Lecture Notes

Generalized Method of Moments

Based on notes and book by Alastair Hall (NCSU)

Lecture outline:

1. Introduction:

2. GMM estimation in correctly specied nonlinear dynamic models - part I.

3. GMM estimation in correctly specied nonlinear dynamic models - part II.

4. Hypothesis testing.

5. Finite sample behaviour and moment selection.

Today:

Why GMM?
Statistical antecedents.
Contemporay example.
IV estimator in linear regression model.

1. Introduction
Hansen (1982, Econometrica) introduced the
Generalized Method of Moments (GMM) estimator.
Two advantages:

Provides a general framework for considering issues of statistical inference because

it encompasses many estimators of interest
in econometrics.

Provides a computationally convenient method

for the estimation of nonlinear dynamic models without complete specication of the
probability distribution of the data.

1.1 Statistical antecedents:

Method of Moments (Pearson,1893, 1894, 1895)
Suppose we wish to estimate the population
mean, , and the population variance, 2, of a
r.v. vt.

These two parameters satisfy the population

moment conditions
E[vt] = 0
E[vt2] ( 2 + 2) = 0
Pearson's method based on analogous sample
moment conditions.

T 1

T
X

t=1

T 1

T
X

t=1

vt
^ = 0

vt2 (^
2 +
^2) = 0
6

This implies

^ = T 1

^2 = T 1

T
X

vt

t=1
T
X

(vt
^)2

t=1

Key idea: population moment conditions provide information upon which estimation of parameters can be based.

Minimum chi-square (Neyman and Pearson, 1928)

Outcome of an experiment lies in one of k

mutually exclusive and exhaustive groups.
pi = P(outcome lies in the ith group).
H0 : pi = h(i; 0)
For given 0, Karl Pearson had shown that under H0
GFT (0) = T

k
X
[^
pi h(i; 0)]2 d

i=1

p^i

! 2
k1

Now suppose 0 unknown.

To perform inference need estimate of 0
^T = argmin GFT ()

- minimum chi-square estimator

Connection to moments:
Let fDt(i); i = 1; 2; : : : k; t = 1; 2; : : : T g satisfy:
Dt(i) = 1 if tth outcome in ith group
Dt(i) = 0 else
)

P (Dt(i) = 1) = h(i; 0)
E[Dt(i)] = h(i; 0)

) k population moment conditions

2

Dt(1) h(1; 0)
6
6 Dt(2) h(2; 0)
6
E6
:
6
6
:
4
Dt(k) h(k; 0)

Sample analogs are given by

2

p^ h(1; )
6 1
6 p
^ h(2; )
6 2
6
:
6
6
:
4
p^k h(k; )

7
7
7
7 = 0
7
7
5

7
7
7
7 = 0
7
7
5
10

GFT () = T
2
6
6
6
6
6
6
4

30 2 1
p^
0
p^1 h(1; )
7 6 1
1
6 0
p^2 h(2; ) 7
p
^
2
7 6
7
6
:
:
7 6 :
7 6
:
:
5 4 :
p^k h(k; )
0
0

:
:
:
:
:

: 0
: 0
:
:
:
:
: p^1
k

32

p^ h(1; )
76 1
76 p
7 6 ^2 h(2; )
76
:
76
76
:
54
p^k h(k; )

) GFT () = a quadratic form in the samplemoment condition.

minimum chi-square estimator = which is

closest to solving the sample moment conditions in the metric of GFT ().

11

Instrumental variables (Wright, 1925, 1928)

Simple agricultural demand and supply system:
qtD = pt + uD
t
qtS = 1nt + 2pt + uS
t

(1)

qtD = qtS = qt

qtD , qtS represent demand and supply in year

t.

pt is the price of the commodity in year t

qt equals quantity produced
Problem: how to estimate given fqt; ptg.
Ordinary Least Squares (OLS) regression of qt
on pt runs into problems.
12

Wright's solution:
Find: ztD such that Cov(ztD ; uD
t ) = 0.
Then (1) )
Cov(ztD ; qt) 1Cov(ztD ; pt) = 0

(2)

E[ztD qt] 1E[ztD pt] = 0

(3)

and so if E[uD
t ] = 0 then

Pearson's Method of Moments principle leads

to:

^1 =

T
X

t=1

ztD qt=

T
X

ztD pt

(4)

t=1

Instrumental variables (IV) estimator with ztD

known as \instrument".

