[R] poisson regression with robust error variance ('eyestudy')

[Thomas Lumley]

On Wed, 2 Jun 2004, Lutz Ph. Breitling wrote:

> Dear all,
>
> i am trying to redo the 'eyestudy' analysis presented on the site
> http://www.ats.ucla.edu/stat/stata/faq/relative_risk.htm
> with R (1.9.0), with special interest in the section on "relative risk
> estimation by poisson regression with robust error variance".
>
> so i guess rlm is the function to use. but what is its equivalent to the
> glm's argument "family" to indicate 'poisson'? or am i somehow totally
> wrong and this is not applicable here?
>

No, no.  You want glm() and then a function to compute the robust
covariance matrix (there's robcov() in the Hmisc package), or use gee()
from the "gee" package or geese() from "geepack" with independence working
correlation.

These are not outlier-resistant estimates of the regression coefficients,
they are model-agnostic estimates of the standard errors.

Stata is unusual in providing these covariance matrix estimates for just
about every regression estimator.  I think R should consider doing
something similar, but haven't got around to it.

	-thomas


> thx a lot-
> lutz
>
>
> =============================
> Lutz Ph. Breitling, CMd
> Unité des Recherches Médicale
> Hôpital Albert Schweitzer
> B.P. 118 Lambaréné (GABON)
>

Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle