Imbens/Wooldridge, Lecture Notes 15, Summer ’07
1
What’s New in Econometrics
NBER, Summer 2007
Lecture 15, Wednesday, Aug 1st, 4.305.30pm
Generalized Method of Moments and Empirical Likelihood
1. Introduction
Generalized Method of Moments (henceforth GMM) estimation has become an important
unifying framework for inference in econometrics in the last twenty years. It can be thought
of as nesting almost all the common estimation methods such as maximum likelihood, or
dinary least squares, instrumental variables and two–stage–least–squares and nowadays it is
an important part of all advanced econometrics text books (Gallant, 1987; Davidson and
McKinnon, 1993; Hamilton, 1994; Hayashi, 2000; Mittelhammer, Judge, and Miller, 2000;
Ruud, 2000; Wooldridge, 2002). Its formalization by Hansen (1982) centers on the presence
of known functions, labelled “moment functions”, of observable random variables and un
known parameters that have expectation zero when evaluated at the true parameter values.
The method generalizes the “standard” method of moments where expectations of known
functions of observable random variables are equal to known functions of the unknown pa
rameters. The “standard” method of moments can thus be thought of as a special case of
the general method with the unknown parameters and observed random variables entering
additively separable. The GMM approach links nicely to economic theory where orthogonal
ity conditions that can serve as such moment functions often arise from optimizing behavior
of agents.
For example, if agents make rational predictions with squared error loss, their
prediction errors should be orthogonal to elements of the information set.
In the GMM
framework the unknown parameters are estimated by setting the sample averages of these
moment functions, the “estimating equations,” as close to zero as possible.
The framework is sufficiently general to deal with the case where the number of moment
functions is equal to the number of unknown parameters, the so–called “just–identified case”,
as well as the case where the number of moments exceeding the number of parameters to be
estimated, the “over–identified case.” The latter has special importance in economics where
This preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Imbens/Wooldridge, Lecture Notes 15, Summer ’07
2
the moment functions often come from the orthogonality of potentially many elements of
the information set and prediction errors. In the justidentified case it is typically possible
to estimate the parameter by setting the sample average of the moments exactly equal to
zero.
In the overidentified case this is not feasible.
The solution proposed by Hansen
(1982) for this case, following similar approaches in linear models such as two– and three–
stage–least–squares, is to set a linear combination of the sample average of the moment
functions equal to zero, with the dimension of the linear combination equal to the number
of unknown parameters.
The optimal linear combination of the moments depends on the
This is the end of the preview.
Sign up
to
access the rest of the document.
 Fall '08
 Staff
 Econometrics, Maximum likelihood, Estimation theory, Generalized method of moments, empirical likelihood, Summer ’07

Click to edit the document details