13

Generalized Method of Moments

Population Moment Condition:
E[f (vt; 0)] = 0

for all t

Generalized Method of Moments Estimator:

^T = argmin2QT ()
where
QT () = T 1

T
X

f (vt; )0WT T 1

t=1

T
X

f (vt; )

t=1
p

and WT is psd and WT ! W , pd.

WT psd ) QT () 0 and QT (^T ) = 0 if
PT
1
^
T
t=1 f (vt; T ) = 0
W pd ) QT (^T ) = 0 i T 1
0 in the limit as T ! 1.

PT
^
t=1 f (vt; T ) =
14

1.2 Contemporary example

Consumption based asset pricing model: Hansen
& Singleton (1982)
A represenative agent chooses consumption to
maximize his/her expected discounted utility
E[

1
X

iU (ct+i)j-t]

i=0

subject to
ct + ptqt = rtqt1 + wt
for all t, where

rt+1 is the return on the asset in period

t+1

pt is the price of the asset in period t

ct is consumption in period t
15

 -t is the information set available to the

agent in period t

Optimal path of consumption and investment

satises
ptU 0(ct) = E[rtU 0(ct+1)j-t]

(5)

for all t.

16

Equation (5) can be rewritten as

E[(rt+1=pt)fU 0(ct+1)=U 0(ct)gj-t] 1 = 0 (6)
Hansen & Singleton (1982) set

U (ct) = ct =
and so (6) becomes
E[(rt+1=pt)(ct+1=ct)1j-t] 1 = 0

(7)

Two parameters to be estimated: (; ).

Model implies following population moment condition:
Let
et(; ) = (rt+1=pt)(ct+1=ct)1 1
then (7) ) for any vector zt 2 -t
E[et(; ) zt] = E [ E[et(; ) j-t] zt] = 0
17

1.3 Instrumental variables estimation of the

static linear model

Population moment condition and identication

Estimator
Identifying and overidentifying restrictions
Asymptotic properties of the estimator
Covariance matrix estimation
Two step estimator
Model specication test
18

Consider:
yt = x0t0 + ut;

t = 1; 2; : : : T

xt is a (p1) vector of observed explanatory

variables;

yt is scalar, observed;
ut is the unobserved error term;
zt is a (q 1) vector of instruments.
Problem: to estimate 0.

19

(i) Population moment condition and identication:

Assumption 2.1: Strict Stationarity
The random vector vt = (x0t; zt0 ; ut)0 is a strictly
stationary process.
Let ut() = yt x0t.
Assumption 2.2: Population Moment Condition
E[ztut(0)] = 0.
Also need: Identication condition:
E[ztut()] 6
= 0 for all 6
= 0
E[ztut()] = E[ztut(0)] + E[ztx0t](0 )
= E[ztx0t](0 )
Assumption 2.3: Identication Condition
rankfE[ztx0t]g = p.
20

Relationship between q and p important:

q < p - 0 is under- (or un- ) identied

q = p - 0 is just identied
q > p - 0 is over- identied

21

(ii) The estimator:

^T = argmin2 QT ()
where
QT () = fT 1u()0ZgWT fT 1Z 0u()g
First order conditions:
(T 1X 0Z)WT (T 1Z 0y) = (T 1X 0Z)WT (T 1Z 0X)^T
)

^T = fX 0ZWT Z 0Xg1X 0ZWT Z 0y

22

(iii) Identifying and overidentifying restrictions:

