Re: [R] problems with optim, "for"-loops and machine precision

From: Ken Beath <>
Date: Wed 10 Jan 2007 - 22:34:10 GMT

Two possibilities for why your 7 parameter model fits worse than the 6 are that you are finding a local maximum, which might suggest using a different parameterisation or the functions are accessing some global data and so aren't behaving as expected. This could be why they work properly when run on their own.

I would also check what happens if convergence fails for the 7 parameter model, in your code this isn't handled well. Also if the constraint on parameters of >=0 is actually >0, it may be better to work with parameters on the log scale, avoiding the constraints.

I have found with simulations it is useful to save the fitted objects, so they can be inspected later, or for the parameters to be extracted after the models are fitted. This method allows restarting in case of computer crashes.


>>> "Simon Ruegg" <> 01/10/07 11:18 PM >>>
Dear R experts,  

I have been encountering problems with the "optim" routine using "for" loops. I am determining the optimal parameters of several nested models by
minimizing the negative Log-Likelihood (NLL) of a dataset.  

The aim is to find the model which best describes the data. To this end, I
am simulating artificial data sets based on the model with the least number
of parameters (6) and the parameters determined with the field data. For each artificial set I estimate the parameters of the model with 6 parameters
and the next more complex model with 7 parameters (two of these parameters
are equal in the 6-parameter model) by minimizing the corresponding NLL with
"optim". In theory the 7-parameter model should fit the data either equally
or better than the 6-parameter model. Therefore the difference of the minimal NLLs should be 0 or larger.

For 500 data sets I use the following code:  






for (s in 1:500)


assign("data",read.table(paste("populations/TE23simset_",s,".txt",sep="")),e nv=MyEnv) #reading a data set  



  if (M23$convergence==0)



























For some strange reason the results do not correspond to what I expect: about 10% of the solutions have a difference of NLL smaller than 0. I have
verified the optimisation of these results manually and found that a minimal
NLL was ignored and a higher NLL was returned at "convergence". To check what was happening I inserted a printing line in the NLL function to print
all parameters and the NLL as the procedure goes on. To my surprise "optim"
then stopped at the minimal NLL which had been ignored before. I have then
reduced the machine precision to .Machine$double.digits=8 thinking, that the
printing was slowing down the procedure and by reducing the machine precision to speed up the calculations. For an individual calculation this
solved my problem. However if I implemented the same procedure in the loop
above, the same impossible results occurred again.  

Can anyone tell me where I should be looking for the problem, or what it is
and how I could solve it?  

Thanks a lot for your help    



Simon Ruegg, PhD candidate

Institute for Parasitology

Winterthurstr. 266a

8057 Zurich


phone: +41 44 635 85 93

fax: +41 44 635 89 07


        [[alternative HTML version deleted]] mailing list PLEASE do read the posting guide and provide commented, minimal, self-contained, reproducible code. mailing list PLEASE do read the posting guide and provide commented, minimal, self-contained, reproducible code. Received on Thu Jan 11 09:44:42 2007

Archive maintained by Robert King, hosted by the discipline of statistics at the University of Newcastle, Australia.
Archive generated by hypermail 2.1.8, at Thu 11 Jan 2007 - 00:30:25 GMT.

Mailing list information is available at Please read the posting guide before posting to the list.