltagi

All three test statistics are based upon j NX�1j � k and, for finite n, the same

exact critical value could be obtained using the fact thatnP

iD1Xi is Gamma

(’ D n, and “ D 1) under Ho, see part (g).

2.10 The Gamma Distribution.

a. Using the MGF for the Gamma distribution derived in problem 2.14f, we get

MX.t/ D .1 � “t/�˛ .

Differentiating with respect to t yields

M0X.t/ D �’.1 � “t/�’�1.�“/ D ’“.1 � “t/�’�1.

Therefore

M0x.0/ D ’“ D E.X/.

Differentiating MX0.t/ with respect to t yields

M00x.t/ D �’“.’ C 1/.1 � “t/�’�2 .�“/ D ’“2.’ C 1/.1 � “t/�’�2.

Therefore

M00x .0/ D ’2“2 C ’“2 D E

�X2� .

Hence

var.X/ D E�X2� � .E .X//2 D ’“2.

b. The method of moments equates

E.X/ D NX D ’“

and

E.X2/ DnX

iD1

X2i =n D ’2“2 C ’“2.

Page 17: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 21

These are two non-linear equations in two unknowns. Substitute ’ D NX=“

into the second equation, one getsnX

iD1

X2i =n D NX2 C NX“

Hence,

O“ D

nP

iD1.Xi � NX/2

n NXand

O’ D n NX2

nP

iD1

�Xi � NX�2

.

c. For ’ D 1 and “ D ™, we get

f .X; ’ D 1, “ D ™/ D 1�.1/™

X1�1e�X=™ for X > 0 and ™ > 0,

D 1™

e�X=™

which is the exponential p.d.f.

d. For ’ D r=2 and “ D 2, the Gamma .’ D r=2, “ D 2/ is ¦2r . Hence, from

part (a), we get

E.X/ D ’“ D .r=2/.2/ D r

and

var.X/ D ’“2 D .r=2/.4/ D 2r.

The expected value of a ¦2r is r and its variance is 2r.

e. The joint p.d.f. for ’ D r=2 and “ D 2 is given by

f.X1, .., Xn; ’ D r=2, “ D 2/ D�

1�.r=2/2r=2

�n

.X1 : : : Xn/r2 �1e

�nP

iD1Xi=2

D h.X1, .., Xn; r/g .X1, .., Xn/

Page 18: 9783642033827-c1

22 Badi Baltagi

where

h.X1, .., Xn; r/ D�

1� .r=2/ 2r=2

�n

.X1 : : : Xn/r2 �1

and g.X1, .., Xn/ D e�

nP

iD1Xi=2

independent of r in form and domain. Hence,

by the factorization theorem .X1 : : : Xn/ is a sufficient statistic for r.

f. Let X1, .., Xm denote independent N(0, 1) random variables. Then, X21, .., X2

m

will be independent ¦21 random variables and Y D

mP

iD1X2

i will be ¦2m. The

sum of m independent ¦21 random variables is a ¦2

m random variable.

2.12 The t-distribution with r Degrees of Freedom.

a. If X1, .., Xn are IIN.�, ¢2/, then NX � N.�, ¢2=n/ and z D . NX��/

¢=p

n is N(0, 1).

b. .n � 1/s2=¢2 � ¦2n�1. Dividing our N(0, 1) random variables z in part (a),

by the square-root of our ¦2n�1 random variable in part (b), divided by its

degrees of freedom, we get

t D� NX � �

�=¢=

pn

q.n�1/s2

¢2 =.n � 1/

D . NX � �/

s=p

n.

Using the fact that NX is independent of s2, this has a t-distribution with (n-1)

degrees of freedom.

c. The 95% confidence interval for � would be based on the t-distribution

derived in part (b) with .n � 1/ D 15 degrees of freedom.

t D NX � �

s=p

nD 20 � �

2=p

16D 20 � �

1=2D 40 � 2�

PrŒ�t’=2 < t < t’=2� D 1 � ’ D 0.95

From the t-tables with 15 degrees of freedom, t0.025 D 2.131. Hence

PrŒ�2.131 < 40 � 2� < 2.131� D 0.95.

rearranging terms, one gets

PrŒ37.869=2 < � < 42.131=2� D 0.95

PrŒ18.9345 < � < 21.0655� D 0.95.

Page 19: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 23

2.13 The F-distribution.

.n1 � 1/ s21=�2

1 D 24.15.6/=�21 � 2

24

also

.n2 � 1/s22=¢2

2 D 30 .18.9/ =¢22 � ¦2

30.

Therefore, under Ho; ¢21 D ¢2

2

F D s22=s2

1 D 18.915.6

D 1.2115 � F30,24.

Using the F-tables with 30 and 24 degrees of freedom, we find F.05,30,24 D 1.94.