Return to rst order conditions:
(T 1X 0Z)WT T 1Z 0u(^T ) = 0
) GMM = MM based on:
E[xtzt0 ]W E[ztut(0)] = 0
q = p: GMM = MM based on E[ztut(0)] =
0
q > p: GMM sets p linear combinations of
E[ztut(0)] equal 0
Rewrite population FOC:
F 0W 1=2E[ztut(0)] = 0
where

F0

0
0
1=2
E[xtzt]W

F (F 0F )1F 0W 1=2E[ztut(0)] = 0
identifying restrictions

23

Remainder is
(Iq F (F 0F )1F 0)W 1=2E[ztut(0)] = 0
overidentifying restrictions.
Sample analogs:

Identifying restrictions satised at ^T

1=2 1 0 ^
T Z u(T )

Overidentifying restrictions: WT
Now,

1=2
QT (^T ) = kWT T 1Z 0u(^T )k

) QT (^T ) measures how far the sample is from

satisfying the overidentifying restrictions.

24

(iv) Asymptotic properties:

It can be shown that:

^T is consistent for 0
q

a
^T;ii=T
N (0; 1) where
(^T;i 0;i)= V

^T = (X 0ZWT Z 0X)1X 0Z SZ
^ 0X(X 0ZWT Z 0X)1
{ V
p
^T !
{ S
limT !1V ar[T 1=2Z 0u]

25

(v) Covariance matrix estimation:

For this model:
V ar[T 1=2Z 0u] = E[fT 1=2

T
X

t=1

T 1

T
X

ztutgfT 1=2
0

E[u2
t ztzt]

t=1

Therefore,
^T =
S

T 1

T
X

u(^T )2ztzt

t=1

26

T
X

t=1

ztutg

(vi) Two step estimator:

Notice that asymptotic variance depends on
weighting matrix.
^1
Optimal choice: WT = S
T
^T
Problem: Need ^T to construct S
Two{step procedure:

1. Estimate with sub{optimal WT ! ^T (1) !

^T (1).
S
^T (1)1.
2. Estimate with WT = S

27

(vii) Model specication test:

Identifying restrictions satised in sample

regardless of whether model is correct.

Overidentifying restrictions not imposed in

sample.

The overidentifying restrictions test

^1 T 1=2Z 0u(^T )
JT = T QT (^T ) = T 1=2u(^T )0Z S
T
Under H0 : E[ztut(0)] = 0
d

JT ! 2
qp

28

Lecture outline:

1. Introduction

2. GMM estimation in correctly specied

nonlinear dynamic models - part I.

3. GMM estimation in correctly specied nonlinear dynamic models - part II.

4. Hypothesis testing.

5. Finite sample behaviour and moment selection.

Today:

Identication
The Estimator
Identifying and overidentifying restrictions
Asymptotic properties
Covariance Matrix estimation

1. Population moment condition and identication

First impose condition on data:

Strict Stationarity
The (r 1) random vectors fvt; 1 < t < 1g
form a strictly stationary process with sample
space V <r .
Population Moment Condition
Let 0 be a vector of unknown parameters
which are to be estimated, vt be a vector of
random variables and f (:) a vector of functions
then a population moment condition takes the
form
E[f (vt; 0)] = 0

for all t:

(1)
4

Global Identication
The parameter vector 0 is globally identied
by the population moment condition in Assumption 3.3 if and only if E[f (vt; )] 6
= 0
for all 2 such that 6
= 0.
\global" ) pmc only holds at one value in the
entire parameter space.
Identication failures can sometimes be diagnozed, but it is often dicult.

