Generalized Method of Moments

1. Population moment condition

E [f (µ0 ; zt )] = 0

where zt is an (r £ 1) vector of observable variables, µ0 2 £ is a (k £ 1)

vector of true value of parameters and f : Rk £ Rr ! Rq vector valued
function.

2. Sample analog
1X
T
g (µ; ZT ) = f (µ; zt )
T t=1
¡ ¢
where ZT = zT0 ; zT0 ¡1 ; ¢ ¢ ¢; z10 : A proper version of law of large numbers
ensures that
1X
T
plim f (µ; zt ) = E [f (µ; zt )]
T t=1

3. Traditional method of moments estimator; Suppose that q = k; i.e., we have

k moment conditions for k unknown parameters. Then, we can …nd b µ such
that
³ ´ 1 X ³b ´
T
g bµ; ZT = f µ; zt = 0 ((*))
T t=1

We expect bµ to have good statistical properties such as consistency and

asymptotic normality. Good news, under some conditions, it is indeed the
case that p ³ ´
p
b b d
µ ! µ0 and T µ ¡ µ 0 ! N (0; V )

4. Generalized method of moments estimator; What if q > k? We have more

equations than unknowns. There does not exist any solution which satis…es
(*). A quick and easy solution is to discard (q ¡ k) equations - moment
conditions- and apply the method of moments technique. Then, which to
discard? The answer should be arbitrary. Moreover, the moment condi-
tions discarded do include valuable information on the parameter. The
generalized method of moments technique uses all q moment conditions by
 weighting them. Suppose that we have a sequence of (q £ q) positive semi
de…nite matrix WT converging to a positive de…nite matrix W0 : Then, GMM
estimator is de…ned as
b
µGMM = arg min J WT (µ; ZT ) = arg min [g (µ; ZT )]0 WT [g (µ; ZT )]
µ µ
" #0 " #
1 XT
1 XT
= arg min f (µ; zt ) WT f (µ; zt )
µ T t=1 T t=1

Since WT is positive semi de…nite, J (µ; ZT ) ¸ 0: The minimization problem

…nds the solution which makes the value of the objective function J (µ; ZT )
as near to zero as possible.

5. Under some regularity conditions,

p
b
µGMM ! µ0
1 X
T
p d
T g (µ0 ; ZT ) = p f (µ0 ; zt ) ! N (0;¡)
T t=1
p ³ ´
T b
d
µGMM ¡ µ0 ! N (0;V )

where

1X
T
@g (µ0 ; ZT )
¡ = plim = plim f (µ0 ; zt )
@µ0 T t=1
¡1 ¡1
V = [¡0 W0 ¡] [¡0 W0 ­W0 ¡] [¡0 W0 ¡]

with
j
X ¡ ¢
­ = lim E [f (µ0 ; zt )] [f (µ0 ; zt¡i )]0
j!1
i=¡j

6. Optimal weighting matrix; we can show that the optimal weighting matrix
which minimizes the asymptotic variance of the estimator is given by

W0 = ­¡1

Then,
p ³ ´
d
³ £ ¤¡1 ´
T bµGMM ¡ µ 0 ! N 0; ¡0 ­¡1 ¡

2
7. How to actually compute the GMM estimator;

1. Construct the sample analog g (µ; ZT ) of the population moments E [f (µ; zt )] :

2. Set WT = I and construct the objective function J I (µ) = [g (µ; ZT )]0 [g (µ; ZT )] :
Find
(1)
b
µGMM = arg min J I (µ)
µ

3. Calculate
l
X h i
b (2) = ¤
­ b (2)
0 + $ (h; l) ¤b (2) + ¤
b (2)0
h h
h=1

where
1 X h ³b(1) ´i0 h ³ (1) ´i
T
b f b
(2)
¤h = f µGMM ; zt µGMM ; zt
T t=h+1

and $ (h; l) is a kernel£for weighs.

¤ For example, Newey-West (Bartlett)
h
suggested $ (h; l) = 1 ¡ l+1 : The number of lags included in the
³ ´
estimation of ­b (2) should increase at a proper rate, say O T 23 ;as the
b (2) :
sample size T grows in order to ensure the consistency of ­
4. Construct next step objective function
h i¡1
b (2)
­ 0 b (2)
J (µ) = [g (µ; ZT )] ­ [g (µ; ZT )]

(1)
and …nd the minimum with b
µGMM as the starting value of iteration.
De…ne
(2) b (2)
b
µGMM = arg min J ­ (µ)
µ

5. Continue the previous two steps until convergence.

6. Asymptotic variance can be calculated as
20 ³ ´ 130 20 ³ ´ 13¡1
@g bµGMM ; ZT h ³ ´i¡1 @g b
µGMM ; ZT
Vb = 4@ A5 ­ b bµ GMM 4@ A5
@µ 0 @µ0

3
 8. Example I ; OLS estimator with i.i.d. error

"t = yt ¡ ¯ 0 x t
¡ ¢
E ("t ) = 0; E "2t = ¾ 2 ; E ("t "s ) = 0 t 6= s

Then,
E [f (µ0 ; zt )] = 0 )E [xt (yt ¡ ¯ 00 xt )] = 0
Therefore,

1X
T
1X
T
0
g (µ; ZT ) = f (µ; zt ) ) [xt (yt ¡ ¯ 0 xt )] [xt (yt ¡ ¯ 0 xt )] = 0
T t=1 T t=1

We have k moment conditions in k unknown parameters.

9. Example II; IV

E [f (µ0 ; zt )] = 0 )E [wt (yt ¡ ¯ 00 xt )] = 0

In case of i.i.d.,
" #0 " #
1X
T
1X
T
min J = [wt (yt ¡ ¯ 0 xt )] 0
[wt (yt ¡ ¯ xt )]
¯ T t=1 T t=1

In case of heterogeneous and correlated data,

" #0 " #
1X
T XT
1
min J = [wt (yt ¡ ¯ 0 xt )] ­¡1 [wt (yt ¡ ¯ 0 xt )]
¯ T t=1 T t=1

10. Example III; ML

" #
_l (µ0 )
E [s (µ0 )] = E =0
l (µ0 )
" #0 " #
1X
T
1X
T
min s (µ) s (µ) = 0
µ T t=1 T t=1

4
11. Example IV; intertemporal optimization

u0 (ct ) = ¯E [(1 + rt+1 ) u0 (ct+1 ) j Xt ]

¡ ¢
where Xt = x0t ; x0t¡1 x0t¡2 ; ¢ ¢ ¢ : Consider

c1¡°
t
u (ct ) = ° > 0 and ° 6= 1
1¡°
Then, £ ¤
c¡°
t = ¯E (1 + rt+1 ) c¡°
t+1 j Xt
" µ ¶¡° #
ct+1
E 1 ¡ ¯ (1 + rt+1 ) j Xt = 0
ct
parameter vector; (¯; °) ; data set (ct ; rt ) :