Sample Answers to Analytical Exercises of Chapter 8

Exercise 1 Suppose the moment conditions are E [g(wi ; )] = 0 with g(w; ) = g1 (w) g2 (w) .
Set up Jn ( ) as in (??) and derive the asymptotic distribution of the corresponding GMM estimator
of .

Solution: This is parallel to the case where g(wi ; ) = zi (yi x0i ). De…ne G = E [g2 (w)] and
=E [g(w; )g(w; )0 ]; then
p d
n b GM M ! N (0; V) ;

where
1 1
V = G0 WG G0 W WG G0 WG :

Exercise 2 In the linear model estimated by GMM with general weight matrix W, the asymptotic
variance of b GM M is V in (??).

1. 1
(i) Let V0 be this matrix when W = Show that V0 = G0 1G .

(ii) We want to show that for any W, V V0 0. To do this, start by …nding matrices A and
B such that V = A0 A and V0 = B0 B.

(iii) Show that B0 A = B0 B and therefore B0 (A B) = 0.

(iv) Use the expressions V = A0 A, A = B + (A B), and B0 (A B) = 0 to show that

V V0 .

Solution:
1 1 1 1
(i) V0 = G0 1 G G0 1 G G0 1 G = G0 1G :
(ii) A = WG (G0 WG) 1 , and B = 1 G G0 1G 1.

(iii)

1 1
B0 A = G0 1
G G0 1
WG G0 WG
1
= G0 1
G
1 1
= G0 1
G G0 1 1
G G0 1
G
= B0 B;
Email: [email protected]

1
so B0 (A B) = 0.
(iv) V V0 = A0 A B0 B = (B+(A B))0 (B+(A B)) B0 B = (A B)0 (A B) 0.

Exercise 3 Show that when a new group of instrumental variables is added in, the optimal asymp-
totic variance matrix V0 will not increase. Discuss when the two asymptotic variance matrices will
be equal. (Hint: use the result in Exercise 2.)

Solution: Denote the original!instruments!as z1 2 Rl1 , the new group of instruments as z2 2 Rl2 ,
g1 (wi ; ) z1 u
and g(wi ; ) = = . When the new group of instruments is added in,
g2 (wi ; ) z2 u
0 1 1 1
the asymptotic variance
! matrix
! changes from V1 = G1 11 G1 to!V = G0 1 G ! , where
0
E[z1 x ] G1 0 2 0 2
E[z1 z1 u ] E[z1 z2 u ] 11 12
G = 0
= , and = 0 2 0 2
. De-
E[z2 x ] G E[z2 z1 u ] E[z2 z2 u ]
! 2 21 22
1
0
…ne W = 11
. We can check that W W = W, and G0 WG = G01 111 G1 . So
0 0
1 1 1
(G0 WG) 1 (G0 W WG) (G0 WG) 1 = G01 111 G1 . From Exercise 2, G0 1 G G01 1
11 G1 .
The equality will hold if 1 = W. From the partitioned matrix inversion formula,

!
1 1 1
1 11:2 11:2 12 22
= 1 1 1
;
22:1 21 11 22:1

1 1 1. A
where ii:j = ii ij jj ji , i; j = 1; 2. Since 22:1 > 0, W cannot be equal to
su¢ cient condition for the equality to hold is E[z2 x0 ] = 0 and E[z1 z02 u2 ] = 01 . Intuitively, if z2 is
not correlated with the existing instruments and the covariates, then it is not correlated with the
existing system at all and so will not provide any further information.
1 1
For a general "if and only if" condition under which G0 1 G = G01 111 G1 , we orthog-
1
onalize the two groups of moment conditions. De…ne r2 (wi ; ) =!g2 (wi ; ) ! 21 11 g1 (wi ; ) =
1 G1 G1
21 11 ; Il2 g(wi ; ); now G changes to 1
and changes to
G2 21 11 G1 D2
" ! # ! !
g1 (wi ; ) 11 0 11 0
E g1 (wi ; )0 ; r2 (wi ; )0 = 1
:
r2 (wi ; ) 0 22 21 11 12 0 22

These di¤erent formulations of moment

! conditions would ! generate the same e¢ cient estimator of
g1 (wi ; ) Il1 0
by noticing that = 1
g(wi ; ), so the original optimal weight
r2 (wi ; ) 21 11 Il 2
matrix 1 applied to g (w ; ) and the new optimal weight matrix
i

" ! !# 1 ! !
1 1
I l1 0 I l1 11 12 Il 1 11 12 1 I l1 0
1
= 1
21 11 Il2 0 Il 2 0 I l2 21 11 I l2
1
In the homoskedastic case, this is equivalent to E[z1 z02 ] = 0.