Example: Eichenbaum's (1989) Model for Inventory Holdings by Firms

Population moment condition takes form
E[ztht+2(0)] = 0
where
ht+1() = It+1 f + ()1gIt + 1It1
+St+1 (1 =) 1St
and = (; ; ; ; ).
(; ; ) are globally unidentied.
Reason: they only appear in pmc via
(; ; ) = 1 =
and
(; ; ) = (k; ; k) = (; k; k)
Notice problem goes away if we set = 1=
and set = (; ; ).
6

In general identication failures arise due to

nonlinearity, properties of data and interaction
of both.
This can be hard to spot via condition for
\global identication" ! local identication.
This concept is based on restricting attention
to suciently small neighbourhood of 0 such
that rst order Taylor series expansion holds
f (vt; ) f (vt; 0) + f@f (vt; 0)=@0g( 0)
Note that f (vt; ) can be approximated by a
linear function of 0.
Taking expectations and using E[f (vt; 0)] = 0
yields
E[f (vt; )] fE[@f (vt; 0)=@0]g( 0)
7

This gives the following condition:

Local Identication
The parameter vector 0 is locally identied by
the population moment condition in Assumption 3.3 if and only if E[@f (vt; 0)=@0g] is of
rank p.
Note:
(i) Condition implies identication fails if there
are fewer moment conditions than parameters,
i.e. q < p.
(ii) Identication in nonlinear models may be
sensitive to the value of 0.
(iii) No local id ) No global id.
This condition is still far from transparent.

Example: Eichenbaum's (1989) Model for Inventory Holdings by Firms (Continued)

First consider version with ; ; included.
E[@f (vt; 0)=@0] = E[ztx
~0t]M(0)
where x
~t = (It; It1; St)0 and for some (3 5)
matrix M (0).
Therefore,
rank(M()) 3 ) rankfE[@f (vt; 0)=@0]g 3 < p
and so 0 is locally unidentied in this model.

Note that if we set = 1 = and =

(; ; )0 then
E[@f (vt; 0)=@0] = E[ztx
~0t]M(0)
where
2

6
M() = 4

(2)1
0
0

+( 2)1
2
2

0
7
0 5
1

This time rankfM()g = 3.

Note that local identication depends on whether
rankfE[ztx
~0t]g = 3.

10

Terminology:

If p > q then 0 is said to be un { identied

If p = q and 0 is local id. then said to be
just { identied

If q > p and 0 is local id. then said to be

over { identied.

11

2. GMM estimation
GMM minimand is:
QT () = fT 1

T
X

t=1

f (vt; )g0WT fT 1

T
X

f (vt; )g

t=1

where WT is the weighting matrix.

GMM estimator of 0 is
^T = argmin2 QT ()
First order conditions for this minimization imply
@QT (^T )=@ = 0
and so
fT 1

T
T
X
X
@f (vt; ^T ) 0
1
g WT fT
f (vt; ^T )g = 0
0
@
t=1
t=1

12

3. Identifying and overidentifying restrictions

GMM based on E[f (vt; 0)] = 0 is MM based
on
F (0)0W 1=2E[f (vt; 0)] = 0
where F (0) = W 1=2E[@f (vt; 0)=@0].
If p = q then GMM = MM based on
E[f (vt; 0)] = 0
If q > p then GMM is MM based on:
F (0)[F (0)0F (0)]1F (0)0W 1=2E[f (vt; 0)] = 0
identifying restrictions
Remainder is:
fIq F (0)[F (0)0F (0)]1F (0)0gW 1=2E[f (vt; 0)] =
overidentifying restrictions
13

So GMM eects a decomposition on the population moment condition:

pmc = id.res + overid. res
id.res ! estimation
overid.res ! remainder
Roles reected in sample analogs.
Sample analog to id.res are satised at ^T by
construction.
Sample analog to overid. res not satised but
1=2

WT

1=2
gT (^T ) = fIq PT (^T )gWT gT (^T )

) QT (^T ) measures of how far the sample is

from satisfying the overidentifying restrictions.
14

4. Asymptotic properties of ^T

p
Consistency: ^T ! 0

Proof based on population analog to QT (),

Q0() = fE[f (vt; )]g0 W fE[f (vt; )]g
p

sup2 jQT () Q0()j ! 0.

pmc ) Q0(0) = 0
W = p.d + ident.
6
= 0.

