Wiji Arulampalam, May 2006

GMM
1. OLS as a Method of Moment Estimator
Consider a simple cross-sectional case
yi = xi β + ui i=1,..,N β true coeff

If E(xi’ ui) = 0 ⇒ E[xi’ (yi -xi β)]=0 [OLS assumption]

The MM estimator solves the sample moment condition:

1
∑
N i
x '
i yi − (
x '
)
i β̂ = 0 giving you β̂= ( X ' X )
−1
(X'y) [OLS]

1
Wiji Arulampalam, May 2006

2. Instrumental Variable Estimation

Now assume that some of the x variables are correlated with the error term.
OLS estimator is inconsistent. We use instrumental variable estimation
using say z as instruments. Assume number of instruments=L and L ≥ K.

The population moment conditions are: E(zi’ui)=0

Then IV estimation solves:

1
∑
N i
z '
i(yi − x )
'
iβ = 0 (1)

2
Wiji Arulampalam, May 2006

The above involves L equations in K unknowns.

Note if L<K, we can’t solve for our estimates.

If L=K, we have K equations and K unknowns and hence have a unique

solution giving you

β= ( Z ' X )−1 ( Z ' y ) . [Simple IVE]

However, if L > K, then we have more equations than unknowns. One

inefficient solution would be to just select K instruments out of the set of
3
Wiji Arulampalam, May 2006

L. But instead, it is better to do something that is more efficient. This is

the GMM estimator.

GMM chooses β to make (1) as small as possible using quadratic loss. i.e.
GMM estimator β minimises

'
⎡1 ⎤ ⎡1 ⎤
QN(β)= ⎢ ∑ z i ( yi − xiβ) ⎥ WN ⎢ ∑ z i ( yi − xiβ) ⎥
' ' ' '

⎣N i ⎦ ⎣N i ⎦

WN is an LxL matrix of weights which is chosen ‘optimally’ [i.e. giving

you the smallest variance GMM estimator].
4
Wiji Arulampalam, May 2006

The solution is

−1
ˆβ = ⎡⎛ X' Z ⎞ W ⎛ Z' X ⎞ ⎤ ⎡⎛ X' Z ⎞ W ⎛ Z' y ⎞ ⎤ (N cancels)
⎢⎜ ∑ i i ⎟ ⎜ ∑ i i ⎟ ⎥ ⎢⎜ ∑ i i ⎟ ⎜ ∑ i i ⎟ ⎥
⎣⎝ i ⎠ ⎝ i ⎠ ⎦ ⎣⎝ i ⎠ ⎝ i ⎠⎦

= ( X'ZWZ'X ) ( X'ZWZ'Y )
−1

This optimal weighting matrix should be a consistent estimate up to a

multiplicative constant of the inverse of the variance of the orthogonality
conditions:

5
Wiji Arulampalam, May 2006

W = Λ−1 where Λ ≡ E(Zi’uiui’Zi) = Var(Zi’ui)

Can show that the GMM estimator is consistent.

The asymptotic variance of the optimal GMM estimator is estimated using

−1
⎡ ⎛ ⎞
−1
⎤
AVAR β̂ = ⎢( X'Z ) ⎜ ∑ Z i 'uˆ i uˆ i ' Z i ⎟ ( Z'X )⎥ (N cancels) (2)
⎢⎣ ⎝ i ⎠ ⎥⎦

6
Wiji Arulampalam, May 2006

−1
⎡1 ⎤
In order to generate the residuals, use GMM with W = ⎢
⎣N
∑ i i ⎥⎦
Z 'Z

This is the GIVE (2SLS) estimator:

( ) ( Z'Y )
−1 −1 −1
= X'Z[Z'Z] Z'X X'Z[ Z'Z ]

This is the same as assuming that

Var(Zi’ui) = E(Zi’uiui’Zi)= σ2 (Zi’Zi)

• Need to be able to invert the above matrices….rank condition!

7
Wiji Arulampalam, May 2006

−1
⎡1 ⎤
• In order to generate the residuals, use GMM with W = ⎢
⎣N
∑ i i ⎥⎦
Z 'Z

• Can think of the equation (2) as giving you a covar matrix under
heteroskedasticity and serial correlation of unknown form.
• When L=K and hence X’Z is a square, then W does not matter.

• Stop with the second step (not a lot of efficiency gains in continuing).

• Simulation studies show - very little efficiency gain in doing 2-step

GMM even in the presence of considerable heteroskedasticity.

8
Wiji Arulampalam, May 2006

• Additionally, since 2-step GMM depends on 1st step coeff est, the std.
error calculations tend to be too small…. Windmeijer provides a
correction (now implemented in some software – PcGive).

9
Wiji Arulampalam, May 2006

Test for over identifying restrictions (Sargan/Hansen)

Moment condition: E(zi’ui)=0 [has L eqns]
'
⎡ ⎤ ⎡ ⎤
Min: Q = ⎢ ∑ Z i u i ⎥ W ⎢ ∑ Z i u i ⎥
' '

⎣ i ⎦ ⎣ i ⎦
• When L=K, Q( β̂ )=0;
• When L>K, then Q( β̂ )>0 although Q( β̂ )→0 in probability.
• So use this to derive the test by comparing the value of the criterion
function Q with its expected value under the null that the restrictions are
valid.
• This is simple when the optimal weighting matrix is used:
N Q( β̂ ) dist asym as a χ2(L-K) under H0. (Only valid under homosk.)
10
Wiji Arulampalam, May 2006

• If you only suspect that L1 are ok but L2 are not (where L=L1 + L2) then
can use N(Q-Q1) is asymp χ2(L2) where Q1 is the minimand when the non-
suspect instruments L1 are used.

11