Re: Glm and problem with a particular value of the categorica

["Southworth, Harry" <Harry.Southworth@astrazeneca.com>]

You might want to look at Firth's method. Also read Heinze & Schemper.

 

There is an S-PLUS library called logistf by Meinhard Ploner for fitting this way. There's also an R version of it (or perhaps its the same version that runs in both packages). There's another R package called brlr by Firth.

 

There are other ways of penalising the likelihood. In the Design library there's a function called lrm. Have a look at it's help file.

 

http://www.akh-wien.ac.at/imc/biometrie/programme/frameset_programme_en.htm

http://www2.warwick.ac.uk/fac/sci/statistics/staff/academic/firth/software/brlr/

D. Firth, Bias reduction of maximum likelihood estimates, Biometrika, 80, 1993, 27-38

G. Heinze & M. Schember, A solution to the problem of separation in logistic regression, Statistics in Medicine, 21, 2002, 2409-2419

____________________________________________

Harry Southworth,
SE2 T/2 Parklands,
Alderley Park,
Cheshire, SK10 4TG, UK
Tel.: +44 1625 518327
Mobile: 07867 676073
____________________________________________

-----Original Message-----
From: Georges KUNSTLER [mailto:kunstler@cefe.cnrs-mop.fr]
Sent: 21 April 2004 10:17
To: s-news@lists.biostat.wustl.edu
Subject: [S] Glm and problem with a particular value of the categorical predictor that unambiguously predicts the output

Dear Splus users,

I'm using Splus 2000 Professional and am having a problem with fitting a glm, with binomial errors and logit link. Y variable is a mortality rate (between 0 and 1) and the predictor variable X is categorical with 8 levels. For the level X4, Y is always 1. When I use the function multicomp to compare the mean for each value of the variables X1, this give really strange result for the comparison with the level X4 (high Std.Error, and no significant differences with the other level).

 As mentioned in a previous mail of Tony Plate in the archive the problems come from the fact that a particular value of one categorical predictor unambiguously predicts the output. So if I replace 1 by 0.9 the multicomp give logical results.

What is the good way to solve this problem? Replace all the value 1 by 0.9 (0.999 didn't seems to work!) or just for the level X4?

Thank you for any help you may be able to offer.
Georges Kunstler
Centre d'Ecologie Fonctionnelle et Evolutive. CNRS
Département Dynamique des Systèmes Ecologiques
1919 Route de Mende F-34293 Montpellier Cedex 5
Tel : 04.67.61.33.05
Fax : 04.67.41.21.38

--
passerelle antivirus du campus CNRS de Montpellier
--