Recompute the coefficient covariance matrix for a panglm fit

With type = NULL (the default) the covariance the object was fit with is returned (set by the vcov argument of panglm()). type = "classical" returns the model-based covariance, also for a fit made with a robust vcov. type = "HC1" (heteroskedasticity-robust / White) and type = "cluster" (cluster-robust sandwich, clustered by the panel individual unless cluster is given) are available for pooled and fixed-effects models. For random-effects models only "cluster" is available: the units of the integrated likelihood are individuals, so the sandwich is built from individual scores (clusters must contain whole individuals) with the factor G / (G - 1). It guards against a misspecified random-effects distribution and is the right choice with sampling weights. For the unweighted Gaussian (Swamy-Arora) fit it is the cluster sandwich of the quasi-demeaned regression, as in Stata's xtreg, re vce(cluster id). For exact conditional binomial models, the independent likelihood units are panel strata rather than observations. Their HC1 covariance therefore uses stratum-level scores. A custom cluster may combine complete strata, but it may not split observations from one conditioned stratum. For the two-way conditional logit (effect = "twoways") the time dummies are parameters of the conditional likelihood: the sandwich is built from the full (slopes + time dummies) stratum scores and bread, its small-sample factors count all of them, and the slope block is returned.

Usage

# S3 method for class 'panglm'
vcov(object, type = NULL, cluster = NULL, ...)

Arguments

object

a "panglm" fit

type

NULL (the covariance stored on the fit) or one of "classical", "HC1", "cluster"

cluster

optional vector giving the clustering variable (one value per original observation, i.e. nrow(data) long); defaults to the panel individual identifier used to fit the model

...

unused

Value

a numeric covariance matrix for the estimated coefficients, with row and column names matching names(coef(object)). For type = "classical" this is the model-based covariance; for "HC1" and "cluster" it is a sandwich (robust or cluster-robust) covariance estimate of the same dimension.

See also

Other panglm methods: coefplot(), confint.panglm(), effectcurve(), fitted.panglm(), fixef(), glance.panglm(), logLik.panglm(), model.matrix.panglm(), partialeffects.panordered(), plot.panglm(), plot.panglm_ape(), plot.panglm_effectcurve(), predict.panglm(), predict.panordered(), residuals.panglm(), summary.panglm(), tidy.panglm()

Examples

data(copd)
fit <- panglm(fev1 ~ crp, data = copd, index = c("id", "visit"),
              model = "within", family = "gaussian")
vcov(fit)
#>             crp
#> crp 3.56989e-05
vcov(fit, type = "cluster")
#>              crp
#> crp 3.686711e-05