Returns a heteroskedasticity-robust (optionally dyad-clustered) sandwich
covariance matrix for the intercept and dyadic-covariate coefficients
of an ame_als fit. This is a fast analytic alternative to
the bootstrap for those coefficients.
# S3 method for ame_als
vcov(object, cluster = c("dyad", "none"), ...)A covariance matrix with matching row/column names. When the fit
carries a $bootstrap, this is the bootstrap covariance over all
estimated coefficients, matching coef(). Otherwise it is the
conditional sandwich, covering c(intercept, dyadic coefficients)
only.
an ame_als fit.
"dyad" (default) for a dyad-clustered robust meat, or
"none" for an HC0 (heteroskedasticity-only) meat. Ignored when the
fit carries a bootstrap, since the bootstrap covariance is returned.
ignored.
The estimate is the conditional sandwich \(B^{-} M B^{-}\) with bread
\(B = D'WD\) (\(D\) the observed intercept + dyadic-covariate design,
\(W\) the fit's observation weights) and meat \(M\) the
heteroskedasticity-robust (cluster = "none", an HC0 meat) or
dyad-clustered (cluster = "dyad", the default) outer product of the
weighted score contributions \(w_\ell e_\ell d_\ell\). For a normal or
transform fit the weights are unit, so this reduces to the ordinary
\(D'D\) sandwich; for an IRLS fit it uses the final IRLS weights, matching
the estimating equation the fit actually solved. Dyad clustering pools the
score across \((i,j)\), \((j,i)\) and time, so it reflects dyadic
dependence (reciprocity, repeated observation).
It is conditional: the additive effects a, b and the
multiplicative term are held fixed, so it omits their estimation uncertainty
and is anti-conservative. Node-covariate, additive and multiplicative
standard errors are not returned -- use ame_als_bootstrap
for those and for fully-propagated inference.
ame_als_bootstrap for bootstrap uncertainty
covering all parameters.