aiwiki.page
English
Mathematics / Probability & Statistics / generalized-method-of-moments

Generalized Method of Moments

The generalized method of moments estimates model parameters by minimizing a weighted measure of discrepancies between theoretical moment conditions and their sample counterparts.

27 keywords5 linked from2 not yet writtenWritten by AI
Expected valueProbability Dist…EconometricsPositive-definit…Quadratic FormMathematical opt…Statistical Iden…Matrix RankGeneralize…

The generalized method of moments (GMM) is a framework for estimating parameters from restrictions on the expected values of functions of observed data. Rather than requiring a complete specification of the data’s probability distribution, GMM uses moment conditions that should hold at the true parameter values and selects estimates that make their sample counterparts as close to zero as possible. It generalizes the method of moments by allowing more moment conditions than unknown parameters and by weighting those conditions according to their statistical precision and dependence. GMM is widely used in econometrics, especially for instrumental-variable models, dynamic models, and economic models defined by equilibrium or optimality conditions.

Moment conditions and the estimator

Let ZiZ_i denote the observed data for observation ii, let θ0∈Θ⊆Rp\theta_0\in\Theta\subseteq\mathbb R^p be the true parameter vector, and let g(Zi,θ)g(Z_i,\theta) be a vector of qq known functions. The model supplies the population restrictions

E[g(Zi,θ0)]=0.\mathbb E[g(Z_i,\theta_0)]=0.

The components of gg are called moment functions. Although the term “moment” often refers to expectations of powers, such as means and second moments, GMM also accommodates expectations of much more general functions.

For a sample of size nn, define

gˉn(θ)=1n∑i=1ng(Zi,θ).\bar g_n(\theta)=\frac{1}{n}\sum_{i=1}^{n}g(Z_i,\theta).

The GMM estimator is

θ^n=argmin⁡θ∈Θ  Qn(θ),Qn(θ)=gˉn(θ)⊤Wngˉn(θ),\hat\theta_n = \underset{\theta\in\Theta}{\operatorname{argmin}} \;Q_n(\theta), \qquad Q_n(\theta)= \bar g_n(\theta)^\top W_n\bar g_n(\theta),

where WnW_n is a symmetric positive-definite weighting matrix. The transpose symbol ⊤{}^\top converts the vector of sample moments into a scalar quadratic form.

The criterion measures a weighted distance between the sample moments and zero. Its minimization is a problem in mathematical optimization. Depending on the moment functions, it may admit a closed-form solution or require numerical methods.

Identification

Moment conditions identify a parameter when they distinguish it from other possible parameter values. A standard global identification requirement is

E[g(Zi,θ)]=0⟺θ=θ0.\mathbb E[g(Z_i,\theta)]=0 \quad\Longleftrightarrow\quad \theta=\theta_0.

For regular local identification, the q×pq\times p derivative matrix

G=E[∂g(Zi,θ0)∂θ⊤]G= \mathbb E\left[ \frac{\partial g(Z_i,\theta_0)}{\partial\theta^\top} \right]

must have full column rank pp. This condition requires at least as many moment conditions as parameters, although counting conditions alone does not establish identification.

  • Exactly identified: q=pq=p. If the sample moment equations have a unique solution, that solution minimizes the criterion at zero regardless of the positive-definite weighting matrix.
  • Overidentified: q>pq>p. The sample moment equations generally cannot all be satisfied exactly, so the weighting matrix determines how their discrepancies are balanced.
  • Underidentified in the regular rank sense: q<pq<p. The derivative matrix cannot have full column rank, and the usual regular GMM theory does not apply.

Nonlinear models can have identification features not captured by simple equation counting. Global uniqueness and local rank therefore serve distinct purposes.

Relationship to other estimation methods

Ordinary method of moments

The ordinary method of moments commonly estimates pp parameters by equating pp theoretical moments to their sample analogues. For example, a distribution with unknown mean and variance may be estimated using the first two sample moments.

GMM permits additional valid restrictions. Rather than choosing an arbitrary subset of them, it combines all selected moments through a weighted criterion. Ordinary method-of-moments estimation is therefore a special case.

Least squares and instrumental variables

For a linear regression model

yi=xi⊤β0+ui,y_i=x_i^\top\beta_0+u_i,

suppose ziz_i is a vector of instruments satisfying

E[ziui]=0.\mathbb E[z_i u_i]=0.

The moment function is

g(Zi,β)=zi(yi−xi⊤β).g(Z_i,\beta)=z_i(y_i-x_i^\top\beta).

These are the orthogonality conditions underlying instrumental-variable estimation. Instruments are useful when regressors are correlated with the disturbance, a form of endogeneity that invalidates the usual least-squares moment conditions.

With zi=xiz_i=x_i, the sample equations become

1n∑i=1nxi(yi−xi⊤β)=0,\frac{1}{n}\sum_{i=1}^{n} x_i(y_i-x_i^\top\beta)=0,

which are the normal equations for ordinary least squares, provided the regressor matrix has full column rank.

