"NicLK" == Nicholas Lewin-Koh <nikko at hailmail.net>
on Sat, 14 Feb 2009 08:34:45 -0800 writes:
NicLK> Hi, For extended glms such as gams, gnm or other
NicLK> distributions such as negative binomial, would there
NicLK> need to be a separate simulate method?
Not necessarily, as I said, the "glm"s are now also dealt with
in simulate.lm() and Heather more or less confirmed that this
gives correct results for "gnm" objects.
For gam(), I'd strongly expect the same to apply, but there
maybe sophisticated gam() models where the result is currently
not correct. That's, BTW, also true for
simulate(lm(...., weights), ...)
NicLK> Or, could the current framework, rather than
NicLK> stopping with an error look for the appropriate model
NicLK> matrix, coefficients, distribution function and
NicLK> family object to simulate from?
What do you mean?
A situation where there's no supported 'family'
or a situation where predict(<obj>) does not work as it's
supposed in the current framework,
or ????
If there are such cases, we'd have to consider them together
with the corresponding package author. It may often make sense
fthen that the author changes his methods {predict(), ..} such
that the (now) extended simulate.lm() will work automatically;
Alternatively, the author can provide simulate.<myclass>().
But I'm not sure I'm answering the question you've asked..
Martin
NicLK> Nicholas
>> Message: 9 Date: Fri, 13 Feb 2009 21:27:57 +0100 From:
>> Martin Maechler <maechler at stat.math.ethz.ch> Subject: Re:
>> [Rd] proposed simulate.glm method To: Heather Turner
>> <Heather.Turner at warwick.ac.uk> Cc: r-devel at r-project.org,
>> Martin Maechler <maechler at stat.math.ethz.ch> Message-ID:
>> <18837.55245.15158.29378 at cmath-5.math.ethz.ch>
>> Content-Type: text/plain; charset=us-ascii
>>
>> Thank you, Heather and Ben,
>>
>> >>>>> "HT" == Heather Turner
>> <Heather.Turner at warwick.ac.uk> >>>>> on Fri, 13 Feb 2009
>> 15:52:37 +0000 writes:
>>
HT> Yes, thanks to Ben for getting the ball rolling. His
HT> code was more streamlined than mine, pointing to further
HT> simplifications which I've included in the extended
HT> version below.
HT> The code for the additional families uses functions from
HT> MASS and SuppDists - I wasn't sure about the best way to
HT> do this, so have just used :: for now.
HT> It appears to be working happily for both glm and gnm
HT> objects (no gnm-specific code used).
>>
>> [....]
>>
>> I have now followed Brian Ripley's suggetion to just
>> extend simulate.lm() to also deal with "glm" objects, but
>> using Heather's suggestions for the different families;
>> I've just commited src/library/stats/R/lm.R with the new
>> code. (get it from svn.r-project.org/R/trunk/ or this
>> night's R-devel tarball).
>>
>> One difference to your propsal: Instead of just
>> object$fitted , the code is using fitted(object)
>> ... something which should properly to the na.action
>> used.
>>
>> For the (MASS and) SuppDists package requirement, I'm
>> using the following
>>
>> if(is.null(tryCatch(loadNamespace("SuppDists"), error =
>> function(e) NULL))) stop("Need CRAN package 'SuppDists'
>> for 'inverse.gaussian' family")
>>
>>
>> I've not yet updated the help page for simulate(), and
>> have only tested relatively few cases for binomial,
>> poisson and Gamma. I've wanted to expose this to you, so
>> you can provide more feedback and possibly even a patch
>> to
>> svn.r-project.org/R/trunk/src/library/stats/man/simulate.Rd
>>
>> Martin
>>
>>
>>