Econometrics II, 2005/06 Solutions Exam, June 23, 2006

Econometrics II - SOLUTIONS EXAM

Answer each question in separate sheets in three hours

1. In the model
0
yt = zt + vt ; t = 1; 2; : : :
suppose that
vt = vt 1 + et ; j j < 1;
2
where et is WN (0; e ).

a. [40%] Study the asymptotic properties of b T ; the OLSE of ; when:

(1) The regressors zt are non-stochastic and bounded.
0
(2) zt = (1; t) .
(3) The regressors zt are stationary and ergodic, and E [zt et ] = 0 but E [zt et 1] 6= 0:
Specify clearly any additional conditions you use for the analysis.
We have that !
T T
b 1X 1X
T = zt z0t zt vt = MT 1 VT ; say:
T t=1 T t=1
In Case (1), we need to assume that MT ! M > 0 as T ! 1: Then we can check that

2 1 XX 0
E kVT k = z zs E [vs vt ]
T2 t s t
C 2 1
XX
sup kzt k j v (t s) j
T t T t s
1
= O T = o(1);

so VT !p 0 and b T is consistent. For the asymptotic normality we need a CLT for T 1=2 VT ;
for which we need further to assume that the et are IID or homoskedastic MD [and/or some
UI conditions], so that
T 1=2 VT !d N (0; S)
where
1
X
S= R (j) v (j)
j= 1

and R ( j) = R (j) ; j = 0; 1; : : : ; and

T j
1 X
R (j) = lim zt z0t+j :
T !1 T
t=1

Then
T 1=2 b T !d N 0; M 1
SM 1
:
Alternatively we could have assumed that,
T
Xj
DT 1 zt z0t+j DT 1 ! R+ (j) ; j = 0; 1; 2; : : :
t=1

1
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

where
T
X
k
DT = diag fdi;T gi=0 ; d2i;T = 2
zi;t
t=1

and for Grenander conditions to hold we only need that

di;T ! 1; 8i as T ! 1;

since by boundedness of zt we have that

jzit j
lim max = 0; 8i:
T !1 1 t T di
Then we can obtain that
DT T VT !d N 0; S+
where
1
X
S+ = R+ (j) v (j) ;
j= 1

and …nally
+
DT b T
1 1
!d N 0; R+ (0) S+ R (0) :

Case (2) is a particular case of this situation, with

P 1=2 ! !
( t 1) 0 T 1=2 0
DT = P 2 1=2 1 3 1=2
:
0 tt 0 3T

In case (3) OLS is inconsistent if 6= 0; because E[zt vt ] 6= 0 since

E [zt vt ] = E [zt et 1] + E [zt et ] = E [zt et 1] 6= 0:

b. [30%] Suppose that is known, and consider

T
! 1 T
!
X X
e = (zt zt 1 )(zt zt 1)
0
(zt zt 1 )(yt yt 1) :
T
t=2 t=2

Find the asymptotic distribution of T 1=2 ( e T ) under the previous conditions, specifying
all the conditions used.
We now have that !
T 1 T
X X
T 1=2
(e T )= zt zt 0
zt et
t=2 t=2

where zt = zt zt 1:

The analysis for case (1) is similar as before, but now we have that

T 1=2 e T !d N 0; 2
M 1
;

where we assume that

T
1X
MT := z z 0!M >0
T t=2 t t
as T ! 1:

2
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

In case (2) we would obtain similar results with the appropriate normalization, that for the
linear trend is not T 1=2 ; but T 3=2 ;

DT e T 2 1
!d N 0; R (0) :

where, j = 0; 1; 2; : : :
T
Xj
1 0 1
R (j) = lim DT zt zt+j DT
T !1
t=1
and
T
X
k 2
DT = diag di;T i=0
; di;T = zi;t2 :
t=1
In Case (3) we have to check whether E[zt et ] = 0:

E [zt et ] = E [zt 1 et ] + E [zt et ] = 0;

because E [zt 1 et ] = 0: In this case, under appropriate conditions on the distribution of zt et

