s-news
[Top] [All Lists]

Re: Use of step.gam

To: Brian Newquist <BNewquist@constellagroup.com>
Subject: Re: Use of step.gam
From: Prof Brian Ripley <ripley@stats.ox.ac.uk>
Date: Wed, 16 Feb 2005 08:03:01 +0000 (GMT)
Cc: s-news@lists.biostat.wustl.edu
In-reply-to: <1BF5A584BBD24645A0524FA419524BCB0AE5F409@banyan.constellagroup.com>
References: <1BF5A584BBD24645A0524FA419524BCB0AE5F409@banyan.constellagroup.com>
On Tue, 15 Feb 2005, Brian Newquist wrote:

Is anyone familiar with the use of the step.gam function?  I am trying to
find the best value for the span parameter for use in the gam function.  The
S-Plus help documentation suggests that step.gam will return the best model
(in terms of AIC criteria) from among those that you have defined in the
scope argument.

Rather, it does a very simple stepwise search amongst them.

The example below illustrates what occurred when I used this function.  Does
anyone know where I went wrong?  Any help will be greatfully appreciated.

fit.gam <- gam( OUTVAR ~ LOGVAR + VAR1 + VAR2 + VAR3, binomial, data =
mydata)

step.gam(fit.gam, scope=list(

  "LOGVAR" = ~ lo(LOGVAR, span=0.1) + lo(LOGVAR, span = 0.15) + lo(LOGVAR,
span = 0.20)))

The following message appears after running step.gam :

Problem in untangle.scope(object$terms, scope): regimen 1 does not appear in
the initial model

You have not allowed the current term, nor is this in increasing order of complexity. Try

"LOGVAR" = ~ 1 + LOGVAR + lo(LOGVAR, span = 0.20) + lo(LOGVAR, span = 0.15)"

and move on from there. All step.gam does is move one step at a time to the left or right in that formula: see the DETAILS section on the help page, or section 7.3.1 of `Statistical Models in S'.

--
Brian D. Ripley,                  ripley@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

<Prev in Thread] Current Thread [Next in Thread>
  • Use of step.gam, Brian Newquist
    • Re: Use of step.gam, Prof Brian Ripley <=