I've received 2 responses (see below). Thanks to Frank Harrell and Greg
Snow. I tried the proposition of Greg Snow and it worked fine for
lm-objects. As for lme-objects (version 3.2) I used
lme.obj$coefficients$fixed instead of coef(lme.obj).
Daniel
****************************************************************************
***********************
Date: Tue, 25 Jan 2000 10:37:17 -0800 (PST)
From: Greg Snow <snow@statsci.com>
To: Daniel Erdin <erdin@inw.agrl.ethz.ch>
Subject: Re: [S] estimates of the untransformed coefficients after the use
of poly()
On Tue, 25 Jan 2000, Daniel Erdin wrote:
>
> Hello,
> I'm working with Splus 4.5 on Windows NT. I have some data with polynomial
> dependencies. As I've understood, the use of poly() has statistical
> advantages in the modelling of polynomial dependencies.
> In the Splus-books I have (for version 3.3, 1995), it is shown, that for a
> model
> lmfit1_lm(y~poly(x1,3))
> the coefficients corresponding to the untransformed values of the
> polynomials of x can be obtained through
> poly.transform(poly(x1,3),lmfit1)
> This works fine.
>
> But I have more than the covariable x1 in my model and if I try
> lmfit2_lm(y~poly(x1,3)+x2+x3)
> I get no result with the following function
> poly.transform(poly(x1,3),lmfit2)
> just an error message
> Error in poly.transform(........): wrong length for coefs
try changing that to:
poly.transform(poly(x1,3),coef(lmfit2)[1:4])
or more generaly, extract the list of coefficients, then give the function
only those coefficients that correspond to the predictor that you are
using (including the intercept).
>
> I understand the reason for the error message, but how do I have to use
> poly.transform() to get what I want? Or is there an other solution to get
> the estimates of the coefficients for the not transformed polynomials? At
> last, I have the same problem in lme-models and would be interested if the
> solution in lme is the same.
It should be the same for lme.
***************************************************************************
Date: Tue, 25 Jan 2000 07:12:08 -0500
From: Frank E Harrell Jr <fharrell@virginia.edu>
Organization: University of Virginia
X-Accept-Language: en
To: Daniel Erdin <erdin@inw.agrl.ethz.ch>
CC: s-news@wubios.wustl.edu
Subject: Re: [S] estimates of the untransformed coefficients after the use
ofpoly()
References: <3.0.32.20000125123456.009d0e60@gessler.ethz.ch>
Sender: owner-s-news@wubios.wustl.edu
X-loop: s-news
I haven't used poly() to be able to check out the error you reported. However
in the days of accurate statistical computing, I don't see any advantage
that justifies the
complexity of orthogonal polynomials. Fitting ordinary polynomials works
fine although
splines are better. Note that ordinary polynomials have highly collinear
terms but this is
of no consequence as this does not affect predicted values, standard errors of
predicted values, or multiple degree of freedom tests of association.
In the Design library I have a little function pol for regular polynomials.
You could easily
strip this function to come up with something like the following.
Speed of the code could be improved
pol <- function(x, poly.degree) {
if(poly.degree < 2) stop('poly.degree must be >= 2')
xd <- matrix(0,nrow=length(x),ncol=poly.degree)
nam <- as.character(substitute(x))
name <- character(poly.degree)
name[1] <- nam
xd[,1] <- x
for(j in 2:poly.degree) {
name[j] <- paste(nam,"^",j,sep="")
xd[,j] <- x^j }
dimnames(xd) <- list(names(x), name)
xd
}
. -Frank Harrell
***************************************************
Daniel Erdin
Swiss Federal Institute of Technology
Institute of Animal Science
Animal Breeding
Research Station Chamau
CH-6331 Huenenberg
Phone: +41 41 780 10 47
Fax: +41 41 780 43 83
e-mail: erdin@inw.agrl.ethz.ch
homepage: http://www.tz.inw.agrl.ethz.ch/~erdin/
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu. To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message: unsubscribe s-news
|