(ergodic MD) we have a similar result to (1), where now M := E [zt zt 0 ] ; if E e2t jzt = 2
and if E et ej jzt ; zj = 0; t 6= j: Otherwise, if zt et is not a MD but satis…es a CLT, we have
that
T 1=2 e T !d N 0; M 1 WM 1
where
1
X
0
W= E et et+j zt zt+j :
j= 1

c. [30%] Consider now the following estimate of ;

T
! 1 T
X X 0
2
bT = vbt vbt vbt 1; vbt = yt T zt ;
t=2 t=2

where T is a consistent estimate of ; and let T be the estimate obtained when replacing
by b T in e T :
Study the relationship between e T and T under (1), (2) and (3). Then deduce if the above
estimates of and of are asymptotically related or not in each case.
The …rst issue is the convergence rate of b T ; which depends on that of the initial estimate
of ; T (this can not be the OLSE in case (3): IV’s estimates can be used instead). If the
latter is T 1=2 consistent (or faster), then b T will be so as well.
Then we have that

T 1=2 e T T
8 ! 1 ! 1
9
< 1X T
1X
T =
0 0
= (zt zt 1 )(zt zt 1 ) (zt b T zt 1 )(zt b T zt 1)
: T T t=2 ;
t=2
0 1
=zt =et
XT z }| { z }| {
B 1 C
@ 1=2 (zt zt 1 ) (vt vt 1 )A
T t=2

T
! 1
1X 0
+ (zt b T zt 1 )(zt b T zt 1)
T t=2
( T T
)
1 X 1 X
(zt zt 1 )(vt vt 1) (zt b T zt 1 )(vt b T vt 1) :
T 1=2 t=2
T 1=2 t=2

3
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

PT
The …rst term in braces can be showed to be op (1) if b T = op (1); and the term T 1=2 t=2 zt et
is Op (1) ; so the …rst term is asymptotically negligible. The leading term in the second factor
converges to M 1 : The key is the second term in braces, which is equal to
T
X T
X
1 1
(zt zt 1 )(vt vt 1) (zt b T zt 1 )(vt b T vt 1)
T 1=2 t=2
T 1=2 t=2
T
X T
X
1 1
= (b T ) (zt zt 1 )vt 1 (b T ) zt 1 (vt b T vt 1)
T 1=2 t=2
T 1=2 t=2

= T 1=2 (b T ) fE [zt vt 1] + op (1)g T 1=2 (b T ) fE [zt 1 et ] + op (1)g :

In Case (1) we have that E [zt vt 1 ] ; E [zt 1 et ] = op (1) ; so T 1=2 e T T = op (1) and both
estimates have the same asymptotic distribution, and therefore estimation of and of are
asymptotically independent.
In Case (2) the same conclusion applies (even despite that the estimates of some elements of
have faster convergence rate).
If in Case (1) the convergence of e T is slower that T 1=2 ; then estimation of is a¤ected if T has
simila properties, but e T and T will be asymptotically equivalent (at the appropriate rate), so
estimation of is not a¤ected by that of (but not the way round)
In Case (3) we have that E [zt 1 et ] = 0; but E [zt vt 1 ] 6= 0; because E [zt et 1] 6= 0: Therefore
T 1=2 e T T = Op (1) ; and in fact we can …nd that

h 0
i 1
T 1=2 e T T = T 1=2 (b T ) E zt zt E [zt 1 et ] + op (1) ;

which shows that estimation of and of are not independent, and their asymptotic distributions
are correlated, since under the given assumptions the zt are not strictly exogeneous, only weakly
exogenous.

4
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

2. Let the scalar dependent variable y be a binary variable, y 2 f0; 1g, and let z be some explanatory
variables. The conditional probability of y given the vector of variables z is given by

Pr(y = 1jz) = (z0 0 );

Pr(y = 0jz) = 1 (z0 0 );

where ( ) is the logit cumulative distribution function,

exp (x)
(x) =
1 + exp (x)
and 0 is a vector of parameters.

a. [15%] Find E (yjz) and V (yjz). Give the objective function for the Nonlinear Least Squares (NLS)
estimates of 0 given a sample of size n of independent observations of (y; z) :

