s-news
[Top] [All Lists]

Problem with nlme

To: S-News <s-news@lists.biostat.wustl.edu>
Subject: Problem with nlme
From: Tristan Lorino <tristan.lorino@lcpc.fr>
Date: Fri, 18 Jul 2008 14:47:42 +0200
Organization: LCPC
Reply-to: Tristan Lorino <tristan.lorino@lcpc.fr>
Hi,

When I fit a nlme model, I obtain this output:

************************************************** 
> res
Nonlinear mixed-effects model fit by maximum likelihood
  Model: Aft ~ SSlogis(Age, A, B, C) 
  Data: afiss 
  Log-likelihood: -1820.924
  Fixed: A + B + C ~ 1 
  A B C 
 60 6 5

Random effects:
 Formula: list(`A ~ 1` , `B ~ 1` , `C ~ 1` )
 Level: IdS
 Structure: General positive-definite
            StdDev  Corr       
       A 76.080776 A     B    
       B 18.139884 0.874      
       C  8.797805 0.448 0.827
Residual 13.925247            

Number of Observations: 399
Number of Groups: 133 
**************************************************

I can see that the model is overparametrized (big correlation between
A-B and B-C), and the command "coef(res, aug=T)" leads to similar
values of A, B and C over all the groups.

When I try to fit the same model with only one random effect (A), I
obtain the following output:

**************************************************
> res2
Nonlinear mixed-effects model fit by REML
  Model: Aft ~ SSlogis(Age, A, B, C) 
  Data: afiss 
  Log-restricted-likelihood: -1850.726
  Fixed: A + B + C ~ 1 
        A        B        C 
 49.73993 4.972476 1.348317

Random effects:
 Formula: A ~ 1 | IdS
               A Residual 
StdDev: 39.51716 17.98275

Number of Observations: 399
Number of Groups: 133 
**************************************************

But the asymptot (A) being common to all groups, I would prefer to put
B as random effect:

**************************************************

> res3 <- nlme(Aft ~ SSlogis(Age, A, B, C), data = afiss, fixed = A +
> B + C ~ 1, random = B ~ 1, start = list(fixed = c(60,
        6, 5)), method = "REML")
Problem in nlme.formula(Aft ~ SSlogis(Age, A, B, C), data = afiss, fixed = ..: 
Step halving factor reduced below minimum in PNLS step 
Use traceback() to see the call stack


**************************************************

This failure message occures with both ML and REML methods, and also
with C introduced as random effect instead of A.

Is there a way to fix this fit? Or is it a problem associated to the
profile of the data?

Thank you,

Tristan Lorino



-- 
Laboratoire central des ponts et chaussées
[Division ESAR — Section AGR]
Route de Bouaye BP 4129
44341 Bouguenais Cedex
France
Tél 33 (0)2 40 84 56 18
Fax 33 (0)2 40 84 59 92


<Prev in Thread] Current Thread [Next in Thread>
  • Problem with nlme, Tristan Lorino <=