2
 !
g1 (wi ; )
applied to would generate the same estimator of . It is not hard to see that the
r2 (wi ; )
inverse of the asymptotic variance matrix of this e¢ cient estimator of is G0 1 G = G01 111 G1 +
1 1
D02 221 D2 . So G0 1 G = G01 111 G1 if and only if D02 221 D2 = 0 or D2 = 0 or G2 =
1 2
21 11 G1 . For more discussions, see Breusch et al. (1999).

Exercise 4 Take the single equation

y = X + u; E[ujZ] = 0:

Assume E[u2i jzi ] = 2. Show that if b is estimated by GMM with weight matrix Wn = (Z0 Z) 1,

then
p d 1
n b ! N 0; 2
G0 M 1
G ;

where G = E[zi x0i ] and M = E[zi z0i ].

Solution: Consider the moment conditions: E [zi ui ] = E [zi (yi x0i )] = 0. Then, the GMM
estimator is written as
! 1
1 1
b= X0 Z Z0 Z Z0 X X0 Z Z0 Z Z0 u
+ :
n n n n n n

Z0 X p Z0 Z p Z0 u p
By the LLN, ! E[zi x0i ] = G; ! E[zi z0i ] = M; ! E[zi ui ] = 0. By the CLT,
n n n
p Z0 u d p 1
n ! N 0; E zi z0i u2i = N 0; 2M . Therefore, b ! + G0 M 1G G0 M 1 0= ,
n
and
p d 1 1
n b ! G0 M 1
G G0 M 1
N 0; 2
M = N 0; 2
G0 M 1
G :

Exercise 5 In Exercise 12 of Chapter 7, …nd the e¢ cient GMM estimator of based on the
moment condition E[zi (yi xi )] = 0. Does this di¤ er from 2SLS and/or OLS?

Solution: Let us use the 2SLS estimator to get the estimate of the optimal weight matrix. Compute
bi = yi xi b 2SLS . Then the e¢ cient GMM estimator
2SLS residuals: u
" n
#" n
!# 1" n !#
X X
x2i x3i X xi
b = arg min (yi xi ) xi x2i b2i
u (yi xi )
i=1 i=1
x3i x4i i=1
x2i
2 " !# 1 n !3 1
Xn n
X X
4 x2i xi 3 2
xi 5
= x2i x3i b2i
u
i=1 i=1
x3i xi 4
i=1
x3i
n
" n
!# 1 n !
X X x2i x3i X xi yi
x2i x3i b2i
u :
i=1 i=1
x3i x4i i=1
x 2y
i i

2
In the homoskedastic case, this is equivalent to E[z2 x0 ] E[z2 z01 ]E[z1 z01 ] 1 E[z1 x0 ] =
E[z2 1 z01 E[z1 z01 ] 1 z1 x0 ] = 0, which is in turn equivalent to 2 = 0 in the …rst stage regression x = 1 z1 + 2 z2 +v
by the FWL theorem.

3
It di¤ers from the 2SLS/OLS. Even under homoskedasticity, it’s di¤erent from them (although
asymptotically equivalent).

bi = yi x0i e where e is consistent

Exercise 6 Take the model yi = x0i + ui with E[zi ui ] = 0. Let u
for (e.g., a GMM estimator with arbitrary weight matrix). De…ne the estimate of the optimal
GMM weight matrix
n
! 1
1X 0 2
Wn = zi zi ubi :
n
i=1
p 1
Show that Wn ! where = E[zi zi u2i ].

bi = ui + x0i
Solution: Note that u b , so

2
b2i = u2i + 2ui x0i
u b + x0 b :
i

and
n n n n
1X 0 2 1X 0 2 2X 0 b +1
X
b
2
bi =
zi zi u zi zi u i + zi zi ui x0i zi z0i x0i :
n n n n
i=1 i=1 i=1 i=1

