can someone enlighten me on the following or point me to the right
references where I can find out more about this?
Imagine, that I have a data set where the response variable is a
proportion (say the proportion of a volume) and with some explanatory
variable(s). The proportion of the response is not made up of several
entities (thus a classical logistic regression with 0s and 1s doesn't
work). [I also know that there are recent special methods to work with
compositional data e.g. developed by Peter Guttorp.] But let's assume for
the moment that I am either only interested in one component or that there
are only two components.
Now, I made up a simple example and tried two things:
Firstly, I fed the data into glm (Y ~ X, family= binomial). This seems to
give a sensible curve (though the tests might be wrong?) but the handbook
is very clear about that the binomial family expects only two different
values in the response.
Secondly, I transformed Y into Z = log (Y/1-Y). Then I fitted a line by
lm (Z ~ X). If I look at the estimated values they are very similar
(though the tests of the coefficients are not).
What I don't see/understand is:
- How exactly are the two methods related (is the former just a least
square and the latter a loglikelihood estimate)?
- Which of the two is (more?) correct?
- Is there a better way to do this with some "standard methods" (I am
trying to give some advice to someone else for this)?
- What happens if I cover the whole range between 0 and 1 (including quite
a few 0s and 1s)? E.g. I expect problems with the second version as
I assume that the resiudals won't be normally distributed anymore.
I am glad about any pointers or comments! Thank you for your time.
Regards, Lorenz
-- Lorenz Gygax| /| LGygax@amath.unizh.ch; room: 36-L-40 / |\ / | \ Department of Applied Mathematics / | \ University of Zuerich-Irchel /____| \ Winterthurerstr. 190; CH-8057 Zurich ______|____\ voice: 41-1-635-58-52 fax: 41-1-635-57-05 / / ~~~~~~~~~~~~~~~~~ Dachslernstr. 141, 8048 ZH, 41-1-432-51-47
----------------------------------------------------------------------- This message was distributed by s-news@wubios.wustl.edu. To unsubscribe send e-mail to s-news-request@wubios.wustl.edu with the BODY of the message: unsubscribe s-news