E (yjz) = (z0 0)
0
V (yjz) = (z 0 ) (1 (z0 0 ))
1 h i
2
Qn ( ) = En (y (z0 )) :
2
b. [35%] Assuming consistency, deduce the asymptotic distribution of the NLS estimates bn and obtain
the Gauss-Newton recursion to approximate the NLS estimates given an initial value of .
Propose a consistent estimate of the asymptotic variance of the NLE bn . State any further
assumption you use.
We have that
@
Qn ( ) = En [(y (z0 )) (z0 )z]
@
@2 h i
Qn ( ) = En (z0 )2 zz0 En (y (z0 )) _ (z0 )zz0 :
@ @ 0
Therefore, u = u ( 0) := y (z0 0 );

@
n1=2 Qn ( 0) = n1=2 En [u (z0 )z]
@
!d N 0; E (z0 0 ) (1 (z0 0 )) (z0 2
0 ) zz
0

:= N (0; D)

and, if n !p 0;

@2 h i
Qn n = En (z0 2
n ) zz
0
En u n _ (z0 n )zz0
@ @ 0 h i
!p En (z0 2
0 ) zz
0
En u _ (z0 0 )zz0
!p E (z0 2
0 ) zz
0
:= E;

so that we obtaint the usual result after linearizing Qn ^n around 0, if the true value 0
is in the interior of the allowed set ,

n1=2 ^n 0 !d N 0; E 1
DE 1
:

We have that
h i
^n
D = E (z0 ^n ) 1 (z0 ^n ) (z0 ^n )2 zz0 !p D:
h i
^n
E = En (z0 ^n )2 zz0 !p E:

5
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

Gauss-Newton recursion:

^(k+1)
n
(k)
= ^n
(k) (k)
Gn 1 ^n gn ^n

where

gn ( ) = En [(y (z0 )) (z0 )z]

Gn ( ) = En (z0 )2 zz0 :

c. [30%] Consider alternatively the estimate obtained through this recursion,

en = bn Fn (bn ) 1
fn ( b n )

where bn is the NLS estimate and

(z0 ) fy (z0 )g
fn ( ) := En z
(z0 ) (1 (z0 ))
" #
2
f (z0 )g
Fn ( ) := En zz0
(z0 ) (1 (z0 ))

2
with (x) = (d=dx) (x) = exp (x) = (1 + exp (x)) : Compare these iterations with Gauss-
Newton method.
en corrects for the conditional variance of u; so is a sort of Weighted or Generalized NLS.
Which objective function is minimizing en asymptotically?
" #
2
1 (y (z0 ))
Qn ( ) = En :
2 (z0 0 ) (1 (z0 0 ))

Deduce the asymptotic distribution of en using that

n1=2 en n1=2 Fn (bn )

1
0 = fn ( 0) + op (1):

We have that
2
(z0 0)
n1=2 fn ( 0) !d N 0; E zz0
(z0 ) (1 (z0 ))
Fn (bn ) !p Fn ( 0)
" #
2
(z0 0) 0
!p E zz
(z0 0 ) (1 (z0 0 ))

and therefore
1
!
2
(z0 0)
n1=2 en 0 !d N 0; E zz0 ;
(z0 0 ) (1 (z0 0 ))

displaying the usual result for e¢ cient estimates. In fact, it can be showed that this is also
the asymptotic distribution of the MLE estimate, and that en is an approximation to it.
Can you say anything about the e¢ ciency of en and bn ?
d. [20%] Suppose that z contains the constant term, so z 0 = 1 + z20 2 . Consider the null hypothesis
that the coe¢ cients of all the other variables in z are zero, H0 : 2;0 = 0. Which will be your
LS estimate of Pr(y = 1jz) under the null?

