R help - Nov 2007 - not R question : alternative to logistic regression

I was just curious if anyone knew of an alternative model to logistic regression
where the probabilities seems pretty linear to the predictor rather than having
that S shape that probit and logit assume.

Maybe there is there some kind of other GLM that could accomplish that. Any
textbook references or suggestions
are appreciated. I have most of the texts but if someone knows of a text that
talks about this, it
would be helpful because I don't even know of the name
for such a model ( if it exists ) so I don't know where to look.

                                 Thanks

> -----Original Message-----
> From: r-help-bounces at r-project.org 
> [mailto:r-help-bounces at r-project.org] On Behalf Of 
> markleeds at verizon.net
> Sent: Thursday, November 15, 2007 12:04 PM
> To: r-help at r-project.org
> Cc: cspark at clemson.edu
> Subject: [R] not R question : alternative to logistic regression
> 
> I was just curious if anyone knew of an alternative model to 
> logistic regression where the probabilities seems pretty 
> linear to the predictor rather than having that S shape that 
> probit and logit assume.
> 
> Maybe there is there some kind of other GLM that could 
> accomplish that. Any textbook references or suggestions
> are appreciated. I have most of the texts but if someone 
> knows of a text that talks about this, it 
> would be helpful because I don't even know of the name
> for such a model ( if it exists ) so I don't know where to look.
> 
>                                  Thanks
> 
Google "linear probablity model" and you will find a lot of
information.

Hope this is helpful,

Dan

Daniel J. Nordlund
Research and Data Analysis
Washington State Department of Social and Health Services
Olympia, WA  98504-5204

MARK LEEDS-3 wrote:
> 
> I was just curious if anyone knew of an alternative model to logistic
> regression where the probabilities seems pretty linear to the predictor
> rather than having that S shape that probit and logit assume.
> 
> 
  Well, the logistic curve is very close to linear over the middle range of
probabilities --
are your probabilities all in the middle, or do they seem to hit 0 and 1
fairly sharply?

  Using the formula interface to mle in the bbmle package (blatant plug),
you
could do something like

  mle2(surv~dbinom(prob=pmin(pmax(a+b*x,0.001),0.999),size=1),start=...)

the pmin/pmax cut the probabilities off before they get to 1 or 0.

  I'm not sure this is a good idea, but it is possible.

  cheers
   Ben Bolker

-- 
View this message in context:
http://www.nabble.com/not-R-question-%3A-alternative-to-logistic-regression-tf4816865.html#a13781110
Sent from the R help mailing list archive at Nabble.com.

Probabilities can only be even approximately linearly related to a
continuous predictor variable for a limited range, otherwise the model
will predict probabilities below 0 or above 1.

At some point, they have to tail off... unless you are modelling
something trivial like 'probability of being above the x'th
quantile...'

But there's no particular reason your logistic regression has to be
based on a linear predictor scale. Take logs or otherwise transform the
predictor if that is justified by the underlying process or the data? 

Steve E
>>> <markleeds at verizon.net> 11/15/07 8:03 PM >>>
I was just curious if anyone knew of an alternative model to logistic
regression where the probabilities seems pretty linear to the predictor
rather than having that S shape that probit and logit assume.