The 1st term goes to in probability by the LLN. As for the 2nd and 3rd term, the consistency
of b implies
! !
n
1X 0 1X
n
p
h i
zi zi ui x0i b kzi k2 kxi k jui j b ! E kzi k2 kxi k jui j 0 = 0;
n n
i=1 i=1

where k k is the Euclidean norm of matrix, the inequality is from the triangle inequality and the
Schwarz matrix
h inequality,i and the convergence is from the WLLN and CMT. To apply the WLLN,
we need E kzi k2 kxi k jui j < 1, which is implied by

h i h i1=2 h i1=2
E kzi k2 kxi k jui j E kzi k4 E kxi k2 jui j2
h i1=2 h i1=4
1=4
E kzi k4 E kxi k4 E jui j4 < 1;
h i
where both inequalities are from the Cauchy-Schwarz inequality, and we assume E kzi k4 <
h i
1; E kxi k4 < 1 and E jui j4 < 1. Similarly, we have

1 Pn 0 b
2
n i=1 zi zi x0i

1 Pn b b
0
n i=1 zi z0i x0i xi
1 Pn 2 2 b
2
n i=1 kzi k kxi k
b 1 Pn kzi k2 kxi k2
2
= n i=1
p
h i
!E kzi k kxi k2 0 = 0;
2

4
 h i Pn 2 p
where E kzi k2 kxi k2 < 1 by the Cauchy-Schwarz inequality. So 1 0 x0i b !
n i=1 zi zi
0, and we have
n
! 1
1X 0 2 p
Wn = bi
zi zi u ! :
n
i=1
Pn
Exercise 7 Suppose we want to estimate 2 in Exercise 4. (i) Show that n 1 b2i
i=1 u is consistent,
bi = yi xi b . (ii) In the structural model
where u

y = X1 1 + X2 2 + u;
X2 = X1 12 + Z2 22 + V;
0
show that n 1 l2 1; b0 Y0 MZ Y 1; b0 is also a consistent estimator of 2, where
2;2SLS 2;2SLS
Y = (y; X2 ).

Solution: (i) This proof is similar to that of Theorem 4 in Chapter 5 with xi = 1.

0 0
(ii) Y 1; b 2;2SLS = X1 b 1;2SLS + u b . Since Z = [X1 ; Z2 ],

0
MZ Y 1; b0 b
= MZ u
2;2SLS

= MZ X1 1 + X2 2 +u X1 b 1;2SLS X2 b 2;2SLS

= MZ X2 b + MZ u:
2 2;2SLS

As usual, MZ X2 b will not contribute to the probability limit, so we need only to

2 2;2SLS
show that
1 p
u0 MZ u ! 2
;
n l2
which is equivalent to show that
1 0 p
u PZ u !0; (1)
n
but which is obvious from the moment condition E[zi ui ] = 0.

Exercise 8 The equation of interest is

yi = m(xi ; ) + ui ; E[zi ui ] = 0:

The observed data is (yi ; xi ; zi ), zi is l 1 and is k 1, l k. Show how to construct an e¢ cient

GMM estimator for .

Solution: The l 1 moment conditions are E[gi ( )] = E [zi (yi m(xi ; ))] = 0. The optimal
weighting matrix is W0 = (E[gi ( )gi ( )0 ]) 1 = (E[zi z0i (yi m(xi ; ))2 ]) 1 which can be consis-
1
1 Pn e ))2
tently estimated by Wn = z i z 0 (y
i i m(xi ; , where e is a consistent estimator of
n i=1
: In summary, we can run a two step procedure:

5
 Step 1. Find an initial consistent estimate b GM M , for example, by

n
! n
!
X X
b = arg min (yi m(xi ; ))z0i (yi m(xi ; ))zi :
GM M
i=1 i=1

Step 2. Find the e¢ cient GMM estimate b by

n
! n
! 1 n
!
b = arg min 1X 1X 1X
(yi m(xi ; ))z0i (yi m(xi ; b )) 2
zi z0i (yi m(xi ; ))zi :
n n n
i=1 i=1 i=1

Exercise 9 Take the linear model

yi = x0i + ui ;
E[zi ui ] = 0;

and consider the GMM estimator b of . Let

0
Jn = ng n b b 1
gn b

d 2
denote the test of overidentifying restrictions. Show that Jn ! l k by demonstrating each of the
following:

(i) Since > 0, we can write 1 = CC0 and = C0 1C 1.

0 1
(ii) Jn = n C0 g n b C0 b C C0 g n b .

