(Re: [R] lme in R-2.0.0: Problem with lmeControl) and parameter specification

[Pavel Khomski]

Hello!

Thanks a lot to Douglas Bates for your advice.

concerning the lme(...) function i wanted to put four other questions.

1.    in the specification of initial values in the pdMat-constructor i 
probably define a standard deviation (sigma_b) and not a variance 
(sigma_b^2). For instance
       in the Rail  example in
       Pinheiro/Bates on p.81 if i specify a random parameter as 
random=pdIdent(value=lambda<-diag(1000,1),form=~1),   (in S-plus),   
then the call to the lme(...)
       with just 0 iterations produces:
 
 > 
fm1Rail.lme<-lme(travel~1,data=Rail,random=pdIdent(value=lambda<-diag(1000,1),form=~1),control=list(msMaxIter=0,msVerbose=TRUE,niterEM=0))
Iteration:  0 ,  1  function calls, F=  66.37359
Parameters:
[1] -3.453878
Warning messages:
  ITERATION LIMIT REACHED WITHOUT OTHER CONVERGENCE in: ms( ~  - 
logLik(lmeSt,
    lmePars), start = list(lmePars = c(coef(lmeSt))),  ....

if i now print out an estimated std.dev.  for sigma_b   i get:

 > (fm1Rail.lme$sigma)*exp(unlist(fm1Rail.lme$modelStruct))
 reStruct.Rail
      107.6767

 > lambda
   [,1]
[1,] 1000

so that the estimated variance would be 107.6767 ^2 = 11594.27 what is 
much grater then 1000.  But hier we know that with  iterations the
value of variance will reduce  ( and at the convergence the StdDev is 
24.80547  )

so i think that lambda=1000 is specified equal to sigma_b as initial value.



2.    What is the meaning for 0-Iteration?



3.    are the parameters    fixed=beta   and   random=sigma    being 
calculated (just on time) only after all iterations have run, or they 
also be updated
       at every iteration with new value of teta ?   if the latter, how 
can i get them for each run ?



4.    Is it in principle possible to hold a variance components 
parameter, say   sigma_b,   as in Rail-example,   fixed (on specified 
value) through all the
        iteration steps (without changing it) and  only optimize  for   
teta=log(sigma_b/sigma_epsilon)  with fixed known value of sigma_b ?  
        how can it be done  ?

       
Thank you for replay