) Q0() > 0 for all

15

d
Asymptotic normality: T 1=2(^T 0) ! N (0; MSM 0)
where

M = (G00W G0)1G00W .
G0 = E[@f (vt; 0)=@0]
S = limT !1 V ar[T 1=2gT (0)]
MVT: gT (^T ) = gT (0)+GT (^T ; 0; T )(^T 0)
Premultiply both sides by GT (^T )0WT , use FOC
and rearrange to give:
T 1=2(^T 0) = MT T 1=2gT (0) + op(1)

16

5. Covariance matrix estimation

Need estimate of long run variance S.

S = 0 +

1
X

(i + 0i)

i=1

where j = E[(ft E[ft])(ftj E[ftj ])0]

Consider three cases.

Case 1: ft serially uncorrelated sequence

^SU = T 1
S

T
X

f^tf^t0

t=1

where f^t = f (vt; ^T ).

17

Case 2: ft invertible VARMA:

Estimate the model
f^t = A1f^t1 + : : : + Ak f^tk + et(k)
by OLS. Then
^V ARM A = fIq
S
^
where (k)
=

k
X

^i(k)g1(k)fI
^
A
q

i=1
PT
1
T
et(k)^
et(k)0.
t=1 ^

k
X

^i(k)g1
A

i=1

Choice of k:

Estimate for k = 0; 1; : : : K and choose using information criterion.

K = O(T 1=3)

18

Case 3: heteroscedasticity autocorrelation covariance estimators

^0 +
^HAC =
S

b(T
X)

^i +
^ 0i)
!iT (

i=1

where
^ i = T 1

T
X

f^tf^ti

t=i+1

choice of kernel: e.g. !i;T = 1i=[b(T )+1]

choice of bandwidth: b(T ) ! 1 and b(T ) =
o(T 1=2)

Evidence suggests HAC estimators do not perform well if ft has slowly decaying autocovariances i.e. a strong autoregressive component.
! rewhitening and recolouring

19

Estimate the model

f^t = A1f^t1 + : : : + Ak f^tk + et(k)
by OLS. Then
^P W RC =
S

k
X

k
X
0
1
^ HAC fIq
^i(k)g
^i(k)g1
fIq
A
A
i=1
i=1

^ HAC is HAC estimator based on ^

where
et(k).

20

Numerical optimization
Three important aspects of to numerical optimization routines.
The starting value for , (1).
The iterative search method by which the
candidate value of ^ is updated on the ith
step.
The convergence criterion used to judge
when the minimum has been reached.
Example of iterative routine: Newton{Raphson
(NR) algorithm
(j)
(j1)
^T = ^T

31
(2)
(j1)
@ 2QT (^T
)
5
4
0
@@
2

(2) (j1)
@QT (^T
)
@

Example of Convergence criterion:

(j)

jj^T

(j1)
^T
jj <
21

Numerical illustration: Hansen & Singleton (1982)

CBAPM

Single asset = equally weighted NYSE index (EWR) or value weighted NYSE index
(VWR)

ct = aggregate per capita consumption

zt = (1; ct=ct1; ct1=ct2; rt=pt1; rt1=pt2)0
Sample: 1960.1{1991.12
Need to choose WT : use 105I5 and (T 1Z 0Z)1.
22

Notice sensitivity to:

1. starting value
2. convergence criterion
3. WT
Solutions:
1. Try dierent starting values
2. Gradually tighten convergence criterion and
try dierent criterion if possible.
3. Characterize \optimal" WT { to do this, we
must rst develop a criterion for optimality
! asymptotic theory.
25

Lecture outline:

1. Introduction

2. GMM estimation in correctly specied nonlinear dynamic models - part I.

3. GMM estimation in correctly specied

nonlinear dynamic models - part II.

4. Hypothesis testing.

5. Finite sample behaviour and moment selection.

Today:

Two step and iterated GMM estimation

Impact of transformations
Continuous updating GMM estimator

1. Two - step and iterated GMM estimation

d
Recall that: T 1=2(^T 0) ! N (0; MSM 0)
where

M = (G00W G0)1G00W .
G0 = E[@f (vt; 0)=@0]
S = limT !1 V ar[T 1=2gT (0)]
If p = q then MSM 0 = (G00S 1G0)1
{ independent of W .