(iii) C0 g n b = Dn C0 g n ( 0) where

1
1 0 0 1 0 b 1 0 1 0 b
Dn = Il C0 ZX XZ 1
ZX XZ 1
C0 1
;
n n n n
1 0
gn ( 0) = Z u:
n

p
(iv) Dn !I l R(R0 R) 1 R0 where R = C0 E[zi x0i ] = C0 G:
d
(v) n1=2 C0 g n ( 0) !Z N (0; Il ).
d
(vi) Jn ! Z 0 Il R(R0 R) 1 R0 Z.

(vii) Z 0 Il R(R0 R) 1 R0 Z 2 .
l k

(Hint: I l R(R0 R) 1 R0 is a projection matrix. Note also the di¤ erence in the notation Z
and Z.)

6
Solution: (i) Since 1 is symmetric, there exists a matrix H such that (s.t.) HH0 = I and
1 = H H0 , where
0 1
1 0
B . .. .. C
=B .
@ . . . C
A
0 k

is the eigenvalue matrix.

Since 1 is positive de…nite, all eigenvalues are positive. Therefore,

0 p 1
1 0
B .. .. .. C
1=2
=B
@ . . . C
A
p
0 k

is well-de…ned. Set C = H 1=2 ; then we have 1 = H 1=2 1=2 H0 = CC0 .

(ii) Jn = ngn ( b )0 CC 1 b 1 (C0 ) 1 C0 gn ( b ) = n(C0 gn ( b ))0 (C0 b C) 1 (C0 gn ( b )):

(iii) Since

1 1
b 1 0 b 11 1 0 b 11 1 0 b 11 1 0 b 1
0 = XZ Z0 X XZ Z0 u = XZ Z0 X XZ gn ( 0 );
n n n n n n n

we have

1 0 1 0h i
gn ( b ) = Z (y Xb) =
Z (y X 0 ) X( b 0)
n n
1
1 0 1 0 b 11 0 1 0 b 1
= gn ( 0 ) ZX XZ ZX XZ gn ( 0 )
n n n n
" #
1
1 0 1 0 b 11 0 1 0 b 1
= Il ZX XZ ZX XZ gn ( 0 )
n n n n
" #
1
1 1 1 1 1 1
) C0 gn ( b ) = C0 C0 Z0 X X0 Z b Z0 X X0 Z b gn ( 0 )
n n n n
" #
1
1 1 1 1 1 1 1
= Il C0 Z0 X X0 Z b Z0 X X0 Z b C0 C0 gn ( 0)
n n n n
= Dn C0 gn ( 0)

(iv)

1
0 1 0 1 0 b 11 0 1 0 b 1 1
Dn = Il C ZX XZ ZX XZ C0
n n n n
p 1 1
! Il C0 E[zi x0i ] E[xi z0i ] 1
E[zi x0i ] E[xi z0i ] 1
C0
1 1 1
= Il C0 E[zi x0i ] E[xi z0i ]CC0 E[zi x0i ] E[xi z0i ]CC0 C0 = Il R R0 R R0 :

7
 p d
(v) Since ngn ( 0) = p1 Z0 u ! N (0; ), where = E[zi z0i u2i ],
n

p d 1
nC0 gn ( 0) ! C0 N (0; ) = N (0; C0 (C0 C 1
)C) = N (0; Il ):

(vi) By the the results of (ii) and (iii),

p 0 1 p
Jn = nC0 gn ( 0) D0n C0 b C Dn nC0 gn ( 0)

d 0 1 1
! N (0; Il )0 Il R(R0 R) 1
R0 C0 C0 C 1
C Il R(R0 R) 1
R0 N (0; Il )
= N (0; Il )0 Il R(R0 R) 1
R0 N (0; Il ):

(vii) Set M = Il R(R0 R) 1 R0 ; then since M is symmetric, there exists a matrix H s.t.
M = H0 M H and HH0 = Il , where M denotes the eigenvalue matrix of M. Therefore,

N (0; Il )0 MN (0; Il ) = N (0; Il )0 H0 M HN (0; Il ) = N (0; HH0 )0 0

M N (0; HH ) = N (0; Il )0 M N (0; Il ):

Now we use the following properties of the trace:

tr(A + B) = tr(A) + tr(B); tr(AB) = tr(BA):

Since R is l k,

tr(M) = tr(Il ) tr(R(R0 R) 1

R0 ) = tr(Il ) tr((R0 R) 1
R0 R) = tr(Il ) tr(Ik ) = l k:

On the other hand,

tr(M) = tr(H0 M H) = tr( M HH0 ) = tr( M );
P
thus we have tr( M ) = li=1 i = l k. Moreover, since M satis…es MM = M, eigenvalues of M
are either 0 or 1. These facts imply that M has the following form:
!
Il k O
M = ;
O Ok

where O represents a zero matrix.

