R: Robust standard errors for regression models

robust {sjstats}

R Documentation

Robust standard errors for regression models

Description

robust() computes robust standard error for regression models. This method wraps the coeftest-function with robust covariance matrix estimators based on the vcovHC-function, and returns the result as tidy data frame.

svy() is intended to compute standard errors for survey designs (complex samples) fitted with regular lm or glm functions, as alternative to the survey-package. It simulates sampling weights by adjusting the residual degrees of freedom based on the precision weights used to fit x, and then calls robust() with the adjusted model.

Usage

robust(x, vcov = c("HC3", "const", "HC", "HC0", "HC1", "HC2", "HC4", "HC4m",
  "HC5"), conf.int = FALSE, exponentiate = FALSE)

svy(x, vcov = c("HC1", "const", "HC", "HC0", "HC2", "HC3", "HC4", "HC4m",
  "HC5"), conf.int = FALSE, exponentiate = FALSE)

Arguments

`x`	A fitted model of any class that is supported by the `coeftest()`-function. For `svy()`, `x` must be `lm` object, fitted with weights.
`vcov`	Character vector, specifying the estimation type for the heteroskedasticity-consistent covariance matrix estimation (see `vcovHC` for details).
`conf.int`	Logical, `TRUE` if confidence intervals based on robust standard errors should be included.
`exponentiate`	Logical, whether to exponentiate the coefficient estimates and confidence intervals (typical for logistic regression).

Value

A summary of the model, including estimates, robust standard error, p-value and - optionally - the confidence intervals.

Note

svy() simply calls robust(), but first adjusts the residual degrees of freedom based on the model weights. Hence, for svy(), x should be fitted with weights. This simulates sampling weights like in survey designs, though lm and glm implement precision weights. The results from svy() are usually more accurate than simple weighted standard errors for complex samples. However, results from the survey package are still more exactly, especially regarding the estimates.

vcov for svy() defaults to "HC1", because standard errors with this estimation type come closest to the standard errors from the survey-package.

Currently, svy() only works for objects of class lm.

Examples

data(efc)
fit <- lm(barthtot ~ c160age + c12hour + c161sex + c172code, data = efc)
summary(fit)
robust(fit)

confint(fit)
robust(fit, conf.int = TRUE)
robust(fit, vcov = "HC1", conf.int = TRUE) # "HC1" should be Stata default

library(sjmisc)
# dichtomozize service usage by "service usage yes/no"
efc$services <- sjmisc::dicho(efc$tot_sc_e, dich.by = 0)
fit <- glm(services ~ neg_c_7 + c161sex + e42dep,
           data = efc, family = binomial(link = "logit"))

robust(fit)
robust(fit, conf.int = TRUE, exponentiate = TRUE)

[Package sjstats version 0.14.2-3 Index]