If p < q then asymptotic variance depends on

W.

Optimal choice of W ? Choice that minimizes

variance.

Optimal weighting matrix: W = S 1

! MSM 0 = (G00S 1G0)1

How can we obtain an estimator that has this

variance?

Two{step procedure:
1. Estimate with sub{optimal WT ! ^T (1) !
^T (1).
S
^T (1)1 ! ^T (1).
2. Estimate with WT = S
Can iterate further
^T (i 1)
1. ^T (i 1) ! S
^T (i 1)1 ! ^T (i).
2. Estimate with WT = S
Continue until k^T (i1)^T (i)k < or i = imax.
6

Numerical illustration: Hansen & Singleton (1982)

CBAPM
pmc : E[et(; ) zt] = 0
where

and

et(; ) = (rt+1=pt)(ct+1=ct)1 1

Single asset = equally weighted NYSE index (EWR) or value weighted NYSE index
(VWR)
ct = aggregate per capita consumption
zt = (1; ct=ct1; ct1=ct2; rt=pt1; rt1=pt2)0
Sample: 1960.1{1991.12
As before for rst step use WT = 105I5 and
(T 1Z 0Z)1.
7

2. Impact of transformations
Consider ve types of transformation:

Units of measurement for vt

Reparameterization
Normalization of the parameter vector
Curvature altering transformations of the
population moment condition

Stationarity inducing transformations of f (:)

10

(i) Units of measurement for vt:

In general, the GMM estimator is not invariant
to changes in the units of measurement of vt.
Example: E[vt] 0 = 0 ! ^T = T 1

PT
t=1 vt.

Data are xt = cvt

E[xt] 0 = 0 ! ~T = T 1

PT
t=1 xt

So ~T = c^T
But interpretation of 0 has changed!

11

(ii) Reparameterization:

0 satises:
globally identied
can be written as 0 = h(0) where h :
<p ! <p is a continuous, dierentiable bijective mapping.

pmc: E[f (vt; h(0))] = E[f (vt; 0)] = 0

What is relationship between:

estimate 0 based on E[f (vt; 0)] = 0

estimate of 0 based on E[f (vt; 0)] = 0?
12

The GMM estimator is invariant to reparameterization in the sense that the two parameterizations yield logically consistent estimators.
Q;T () = GMM minimand associated with the
reparameterized model, i.e. Q;T () = QT (h()).
^
T = argmin Q;T ()
Can calculate ^
T as follows.
min QT (h()) wrt h() ! ^
hT
^
hT = h(^
T ) ! ^
T .
But ^
hT = ^T and so by construction
^T = h(^
T )
13

Similar result does not extend to the estimated

asymptotic standard errors.
It can be shown that
^;T = [H(^
^;T [H(^
V
T )]1V
T )0]1
where H(:) = @h(:)=@ 0.
So inferences may be sensitive.
This sensitivity is a potential source of concern, and motivates an alternative method for
the construction of condence intervals covered later.
\natural parameterization" argument

14

(iii) The GMM estimator and normalization of

the parameter vector
In some cases, 0 may only be identied up
to some scaling factor and so it is necessary
to impose some normalization on 0, such as
0;1 = 1, in order to achieve identication.
In general, the GMM estimators associated with
dierent normalizations of the parameter vector do not exhibit a logical consistency in nite
samples. However, they do exhibit a logical
consistency in the limit.
Example: Suppose that
et() = 1R1;t + 2R2;t + 3It + 4St
where
E[ztet(0)] = 0
0 is unidentied
15

Normalize to achieve identication:

Divide et(0) by 0;1:

~
et(0) = R1;t + 0;1R2;t + 0;2It + 0;3St
where 0;i = 0;i+1=0;1.
Divide et(0) by 0;4:

et(0) = 0;1R1;t + 0;2R2;t + 0;3It + St

where 0;i = 0;i=0;4.