Consequently, we get:

l k
X
d
Jn ! N (0; Il )0 M N (0; Il ) = Zj2 2
l k;
j=1

i:i:d:
where Zj N (0; 1).

b0 P e u
u b
Exercise 10 In the homoskedastic linear model (??), show that (i) Jn = nRu2 = ub 0 ubZ=n
2
, where Ru2
e 2 = MX Z 2 ; (ii) l2 F has the same asymptotic
is the uncentered R2 in the regression (??), and Z 1

8
distribution as Jn .

Solution: (i) In the homoskedastic model,

n
! n
! 1 n
!
1X 1X 0 1X
Jn = n bi z0i
u b2u zi zi bi
zi u
n n n
i=1 i=1 i=1
b Z (Z0 Z) 1 Z0 u
u0 b
= = nRu2 :
b0u
u b =n

The numerator is the projection of ub on span(Z). Since u b is orthogonal to X1 , this is equivalent

b0 P u
u b
e 2 = MX Z2 . In other words, Jn = 0 Ze 2 . Note also that u
to the projection on Z b 0 Z (Z0 Z) 1 Z0 u
b
1 bu
u b =n
is the sum square residuals of the "second-stage" regression of the …tted y (PZ y) on the …tted X
(PZ X).
(ii) From Theorem 4 of Chapter 3,

(SSRR SSUU ) u0 u
(b b ub 0 MZ u
b) b 0 PZ u
u b
l2 F = = 0 = 0 :
SSU=(n l) b MZ u
u b =(n l) b MZ u
u b =(n l)

The numerator of l2 F is the same as Jn , so as long as u b 0 MZ u

b =(n l) has the same probability
0 0 d
bu
limit as u b =n, we are done. This is equivalent to show that ub PZ ub =n ! 0, which is exactly (1).

Exercise 11 We use the same setup and notations as in Exercise 9 except that E[z(y x0 n )] =
p
E [zu] = = n. Show the following results.
p d 1 1
(i) n b n !N G0 1G G0 1 ; V , where V = G0 1G .

d 0
(ii) Jn ! 2 (
l k ), where = 1( GVG0 ) 1 .

(iii) = 0 if and only if 2 span(G).

Solution: (i) Note that

1
p 1 0 b 11 1 0 b 1 1 0
n b n = XZ Z0 X XZ p Zu
n n n n
Xn
1 1 1
= G0 1
G G0 p f(zi ui E [zi ui ]) + E [zi ui ]g + op (1)
n
i=1
d 0 1 1 0 1 1
! G G G N (0; ) + G0 1
G G0 1

1 1
= N G0 1
G G0 1
; G0 1
G ;

p
where the second equality is due to b ! even under the local alternative (see Exercise 18 of
p
Chapter 5), and the third equality is because E [zi ui ] = = n.

9
 (ii) Repeating the analysis in Exercise 9, we can see

0 1
Jn = n C0 g n b C0 b C C0 g n b ;
p
C0 g n b = Dn C0 g n ( n ) ; Dn ! Il R(R0 R) 1
R0 with R = C0 G

still hold, and the only di¤erence is that

d
n1=2 C0 g n ( n) ! N C0 ; Il :

As a result,
d
Jn ! N C0 ; Il Il R(R0 R) 1
R0 N C0 ; Il = 2
l k( );

where = 0 C Il R(R0 R) 1 R0 C0 = 0 1 ( GVG0 ) 1 .

2
(iii) Given that = Il R(R0 R) 1 R0 C0 , = 0 is equivalent to Il R(R0 R) 1 R0 C0 =
0. Because rank Il R(R0 R) 1 R0 C0 = l k, the dimenion of such that = 0 must be k.
Because Il R(R0 R) 1 R0 C0 G = 0 and dim (span(G)) = k, the result follows.

