Scribd
Upload
609 views

Solutions Advanced Econometrics 1 Midterm 2014

1. The document provides solutions to exam questions on econometrics topics including ordinary least squares estimation, instrumental variables estimation, and nonlinear least squares estimation. 2. For instrumental variables estimation, the document shows that the two-stage least squares estimator is equivalent to the instrumental variables estimator when the instruments are a subset of the regressors. It also discusses the weak instrument problem. 3. For nonlinear least squares estimation, the document derives the score function and information matrix and shows the conditions for consistency and discusses why efficiency is not guaranteed compared to maximum likelihood.

Uploaded by

Esmée Winnubst
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
609 views

Solutions Advanced Econometrics 1 Midterm 2014

1. The document provides solutions to exam questions on econometrics topics including ordinary least squares estimation, instrumental variables estimation, and nonlinear least squares estimation. 2. For instrumental variables estimation, the document shows that the two-stage least squares estimator is equivalent to the instrumental variables estimator when the instruments are a subset of the regressors. It also discusses the weak instrument problem. 3. For nonlinear least squares estimation, the document derives the score function and information matrix and shows the conditions for consistency and discusses why efficiency is not guaranteed compared to maximum likelihood.

Uploaded by

Esmée Winnubst
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Solutions to Midterm Exam Advanced Econometrics 1, 22 October 2014

1.

(a) E[u] = E[E(u|x)] = E[0] = 0, so V[u] = E[u2 ] = E[E(u2 |x)] = E[x] = 21 .


(b) Write
P
N 1 N
i=1 xi ui
b
=+
.
P
2
N 1 N
i=1 xi
Because (ui , xi ) is iid, so is (xi ui , x2i ), so that that the LLN for iid sequences with finite
mean implies
N
1 X
xi ui = E[xu] = E[xE(u|x)] = 0,
N

plim

plim

i=1
N
X

1
N

x2i = E[x2 ] = 13 ,

i=1

b ) = 0.
which implies, by Slutskys Theorem, plim(
(c) Because E[xu] = 0 and V[xu] = E[x2 u2 ] = E[x2 E(u2 |x)] = E[x3 ] = 41 , the CLT for
iid observations implies
N


1 X
d

xi ui N 0, 14 .
N i=1
Using

together with plim N 1

b ) =
N (

N 1/2
N 1

PN

i=1
PN

xi ui

2
i=1 xi

PN

= 13 , and the product normal limit rule, we find


"
#
PN
1
1/2




x
u
N
d
i
i
i=1
b ) =
N 0, 41 2 N 0, 94 ,
N (
PN
2
1
(3)
N
i=1 xi
2
i=1 xi

b N [, 9 /N ].
or
4
(d) The (infeasible) GLS estimator is given by
0

e = (x

1 0

x)

PN
y=

2
i=1 xi yi / i
PN 2 2
i=1 xi / i

PN
=

xi yi /(xi )
Pi=1
N
2
i=1 xi /(xi )

PN

i=1
= PN

yi

i=1 xi

y
.
x

Note that drops out of the formula, so that the GLS estimator is in fact feasible. Using
y = x
+u
, and



d
Nu
N 0, 12 ,
we have

"
#

N
u

d
e ) =
N (
N 0, 21 2 N [0, 2] ,
x

(2)

e = 2/N . So the GLS estimator has a slightly smaller variance than the OLS
or V[]
estimator.
2.

(a) Because Z = [z1 : X2 ], it follows that X2 = ZA0 with A = [0 : IK1 ], and hence
PZ X2 = Z(Z0 Z)1 Z0 ZA0 = ZA0 = X2 ,

so that PZ X = PZ [x1 : X2 ] = [b
x1 : X2 ]. Therefore, the 2SLS estimator is
0 0
1 0 0
b

2SLS = (X PZ PZ X) X PZ PZ y

= [X0 Z(Z0 Z)1 Z0 X]1 X0 Z(Z0 Z)1 Z0 y


1

= (Z0 X)1 Z0 Z(X0 Z)

X0 Z(Z0 Z)1 Z0 y

= (Z0 X)1 Z0 y
b .
=
IV

The second equality follows from the fact that P0Z PZ = PZ because PZ is symmetric and
idempotent.
(b) The assumptions imply that u is independent of Z, hence E[Z0 u] = 0. This can be tested
only if the number of instruments exceeds the number of regressors. In this case it cannot be
0 b = 0 by definition.
b
b = y X
tested, since the 2SLS residuals u
2SLS satisfy Z u
b 2 . The Frisch-Waugh-Lovell theorem implies that
b1 = z1
(c) Write x
b 1 + X2
b
b01 M2 y
b1 )1 x

x01 M2 x
1,2SLS = (b
= (b
1 z01 M2 z1
b1 )1
b1 z01 M2 y,
where M2 = IN X2 (X02 X2 )1 X02 . This shows that a consistent estimator requires that
plim
b1 = 1 6= 0; so if this condition is violated, then there is a weak instrument problem
(or equivalently, a problem with instrument relevance).
3.

(a) We find
LN () =

N
X

ln f (yi |xi ; ) =

i=1

N 
X
i=1

2yi
ln 4 + ln yi 2 ln i ()
i ()


,

and taking first derivatives, using the chain rule and


exp(x0i )
i ()
=
= xi exp(x0i ) = xi i (),

we obtain



N 
N
X
LN () X
2
2yi
yi i ()
sN () =
=
+
xi i () =
2xi
.

i () i ()2
i ()
i=1

i=1

(b) Because of independent observations,


"
0

IN = E[sN ( 0 )sN ( 0 ) ] = E

N
X

(yi
4xi x0i

i=1

#
i ())2
.
i ()2

Using the law of iterated expectations, this reduces to


4

N
X
i=1


 


N
N
2
X
X


1
0
0
0 (yi i ())
x
=
4
E
x
x
V(y
|x
)
=
2
E
x
x
E E xi xi
.
i
i
i
i
i
i
i

i ()2
i ()2
i=1

Next, using
N
N
X
X
yi
sN ()
yi
0
=
2xi
i ()xi = 2
xi x0i ,
0
2
i ()
i ()

i=1
i=1

i=1

we have






N
N
X
X
sN
yi
yi
0
0
E
( ) = 2
xi xi = 2
E
E xi xi E
xi
i ()
i ()
0 0
i=1
i=1


= 2

N
X



E xi x0i = IN .

i=1

(c) Let
N
QN ()
1 X
hN () =
=
(yi i ())i ()xi .

N
i=1

The essential condition for consistency is E[hN ( 0 )] = 0, and using the law of iterated
expectations it is easy to see that this condition is satisfied, as
E[(yi i ())i ()xi ] = E[i ()xi E(yi i ()|xi )] = 0.
Regarding efficiency, we know that maximum likelihood is asymptotically efficient in the
1
sense that its asymptotic variance matrix is IN
, which is a lower bound for the variance
matrix of consistent and asymptotically normal estimators. Because the NLS estimator is
not equal to the maximum likelihood estimator (it solves other first-order conditions), we do
not expect it to be asymptotically efficient (which would mean asymptotically equivalent to
maximum likelihood). A formal analysis of the degree of inefficiency would require deriving
1
b
an expression for the asymptotic variance of
NLS and comparing it with IN .

You might also like