[R] GLM for logistic regression and WEIGHTS

[Prof Brian Ripley]

There is no restriction to integer weights in R.  Here is a (silly) 
example.

library(MASS)
fit1 <- glm(cbind(pm.y, pm.tot - pm.y) ~ density, binomial, data=rotifer)
wt <- runif(20)
update(fit1, weights=wt)

or

glm(pm.y/pm.tot ~ density, binomial, data=rotifer, weights=pm.tot)
glm(pm.y/pm.tot ~ density, binomial, data=rotifer, weights=pm.tot*wt)

which gives an harmless warning (not an error message).

I use this sort of thing for multiple imputation quite frequently.


On Sun, 28 Mar 2004, Marie-Pierre Sylvestre wrote:

> Hi all,
> 
> I want to use weights for a logistic regression. In SAS, all I have to
> do is to specify my weight vector (they are fractions) and use proc
> logistic on my binary output.

That is all you do in R, too.  See the example above.

> When I tried to do the same in R, I got an error message because my
> weights were not integer. 

Please read the posting guide and supply a reproducible example of how you 
got an *error* message here.

> I understand that the weight option in R is to
> be used when the dependent variable is a proportion so that the weight
> is the total from which this proportion is derived.

You `understand' incorrectly.  Are you familiar with the theory of
generalized linear models -- weights are part of the definition of a glm?

> So what should I do if I want to use logistic regression but want to use
> weight to give more importance to certain observations (e.g.
> weight=0.87) and less to others (e.g. weight=.45) ? Should I
> reparametrize everything in terms of counts or is there an easier way
> out?

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595