[R] nlme function in S+ and R

[Bert Gunter]

Derrick:

1. There are lots of people "out here" with such experience -- including the
author of the software!

2. nlme fitting involves iterative procedures whose final value (if indeed,
convergence to a final value occurs -- it certainly need not) depends on a)
the initial value you start from; (b) various control parameters
(tolerances, step sizes, direction), most of which are user-specifiable with
certain defaults; (c) the details of the iterative algorithms; (d)the exact
data: overspecified models can lead to great sensitivity to the other
aspects; and (e) the details of the numeric software in which the algorithms
are implemented.

afaics, you have only guaranteed that a) and d) are the same between S+ and
R (versions??? -- as requested by the posting guide!). However, in my
experience, it is very easy to overspecify nonlinear models (since this
depends in very complex ways on the exact data that you have), yielding
sensitivity to all of b,c, and e. So the differences that you see are not
surprising. 

-- Bert Gunter
Genentech


 -----Original Message-----
From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On
Behalf Of d.lee at stat.ubc.ca
Sent: Thursday, May 15, 2008 12:19 PM
To: r-help at r-project.org
Subject: [R] nlme function in S+ and R

Greetings and Salutations,

I was curious if anyone out there has had much experience with the nlme
function
in both S+ and R. Right now I am working on an HIV dataset and trying to fit
a
NLME model to the data, and the coefficients I get in S+ differ slightly
from
the coefficients in R:

R model    : 12.44  0.84 39.07 10.70  4.48  7.48  0.60  1.85  0.18
S model    : 12.45  0.59 36.65  1.07  5.46  7.59  0.35  1.92  0.20

Since there is a random component, I expect there to be some deviations from
one
another, but as you can see the coefficients for the 4th variable are just
too
different from one another and hence my concern. The S+ code I am using is:

fit4 <- try(nlme(lgcopy ~ logexp2(p1,b1,p2,b2,day),
                 fixed = list(p1 ~tnf, b1 ~tnf+cd4coun, p2~tnf,b2~basech50),
                 random = list(p1 ~ .,b1 ~.,p2~.,b2~.),
                 cluster = ~ patid,
                 data = completedata,
                 start = list(fixed=start0)
                 )
           )

and, I believe, I made the appropriate adjustments for R:

fit4 <- try(nlme(lgcopy ~ logexp2(p1,b1,p2,b2,day),
                 fixed = list(p1 ~tnf, b1 ~tnf+cd4coun, p2~tnf,b2~basech50),
                 random = list(p1+b1+p2+b2~1),
                 groups = ~ patid,
                 data = completedata,
                 start = list(fixed=start0)
                 )
            )

If anyone could lend some insight to this difference, it would be very much
appreciated. Cheers.


Sincerely

 - Derrick


 Derrick Lee, MSc Candidate
 Department of Statistics
 The University of British Columbia
 LSK-314A | 604 - 822 - 1299 x532
 d.lee at stat dot ubc dot ca | dgylee at mun dot ca
 www.stat.ubc.ca/~d.lee/ | www.math.mun.ca/~derrick0/

______________________________________________
R-help at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.