Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Motivation: The Hansen-Singleton problem

The non-linear GMM estimator


Asymptotic properties

GMM for non-linear models


Walter Sosa-Escudero
Econ 507. Econometric Analysis. Spring 2009

April 30, 2009

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Prelude: Hansen-Singletons consumer problem


Consider the following optimization problem for a representative
consumer:
maxct+i ,At+i

X
U (ct+i )
Et
(1 + )i
i=0

subject to:
At+i = (1 + r)At+i1 + yt+i ct+i
lim Et At+i (1 + r)i = 0

y = labor income, c = consumption of non-durables, A = wealth


(a financial asset) with return r. U is a utility function. is the
rate of intertemporal preferences.
Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Assume that:
U (ct+i ) =

c1
t+i
1

(CRRA specification). represents consumers risk aversion.


First order (Euler) conditions are:


1 + r

c
ct
Et
=0
1 + t+1
We will use this knowledge to obtain consistent estimates for and
, the deep parameters of the problem.

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

We can write the FOC as follows:


Et [g(xt+i , 0 )zt ] = 0

0
1+r 0
where g(xt+i , 0 ) 1+
c

c
, = (, )0 and zt is a
t
0 t+1
vector of n variables that are orthogonal to g(xt+i , 0 ).


This is a set of non-linear momment conditions.


If there is available a sample of size T , and n > p, we will use a
non-linear GMM strategy, that is, solve:
min

T
X

!0
g(xt+i , )zt

t=1

Walter Sosa-Escudero

T
X

!
g(xt+i , )zt

t=1

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Structure and assumptions

1
2

Random Sample: vi is an i.i.d. sequence of random variables.


Regularity: f (vi , ) : V 7 <q is a vector function that
satisfies: i) it is continuous on for each vi V ; ii)
E[f (vi , )] exists and is finite ; iii) E[f (vi , )] is
continuous on .
Moment condition: E[f (vi , 0 )] = 0

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Identification:
1
2

6= 0 for all 6= in .
Global identification: E[f (vi , )]
Regularity on derivatives: f (vi ; )/0 is p q matrix that
exists and is continuous on for each vi V ; ii) 0 is an
interior point of ; iii) E[f (vi ; )/]=0exists and is finite.
Local identification: E[f (vi ; )/]=0 = p.

Weighting: Wn is psd which converges in probability to a pd


matrix W .

Compactness: is compact.

Domination: E[sup ||f (vi , )||] < .

i) E[f (vi , 0 )f (vi , 0 )0 ] < , ii) limn


V ar[ g n (0 )] = S,
P
a finite pd matrix with gn (0 ) n1 nt=1 f (vi ; 0 ).
f (vi ; )/0 is continuous on some neighborhood of 0 .

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Identification
6= 0 for all 6= 0 is
The global identification condition E[f (vi , )]
difficult to characterize. Remember that in the IV case we tied
identification to a rank condition.
Local identification refers to conditions that hold in a small
neighborhood of 0 .
Under the assumed regularity conditions, we can expand f (.)
around 0 in a small neighborhood of 0 :
f (vi , ) ' f (vi , 0 ) +

Walter Sosa-Escudero

f (vi , 0 )
( 0 )
0

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

f (vi , 0 )
( 0 )
0
Taking expectations and using the moment conditions


f (vi , 0 )
E[f (vi , )] ' E
( 0 )
0
f (vi , ) ' f (vi , 0 ) +

which under the local identification condition is zero whenever


6= 0 .
Note the close similarity of the local identification condition and
the rank condition in the IV case.

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

The GMM estimator


Recall that the GMM objective function is:
 Pn
0

 Pn
t=1 f (vi , )
t=1 f (vi , )
Wn
Qn () =
n
n
and the GMM estimator is defined as:
n = argmin Qn ()
The FOCs for this problem are:
( n
)0
( n
)
1 X f (vi , n )
1X
Wn
f (vi , n ) = 0
n
0
n
t=1

t=1

a possibly non-linear system of q equations with p unknowns.


Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Asymptotics 1: Consistency

The GMM estimator is based on minimizing:


 Pn
0
 Pn

t=1 f (vi , )
t=1 f (vi , )
Qn () =
Wn
n
n
Define the population version of this function as:
Q0 () = E[f (vi , )]0 W E[f (vi , )]

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Result 1: Q0 () achieves a unique minimum at 0


By the moment condition E[f (vi , 0 )] = 0 and by the
6= 0 for
identification condition and the pd of W , E[f (vi , )]

any 6= 0 .
p

Result 2: Q0 () Qn () uniformly on .
Intuition:
if we fix at some point, by the LLN
Pn
p
t=1 f (vi , ) E[f (vi , )] and by assumption Wn W ,
then by continuity the result follows.
It is more difficult to make the uniform statement.

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Intuition of consistency:
n minimizes Qn ().
p

Qn () Q0 () uniformly.
0 minimizes Q0 ()
p
Then n 0
We will work out through a detailed proof in the homework.

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Asymptotics 2: Normality
Now we are in much more familiar territory...

P
P
Let gn () n1 nt=1 f (vi , ) and Gn () n1 ni=1 f (vi , )/0 .
Then, the FOCs of the GMM problem are:
Gn (n )0 Wn gn (n ) = 0
at 0 :
Now take a mean value expansion of gn (n)

gn (n ) = gn (0 ) + Gn ( ) n 0
where lies between and 0 . Now replace above:

Gn (n )0 Wn gn (0 ) + Gn (n )0 Wn Gn ( ) n 0 = 0

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties


Gn (n )0 Wn gn (0 ) + Gn (n )0 Wn Gn ( ) n 0 = 0


Multiply by n and solve for n n 0






n n 0 = Gn (n )0 Wn Gn ( ) Gn (n )0 Wn n gn (0 )

n n gn (0 )
= M


n Gn (n )0 Wn Gn ( ) Gn (n )0 Wn . Sounds familiar?.
with M

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Now we are definitely at home.

n n gn (0 )
n n 0 = M

n does not explode and that n gn (0 ) is


We will show that M
asymptotically normal
n . Under the continuity assumptions we get by
We start the M
an appropriate LLN:
p
p
Gn (n ) G0 and Gn ( ) G0

where G0 E[f (vi ; 0 )/0 ]. Then:







p
Gn (n )0 Wn Gn ( ) Gn (n )0 Wn G00 W G0 G00 W M0

which is a finite matrix.


Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Regarding ngn (0 ), by the iid and regularity assumptions, we can


apply the CLT to show:
Pn

f (vi ; 0 ) d
N (0, S)
ngn (0 ) = n t=1
n
Then by Slutzkys theorem

d
n n gn (0 )
n n 0 = M
N (0, M0 SM00 )

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

We can write the FOC as follows:


Et [g(xt+i , 0 )zt ] = 0

0
1+r 0
where g(xt+i , 0 ) 1+
c

c
, = (, )0 and zt is a
t
0 t+1
vector of n variables that are orthogonal to g(xt+i , 0 ).


This is a set of non-linear momment conditions.


If there is available a sample of size T , and n > p, we will use a
non-linear GMM strategy, that is, solve:
min

T
X

!0
g(xt+i , )zt

t=1

Walter Sosa-Escudero

T
X

!
g(xt+i , )zt

t=1

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

Empirical example: Hansen-Singleton reloaded

Data:
x1t = ct1 /ct
x2t = 1 + rt
In these terms, the moment conditions can be written as:


Et x1,t+1 x2,t+1 1 zt = 0
with (1 + )1 . Following Hansen and Singleton, for zt we will
take a constant, rt and ct1 /ct .

Walter Sosa-Escudero

GMM for non-linear models

Motivation: The Hansen-Singleton problem


The non-linear GMM estimator
Asymptotic properties

The estimated discount factor is: 0.998 (exponential)


Risk aversion parameter: 0.89 (risk adverse), careful, not
significant.
Walter Sosa-Escudero

GMM for non-linear models

You might also like