6
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

In the case that z0 = 1 we have that our model is simply that E [y] = E [yjz] = Pr(y =
1jz) = ( 1 ) := p; and the LS of a model with only an intercept is the sample mean, p^ = yn ;
and the restricted estimate is de…ned by ^1n = yn :
Propose a LM test for H0 .
0
We need to calculate, where the restricted estimates are ^n = ^1n ; 00 ; [here I work with
the general case of vector z2 ]

@ h i
Qn ^n = En y (^1n ) (^1n )z2
@ 2
h i
= En (y yn ) (^1n )z2 ;

which looks like a covariance between the errors of the restricted model, y yn ; and the new
regressors, z2 :
and we have that the LM statistic is given by (using Qn )

@ 1 @
LMn = n Qn ^n Hn ^n Qn ^n
@ 0 @
@ 1 @
= n 0 Qn ^n Hn ^n Qn ^n
@ 2 2;2 @ 2

where, for example,

@2 h i
Hn ( ) = Qn ( ) = En (z0 )2 zz0 En (y (z0 )) _ (z0 )zz0 :
@ @ 0
2
The asymptotic distribution is k where k is the dimension of z2 :

7
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

3. Consider the following system, where the …rst equation is a labour supply function for working,
married women, and the second equation is a wage o¤ er function, with equilibrium conditions
imposed,

hours = 12 log(wage) + 10 + 11 educ + 12 age

+ 13 kidslt6 + 14 kidsge6 + 15 mwif einc + u1

2
log(wage) = 21 hours + 20 + 21 educ + 22 exper + 23 exper + u2 ;

where kidstl6 is number of children less than 6, kidsge6 is the number of children between 6 and
18 and mwif einc is income other than the woman’s labor income. We assume that u1 and u2
have zero mean conditional on educ; age; kidslt6; kidsge6, nwif einc and exper.

a. [30%] Study the identi…cation of the labor supply function when exper and exper 2 have not direct
e¤ ect on current annual hours ( hours). Is this equation no, just or over-identi…ed?
It is over-identi…ed as far as 22 6= 0 and 23 6= 0 :
!
0 0 0 0 0 0 0 0 1 0
R1 = ;
0 0 0 0 0 0 0 0 0 1

which is of rank J1 = 2 > G 1 in this case, and

!
0 22
R1 B =
0 23

which is of rank 1 = G 1:
Is the wage o¤ er function identi…ed with the exclusion restrictions imposed?
Yes because we have many potential IV’s available: age; kidslt6; kidsge6; mwif einc:
b. [50%] We obtain the following output from Eviews.

(1) Dependent Variable: HOURS

Method: Two-Stage Least Squares
Included observations: 428 after adjusting endpoints
Instrument list: EDUC AGE KIDSLT6 KIDSGE6 NWIFEINC EXPER EXPER^2

Variable Coe¢ cient Std. Error t-Statistic Prob.

C 2432.198 594.1719 4.093425 0.0001
LWAGE 1544.819 480.7387 3.213426 0.0014
EDUC -177.4490 58.14260 -3.051961 0.0024
AGE -10.78409 9.577347 -1.125999 0.2608
KIDSLT6 -210.8339 176.9340 -1.191596 0.2341
KIDSGE6 -47.55708 56.91786 -0.835539 0.4039
NWIFEINC -9.249121 6.481116 -1.427088 0.1543
Prob(F-statistic) 0.002692

(2) Dependent Variable: HOURS

Method: Two-Stage Least Squares

8
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

Included observations: 428 after adjusting endpoints

Instrument list: EDUC AGE KIDSLT6 KIDSGE6 NWIFEINC EXPER

Variable Coe¢ cient Std. Error t-Statistic Prob.

C 2478.435 655.2070 3.782674 0.0002
LWAGE 1772.323 594.1850 2.982781 0.0030
EDUC -201.1870 69.91013 -2.877795 0.0042
AGE -11.22885 10.53692 -1.065668 0.2872
KIDSLT6 -191.6588 195.7609 -0.979046 0.3281
KIDSGE6 -37.73248 63.63485 -0.592953 0.5535
NWIFEINC -9.977747 7.174493 -1.390725 0.1650
Prob(F-statistic) 0.009204

(3) Estimation Method: Iterative Three-Stage Least Squares

Included observations: 428
Instruments: EDUC AGE KIDSLT6 KIDSGE6 NWIFEINC EXPER EXPER^2 C
Convergence achieved after: 5 weight matricies, 6 total coef iterations

Coe¢ cient Std. Error t-Statistic Prob.

