Sandwich covariance of a GLM — robust

Heteroskedasticity- or cluster-robust covariance of the coefficients of a glm_fit() model, valid when the variance function or dispersion is wrong as long as the mean is right. The dispersion cancels. Results match statsmodels' cov_type = "HC0" and "cluster" (validation/scripts/statsmodels_glm_robust.py). For a non-canonical link (a log-link gamma, say) the bread is the observed information, as in statsmodels; sandwich::vcovHC() uses the expected information, so the two differ slightly there and agree for canonical links.

Usage

robust_vcov(object, type = c("HC0", "HC1"), cluster = NULL)

Arguments

object: A glm_model.
type: "HC0" (White's estimator) or "HC1" (scaled by n / (n - p)). Ignored when cluster is given.
cluster: Optional cluster labels, one per row the model was fitted on (a policy or an event, say). The scores are summed within each cluster and the result scaled by G / (G - 1) * (n - 1) / (n - p) for G clusters.

Value

A named covariance matrix; sqrt(diag(.)) gives the robust standard errors.

Examples

d <- data.frame(y = c(1, 2, 6, 1, 4, 2), x = c(0, 0, 0, 1, 1, 1),
                policy = c(1, 1, 2, 2, 3, 3))
m <- glm_fit(y ~ x, d, family = "poisson")
robust_vcov(m)
#>             (Intercept)          x
#> (Intercept)   0.1728395 -0.1728395
#> x            -0.1728395  0.2680776
sqrt(diag(robust_vcov(m, cluster = d$policy)))
#> (Intercept)           x 
#>   0.6454972   0.8892785