Exercise 12 Take the linear model

yi = x0i + ui ;
E[zi ui ] = 0;

and consider the unrestricted GMM estimator b and restricted GMM estimator e of under the
linear constraints R0 = c. De…ne

Jn ( ) = n g n ( ) b 1
gn ( ) ;

and then b = arg minJn ( ) and e = arg min

0
Jn ( ). De…ne the Lagrangian
R =c

1 0
L( ; ) = Jn ( ) + R0 c :
2

(i) Show that

1 1 1
e = b X0 Z b 1
Z0 X R R0 X0 Z b 1
Z0 X R R0 b c ;
1
b = 1 1
R0 X0 Z b 1
Z0 X R R0 b c :
n

(ii) Derive the asymptotic distribution of e under the null.

(iii) Show that

1 b 0
Jn e = Jn b + e X0 Z b 1
Z0 X b e :
n

10
(iv) Show that the distance statistic is equal to the Wald statistic.

Solution: (i) The FOCs for the Lagrange problem are

1 0 1
n XZ b 1
Z0 y Z0 X e + R = 0;
n n

so
1
e=b 1 1 0 b 1 1 0
XZ ZX R :
n n n

Substitute this into the constraint R0 = c to obtain the expression for b in the question. Then
substitute this expression for b into the FOCs to obtain the expression for e in the question.
(ii) When R0 = c,

1
p p 1 1 p
n e = n b X0 Z b 1
Z0 X R R0 X0 Z b 1
Z0 X R R0 n b
( )
1
1 1 p
= Ik X0 Z b 1
Z0 X R R0 X0 Z b 1
Z0 X R R0 n b
h i 1 p n o
d 1
= Ik b
VR b
R0 VR R0 n b ! Ik VR R0 VR R0 N (0; V) ;

1 p 1
where V b = 1 0 b 1 1 Z0 X ! V = E [xi z0i ] 1E [zi x0i ] .
nX Z n
(iii)
0
1 0 1 0 e 1 0 1 0 e
Jn e = n Zy ZX b 1
Zy ZX
n n n n
0
1 1 0 1 1 0
= n Z0 y Z0 X b + ZX b e b 1
Z0 y Z0 X b + ZX b e
n n n n
1 b 0
= Jn b + e X0 Z b 1
Z0 X b e :
n

where the third equality uses the FOCs for the unrestricted GMM:

1 0 1
XZ b 1
Z0 y Z0 X b = 0:
n n
0
(iv) What needs to be shown is that 1 b e X0 Z b 1 Z0 X b e equals the Wald statistic.
n
But this is immediate from substitution of the expression for e in (i):

1 b e 0 0 b 1 0
XZ ZX b e
n
1
1 0 1 1
= R0 b c R0 X0 Z b 1 Z0 X R R0 X0 Z b 1
Z0 X X0 Z b 1
Z0 X
n
1 1 1
X0 Z b 1
Z0 X R R0 X0 Z b 1
Z0 X R R0 b c

11
 1
1 0 1
= R0 b c R0 X0 Z b 1
Z0 X R R0 b c
n
0h i 1
= n R0 b c b
R0 VR R0 b c :

Exercise 13 Suppose 2 R, H0 is c = 0, and the curvature of Qn ( ) at e is e and at b is

b . Show that (i) Wn is twice the di¤ erence in Qn ( ) at e and b, using a quadratic approximation
to Qn ( ) at b; (ii) LMn is twice the di¤ erence in Qn ( ) at e and b, using a quadratic approximation
to Qn ( ) at e.

Solution: (i) The quadratic approximation to Qn ( ) at b is

1b b
2
+ Qn b :
2
2
The height di¤erence of this approximation at e = c and b is 1b
2 c b , which is equal to Wn =2
by noticing that r b = b c and R(b) = 1.
(ii) The quadratic approximation to Qn ( ) at e is
" #
1e e
2 @Qn (e) @Qn (e) e
+ + Qn (e) :
2 @ @

e
. 2
@Qn (e)
The height di¤erence of this approximation at e and the highest point e+ @Q@n ( ) e is 1 1
2e @ ,
which is equal to LMn =2 by noticing that e 1 = 1= e when e is a scalar.

Exercise 14 When 2u E u2 6= 1, what is the form of the Wald statistic, and what is its
asymptotic distribution?

