st: RE: RE: Logit using geometric means

[Nick Cox <n.j.cox@durham.ac.uk>]

My first take on this question was, I guess, similar to Kieran's, that the geometric mean of a binary variable which is partly 0 and partly 1 will be 0 regardless, so that the idea just does not fit with logit regression in any natural way. 

But the second take which makes more sense of the question is that Natacha is supposing that it's an assumption of logit regression that the _predictors_ are normally distributed. Not so, but nothing you stops you transforming (e.g. by taking logarithms) if that improves the model. 

On the other hand, Natacha may be using logit on continuous proportions bounded by 0 and 1, or 0 or 100%. But such proportions can't be lognormally distributed.  

Natacha: Your previous question, twice posted, on specific aweights evoked a range of puzzled replies from Steven Samuels, Austin Nichols and myself. You didn't reply to ease any of that puzzlement. 

If answers to this question do not solve your problem, you should expand on what you mean, as I think it's very unclear.

Nick 
n.j.cox@durham.ac.uk 

Kieran McCaul

If you are using logistic regression your dependent variable is binary.

N.M.Postel-Vinay@lse.ac.uk

I would like to know whether it is possible to make Stata use geometric
means instead of arithmetic means in a logistic regression.

I know that for a normal regression one can use a log transformation of
the dependent variable to achieve that aim.
Of course this is not possible with logistic regression. What solutions
would you suggest? Most of my variables are not normally distributed,
but are in fact lognormally distributed. But for my purposes only
logistic regression is of interest. How can I reconcile the two?


*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/