Sample Answers to Analytical Exercises of Chapter 7

Exercise 1 Suppose the farmer could observe i as well as Ai before deciding on labor input,
how does the demand equation for labor (??) change? Show that log Qi and log(Li ) are perfectly
correlated.
1
1
w 1 1
Solution: Li = p (Ai exp( i ) 1)
1 1 , so

1
log Li = 0 + (ui + vi ) ;
1 1

1 w
where 0 = 1 0 + log( 1) log p . As a result,
1

log Qi = log Ai + 1 log(Li ) + i

= 0 + 1 log(Li ) + ui + vi
= 0 + 1 log(Li ) + (1 1 ) (log(Li ) 0)

= 0 (1 1) 0 + log(Li );

that is, log Qi and log(Li ) are perfectly correlated.

Exercise 2 (i) Show that E [vi ui ] 6= 0. (ii) Suppose k = k2 = l = 1, and all variables are
demeaned. Can the correlation between ui and vi be 1?

Solution: (i) If E [vi ui ] = 0, E [xi ui ] = E [( 0 zi + vi ) ui ] = 0, and these is no endogeneity

problem. (ii) In this simple case, y = x + u and x = z + v. Since E[zu] = 0, z 2 span? (u). If
v = vu + vu? , where vu 2 span(u), and vu? 2 span? (u), then

Cov (u; vu + vu? ) Cov (u; vu )

Corr(u; v) = p =p p 1;
V ar (vu + vu? ) V ar(u) V ar (vu ) + V ar (vu? ) V ar(u)

where the equality holds only if V ar (vu? ) = 0, that is, vu? = 0 almost surely. In summary, z 2
span? (u) and vu 2 span(u). More speci…cally, vu = uE[u2 ] 1 E[xu], and z = x uE[u2 ] 1 E[xu].
In other words, z includes all variation in x that is orthogonal to u, which seems unrealistic in
practice.
Email: [email protected]

1
Exercise 3 Suppose y = X + u and X = Z + V, where E[ujZ] = 0 and E[VjZ] = 0 but
E[V0 u] 6= 0. Derive the plim of b OLS . When = 0, what will plim b OLS degenerate to?

Solution: Recall that b OLS = (X0 X) 1 X0 u = ( 0 Z0 Z + 0 Z0 V + V0 Z + V0 V) 1 ( 0 Z0 u + V0 u).

p p
Note that n 1 Z0 V !0, and n 1 Z0 u !0, so the denominator divided by n converges in probability
to 0 E [zz0 ] + E [vv0 ], and the numerator converges to E [vu0 ], so

b p 0 1
OLS ! E zz0 + E vv0 E vu0
1=2 1=2 1=2 1 1=2
= E vv0 E vv0 0
E zz0 E vv0 + Ik E vv0 E vu0
1=2 1 1=2
= E vv0 ( + I) E vv0 E vu0 ;

where = E [vv0 ] 1=2 0 E [zz0 ] E [vv0 ] 1=2 is called the concentration parameter which measures
the relative variation explained by z with respect to the unexplained part. 0 E [zz0 ] is also the
probability limit of n 1 XPZ X, while when k = 1, b = n 1 XPZ X=s2v is the F -statistic to test
p
= 0 in X = Z + V. When = 0, b OLS ! E [vv0 ] 1 E [vu0 ].

Exercise 4 In the reduced form between the regressors xi and instruments zi (??), the parameter
is de…ned by the population moment condition

E zi vi0 = 0:

Show that the MoM estimator for is b = (Z0 Z) 1

Z0 X.

Solution: The moment condition is

0 = E zi vi0 = E zi x0i z0i ;

whose sample analog is

n
1X
0= zi x0i z0i = Z0 X Z0 Z :
n
i=1

so the MoM estimator b is

b = Z0 Z 1
Z0 X:

Exercise 5 In the structural model

y = X + u;
X = Z + V;

with l k, l k, we claim that is identi…ed (can be recovered from the reduced form) if
rank( ) = k. Explain why this is true. That is, show that if rank( ) < k then cannot be
identi…ed.

2
Solution: We have
y =X +u=Z +V +u=Z + e;

where e = V + u. Since z is orthogonal to both v and u, we have E [zi (yi z0i )] = 0, that is,

1
= E zi z0i E [zi yi ] : (1)

If rank( ) < k, then there are less than k linearly independent equations with k unknowns, thus,
we cannot identify , that is, there is no unique solution to (1). Provided that there is satisfying
= and is of full column rank, then the solution to (1) is unique. As long as we could …nd
one solution, we are done. For any positive de…nite matrix A;

0 0
A = A

0W 1
since is nonsingular if rank( ) = k, we have = ( 0W ) 0W .