Solution: If E u2 6= 1, then

1 1 1 1
e2 Z
u0 Z e0 Z
e e 0 Y2 Y 0 Z
Z e e0 e e 0 Y2 e2 Z
Y20 Z e0 Z
e e0 u
2 2 2 2 2 Z2 Z2 Z 2 2 2 Z2
Wn = ;
b2u

where b2u = 1
b0u
b b=y
with u e e 0 b 2, y e 2 = MZ Y2 , and
e = MZ1 y, Y
n ku Y 2 1

1 1 1
b = e2
Y20 Z e0 Z
Z e e 0 Y2
Z e2 Z
Y20 Z e0 Z
e e0 y :
Z
2 2 2 2 2 2 2

Now,

1 e0 d
p Z u ! Q1=2 0 u ;
n 2
1 e0 1 e0 e 1 e0 d
p Z Y2 = Z2 Z2 C + p Z V2 ! Q1=2 C 1=2
+ Q1=2 2
1=2
;
n 2 n n 2

12
where !
0
N 0; I Il 2 ;
vec ( 2 )
!
1 0
1=2 2 1=2
where I = , and = E [v2 u] u . So
Ik 2

1 1 1
b = e2 Z
Y20 Z e0 Z
e e 0 Y2
Z e2 Z
Y20 Z e0 Z
e e0 u
Z
2 2 2 2 2 2 2 2
h i 1
d 1=2 0 0
! u C+ 2 C+ 2 C+ 2 0
1=2 1
u 1 2:

As a result,

b=y
u e e 0 b 2 = MZ u
Y e 2 b2
Y 2 = MZ1 u e 2n
Z 1=2
C + MZ1 V2 b
2 2 ;
2 1

and

1
b2u = b0u
u b
n k
1 0
= u V2 b 2 2 u V2 b 2 2 + op (1)
n
d 2 1=2 1 2 0 1 1=2 1=2 1
! u 2 uE [uv2 ] 1 2 + u 2 1 1 2:

Combining all these elements together, we have

0 1
d 2 1 2
Wn ! 0 1 0 2 :
1 2 1 2+ 2 1 2

Exercise 15 In the LIML estimator, what is the asymptotic distribution of n (b 1) when 22 =

n 1=2 C?

Solution: From the analysis in Appendix B of Chapter 7, we can see n (b 1) is the smallest root
1
e2 Z
of Y0 Z e0 Z
e e0 Y
Z 1 0
= 0, and
2 2 2 n Y MZ Y

MZ Y = MZ u, M1 Y = M1 u;

0 0
where = 1; 2 . Note that this smallest root is the same as the smallest root of
! ! ! !
1 0
u 00 1 2 0e e
e0 Z
1
e0 Y 1 0 1 00 u
1
00
1=2
Y Z2 Z2 2 Z2 Y MZ Y 1=2
(2)
0 0 I k2 n 2 Ik 2 0
! !
1
u 00 e2 Z
0 e0 Z
e
1
e 0 (u; Y2 ) 1 0 u
1
00
= 1=2
(u; Y2 ) Z 2 2 Z 2 (u; V2 ) MZ (u; V2 ) 1=2
= 0:
0 n 0

13
From the main text and the last exercise,

1 e0 e p
Z Z2 ! Q;
n 2
1 e0 d
p Z u ! Q1=2 0 u ;
n 2
1 e0 d
p Z Y2 ! Q1=2 C + 2
1=2
:
n 2

Also,
1
1 0 1 0 1 0 1 0 1 0 p 2
u MZ u = uu uZ ZZ Zu ! u;
n n n n n
1
1 0 1 0 1 0 1 0 1 0 p
u MZ V2 = u V2 uZ ZZ Z V2 ! E [uv2 ] ;
n n n n n
1
1 0 1 0 1 0 1 0 1 0 p
V MZ V2 = V V2 V Z ZZ Z V2 ! :
n 2 n 2 n 2 n n

As a result,

e2 Z
e0 Z
e
1
e 0 (u; Y2 ) 1
(u; Y2 )0 Z 2 2 Z 2 (u; V2 )0 MZ (u; V2 )
n !
0 2 E [uv2 ]
d 1=2 1=2 u
! 0 u; C+ 2 0 u; C+ 2 ;
E [v2 u]

and the expression in (2) converges in distribution to

0
0; C + 2 0; C + 2 I ;

where I is de…ned in the last exercise. By the CMT, n (b 1) converges in distribution to the
0
smallest root of 0; C + 2 0; C + 2 I = 0.

