[Rd] standardized residuals (rstandard & plot.lm) (PR#8468)

David Firth d.firth at warwick.ac.uk
Wed Jan 11 12:56:38 CET 2006


On 10 Jan, 2006, at 14:58, maechler at stat.math.ethz.ch wrote:

>>>>>> "Heather" == Heather Turner <Heather.Turner at warwick.ac.uk>
>>>>>>     on Tue, 10 Jan 2006 14:30:23 +0100 (CET) writes:
>
>     Heather> This bug is not quite fixed - the example from my
>     Heather> original report now = works using R-2.2.1, but
>
>     Heather> plot(Uniform, 6)
>
>     Heather> does not. The bug is due
>
>     .........
>      g <- hatval/(1 - hatval) # Potential division by zero here
>
>      plot(g, cook, xlim = c(0, max(g)), ylim = c(0, ymx),
>      ..........
>
>     Heather> All other values of 'which' seem to work
>     Heather> fine. Sorry not to have checked this in the beta
>     Heather> version,
>
> (indeed; that would have been useful)
>
>
> Hmm, it's not clear what *should* be drawn in such a
> case. Leaving away all the observations with  h_ii = 1
> seems a particularly bad idea, since these are the ones that
> you'd definitely should remark.
> OTOH, for h_ii = 1, the cook distance is 'NaN'
> (or should that be changed; to  "very large" instead ???)
> and plot number 6 doesn't seem to make any sense to me
>
> When 'which = 6' was proposed
> [ on R-devel as well, last April,
>   http://tolstoy.newcastle.edu.au/R/devel/05/04/0595.html
>   ah, I see, by David Firth, from your place, so, Heather, can you
>   make sure he sees this e-mail ?
> ]
> I actually had wondered a bit about it's general usefulness,..

Yes, I remember that there was some discussion of this last year, and 
my recollection is that it was mostly luke-warm at best in regard to 
including this plot.

The "h_ii = 1" problem can of course be taken care of by leaving out 
such points if they can be reliably detected, but I share Martin's 
unease about this.  We should also worry about for example h_ii = (1 - 
epsilon), with epsilon small, as plotting such a point would 
effectively make the rest of the graph useless.

Maybe it would be safest to remove the which=6 option?

David



More information about the R-devel mailing list