Exercise 6 Show that the true value of 2 in (??).

Solution: By the formula of linear projection in Chapter 2,

! " ! # 1 " ! #
1 x1 x1
= E x1 z E x2
2 z z
! 1 !
E x21 E x21 E [x1 x2 ]
=
E x21 E x21 + E "2 E [x1 x2 ]
! !
1 E x21 + E "2 E x21 E [x1 x2 ]
=
E x21 E ["2 ] E x21 E x21 E [x1 x2 ]
0 1
E[x1 x2 ]
E [x21 ]
= @ A;
0

where note that E [x1 z] = E x21 , E z 2 = E x21 +E "2 and E [zx2 ] = E [(x1 + ") x2 ] = E [x1 x2 ]
because " is independent of x2 . So 1 is actually the coe¢ cient of x1 in the regression of x2 on x1 ,
i.e., 1 + 2 , while 2 = 0. In other words, the e¤ect of z is absorbed in that of x1 .
Alternatively, by the FWL, 2 is the coe¢ cient of " in the regression of x2 on ",

E ["x2 ]
2 = = 0;
E ["2 ]

because " is independent of x2 so that E ["x2 ] = 0, where " is the residual in the regression of z on
x1 .

3
Exercise 7 Consider the linear demand and supply system:

Demand: qi = 0 + 1 pi + 2 yi + ui ;
Supply: qi = 0 + 1 pi + 2 wi + vi :

where income (y) and wage (w) are determined outside the market. In this model, are the parameters
identi…ed?

Solution: Yes, just-identi…ed. Each equation has one endogenous variable on the RHS, but we
have IVs, y and w, for each equation, so we have three IVs and three unknowns for each equation
and the system is just-identi…ed. To be speci…c, the moment conditions are

0
E 1 y w (q a0 a1 p a2 y) = 0;
0
E 1 y w (q b0 b1 p b2 w) = 0:

These conditions enable us to just-identify the parameters.

Exercise 8 In the linear model,

yi = x0i + ui ;
E[ui jxi ] = 0:

Suppose 2i = E[u2i jxi ] is known. Show that the GLS estimator of with the weight matrix
2 2
diag 1 ; ; n can be written as an IV estimator using some instrument zi . (Find an ex-
pression for zi .)

Solution: zi = x2i :
i

Exercise 9 Suppose y = x + u, x = " + 1 u, and z = v + ", where "; u and v are independent.
P
Find the probability limits of b OLS and b IV . Show that if = 0; n1 ni=1 vi "i = 0, and 2u is large,
the two probability limits are the same.

Solution: Note that

1 Pn 1 2 2
b n i=1 xi ui p u u
OLS = 1 Pn 2
! 2 2 2 = 2 2 2
;
n i=1 xi " + u " + u

p Pn
which is roughly when 2
u is large, and b IV ! if > 0. If = 0 and 1
n i=1 vi "i = 0, then

1 Pn 1 Pn
b zi u i i=1 zi ui
IV = n
1 Pi=1
n = 1 Pn n
1 Pn 2 11 Pn
n i=1 zi xi n i=1 vi "i + n i=1 "i + n i=1 zi ui
1 Pn
n i=1 zi ui
= 1 1 Pn = ;
n i=1 zi ui

4
which is close to the probability limit of b OLS .

Exercise 10 Show that if l = k, then b 2SLS = b IV , that is, no matter we use x

bi or zi as IVs, we
get the same results.

Solution:

1 1 1 1
b = X0 Z Z0 Z Z0 X X0 Z Z0 Z Z0 y = Z0 X Z0 y = b IV :
2SLS

Exercise 11 Show that b 2SLS satis…es

!
b e1 e0 1
S 1 = 0;
b
2SLS

where e1 = (1; 0; ; 0)0 is the …rst (k + 1) 1 unit vector, S = (y; X)0 P(y; X), and b =
0 0
(1; b 2SLS )S(1; b 2SLS )0 . Further show that when the model is just identi…ed, b = 0 and S is
singular.

Solution: Note that b = u

b 0 Pb
u, so the left hand side is equal to

(y; X)0 Pb
u b 0 Pb
u ue1 ;

b = y X b 2SLS . Since X0 Pb
where u u = X0 Py X0 PX b 2SLS = 0 by the de…nition of b 2SLS , we
need only to show that
y0 Pb b 0 Pb
u=u u:

However, y = X b 2SLS + u
b , and given that X0 Pbu = 0, this is obvious.
When the ! 0
model is just identi…ed, Z u b = 0, so b = u
b 0 Pb
u=u b 0 Z(Z0 Z) 1 Z0 u
b = 0. As a result,
1
S = 0, so S is singular.
b
2SLS

Exercise 12 (i) Show that E [ujv; z] = E [ujv] if E[zu] = 0 and E[zv0 ] = 0. (ii) Show that (??)
generates the same formula of b 2SLS as (??).

Solution: (i)
" ! # 1 " ! #
0 0 v 0 0 v
E [ujv; z] = (v ; z )E (v ; z ) E u
z z
" !# 1 !
vv0 0 E [vu]
= (v0 ; z0 )E
0 zz0 0
1
= v0 E vv0 E [vu] = E [ujv]:

Note that both E[zu] = 0 and E[zv0 ] = 0 are necessary for this result to hold. Intuitively, both u
and v are orthogonal to z, so u v0 is orthogonal to z. As a result, in the regression u on v and

5
z, the coe¢ cient of z is the same as in the regression u v0 on z by the FWL theorem, which is
0.
(ii) Write the original structural equation as

b + u
y =X +V b
V ;

where Vb = MZ X are the residuals from the …rst step regression. By the FWL theorem, b can be
achieved from regressing y on MV
b X, but MV b X = PZ X, so
b is exactly the 2SLS estimator.

Exercise 13 Take the linear model

yi = xi + ui ; E[ui jxi ] = 0;

where xi and are scalars.

(i) Show that E[xi ui ] = 0 and E[x2i ui ] = 0. Is zi = (xi ; x2i )0 a valid instrumental variable for
estimation of ?

(ii) De…ne the 2SLS estimator of , using zi as an instrument for xi . How does this di¤ er from
OLS?

Solution: (i) E [xi ui ] = 0 and E x2i ui = 0 follow directly from E [ui jxi ] = 0: Therefore, zi =
(xi x2i )0 is a valid instrument provided that E x2i or E x3i are nonzero.
(ii) In this case, ^ 2SLS = ^ OLS : To see this, note that PZ X = X; since xi is included in zi .
Therefore,
^ 0 1 0 0 1 0 ^
2SLS = (X PZ X) X PZ y = (X X) X y = OLS :

Note: Sometimes 2SLS is unable to exploit extra moment conditions (which would bene…t us
if the heteroskedasticity exists) because of a rigid form of the weight matrix. The e¢ cient GMM
(see the next chapter) sets the weight matrix more wisely and gives us e¢ ciency gains.

Exercise 14 Suppose y = m(x) + u, E[xu] 6= 0 and E[zu] = 0.

(i) Derive the probability limit of b 2SLS .

(ii) Is b 2SLS is the best linear predictor of m(x) in the sense that
h i
2
plim b 2SLS = arg minE m(x) x0 ?

p 1
Solution: (i) b 2SLS
h = (X0 PX) i 1 (X0 Py) ! E[xz0 ]E[zz0 ] 1 E[zx0 ] E[xz0 ]E[zz0 ] 1 E[zm(x)].
(ii) arg minE (m(x) x0 )2 = E[xx0 ] 1 E[xm(x)] which is generally not equal to plim b 2SLS .

To see why, check the just-identi…ed case. In this case, plim b 2SLS = E[zx0 ] 1 E[zm(x)]. Given
that z 6= x, E[xx0 ] 1 E[xm(x)] 6= E[zx0 ] 1 E[zm(x)] unless E [zjx] = Ax for a nonsingular matrix

6
A. Recall that x = 0 z + v, so z = 0 1 (x v). E [zjx] = Ax implies that E [vjx] = Bx for some
constant matrix B. A special case is z = Ax, which implies x = A 1 z, i.e., x is exogenous at the
beginning.

Exercise 15 In the structural model

y = X1 1 + X2 2 + u;
X2 = X1 12 + Z2 22 + V;
1
(i) show that b 2;2SLS = X02 PZ? X2 X02 PZ? y, where Z? 2 = MX1 Z2 ; (ii) Given that x1i are the
2 2

included exogenous variables with E[x1i ui ] = 0, does X01 u

b equal 0? where u b = y X b 2SLS ; (iii)
Does b 1;2SLS equal b 1;OLS ?