Exercise 16 (i) Write out the objective function of the CUE in the linear homoskedastic endoge-
nous model. (ii) Show that this CUE is equivalent to the LIML estimator. (iii) Show that if gi ( )
is replaced by gi ( ) in Jn ( ), then the new objective function Jen ( ) = Jn ( )=(1 + Jn ( )).

Solution: (i) and (ii). In the linear homoskedastic endogenous model, b ( ) = b ( ) 2

(Z0 Z) 1
=
(Z0 Z) 1 = n 1 (y X0 )0 (y X0 ) , so

(y X0 )0 PZ (y X0 )
Jn ( ) =
n 1 (y X0 )0 (y X0 )
1
1
= n 1 + AR ( ) ;

(y X0 )0 PZ (y X0 )
where AR ( ) = (y X0 )0 MZ (y X0 )
. So the minimizer Jn ( ) is the same as the minimizer
0
0
(y X ) PZ (y X )0 (y X0 )0 PZ (y X0 )
of (y X0 )0 MZ (y X0 )
, which is in turn the same as the minimizer of (y X0 )0 MZ (y X0 )
+1 =

14
 (y X0 )0 (y X0 )
(y X0 )0 MZ (y X0 )
, that is, the LIML estimator.
1 1 BDA 1 .
(iii) Recall that If T = A+CBD, then T 1 = A 1 A 1 CB(B + BDA CB) Let
Pn
A = n1 gi ( )gi ( )0 , C = g n ( ), D = g n ( )0 and B = 1; then
i=1

n
! 1
1X 1
gi ( )gi ( )0 =A 1
+A 1
gn( ) 1 g n ( )0 A 1
gn( ) g n ( )0 A 1
:
n
i=1

As a result,
1
P
n
g n ( )0 1
n gi ( )gi ( )0 gn ( )
Jn ( ) i=1
= 1
n + Jn ( ) 1 P
n
1 + g n ( )0 n gi ( )gi ( )0 gn ( )
i=1
g n ( )0 A 1 g n ( )g n ( )0 A 1 g n ( )
g n ( )0 A 1g
n( )+ 1 g n ( )0 A 1 g n ( )
=
g n ( )0 A 1 g n ( )g n ( )0 A 1 g n ( )
1 + g n ( )0 A 1g
n( ) + 1 g n ( )0 A 1 g n ( )
Jen ( )2
Jen ( ) +
1 Jen ( )
= e 2
= Jen ( ):
1 + Jen ( ) + Jne( )
1 Jn ( )

z0i b Pii x0i

Exercise 17 De…ne Pij as the ijth element of PZ . Show that (i) z0i b ( i) = 1 Pii ; (ii) b JIVE2
P 0
1 P
0
can be rewritten as i6=j xi Pij xj i6=j xi Pij yj .

Solution: (i) From Exercise 14 of Chapter 3, b ( i) =b (1 Pii ) 1

(Z0 Z) 1
zi x0i z0i b . So

Pii z0i b Pii z0i b Pii x0i z0i b z0i b Pii x0i
z0i b ( i) = z0i b x0i z0i b = = :
1 Pii 1 Pii 1 Pii
0y
P P
(ii)We only show that X = i6=j x0i Pij yj since (X 0 X) = i6=j x0i Pij xj can be similarly
shown.
n
X n
X
0 e0 1 0
X y = ( i) zi yi = X0 i Z i Z0 Z zi y i
i=1 i=1
n
X X
1 0
= X0 Z xi z0i Z0 Z zi y i = x0i Pij yj :
i=1 i6=j

Exercise 18 (Emprical)

Exercise 19 Show that N (0; V )0 1N (0; V ) = 2 .

l k

15
 1
Solution: Recall that V = G G0 1G G0 . So

N (0; V )0 1
N (0; V )
0
1=2 1=2
= N 0; V N 0; V
1 0 1
1=2
= N 0; I G G0 1
G G0 1=2
N 0; I 1=2
G G0 1
G G0 1=2

1
= N (0; Il )0 I 1=2
G G0 1
G G0 1=2
N (0; Il ) :

1=2 G G0 1G 1
Note that G0 1=2 is a projection matrix on span 1=2 G , so by the
arguments in Exercise 9(vii),

N (0; V )0 1
N (0; V ) = 2
l k

1=2 G 1 1
since tr(I G0 1G G0 1=2 ) = l tr G0 1G G0 1=2 1=2 G =l k.

16