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

Lutz Ph. Breitling breitling at lb-netz.de
Wed Jun 9 13:13:50 CEST 2004

okay, so now the bootcov works fine.

aren't the lower bootstrap variances just what Karla is talking about when 
she writes on the website describing the eyestudy that i was trying to redo 
in the first place:
"Using a Poisson model without robust error variances will result in a 
confidence interval that is too wide."
(Karla Lindquist, Senior Statistician in the Division of Geriatrics at UCSF)

but one more question: so i cannot get SANDWICH estimates of the standard 
error for a [R] glm or glmD?

many thanks for all your efforts!

At 13:46 05.06.2004, Frank E Harrell Jr wrote:

>Sorry I didn't think of that sooner.  robcov needs the residuals method 
>for the fitter to allow a type="score" or type="hscore" (for Efron's 
>method) argument.  Until someone adds score residuals to residuals.glm 
>robcov will not work for you.  residuals.lrm and residuals.coxph are 
>examples where score residuals are computed.  You can get robust 
>variance-covariance estimates with the bootstrap using bootcov for glmD 
>fits.  Oddly in your example I am finding that the bootstrap variances are 
>lower than the information-matrix-based ones.
>Frank Harrell

Lutz Ph. Breitling
Unité des Recherches Médicale
Hôpital Albert Schweitzer
B.P. 118 Lambaréné (GABON) 
-------------- next part --------------


More information about the R-help mailing list