C(1) 2511.719 566.1989 4.436106 0.0000
C(2) 1689.389 455.0691 3.712378 0.0002
C(3) -207.6063 54.69267 -3.795870 0.0002
C(4) -12.42447 8.683959 -1.430738 0.1529
C(5) -199.5608 139.0501 -1.435172 0.1516
C(6) -48.71302 36.00547 -1.352934 0.1764
C(7) 1.259363 3.309783 0.380497 0.7037
C(8) -0.705146 0.308275 -2.287395 0.0224
C(9) 0.000201 0.000213 0.943669 0.3456
C(10) 0.112975 0.015321 7.373960 0.0000
C(11) 0.020856 0.014085 1.480770 0.1390
C(12) -0.000293 0.000243 -1.207872 0.2274

Determinant residual covariance 98640.45

Equation 1: HOURS = C(1)+C(2)* LWAGE+C(3)* EDUC +C(4)*AGE +C(5)*KIDSLT6 +C(6)*KIDSGE6
+C(7)*NWIFEINC
Equation 2: LWAGE =C(8) +C(9)*HOURS+C(10)* EDUC +C(11)*EXPER +C(12)* EXPER^2

(4) Estimation Method: Generalized Method of Moments

Included observations: 428
Instruments: EDUC AGE KIDSLT6 KIDSGE6 NWIFEINC EXPER EXPER^2 C
White Covariance

9
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

Coe¢ cient Std. Error t-Statistic Prob.

C(1) 2688.763 527.2528 5.099570 0.0000
C(2) 1937.325 492.1760 3.936245 0.0001
C(3) -230.8447 54.08686 -4.268037 0.0000
C(4) -15.31151 9.043993 -1.693003 0.0908
C(5) -231.0814 157.5143 -1.467050 0.1427
C(6) -52.79678 38.84534 -1.359154 0.1745
C(7) -1.782789 3.457134 -0.515684 0.6062
C(8) -0.559507 0.364794 -1.533763 0.125
C(9) 0.000106 0.000247 0.430531 0.6669
C(10) 0.111260 0.014364 7.745982 0.0000
C(11) 0.020723 0.014140 1.465522 0.1432
C(12) -0.000261 0.000268 -0.975978 0.3294

Determinant residual covariance 143243.9

J-statistic 0.053546

Equation 1: HOURS = C(1)+C(2)* LWAGE+C(3)* EDUC +C(4)*AGE +C(5)*KIDSLT6 +C(6)*KIDSGE6

+C(7)*NWIFEINC
Equation 2: LWAGE =C(8) +C(9)*HOURS+C(10)* EDUC +C(11)*EXPER +C(12)* EXPER^2

(5) Dependent Variable: LWAGE

Method: Least Squares
Included observations: 428 after adjusting endpoints

Variable Coe¢ cient Std. Error t-Statistic Prob.

C 0.903705 0.267473 3.378678 0.0008
EXPER 0.046301 0.014007 3.305528 0.0010
EXPER^2 -0.000867 0.000421 -2.057972 0.0402
AGE -0.007040 0.005655 -1.244817 0.2139
KIDSLT6 0.015490 0.092304 0.167816 0.8668
KIDSGE6 -0.041348 0.029031 -1.424277 0.1551
NWIFEINC 0.012434 0.003316 3.749734 0.0002

R-squared 0.076983 Mean dependent var 1.190173

Log likelihood -450.9589 F-statistic 5.852191
Durbin-Watson stat 1.940720 Prob(F-statistic) 0.000007

(6) Dependent Variable: LWAGE

Method: Least Squares
Included observations: 428 after adjusting endpoints

10
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

Variable Coe¢ cient Std. Error t-Statistic Prob.

