Clustered covariance matrix of the coefficients of an lm or glm fit. Equivalent to R's
sandwich::vcovCL().
The meat sums the estimating functions within each cluster (Zeileis, Köll and Graham 2020), times G / (G - 1) with
cadjust; HC1 multiplies by (n - 1) / (n - k). Several clustering dimensions are combined by inclusion and exclusion
over all their intersections, each term with its own G (Cameron, Gelbach and Miller 2011); multi0 uses clusters of
one observation for the full intersection. HC2 and HC3 adjust each cluster's residuals by (I - H_gg)^-1/2 or
(I - H_gg)^-1 in the symmetric form of Bell and McCaffrey (2002) (for HC3 the leave-one-cluster-out jackknife),
which is sandwich's for an unweighted lm; sandwich transposes the adjustment for a weighted lm or a glm. A singular
block (cluster fixed effects) is an error. fix sets negative eigenvalues to 0.
The default type is HC1 for an lm and HC0 for a glm, as sandwich. Without cluster each observation is its own
cluster. G is the number of distinct labels among the rows used (labels are compared as text, numbers with 15
significant digits as R's as.character: 1 and 1 + 2^-52, or 0.1 + 0.2 and 0.3, are one label); intersections of
dimensions are tuples of labels; at least two clusters are needed. Rows of prior weight 0 are left out. HC1 needs
more observations than coefficients. Data far from 1 as vcovHC (the same record of the units, lost by a copy
of the matrix); with fix, a negative eigenvalue is set to 0 in the units of the largest coefficient's variance, so
that coefficients whose scale is about 2^537 or more below it then get variance 0. See docs/differences-from-r.md.
Clustered covariance matrix of the coefficients of an
lmorglmfit. Equivalent to R'ssandwich::vcovCL().The meat sums the estimating functions within each cluster (Zeileis, Köll and Graham 2020), times G / (G - 1) with
cadjust; HC1 multiplies by (n - 1) / (n - k). Several clustering dimensions are combined by inclusion and exclusion over all their intersections, each term with its own G (Cameron, Gelbach and Miller 2011);multi0uses clusters of one observation for the full intersection. HC2 and HC3 adjust each cluster's residuals by (I - H_gg)^-1/2 or (I - H_gg)^-1 in the symmetric form of Bell and McCaffrey (2002) (for HC3 the leave-one-cluster-out jackknife), which is sandwich's for an unweighted lm; sandwich transposes the adjustment for a weighted lm or a glm. A singular block (cluster fixed effects) is an error.fixsets negative eigenvalues to 0.The default type is HC1 for an lm and HC0 for a glm, as sandwich. Without
clustereach observation is its own cluster. G is the number of distinct labels among the rows used (labels are compared as text, numbers with 15 significant digits as R'sas.character: 1 and 1 + 2^-52, or 0.1 + 0.2 and 0.3, are one label); intersections of dimensions are tuples of labels; at least two clusters are needed. Rows of prior weight 0 are left out. HC1 needs more observations than coefficients. Data far from 1 as vcovHC (the same record of the units, lost by a copy of the matrix); withfix, a negative eigenvalue is set to 0 in the units of the largest coefficient's variance, so that coefficients whose scale is about 2^537 or more below it then get variance 0. See docs/differences-from-r.md.