thr3ads.net - R help - [R] nlme iteration process, few questions [Aug 2006]

If this information is useful, please help other people find it:
Share via:

Vaidotas Zemlys

2006-Aug-09 07:58 UTC

[R] nlme iteration process, few questions

Hi all,

Recently I started using nlme intensively, and since it is all new for
me, I have some questions. I am running nlme with
control=list(verbose=TRUE) and during one lengthy fitting, I started
watching the output for some clues, how to speed up the process. I
noticed that in one case, the iteration process is alternating between
two solutions. Here is an output of 5 iterations, after which I
stopped the calculation:

**Iteration 22
LME step: Loglik: -1580.5 , nlm iterations: 50
reStruct  parameters:
        id1         id2         id3         id4         id5         id6
 -2.6387745   5.7964164   3.4153271   4.1734349  48.7541909  -0.1250377
        id7         id8         id9        id10
-36.7011651  30.1780697 -56.0266217 127.3234221

PNLS step: RSS =  137743.3
 fixed effects:24.0656  0.105708  -2.86501  0.542384
 iterations: 7

Convergence:
    fixed  reStruct
0.1353935 3.3655468

**Iteration 23
LME step: Loglik: -1579.346 , nlm iterations: 50
reStruct  parameters:
       id1        id2        id3        id4        id5        id6        id7
 -2.285671   6.380246   3.583532   4.248227  -1.398577  -1.318335 -47.773224
       id8        id9       id10
 12.834766 -63.669545 136.063383

PNLS step: RSS =  151492.2
 fixed effects:20.8075  0.105065  -2.85908  0.544615
 iterations: 7

Convergence:
    fixed  reStruct
0.1565794 1.5396836

**Iteration 24
LME step: Loglik: -1580.500 , nlm iterations: 50
reStruct  parameters:
        id1         id2         id3         id4         id5         id6
 -2.6387708   5.7964464   3.4153368   4.1734411  48.7509855  -0.1246523
        id7         id8         id9        id10
-36.7012641  30.1788064 -56.0254345 127.3242285

PNLS step: RSS =  137743.5
 fixed effects:24.0665  0.105706  -2.86503  0.542385
 iterations: 7

Convergence:
    fixed  reStruct
0.1354136 3.3647287

**Iteration 25
LME step: Loglik: -1579.346 , nlm iterations: 50
reStruct  parameters:
       id1        id2        id3        id4        id5        id6        id7
 -2.285674   6.380234   3.583527   4.248225  -1.399981  -1.318104 -47.772885
       id8        id9       id10
 12.835424 -63.669219 136.063111

PNLS step: RSS =  151495.6
 fixed effects:20.8062  0.105066  -2.85907  0.544616
 iterations: 7

Convergence:
    fixed  reStruct
0.1566962 1.5398072

**Iteration 26
LME step: Loglik: -1580.500 , nlm iterations: 50
reStruct  parameters:
        id1         id2         id3         id4         id5         id6
 -2.6387698   5.7964218   3.4153246   4.1734379  48.7501373  -0.1249832
        id7         id8         id9        id10
-36.7020319  30.1775554 -56.0315196 127.3225180

PNLS step: RSS =  137744.8
 fixed effects:24.066  0.105707  -2.86501  0.542384
 iterations: 7

Convergence:
    fixed  reStruct
0.1354528 3.3650687


Is there any particular strategy I can pursue in such case? Besides
the usual ones: changing the starting values, changing the model,
using different scaling. I am not familiar with inner workings of nlme
algorithm, maybe this an indication of some known problem?

Thanks for all answers.

Vaidotas Zemlys
--
Doctorate student, Vilnius University
http://www.mif.vu.lt/katedros/eka/katedra/zemlys.php

Ken Beath

2006-Aug-10 11:45 UTC

head link

[R] nlme iteration process, few questions

> Recently I started using nlme intensively, and since it is all new for
> me, I have some questions. I am running nlme with
> control=list(verbose=TRUE) and during one lengthy fitting, I started
> watching the output for some clues, how to speed up the process. I
> noticed that in one case, the iteration process is alternating between
> two solutions. Here is an output of 5 iterations, after which I
> stopped the calculation:
>
I have sometimes fixed this problem by increasing the number of nlm  
and PNLS iterations. It can be due to the random effects estimates  
being too small, so not necessary in the model, or I suspect that a  
high correlation between the random effects would also produce the  
problem.

The usual method is to try fitting simpler models by removing random  
effects or with structured covariance matrix and then use the  
estimates of the fixed effects as starting values for more complex  
models.

HTH
Ken

R help - Aug 2006 - nlme iteration process, few questions

[R] nlme iteration process, few questions

[R] nlme iteration process, few questions