= Sandwich covariance matrix
{title2=$A^{-1}BA^{-T}$}
Suppose an <estimating equation> $n^{-1}\sum_i\psi(X_i,\widehat\theta)=0$ has the expansion
$$
\sqrt n(\widehat\theta-\theta_0)=-A^{-1}\frac1{\sqrt n}\sum_i\psi(X_i,\theta_0)+o_{\mathbb P}(1),
$$
where $A=\mathbb E[\partial\psi(X,\theta_0)/\partial\theta^T]$ is nonsingular and the centered estimating function has <covariance matrix> $B$. A <central limit theorem> gives asymptotic covariance $A^{-1}BA^{-T}$ for the rescaled error, or approximately $n^{-1}A^{-1}BA^{-T}$ for $\widehat\theta$.
The transpose is essential when $A$ is not symmetric; the simplified form $A^{-1}BA^{-1}$ applies only in the symmetric case. Plugging in consistent estimates gives a sandwich covariance estimate. Regression <heteroskedasticity-consistent standard errors> and cluster-level <generalized estimating equations> use instances of this construction.
Back to article page