R devel - Feb 2019 - mle (stat4) crashing due to singular Hessian in covariance matrix calculation

Hi, R developers.
when running mle inside a loop I found a nasty behavior. From time to 
time, my model had a degenerate minimum and the loop just crashed. I 
tracked it down to "vcov <- if (length(coef)) solve(oout$hessian)"
line,
being the hessian singular.
Note that the minimum reached was good, it just did not make sense to 
calculate the covariance matrix as the inverse of a singular Hessian. In 
my case i am just interested on the value of the log-likelihood. For my 
application, I patched it easily in a local version of mle just removing 
this call since I am not using vcov at all, but i wonder if it can be 
improved in the official release. I can imagine of two simple solutions, 
either including vcov calculation as an option or avoiding the call to 
solve if the hessian is singular (setting vcov to NA). I am willing to 
write a few lines of coded if you think it is worth.

regards

Francisco Matorras
Instituto de F?sica de Cantabria
Universidad de Cantabria

I don't know if this will get much response from the R developers;
they might just recommend that you protect your mle() call in a try() or
tryCatch() to stop it from breaking your loop.  Alternatively, you could
try mle2() function in the bbmle package, which started out long ago as
a slightly more flexible and robust version of stats4::mle(); I don't
remember/can't promise that it handles fits with singular Hessians, but
I'm guessing it does ...

  cheers
   Ben Bolker


On 2019-02-19 12:02 p.m., Francisco Matorras wrote:
> Hi, R developers.
> when running mle inside a loop I found a nasty behavior. From time to
> time, my model had a degenerate minimum and the loop just crashed. I
> tracked it down to "vcov <- if (length(coef))
solve(oout$hessian)" line,
> being the hessian singular.
> Note that the minimum reached was good, it just did not make sense to
> calculate the covariance matrix as the inverse of a singular Hessian. In
> my case i am just interested on the value of the log-likelihood. For my
> application, I patched it easily in a local version of mle just removing
> this call since I am not using vcov at all, but i wonder if it can be
> improved in the official release. I can imagine of two simple solutions,
> either including vcov calculation as an option or avoiding the call to
> solve if the hessian is singular (setting vcov to NA). I am willing to
> write a few lines of coded if you think it is worth.
> 
> regards
> 
> Francisco Matorras
> Instituto de F?sica de Cantabria
> Universidad de Cantabria
> 
> 
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>