Skip to content

Problem with mle in stats4 (R 1.9.1)

4 messages · Daniel Hoppe, Brian Ripley, Peter Dalgaard

Original
Daniel Hoppe
# 
Hi!

This is a repost of an earlier message (with a clearer example
demonstrating the problem I ran into). If you run the mle example in
stats4

     library(stats4)
     x <- 0:10
     y <- c(26, 17, 13, 12, 20, 5, 9, 8, 5, 4, 8)
     ll <- function(ymax=15, xhalf=6)
         -sum(stats::dpois(y, lambda=ymax/(1+x/xhalf), log=TRUE))
     (fit <- mle(ll))
     plot(profile(fit), absVal=FALSE)

everything works fine. Now run 

     (fit <- mle(ll, method="BFGS", control=list(ndeps=c(1e-3, 1e-3))))
     plot(profile(fit), absVal=FALSE)

and you will get

Error in interpSpline.default(obj[[i]]$par.vals[, i], obj[[i]]$z,
na.action = na.omit) : 
        only 0's may mix with negative subscripts

This happens because optim gets the old ndeps but just one variable to
optimize. The optim error is not reported, so the problem is rather hard
to trace. Would you agree that this would be worth fixing?

Best regards,

Daniel

(original posting with quick workaround
http://tolstoy.newcastle.edu.au/R/help/04/07/1778.html)

--
Daniel Hoppe
Department of Marketing
University of Vienna
Bruenner Strasse 72
1210 Vienna
Austria
Brian Ripley
# 
We are currently on R 2.0.0 alpha.  If you submit a bug fix, preferably 
this week and at the latest next week, it will be considered for 2.0.0.

Please take seriously that R is an open project and does not have `staff' 
to fix bugs.  If you think it is worth fixing, please fix it for us
by providing a complete analysis and fixes for all the possible options 
that might be affected.

The best place for a patch to be sent is R-bugs.
On Mon, 13 Sep 2004, Daniel Hoppe wrote:

            

      
      
      
    

        

      
      
    

        

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Peter Dalgaard
    

    

      
      #
    

  


  

      
"Daniel Hoppe" <daniel.hoppe@univie.ac.at> writes:

            

      
      
      
    

        
....

            

      
      
      
    

        
Definitely (and the problem is still in 2.0.0 alpha). Thanks for the
reminder.

            

      
      
      
    

        
Could well be fix, not just workaround.... Given that optim wasn't
designed to be able to hold arguments constant, any fix is going to be
somewhat kludgy.

  
    

      
      
    

  


  


      

    

    

      
      

        


  

        

      Peter Dalgaard
    

    

      
      #
    

  


  

      
Peter Dalgaard <p.dalgaard@biostat.ku.dk> writes:

            

      
      
      
    

        
Had another look and your workaround turned out to be a bit too
kludgy. This seems to work (the eval() is important):

        if (!is.null(call$control$ndeps))
           call$control$ndeps <- eval(call$control$ndeps)[-i]
        if (!is.null(call$control$parscale))
           call$control$parscale <- eval(call$control$parscale)[-i]

and a bit further down, I think you need to move the assignment to
"call" inside the "for (i in which)" loop, so that the changes don't
accumulate. 

Leaves one issue, namely *where* to eval, and the above code is not
quite right (you might capture a local variable of profile.mle, and if
you called mle from a function, references to local variables inside
that function can not be passed). Then again it actually never was
right since the try(eval(call),....) call has the same issue. We
really ought to save the environment of the original call to mle to be
safe. How best to do that is a bit unclear to me. Could add a slot of
just set the environment of the call (but that seems not to be
documented to work...).