Solution: (i) Note that X2 = X1 e 12 + Z? e 0 1 0

2 22 + V, where 12 = 12 + (X1 X1 ) X1 Z2 22 , and
X01 Z? ? b
2 = 0. Projection of X2 on (X1 ; Z2 ) is the same as the projection on X1 ; Z2 . 2;2SLS is the
b b
coe¢ cient of X2 = PZ X2 = PX1 X2 + PZ? X2 in regressing y on X1 and X2 . Since
2

b = X1 b 1;2SLS + PX1 X2 + PZ? X2 b 2;2SLS

y
2

= X1 b 1;2SLS + (X01 X1 ) 1
X01 X2 b 2;2SLS + PZ? X2 b 2;2SLS ;
2

b is the coe¢ cient of PZ? X2 in regressing y on X1 ; PZ? X2 . Since X01 PZ? X2 = 0, this is
2;2SLS 2 2 2
exactly the coe¢ cient of regressing y on PZ? X2 solely.
2
(ii) This is obvious since b2SLS solves
!
X01
y X b 2SLS = 0:
b0
X 2

(iii) b 2SLS is obtained by regressing y on X1 and Xb 2 , while b OLS is obtained by regressing y

on X1 and X2 , where X b 2 = PX X2 + P ? X2 and X2 = PX X2 + P ? X2 with X? span? (X1 ).
1 Z2 1 X1 1
Because X2 and X2 are di¤erent mixtures of the elements in span (X1 ) and span? (X1 ), b 1;2SLS
b
need not equal b 1;OLS . Actually, because b is consistent while b
1;2SLS need not be, they
1;OLS
need not be equal even asymptotically. There is a special case where they are indeed equal:
span (X1 ) ? span (X2 ).

Exercise 16 Consider the single equation model

yi = xi + ui ;

where yi and xi are both real-valued. Let b denote the IV estimator of using as instrument a
dummy variable di (takes only the values 0 and 1). Find a simple expression for the IV estimator
in this context and derive its probability limit. What is the di¤ erence between this probability limit
and the probability limit of the Wald estimator?

7
Solution: Recall that the IV estimator is as follows

b = Z0 X 1
Z0 y.

When z is one-dimensional, we have

0 1
b = (Z Z) Z0 y
1 ;
(Z0 Z) Z0 X

that is, ratio of coe¢ cient on Z in regression of y on Z and coe¢ cient on Z in regression of X on
Z. When z is binary,
P P ,
X X
b = Pyi 1(di = 1) P
xi 1(di = 1)
= yi xi :
1(di = 1) 1(di = 1)
di =1 di =1

.
so plim b = E[y1(d=1)]
P (d=1)
E[x1(d=1)]
P (d=1)
E[yjd=1]
= E[xjd=1] , while plim b W ald = E[xjd=1]
E[yjd=1] E[yjd=0]
E[xjd=0] . So only
if E [yjd = 1] E [xjd = 0] = E [yjd = 0] E [xjd = 1], they are equal. This includes some interesting
special cases, e.g., E[yjd=1] E[xjd=1]
E[yjd=0] = E[xjd=0] if E [yjd = 0] 6= 0 and E [xjd = 0] 6= 0 or E [yjd = 0] =
E [xjd = 0] = 0 or E [yjd = 0] = E [yjd = 1] = 0.

Exercise 17 (*) Suppose

y = 0 + 1x + u, Cov(x; u) 6= 0;
x = 0 + 1z + v; E[ujz] = 0; E[zv] = 0;

where x is binary. Unless z is binary, E[xjz] cannot be a linear function. Suppose we run a Probit
b = (b0 + b1 z).
regression in the …rst stage and get x
0
(i) Show that if E[xjz] = ( 0 + 1 z), then b b ;b
0 1 b is consis-
based on regressing y on 1; x
tent.

(ii) Show that if E[xjz] 6= ( 0 + 1 z), then b based on regressing y on 1; x

b is not consistent.

