[R] lme, best model without convergence

[Spencer Graves]

	  I'm sorry.  My post was in error on that point.  I did "rm(fm1.)", 
then tried again and got the same error message you got.

	  Conclusion:  Either the 'returnObject' argument doesn't work in the 
version(s) of lme that you and I are using or we've flunked another 
literacy exam while RTFM.  With luck, someone else will read this post 
and answer this question for both of us.

	  If I had time to work on this, I'd make local copies of the code and 
walk through them line by line using 'debug' until I figured out how to 
solve the problem.  I've outlined how to do this in previous posts, and 
RSiteSearch("graves debug lme") led me to 
"http://finzi.psych.upenn.edu/R/Rhelp02a/archive/72784.html", in case 
you want to try that.

	  Hope this helps,
	  Spencer Graves

 > sessionInfo()
Version 2.3.0 (2006-04-24)
i386-pc-mingw32

attached base packages:
[1] "methods"   "stats"     "graphics"  "grDevices" "utils"     "datasets"
[7] "base"

other attached packages:
     nlme
"3.1-73"

#########################
 > Dear Spencer Graves,
 >
 > thank you for your reply and the hint to the 'control' argument of the
 > lme function.
 > Thank you for your self-contained example. Running this example I get
 > the verbose output. So I am able to trace the iterations and maybe I am
 > able to detect ridges.
 > However, I just updated the nlme-package at the Mainz CRAN mirror (on
 > R-windows-2.2.1 version), but I do not get the result object from the
 > not-converging fit:
 >
 >  > lmeCtlList <-
 > + lmeControl(maxIter=2,
 > +             msMaxIter=3, tolerance=1e-4, niter=4,
 > +  msTol=1e-5, nlmStepMax=5,
 > +   ,msVerbose=TRUE
 > +   ,returnObject=TRUE
 > + )
 >  >
 >  > fm1. <- lme(distance ~ age, data = Orthodont,
 > +                   control=lmeCtlList) # random is ~ age
 >   0      318.310: -0.202572 0.0501133  2.17193
 >   1      318.297: -0.202070 0.0662548  2.16740
 >   2      318.286: -0.207284 0.0593806  2.15302
 >   3      318.257: -0.215916 0.0783917  2.12677
 >   3      318.257: -0.215916 0.0783917  2.12677
 > Error in lme.formula(distance ~ age, data = Orthodont, control =
 > lmeCtlList) :
 >         iteration limit reached without convergence (9)
 >  >
 >  > fm1.
 > Error: object "fm1." not found
 >  >

##########################
lmeControl               package:nlme               R Documentation

Control Values for lme Fit

Description:

      The values supplied in the function call replace the defaults and
      a list with all possible arguments is returned. The returned list
      is used as the 'control' argument to the 'lme' function.
##################

Note the last sentence.  Now consider the following example, combining
modifications of your example with the example from "?lme":

> lmeCtlList <-
+ lmeControl(maxIter=2,
+            msMaxIter=3, tolerance=1e-4, niter=4,
+ msTol=1e-5, nlmStepMax=5,
+  ,msVerbose=TRUE
+  ,returnObject=TRUE
+ )
>
>      fm1. <- lme(distance ~ age, data = Orthodont,
+                  control=lmeCtlList) # random is ~ age
   0      318.310: -0.202572 0.0501133  2.17193
   1      318.297: -0.202071 0.0662548  2.16740
   2      318.286: -0.207284 0.0593805  2.15302
   3      318.257: -0.215916 0.0783916  2.12677
   3      318.257: -0.215916 0.0783916  2.12677
Error in lme.formula(distance ~ age, data = Orthodont, control =
lmeCtlList) :
	iteration limit reached without convergence (9)
> fm1.
Linear mixed-effects model fit by REML
   Data: Orthodont
   Log-restricted-likelihood: -221.3183
   Fixed: distance ~ age
(Intercept)         age
  16.7611111   0.6601852

Random effects:
  Formula: ~age | Subject
  Structure: General positive-definite
             StdDev    Corr
(Intercept) 2.3270360 (Intr)
age         0.2264279 -0.609
Residual    1.3100396

Number of Observations: 108
Number of Groups: 27
>
	  Hope this helps.
	  Spencer Graves
p.s.  First thanks for this example.  I didn't know until I read your
question that one could actually get lme to return something when it
didn't converge.  Second, you might have gotten a quicker reply if you
had included a simple, self-contained example, as suggested in the
posting guide! "www.R-project.org/posting-guide.html".  Instead, since I
didn't know the answer, I had to invent one.  If your example had been
self contained, you might have gotten an earlier reply from someone who
didn't have the time or inclination to invent an example like I did but
who might have otherwise been able to solve the problem.

Thomas Wutzler wrote:
> Dear R-help list readers,
> 
> I am fitting mixed models with the lme function of the nlme package.
> If I get convergence depends on how the method (ML/REM) and which (and 
> how much) parameters will depend randomly on the cluster-variable.
> 
> How get the bist fit without convergence?
> 
> 
> 
> 
> I set the parameters msVerbose and returnObject to TRUE:
> 
> lmeControl(maxIter=50000, msMaxIter=200, tolerance=1e-4, niter=50, 
> msTol=1e-5, nlmStepMax=500, 	
> 	,msVerbose=TRUE
> 	,returnObject=TRUE
> )
> 
> However, the lme-functions does not produce verbose output, nor does it 
> return the best fit if lme is not converging.
> It returns only an error:
> Error in lme.formula(y ~ lndbh + I(lndbh^2) + lnh + I(lnh^2), random = 
> ~lndbh +  :
>          iteration limit reached without convergence (9)
> 
> 
> 
> Best regards
> Thomas
> 
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html