Since the observed F-statistic 1.2115 is less than 1.94, we do not reject Ho that

the variance of the two shifts is the same.

2.14 Moment Generating Function (MGF).

a. For the Binomial Distribution,

Mx.t/ D E.eXt/ DnX

XD0

nX

!

eXt™X .1 � ™/n�X

DnX

XD0

nX

!�™et�X

.1 � ™/n�X

D �.1 � ™/ C ™et�n

where the last equality uses the binomial expansion .a C b/n DnP

XD0

nX

!

aXbn�X with a D ™et and b D .1 � ™/. This is the fundamental

relationship underlying the binomial probability function and what makes it

a proper probability function.

b. For the Normal Distribution,

MX .t/ D E�eXt� D

Z C1

�1eXt 1

�p

2e� 1

2�2 .X��/2dx

D 1�

p2

Z C1

�1e� 1

2�2 fX2�2�XC�2�Xt2�2gdx

D 1¢

p2 

Z C1

�1e� 1

2¢2 fX2�2.�Ct¢2/XC�2gdx

Page 20: 9783642033827-c1

24 Badi Baltagi

completing the square

Mx.t/ D 1¢

p2 

Z C1

�1e� 1

2¢2

nŒx�.�Ct¢2/�

2�.�Ct¢2/2C�2

o

dx

D e� 12¢2 Œ�2��2�2�t¢2�t2¢4�

The remaining integral integrates to 1 using the fact that the Normal den-

sity is proper and integrates to one. Hence Mx.t/ D e�tC 12 ¢2t2 after some

cancellations.

c. For the Poisson Distribution,

Mx.t/ D E.eXt/ D1X

XD0

eXt e���X

XŠD

1X

XD0

e��.�et/X

D e�œ

1X

XD0

.œet/X

XŠD e�et � œ D eœ.et�1/

where the fifth equality follows from the fact that1P

XD0

aX

XŠD ea and in this

case a D œet. This is the fundamental relationship underlying the Poisson

distribution and what makes it a proper probability function.

d. For the Geometric Distribution,

MX.t/ D E.eXt/ D1X

XD1

™.1 � ™/X�1eXt D ™

1X

XD1

.1 � ™/X�1e.X�1/tet

D ™et1X

XD1

�.1 � ™/ et�X�1 D ™et

1 � .1 � ™/et

where the last equality uses the fact that1P

XD1aX�1 D 1

1�a and in this case

a D .1�™/et. This is the fundamental relationship underlying the Geometric

distribution and what makes it a proper probability function.

Page 21: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 25

e. For the Exponential Distribution,

MX.t/ D E.eXt/ DZ 1

0

1™

e�X=™eXtdx

D 1™

Z 1

0e�XŒ 1

™�t�dx

D 1™

Z 1

0e�XŒ 1�™t

™ �dx

D 1™

� �™

.1 � ™t/

he�X. 1�™t

™ /i1

0D .1 � ™t/�1

f. For the Gamma Distribution,

MX.t/ D E.eXt/ DZ 1

0

1� .’/ “’

X’�1e�X=“eXt dx

D 1� .’/ “’

Z 1

0X’�1e�X

�1“

�t

dx

D 1�.’/“’

Z 1

0X’�1e�X

�1�“t

dx

The Gamma density is proper and integrates to one using the fact thatR1

0 X’�1e�X=“dx D �.’/“’. Using this fundamental relationship for the

last integral, we get

Mx.t/ D 1� .’/ “’

� �.’/

�“

1 � “t

�’

D .1 � “t/�’

where we substituted “=.1 � “t/ for the usual “. The ¦2r distribution is

Gamma with ’ D r2 and “ D 2. Hence, its MGF is .1 � 2t/�r=2.

g. This was already done in the solutions to problems 5, 6, 7, 9 and 10.

2.15 Moment Generating Function Method.

a. If X1, .., Xn are independent Poisson distributed with parameters .œi/ respec-

tively, then from problem 2.14c, we have

MXi.t/ D e�i.et�1/ for i D 1, 2, : : : , n

Page 22: 9783642033827-c1

26 Badi Baltagi

Y DnP

iD1Xi has MY.t/ D

nQ

iD1MXi.t/ since the Xi

0s are independent. Hence

MY.t/ D enP

iD1�i.et�1/

which we recognize as a Poisson with parameternP

iD1œi.

b. If X1, .., Xn are IIN��i, ¢2

i�, then from problem 2.14b, we have

MXi.t/ D e�itC 12 ¢2

i t2 for i D 1, 2, .., n

Y DnP

iD1Xi has MY.t/ D

nQ

iD1MXi.t/ since the Xi

0s are independent. Hence

MY.t/ D e

nP

iD1�i

!

tC 12

nP

iD1¢2

i

!

t2

which we recognize as Normal with meannP

iD1�i and variance

nP

iD1¢2

i .