Both normalizations are logically consistent.

16

These normalizations lead to pmc's:

E[zt~
et(0)] = 0
E[zt
et(0)] = 0
Corresponding estimators are:
2

^T = 4(T 1

T
X

^T = 4(T 1

x1;tzt)WT

(T 1

t=1

ztx1;t

t=1

(T 1

T
X

x1;tzt)WT

(T 1

T
X

x2;tzt)WT

(T 1

(T 1

T
X

t=1

T
X

x2;tzt)WT

ztx2;t

t=1
T
X

31
)5

ztR1;t)

t=1

t=1

t=1

T
X

(T 1

31
)5

T
X

ztSt)

t=1

^T and
^T are not logically consistent but ex
hibit this property in the limit.
17

(iv) Curvature altering transformations of the

population moment condition
Model implies E[f (vt; 0)] = 0 but estimation
is based on c(0)E[f (vt; 0)] = 0.
GMM is invariant to curvature altering transformations if p = q but only in the limit if p < q.
p = q: ^T solves c(^T )T 1

PT
^
t=1 f (vt; T ) = 0.

p < q: FOC are

9
8"
#
T
=
< @c(
^T ) 1 X
0
0
T
f (vt; ^T ) + c(^T )GT (^T )
;
:
@
t=1
T
X
1
WT T
f (vt; ^T ) = 0
t=1

18

(v) Stationarity inducing transformations of f (:)

Model implies E[f (v~t; 0)] = 0 but ~
vt is nonstationary.
Seek H(~
vt1; 0) such that:
H(~
vt1; 0)f ((~
vt; 0) = h(vt; 0) where vt is
stationary
E[h(vt; 0)] = 0.
So that GMM estimation can be based on the
scaled moment condition E[h(vt; 0)] = 0.
If one stationarity inducing transformation of
f (:) then many.
GMM estimator is sensitive to the choice of
transformation in nite samples, but is consistent no matter which transformation is used.
19

Example: Consumption based asset pricing model

Recall that FOC

1

ptct 0

0
= 0E[rt+1ct+1
j-t]

Since ptct 0
2 -t, both sides of this equation
1
were divided by ct 0 pt to give
E[0(rt+1=pt)(ct+1=ct)01 1j-t] = 0
However FOC also implies
1

0
E[0rt+1ct+1
ptct 0

j-t] = 0

Take rst approach because x1;t+1 = ct+1=ct

and x2;t+1 = rt+1=pt are stationary but (ct; rt; pt)
are not.

20

Transformation not unique.

If wt 2 -t is a stationary random variable, then
1
division of FOC by wtrtct 0
yields
E[0wt1(rt+1=pt)(ct+1=ct)01 wt1j-t] = 0
(1)
Now consider pmc: E[ut(0)zt] = 0. If apply
similar argument to (1) then obtain
E[ut(0)~
zt] = 0
where z~t = wt1zt.

21

3. Continuous updating GMM estimator

Recall optimal choice of weighting matrix W =
S 1.
Recall also that S = limT !1 V ar[T 1=2gT (0)]
and so write S = S(0).
Population analog to the GMM minimand is
Qpop() = E[f (vt; )]0S()1E[f (vt; )]
In the iterated estimation,
^T (i 1)1gT ()
Qiter;T () = gT ()0S
Alternative is to minimize
Qcont;T () = gT ()0ST ()1gT ()
The continuous updating GMM estimator is
dened to be,
^cont;T = argmin2 Qcont;T ()
22

Properties of Continuous updating GMM estimator:

numerically diferent from two-step or iterated estimator

has same limiting distribution as two-step

or iterated estimator

invariant to curvature altering transformations

23

Condence sets that are invariant to reparameterization.

d

T Qcont;T (0) ! 2
q
Therefore an asymptotically valid 100(1 )%
condence set for 0 is then given by
f : T Qcont;T () < cq () g
where cq () is the 100(1 )% percentile of
2
q distribution.

24