R help - Sep 2008 - question about GLM

Hello R-users,

I do not have much knowledge about generalized linear models therefore my
question maybe quite stupid.

I have data from 20 towns with their population and number of people with an
illness from those towns. I would like to use glm function in R so I
can calculate proportions of ill people (and later on produce confidence
intervals). I also want to compare those with original proportions of ill
people.

If I use:

  model1 <- glm(ill ~ offset(log(total)), family = poisson)
# ill - number of people with illness
#total - total number of people

with predict.glm I could get number of people (count data), but not the
proportions. If the obtained number I divide by 'total', I get the same
proportion for everyone. But what I want is a way of modeling proportions.
This probably requires to fit a different model but my lack of knowledge
isn't helping here.

I would appreciate any help.


Aiste

	[[alternative HTML version deleted]]

Aiste Aistike <aiste.aistike <at> gmail.com> writes:
> 
> Hello R-users,
> 
> I do not have much knowledge about generalized linear models therefore my
> question maybe quite stupid.
> 
> I have data from 20 towns with their population and number of people with
an
> illness from those towns. I would like to use glm function in R so I
> can calculate proportions of ill people (and later on produce confidence
> intervals). I also want to compare those with original proportions of ill
> people.
> 
> If I use:
> 
>   model1 <- glm(ill ~ offset(log(total)), family = poisson)
> # ill - number of people with illness
> #total - total number of people
> 
> with predict.glm I could get number of people (count data), but not the
> proportions. If the obtained number I divide by 'total', I get the
same
> proportion for everyone. But what I want is a way of modeling proportions.
> This probably requires to fit a different model but my lack of knowledge
> isn't helping here.
> 
  Not stupid -- but -- wouldn't a binomial model

 glm(cbind(ill,total-ill) ~ 1, family=binomial)

 make more sense?

  Read ?predict.glm carefully to determine whether you
are predicting responses on the linear predictor (=log-odds)
scale or the original scale

  Ben Bolker