(iii) Show that if E[xjz] = ( 0 + 1 z), plim b is the same as the plim of the IV estimator using
b) as the instrumental variables, but if E[xjz] 6= ( 0 + 1 z), they are generally di¤ erent.
(1; x

(iv) Show that the IV estimator using (1; xe) as the instrumental variables is consistent, where x
e is
the linear projection of x on (1; z).
(Hint: if E[xjz] = ( 0 + 1 z), b = (b0 ; b1 )0 is consistent to the true value; otherwise, it is
inconsistent.)

Solution: (i) First,

P ! 1 P !
1 n 1 bi
x n 1 yi
b= P P 2 P :
n 1 bi
x n 1 bi
x n 1 xbi yi

8
 bi by
We can show replacing x ( 0 + 1 zi ) will not a¤ect the probability limit of b . For example,
X X
1 1
n bi
x n ( 0 + 1 zi )
X
1
= n [ (b0 + b1 zi ) ( 0 + 1 zi )]
X
1
= n (ai ) [(b0 0) + (b1 1 ) zi ] :

From the hint, if E[xjz] = ( 0 + 1 z), plim(b ) = . So

X X p
1 1
n bi
x n ( 0 + 1 zi ) C (jb0 0j + jb0 0 j jzj) ! 0;

where C is the supremum of the standard normal density ( ) and is …nite. In conclusion,
! 1 !
b ! p 1 E[ ( 0+ 1 zi )] E [yi ]
2 (2)
E[ ( 0 + 1 zi )] E[ ( 0+ 1 zi ) ] E [ ( 0 + 1 zi ) y i ]
! 1 ! !
1 E[ ( 0+ 1 zi )] 1 E[ ( +
0 1 zi )] 0
= 2 2
E[ ( 0 + 1 zi )] E[ ( 0+ 1 zi ) ] E[ ( 0 + 1 zi )] E[ ( 0+ 1 zi ) ] 1
!
0
= ;
1

where the second equality uses

yi = 0 + 1x +u= 0 + 1 ( 0 + 1 zi ) +e (3)

with e = 1 (x ( 0 + 1 zi )) + u and E[ejz] = 0.

(ii) If E[xjz] 6= ( 0 + 1 z), the analysis is more complicated. Basically, we need to interpret
as the pseudo-true (rather than true) value now, and y cannot be expressed as (3) with E[ejz] =
0 anymore. So the second equality in (2) does not hold (although we can still replace x bi by
p
( 0 + 1 zi )). As a result, b 9 .
(iii) Note that

P ! 1 P !
b 1 n 1 xi n 1 yi
IV = 1
P P P :
n bi
x n 1 xbi xi n 1 xbi yi

bi by
We can still replace x ( without a¤ecting the plim of b IV . To prove plim b IV =plim b ,
0 + 1 zi )
P p P p
we need only to show that n 1 i ! 0 and n 1 ( 0 + 1 zi ) i ! 0, where i = xi
( 0 + 1 zi ). If E[ i jzi ] = 0, this is obvious. But if E[ i jzi ] 6= 0, the FOCs for b guarantee that

xi ( 0 + 1 zi )
E ( 0 + 1 zi ) = 0;
( 0 + 1 zi ) (1 ( 0+ 1 zi ))
xi ( 0 + 1 zi )
E zi ( 0 + 1 zi ) = 0;
( 0 + 1 zi ) (1 ( 0+ 1 zi ))

9
 p
which generally does not imply E[ i ] = 0 and E[ ( 0 + 1 zi ) i ] = 0. Also, b IV 9 in general.
The exercise shows that in the just-identi…ed case, the nonlinear 2SLS estimator is not only
nonequivalent to the IV estimator numerically but nonequivalent to the IV estimator asymptotically
under misspeci…cation of the …rst-stage conditional mean. Also, under misspeci…cation, both the
IV estimator and the 2SLS estimator are generally inconsistent.
(iv) This is the usual IV estimator and the consistency is straightforward. This exercise shows
that the consistency of the 2SLS estimator does not rely on whether the …rst-stage conditional
mean is speci…ed correctly or not as long as a "linear" projection is conducted in the …rst stage.
This is essentially because the structural equation is linear.

Exercise 18 (*) Show (??) in the case of J = 2.