C -0.357997 0.318296 -1.124729 0.2613
EDUC 0.099884 0.015097 6.615970 0.0000
AGE -0.003520 0.005415 -0.650176 0.5159
KIDSLT6 -0.055873 0.088603 -0.630591 0.5287
KIDSGE6 -0.017648 0.027891 -0.632765 0.5272
NWIFEINC 0.005694 0.003320 1.715373 0.0870
EXPER 0.040710 0.013372 3.044344 0.0025
EXPER^2 -0.000747 0.000402 -1.860055 0.0636

R-squared 0.164098 Mean dependent var 1.190173

Log likelihood -429.7438 F-statistic 11.77879
Durbin-Watson stat 1.966359 Prob(F-statistic) 0.000000

i. Does the data con…rm your identi…cation analysis for the labor supply function in a.?
The key is to check identi…cation is the joint signi…cance of exper and exper 2 in the RF
for lwage; using a Wald test,

Ru2 Rr2 0:164098 0:131258

W =n = 428 = 15: 228;
1 Ru2 1 0:076983

which is signi…cative compared to a 22 (though exper 2 is not signi…cative). Note that

2SLS and 3SLS/GMM …tted models have to be taken with care because they need the
equation or the system to be identi…ed to the t- and F-tests to be valid, but they also dis-
play some doubts. In principle adding instruments makes a di¤erence in 2SLS, so under
the assumption of valid estimates, it could be some room to imagine overidenti…cation
(see also the J test below; also a Hausman test could be used, but not OLS estimation
to compare is available).
ii. Why do you expect that estimates in (1) to be more e¢ cient than those of (2)?
Because we are using more instruments (overidentifying restrictions).
Is always this the case?
Not always: e.g. when in the RF for lwage, exper 2 is not signi…cative ((6) indicates that
this might be the case). In principle we don’t have information about the conditional
variances to design more e¢ cient estimates with optimal instruments.
iii. Why do you think that in (3) are needed iterations?
Because we can use estimates of from intial 3SLS iterations, not just from e.g. initial
2SLS.
iv. Which indicates the J statistic in (4) [usual de…nitions requires to multiply this number
given by Eviews by n]? Give an example where J should be zero.
This is a test of speci…cation of the model (of the moment conditions), after multiplication
by n:
n J = 428 0:053546 =22:918;
it should be distributed as 2q ; where q is the number of overidentifying restrictions. In
this case we have that the …rst equation has 1 o.r. and the second equation has 3 o.r.,
so we should compare with a 24 : it appears to be signi…cative, and some speci…cation
problem could be present.

11
Econometrics II, 2005/06 Solutions Exam, June 23, 2006

v. Which estimates are more e¢ cient asymptotically, those in (3) or (4)? Specify clearly
your assumptions.
In principle (e¢ cient) GMM estimates are always e¢ cient given the moment conditions
(which …x the instruments to be used). But 3SLS can also be e¢ cient given conditional
homoskedasticity conditions.
2
c. [10%] Discuss identi…cation of the …rst equation when we add on the rhs two new terms: 13 (log(wage)) +
3
14 (log(wage)) :
In this case we don’t have enough linear instruments: we should rely on nonlinear transfor-
mations of the IV’s, which provide valid instruments for the quadratic and cubic terms in
the equation. Even if 13 = 14 = 0; the equation remains identi…ed because the reduced
2 3
forms for (log(wage)) and (log(wage)) contains cuadratic and cubic terms of the x0 s and
we have two linear instruments for log(wage):
d. [10%] Assume that educ is endogenous in the second equation, but exogenous in the …rst. Provide
conditions which guarantee the identi…cation of the system.
We need to provide a RF for educ; but at least in principle we have enough instruments
(age; kidslt6; kidsge6; mwif einc) from the …rst equation to instrumentalize both educ and
hours; so that the respective linear projections are not collinear.

2 2 2 2
(1) (2) (3) (4)
2 2 2 2
Pr (1) > 6:63 = 0:01 Pr (2) > 9:21 = 0:01 Pr (3) > 11:34 = 0:01 Pr (4) > 13:28 = 0:01
2 2 2 2
Pr (1) > 3:84 = 0:05 Pr (2) > 5:99 = 0:05 Pr (3) > 7:81 = 0:05 Pr (4) > 9:49 = 0:05
2 2 2 2
Pr (1) > 2:71 = 0:10 Pr (2) > 4:61 = 0:10 Pr (3) > 6:25 = 0:10 Pr (4) > 7:78 = 0:10

12