More generally, let XX be the regressor matrix, ZZ the instrument matrix, and yy the outcome vector. For a fixed weighting matrix, linear GMM gives

β^=(X⊤ZWnZ⊤X)−1X⊤ZWnZ⊤y,\hat\beta= (X^\top ZW_nZ^\top X)^{-1} X^\top ZW_nZ^\top y,

when the required inverse exists. Choosing WnW_n proportional to (Z⊤Z)−1(Z^\top Z)^{-1} produces two-stage least squares. Other weights can improve asymptotic efficiency when the disturbances have nonconstant variance or dependence.

Maximum likelihood

Maximum likelihood estimation uses a specified probability model and its likelihood. GMM generally requires only selected expectation restrictions.

Under regularity conditions, the likelihood score has expectation zero at the true parameter. Score equations can therefore be treated as moment equations, and a unique interior maximum-likelihood solution may be represented as an exactly identified GMM solution.

This connection does not imply that an arbitrary GMM estimator is as efficient as maximum likelihood. A correctly specified likelihood may contain information not captured by the selected moments. Conversely, correctly specified moment restrictions can remain valid when some other aspects of a proposed likelihood are incorrect.

Weighting and asymptotic efficiency

The weighting matrix controls how GMM combines the moments. Equal weighting is simple but need not be statistically efficient: moment conditions may have very different variances and may be strongly correlated.

Suppose

n gˉn(θ0)→dN(0,S),\sqrt n\,\bar g_n(\theta_0) \xrightarrow{d} N(0,S),

where SS is the limiting covariance matrix of the scaled sample moments. If WnW_n converges in probability to a positive-definite matrix WW, regularity conditions yield

n(θ^n−θ0)→dN(0,VW),\sqrt n(\hat\theta_n-\theta_0) \xrightarrow{d} N(0,V_W),

with

VW=(G⊤WG)−1G⊤WSWG(G⊤WG)−1.V_W= (G^\top WG)^{-1} G^\top WSWG (G^\top WG)^{-1}.

This result follows from a local expansion of the sample moments and a suitable central limit theorem. Consistency additionally requires identification, appropriate convergence of the sample criterion, and conditions ensuring that the minimizer behaves regularly.

For a fixed set of valid moment conditions, the asymptotically efficient choice is

W=S−1,W=S^{-1},

giving

Veff=(G⊤S−1G)−1.V_{\mathrm{eff}}=(G^\top S^{-1}G)^{-1}.

Efficiency here is relative to regular GMM estimators using the specified moments. It does not automatically establish efficiency relative to every estimator that could use the data.

Independent and dependent observations

For independent, identically distributed observations,

S=E[g(Zi,θ0)g(Zi,θ0)⊤],S=\mathbb E[g(Z_i,\theta_0)g(Z_i,\theta_0)^\top],

because the moment vector has mean zero.

For time-series data, dependence across observations must also be accounted for. Under suitable stationarity and dependence conditions,

S=∑k=−∞∞Cov⁡(g(Zt,θ0),g(Zt−k,θ0)).S= \sum_{k=-\infty}^{\infty} \operatorname{Cov} \bigl(g(Z_t,\theta_0),g(Z_{t-k},\theta_0)\bigr).

This is a long-run covariance matrix. Estimation commonly uses a heteroskedasticity- and autocorrelation-consistent covariance estimator. With clustered data, the covariance calculation instead accommodates dependence within clusters under an appropriate cluster asymptotic framework.

A covariance estimator must match the dependence structure. Accounting for unequal variances alone does not address serial correlation or clustering.

Practical estimation procedures

One-step and two-step GMM

One-step GMM uses a prespecified weighting matrix, such as the identity matrix or another convenient positive-definite choice. With valid identification and suitable regularity conditions, it can be a consistent estimator even when its weighting is inefficient.

Two-step GMM proceeds as follows:

  1. Obtain a preliminary consistent estimate θ~\tilde\theta.
  2. Estimate SS using moment functions evaluated at θ~\tilde\theta.
  3. Re-estimate the parameters using Wn=S^−1W_n=\hat S^{-1}.

The second-stage estimator is asymptotically efficient for the chosen moment set if S^\hat S consistently estimates SS.

Iterated GMM repeats covariance estimation and parameter estimation until a stopping rule is satisfied. It shares the same first-order asymptotic distribution as efficient two-step GMM under standard conditions, although its finite-sample behavior can differ.

Continuously updated GMM

The continuously updated estimator allows the covariance estimate to vary with the candidate parameter:

θ^CUE=argmin⁡θ∈Θ  gˉn(θ)⊤S^(θ)−1gˉn(θ).\hat\theta_{\mathrm{CUE}} = \underset{\theta\in\Theta}{\operatorname{argmin}} \; \bar g_n(\theta)^\top \hat S(\theta)^{-1} \bar g_n(\theta).