Solution: Note that

y = y0 + (y1 y0 )1(S 1) + (y2 y1 )1(S 2);

so

E [yjz = 1] E [yjz = 0] = E [(y1 y0 ) (1(S1 1) 1(S0 1))] + E [(y2 y1 ) (1(S1 2) 1(S0 2))]
= E [(y1 y0 )1(S1 1 > S0 )] + E [(y2 y1 )1(S1 2 > S0 )] :

Similarly, since
S = 1(S 1) + 1(S 2);

i.e., in the expression of y, let yj = j, we have

E [Sjz = 1] E [Sjz = 0] = E [1(S1 1 > S0 )] + E [1(S1 2 > S0 )] :

Exercise 19 (*) S1 S0 is called the compliance intensity which need not be 1, so in the expression
of in (??), the sets of compliers fS1 j > S0 gJj=1 are overlapping subpopulations. Denote the
complier subpopulation with S0 = k and S1 = l as Ckl , where k < l, k; l = 0; 1; ; J. Show that
can be re-expressed in terms of nonoverlapping subpopulations as

J
X1 X
= wkl E [yl yk jCkl ] ;
k=0 l>k

P (Ckl )
where wkl = PJ 1 P .
k=0 l>k (l k)P (Ckl )

Sj 1 SJ
Solution: Note that fS1 j > S0 g = k=0 l=j Ckl . So

j 1X
X J
P (S1 j > S0 ) = P (Ckl ) ;
k=0 l=j

10
and
j 1X
X J
P (S1 j > S0 )E [yj yj 1 jS1 j > S0 ] = E [(yj yj 1 ) Ckl ] :
k=0 l=j

As a result,
J
X X j 1X
J X J J
X1 X
P (S1 j > S0 ) = P (Ckl ) = (l k)P (Ckl ) ;
j=1 j=1 k=0 l=j k=0 l>k

and
P
J
P (S1 j > S0 )E [yj yj 1 jS1 j > S0 ]
j=1
J jP1 P
P J
= E [(yj yj 1 ) Ckl ]
j=1 k=0 l=j
JP1 P
= E [(yl yk ) Ckl ]
k=0 l>k
JP1 P
= E [yl yk jCkl ] P (Ckl ) ;
k=0 l>k

where the second equality can be understood by letting yj yj 1 = 1 so that E [(yl yk ) Ckl ] =
P
(l k)P (Ckl ) as in the expression of Jj=1 P (S1 j > S0 ).

Exercise 20 (Empirical)

0 b (1 b)
Exercise 21 (*) De…ne b = min (y(y XX) P Z (y X )
)0 (y X )
and e = n l
b) . Show that the LIML
l(
1 n
1
estimator can be re-expressed as
1
b = X 0 PZ X b X0 X X 0 PZ y b X0 y ;

and Fuller’s estimator can be re-expressed as

1
b = X 0 PZ X e X0 X X 0 PZ y e X0 y : (4)

Solution: Repeat the analysis in the technical appendix, but we …rst concentrate out 12 and 22
rather than 1 in `n ( ). We now need to minimize

0 0
Y Y Y (X2 Z1 12 Z2 22 )
0 0
(X2 Z1 12 Z2 22 ) Y (X2 Z1 12 Z2 22 ) (X2 Z1 12 Z2 22 )

0 0
with respect to ( 12 ; 22 ), where we de…ne Y = [y; X] and = (1; ). From (33), we can get a
similar result as (34):

0 0
Y Y (X2 Z1 12 Z2 22 ) Mv (X2 Z1 12 Z2 22 ) ;

where v = Y . The determinant is minimized by OLS, so it is equal to jX02 Mv;Z X2 j as in (35).

11
By a similar technique as in (36), we can show

v0 MZ v v0 MZ X2
= Y0 MZ Y = v0 MZ v X02 Mv;Z X2 ;
X02 MZ v X02 MZ X2

so the concentrated likelihood is

jY0 MZ Yj 0 Y0 Y
0
Y Y = Y0 MZ Y ;
v0 MZ v 0 Y0 M Y
Z

0 Y0 Y
and we estimate by minimizing 0 Y0 M Y
. The FOCs are
Z

0 0
YY Y MZ Y = 0;

which is equivalent to
0 0
Y PZ Y Y Y = 0; (5)

where = 1 1= .1 Note that (5) is exactly the FOC for b. Finally, b can be derived from (5)
straightforwardly.
As to e, note that

1
I b MZ = I MZ
n l 1 b n l
!
1 b
= PZ I
1 b n l 1 b n l
1 h i
= PZ eI :
1 b n l

1
Since the constant will not a¤ect the estimator, so Fuller’s estimator can be expressed
1 b n l
as in (4).

1
So this new formulation cannot include the least squares estimator since its = 0.

12