On Fri, 9 Oct 2020 at 12:28, Martin Maechler <maech...@stat.math.ethz.ch> wrote: > > >>>>> Steven Yen > >>>>> on Fri, 9 Oct 2020 05:39:48 +0800 writes: > > > Oh Hi Arne, You may recall we visited with this before. I > > do not believe the problem is algorithm specific. The > > algorithms I use the most often are BFGS and BHHH (or > > maxBFGS and maxBHHH). For simple econometric models such > > as probit, Tobit, and evening sample selection models, old > > and new versions of R work equally well (I write my own > > programs and do not use ones from AER or > > sampleSekection). For more complicated models the newer R > > would converge with not-so-nice gradients while R-3.0.3 > > would still do nicely (good gradient). I use numerical > > graduent of course. I wonder whether numerical gradient > > routine were revised at the time of transition from > > R-3.0.3 to newer. > > As R-core member, particularly interested in numerical accuracy > etc, I'm also interested in learning what's going on here. > > I think we (R core) have never heard of anything numerically deteriorating > going from R 3.0.x to R 4.0.x, and now you are claiming that in > public, you should really post *reproducible* code giving > evidence to your claim. > > As was mentioned earlier, the difference may not be in R, but > rather in the versions of the (non-base R, but "extension") R > packages you use; and you were saying earlier you will check > that (using the old version of the 'maxLik' package with a newer > version of R and vice verso) and tell us about it. > > Thank you in advance on being careful and rational about such > findings.
I totally agree with Martin: It would be good if Steven could run his code with different combinations of versions of R and maxLik so that we know whether the problem is caused by newer versions of R or by newer versions of maxLik (and hopefully also which version introduced the problem). Steven wrote that the problem appeared both in maxBFGS() and in maxBHHH(). This is somewhat surprising because maxBFGS() uses for the optimisation optim() which is implemented in R's base package "stats", while maxBHHH() uses for the optimisation maxNR() which is implemented in the maxLik package. Hence, it would be a very unlikely coincidence if the optimisation routines in optim() and maxNR() would become worse at the same time. I remember that we slightly changed the user interface and the default values of some arguments of some function in the maxLik package (e.g., maxBFGS, maxBHHH, maxNR, ...) a long time ago. I suggest that Steven checks whether he needs to update his code to reflect the changes in the user interface and/or in the default values of some arguments. Best wishes, Arne -- Arne Henningsen http://www.arne-henningsen.name ______________________________________________ R-help@r-project.org mailing list -- To UNSUBSCRIBE and more, see https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide http://www.R-project.org/posting-guide.html and provide commented, minimal, self-contained, reproducible code.