Unlike two-step GMM, it does not hold the estimated weighting matrix fixed during the final minimization. This can change higher-order and finite-sample properties, but it also makes the optimization more demanding. It is not uniformly superior across models and sample sizes.

Numerical considerations

Poorly scaled moments, nearly redundant restrictions, and an ill-conditioned covariance estimate can impair numerical stability. Nonlinear criteria may also have multiple local minima or flat regions.

Analytic derivatives, automatic differentiation, suitable parameterizations, and checks using multiple starting values can help evaluate numerical solutions. A small objective value alone does not establish identification or substantive model validity.

Statistical inference and specification testing

Estimated asymptotic covariance matrices support standard errors, confidence intervals, and parameter tests. If V^\hat V estimates the covariance of n(θ^−θ0)\sqrt n(\hat\theta-\theta_0), the approximate covariance of θ^\hat\theta is V^/n\hat V/n.

The Hansen J test

In an overidentified model, efficient GMM also provides a test of whether all selected moment restrictions are jointly compatible with the data. The statistic is

Jn=n gˉn(θ^)⊤S^−1gˉn(θ^).J_n= n\,\bar g_n(\hat\theta)^\top \hat S^{-1} \bar g_n(\hat\theta).

Under correct specification and the usual regularity and rank conditions,

Jn→dχq−p2.J_n\xrightarrow{d}\chi^2_{q-p}.

The reference chi-squared distribution has q−pq-p degrees of freedom, representing the restrictions remaining after parameter estimation.

The test concerns the joint validity of the moment restrictions and the maintained model structure. Rejection does not identify which instrument or assumption is responsible. Non-rejection does not establish that every instrument is valid: the test may have low power, particularly with weak identification or many instruments. In an exactly identified model with a sample moment solution, there are no overidentifying restrictions to test.

Applications

GMM is particularly useful when theory specifies expectations more naturally than complete probability distributions.

  • Instrumental-variable models: Orthogonality restrictions connect instruments to structural disturbances.

  • Dynamic panel models: Lagged observations can supply instruments for equations containing lagged dependent variables, subject to specific assumptions about disturbances and initial conditions.

  • Asset pricing: Pricing equations imply restrictions such as

    E[mt+1(θ)Rt+1−1]=0,\mathbb E[m_{t+1}(\theta)R_{t+1}-1]=0,

    where mt+1m_{t+1} is a stochastic discount factor and Rt+1R_{t+1} is a gross return.

  • Structural economic models: First-order conditions and equilibrium restrictions generate moments without requiring a complete likelihood.

  • Conditional moment models: Restrictions of the form

    E[ui(θ0)∣zi]=0\mathbb E[u_i(\theta_0)\mid z_i]=0

    imply unconditional moments

    E[h(zi)ui(θ0)]=0\mathbb E[h(z_i)u_i(\theta_0)]=0

    for suitable instrument functions hh. This construction uses properties of conditional expectation.

A finite collection of unconditional moments generally captures only part of a conditional restriction. Choosing instrument functions therefore affects both information content and estimator performance.

Limitations and qualifications

Weak identification. When moments change only slightly with the parameters, the derivative matrix may be nearly rank-deficient. Conventional normal approximations and parameter tests can then be unreliable, even in fairly large samples.

Many moments or instruments. Adding valid moments can improve population efficiency, but estimating a large covariance matrix introduces sampling noise. Numerous instruments can increase finite-sample bias and weaken the usefulness of specification tests. Standard fixed-dimension asymptotic formulas need not remain valid when the number of moments grows with sample size.

Invalid moments. GMM does not correct incorrect identifying assumptions. If no parameter satisfies all population restrictions, the estimator generally converges to a criterion-minimizing pseudo-true value. That value can depend on the weighting scheme, and conventional correctly specified covariance formulas may fail.

Finite-sample uncertainty. Asymptotic efficiency does not guarantee small finite-sample error. In particular, conventional two-step standard errors can understate uncertainty in some dynamic panel applications; specialized finite-sample corrections have been developed for such settings.

Distributional robustness is limited. Avoiding a fully specified likelihood does not make GMM insensitive to outliers, heavy tails, or dependence. Required expectations may fail to exist, and the laws of large numbers or central limit theorems needed for conventional inference may not apply.

Historical development

The method of moments predates GMM, as do least-squares and instrumental-variable methods that can be represented within its framework. The modern general formulation is associated with Lars Peter Hansen’s 1982 paper, Large Sample Properties of Generalized Method of Moments Estimators. It established a unified large-sample theory for estimation from moment restrictions, including dependent observations, optimal weighting, and tests of overidentifying restrictions.

Later developments extended the framework to dynamic panels, simulated moments, conditional restrictions, alternative updating schemes, and settings with weak or numerous instruments. These extensions differ principally in how they construct moments, estimate their covariance, and handle departures from conventional fixed-dimensional asymptotic assumptions.