[Rd] Standardized Pearson residuals
presnell at stat.ufl.edu
Tue Mar 15 04:40:29 CET 2011
Thanks Peter. I have just a couple of minor comments, and another
possible feature request, although it's one that I don't think will be
peter dalgaard <pdalgd at gmail.com> writes:
> On Mar 14, 2011, at 22:25 , Brett Presnell wrote:
>> Is there any reason that rstandard.glm doesn't have a "pearson" option?
>> And if not, can it be added?
> Probably... I have been wondering about that too. I'm even puzzled why
> it isn't the default. Deviance residuals don't have quite the
> properties that one might expect, e.g. in this situation, the absolute
> residuals sum pairwise to zero, so you'd expect that the standardized
> residuals be identical in absolute value
>> y <- 1:4
>> r <- c(0,0,1,1)
>> c <- c(0,1,0,1)
> 1 2 3 4
> -0.2901432 0.2767287 0.2784603 -0.2839995
> in comparison,
>> i <- influence(glm(y~r+c,poisson))
> 1 2 3 4
> -0.2817181 0.2817181 0.2817181 -0.2817181
> The only thing is that I'm always wary of tampering with this stuff,
> for fear of finding out the hard way why thing are the way they
I'm sure that's wise, but it would be nice to get it in as an option,
even if it's not the default
>> Background: I'm currently teaching an undergrad/grad-service course from
>> Agresti's "Introduction to Categorical Data Analysis (2nd edn)" and
>> deviance residuals are not used in the text. For now I'll just provide
>> the students with a simple function to use, but I prefer to use R's
>> native capabilities whenever possible.
> Incidentally, chisq.test will have a stdres component in 2.13.0 for
> much the same reason.
Thank you. That's one more thing I won't have to provide code for
anymore. Coincidentally, Agresti mentioned this to me a week or two ago
as something that he felt was missing, so that's at least two people who
will be happy to see this added.
It would also be nice for teaching purposes if glm or summary.glm had a
"pearsonchisq" component and a corresponding extractor function, but I
can imagine that there might be arguments against it that haven't
occured to me. Plus, I doubt that anyone wants to touch glm unless it's
to repair a bug. If I'm wrong about all that though, ...
BTW, as I go along I'm trying to collect a lot of the datasets from the
examples and exercises in the text into an R package ("icda"). It's far
from complete and what is there needed tidying up, but I hope to
eventually to round it into shape and put it on CRAN, assuming that
Agresti approves and that there are no copyright issues.
>> I think something along the following lines should do it:
>> rstandard.glm <-
>> infl=influence(model, do.coef=FALSE),
>> type=c("deviance", "pearson"), ...)
>> type <- match.arg(type)
>> res <- switch(type, pearson = infl$pear.res, infl$dev.res)
>> res <- res/sqrt(1-infl$hat)
>> res[is.infinite(res)] <- NaN
More information about the R-devel