[R] GLM weights for the Poisson family

[Milan Bouchet-Valat]

I think you should have a look at svyglm() from the survey package.


My two cents


Le mercredi 05 février 2014 à 14:41 +1300, Rolf Turner a écrit :
> You should direct your inquiry to R-help, not to me personally.  I am 
> taking the liberty of cc-ing my reply back to the list.
> 
> I really haven't the time at the moment to think the issue through 
> thoroughly, but off the top of my head:  If you are going to use 
> weighted log likelihoods then any comparison of models that you engage 
> in should involve the *same* weights, otherwise you doing the good old 
> apples-with-oranges thing.
> 
> So yes, the weights will change the log-likelihood and the AIC.  And so 
> they should.  And if you use AIC to compare models which are 
> meaningfully comparable (id est, have the same weights) this is not a 
> problem.
> 
> As I say, this is off the top of my head.  Others older (???) and wiser 
> than I may correct me.
> 
> cheers,
> 
> Rolf Turner
> 
> On 05/02/14 11:56, IamRandom wrote:
> > I am trying to do weighted Poisson regression.  I have count data.
> >
> > Simple example:
> > set.seed(50)
> > x=seq(0,1,length=100)
> > y=numeric(100)
> > y[seq(1,100,by=2)]=round(exp(1.5*x[seq(1,100,by=2)]+rnorm(50,0,.1)),0)
> > y[seq(2,100,by=2)]=round(exp(1.5*x[seq(1,100,by=2)]+rnorm(50,0,1)),0)
> > weigh1=numeric(100)
> > weigh1[seq(1,100,by=2)]=rep(5,50)
> > weigh1[seq(2,100,by=2)]=rep(1,50)
> >
> >
> > The -2*loglikelihood of both of these regressions is the same with lm,
> > which makes sense. The scaling of the weights does not affect the
> > log-likelihood.
> >  >-2*logLik( lm(y~x, weights=weigh1))[1]
> >  >-2*logLik( lm(y~x, weights=weigh1/3))[1]
> >
> > The -2*loglikelihood of these two regressions are different with glm:
> >  > -2*logLik(glm(y~x, family="poisson", weights=weigh1))[1]
> >  > -2*logLik(glm(y~x, family="poisson", weights=weigh1/3))[1]
> >
> > This means that the AIC and other model comparison techniques with this
> > weighted Poisson regression are dependent on the scaling of the
> > weights.  So I assume I misunderstand what the "weights" are doing in
> > the glm function.
> >
> > -Tracy
> >
> >
> >
> > On 2/4/2014 12:56 PM, Rolf Turner wrote:
> >>
> >> On 04/02/14 20:12, IamRandom wrote:
> >>
> >>> I am running a simple example of GLM.  If I include weights when
> >>> family="poisson" then the weights are calculated iteratively and
> >>> $weights and $prior.weights return different values.  The $prior.weights
> >>> are what I supplied and $weights are the "posterior" weights of the
> >>> IWLS.  If I include weights with family="gaussian" then the weights are
> >>> static and $weights and $prior.weights return the same values; it seems
> >>> to ignore IWLS algorithm procedure.  I really want the family="poisson"
> >>> to behave like the family="gaussian" and use the static weights.
> >>> Thoughts?
> >>
> >> As far as I understand things, your desideratum makes no sense. The
> >> prior weights and the just-plain-weights are very different creatures.
> >> The reason they wind up being the same for the gaussian family is that
> >> for the gaussian family the likelihood is maximized by least squares;
> >> there is no need for iteration or for re-weighting.
> >>
> >> The poisson family cannot behave like the gaussian family because for
> >> the poisson family (or any family *other* than gaussian) iteration is
> >> necessary in order to maximize the likelihood.
> >>
> >> You might get some insight into what's going on if you were to read
> >> Annette Dobson's book "An Introduction to Generalized Linear Models"
> >> (Chapman and Hall, 1990).
> >>
> >> cheers,
> >>
> >> Rolf Turner
> >>
> >>
> >>
> >
> 
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.