R help - Aug 2005 - How to collect better estimations of a logistic model parameters, by using bootstrapping things ?

Dear all,

I know that when using R, people should have a sufficient level in  
statistics.
As well, I'm not a genius, when dealing with logistic regressions.
I would like to construct ICs, IPs,  for a logistic regression, but  
the point is I have just 41 observations.
I had a look at the Design package and noticeably the lrm function,  
but I'm still not able to reduce the IC's, as I was trying to do this  
in SPSS (but do not like it).
I have heard of a mean to do this by using bootstrap, but I'm still  
waiting to find the right way to use it.

As well I would like to find a fine way to do stepwise forward  
selection In R, as I am not sure wich kind of variable may be  
involved in the model, which is composed with 13 numeric variables,  
and a dichotomic variable named "expatriation". I have got a total of
41 observations, as mentionned above.

I'm using R on macintosh, I have used the function lrm, brlr,  
bootstrap (but for others uses than logistic regressions), and I am  
looking for a great and paved way to do Confiance Intervals, and to  
compute significance values for each of the logistic model's  
parameters, by using R ans bootstrapping, of course.


Any ideas ?

Laurent.

--
--~~ Toulouse, Grenoble, Auch, Arcachon, B??ziers, Paris,
Saragosse, L??vignac Sur Save, habitats naturel du Valdo. ~~--
< http://www.le-valdo.com>

I presume an `IC' is a confidence interval, but what is an `IP'?

I think you need to think hard about the assumptions you want to make. The 
usual way to do logistic regression is via glm(), and the confint() 
function will give you confidence intervals based on profile likelihoods 
that are rather accurate (you need package MASS for this).

If you want to bootstrap you have to decide how.  Case-based resampling is 
the only easy way, and would be appropriate only if the 41 cases were a 
sample and not a design (and even then experts would argue for conditional 
inference).  Bootstrapping logistic regression for a design involves a lot 
of assumptions, and there is not much to suggest that bootstrapping will 
better than using confint().

As for model selection, step() will do it, but given your problem sizes it 
is really _at best_ an exploratory procedure for what extra data might be 
worth collecting.

On Wed, 24 Aug 2005, Laurent Valdes wrote:
>
>
> Dear all,
>
> I know that when using R, people should have a sufficient level in
> statistics.
> As well, I'm not a genius, when dealing with logistic regressions.
> I would like to construct ICs, IPs,  for a logistic regression, but
> the point is I have just 41 observations.
> I had a look at the Design package and noticeably the lrm function,
> but I'm still not able to reduce the IC's, as I was trying to do
this
> in SPSS (but do not like it).
> I have heard of a mean to do this by using bootstrap, but I'm still
> waiting to find the right way to use it.
>
> As well I would like to find a fine way to do stepwise forward
> selection In R, as I am not sure wich kind of variable may be
> involved in the model, which is composed with 13 numeric variables,
> and a dichotomic variable named "expatriation". I have got a
total of
> 41 observations, as mentionned above.
>
> I'm using R on macintosh, I have used the function lrm, brlr,
> bootstrap (but for others uses than logistic regressions), and I am
> looking for a great and paved way to do Confiance Intervals, and to
> compute significance values for each of the logistic model's
> parameters, by using R ans bootstrapping, of course.
>
>
> Any ideas ?
>
> Laurent.
>
> --
> --~~ Toulouse, Grenoble, Auch, Arcachon, B?ziers, Paris,
> Saragosse, L?vignac Sur Save, habitats naturel du Valdo. ~~--
> < http://www.le-valdo.com>
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide!
http://www.R-project.org/posting-guide.html
>
-- 
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