# Whether to use robust linear regression or bootstrapping when there is heteroscedasticity?

I have a dataset where I need to do linear regression. Unfortunately there is a problem with heteroscedasticity. I´ve rerun the analysis using robust regression with the HC3 estimator for the variance and also done bootstrapping with the bootcov function in Hmisc for R. The results are quite close. What is generally recommended?

-
 What R package did you use for HC3 estimation? sandwich, contrast? – chl♦ Sep 15 '10 at 19:46 Another question while we are in: What is the design you are considering, I mean is there any clustering or multiple predictors, or is it a simple linear regression? This may help the reader to better understand the context of your study. – chl♦ Sep 15 '10 at 20:47 Have you tried re-expressing the dependent variable to stabilize variance? – whuber♦ Sep 15 '10 at 21:34 I´m using the sandwich package for the HC3 estimator. I´m using just simple linear regression. Intuitively I feel more comfortable with the bootstrap version and I guess I´ll stick with that. //thx for the input – Misha Sep 16 '10 at 5:19

In economics, the Eicker-White or "robust" standard errors are typically reported. Bootstrapping (unfortunately, I'd say) is less common. I'd say that the robust estimates are the standard version.

-

You could use generalized least squares, such as the gls() function from the nlme package, which allows you to specify a variance function using the weight argument.

-