c. If X1, .., Xn are IIN.�, ¢2/, then Y DnP

iD1Xi is N.n�, n¢2/ from part b using

the equality of means and variances. Therefore, NX D Y=n is N.�, ¢2=n/.

d. If X1, .., Xn are independent ¦2 distributed with parameters .ri/ respectively,

then from problem 2.14f, we get

MXi.t/ D .1 � 2t/�ri=2 for i D 1, 2, .., n

Y DnP

iD1Xi has MY.t/ D

nQ

iD1MXi.t/ since the Xi

0s are independent. Hence,

MY.t/ D .1 � 2t/�

nP

i�1ri=2

which we recognize as ¦2 with degrees of freedomnP

iD1ri.

2.16 Best Linear Prediction. This is based on Amemiya (1994).

a. The mean squared error predictor is given by MSE D E.Y � ’ � “X/2 DE.Y2/ C ’2 C “2E.X2/ � 2’E.Y/ � 2“E.XY/ C 2’“E.X/ minimizing this

Page 23: 9783642033827-c1

Chapter 2: A Review of Some Basic Statistical Concepts 27

MSE with respect to ’ and “ yields the following first-order conditions:

@MSE@’

D 2’ � 2E.Y/ C 2“E.X/ D 0

@MSE@“

D 2“E.X2/ � 2E.XY/ C 2’E.X/ D 0.

Solving these two equations for ’ and “ yields O’ D �Y � O“�X from the

first equation, where �Y D E.Y/ and �X D E.X/. Substituting this in the

second equation one gets

O“E.X2/ � E.XY/ C �Y�X � O“�2X D 0

O“var.X/ D E.XY/ � �X�Y D cov.X, Y/.

Hence, O“ D cov.X, Y/=var.X/ D ¢XY=¢2X D ¡¢Y=¢X, since ¡ D

¢XY=¢X¢Y. The best predictor is given by OY D O’ C O“X.

b. Substituting O’ into the best predictor one gets

OY D O’ C O“X D �Y C O“.X � �X/ D �Y C ¡¢Y

¢X.X � �X/

one clearly deduces that E. OY/ D �Y and var. OY/ D ¡2 ¢2Y

¢2X

var.X/ D ¡2¢2Y.

The prediction error Ou D Y � OY D .Y � �Y/ � ¡ ¢Y¢X

.X � �X/ with E.Ou/ D 0

and var.Ou/ D E.Ou2/ D var.Y/ C ¡2 ¢2Y

¢2X

var.X/ � 2¡ ¢Y¢X

¢XY D ¢2Y C ¡2¢2

Y �2¡2¢2

Y D ¢2Y.1 � ¡2/.

This is the proportion of the var(Y) that is not explained by the best linear

predictor OY.

c. cov. OY, Ou/ D cov. OY, Y � OY/ D cov. OY, Y/ � var. OY/

But

cov. OY, Y/ D E� OY � �Y

.Y � �Y/ D E

¡¢Y

¢X.X � �X/ .Y � �Y/

D ¡¢Y

¢Xcov .X, Y/ D ¡2 ¢X¢2

Y¢X

D ¡2¢2Y

Hence, cov. OY, Ou/ D ¡2¢2Y � ¡2¢2

Y D 0.

Page 24: 9783642033827-c1

28 Badi Baltagi

2.17 The Best Predictor.

a. The problem is to minimize EŒY � h.X/�2 with respect to h(X). Add and

subtract E(Y/X) to get

EfŒY � E.Y=X/� C ŒE.Y=X/ � h.X/�g2

D EŒY � E.Y=X/�2 C EŒE.Y=X/ � h.X/�2

and the cross-product term EŒY �E.Y=X/�ŒE.Y=X/�h.X/� is zero because

of the law of iterated expectations, see the Appendix to this chapter or

Amemiya (1994). In fact, this says that expectations can be written as

E D EXEY=X

and the cross-product term given above EY=XŒY�E.Y=X/�ŒE.Y=X/�h.X/�

is clearly zero. Hence, EŒY � h.X/�2 is expressed as the sum of two positive

terms. The first term is not affected by our choice of h(X). The second term

however is zero for h.X/ D E.Y=X/. Clearly, this is the best predictor of Y

based on X.

b. In the Appendix to this chapter, we considered the bivariate Normal dis-

tribution and showed that E.Y=X/ D �Y C ¡ ¢Y¢X

.X � �X/. In part (a), we

showed that this is the best predictor of Y based on X. But, in this case, this

is exactly the form for the best linear predictor of Y based on X derived in

problem 2.16. Hence, for the bivariate Normal density, the best predictor is

identical to the best linear predictor of Y based on X.

Page 25: 9783642033827-c1

http://www.springer.com/978-3-642-03382-7