[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
Re: st: possible bug in Nlogit
Bo Cutter <email@example.com> asks about differences in -nlogit- results
between a fully up-to-date Stata and one just off the installation CD:
> I have Stata 8 (SE) on the desktop and laptop. The desktop version is
> recently updated (update all) including swapping the executable. The
> notebook stata 8 is just what came off the CD. I ran the exact same
> two-level Nlogit (7 nests, 43 choice) on both machines, with the
> maximization options d1 and difficult. On the laptop it converges
> nicely to a LLF=-1810. On the desktop there are no warning messages but
> it converges to LLF=-1811 and several standard errors are missing. The
> files are exactly the same on each machine (I copied the files from
> laptop to desktop and ran them without any changes to make sure). Can
> anybody explain why there is this difference and tell me whether the
> convergence results on the laptop are trustworthy?
Prior to version 8.1, -nlogit- used the Newton-Raphson optimization algorithm
("technique") when supplied with the -d1- option.
Now -nlogit- uses the BFGS algorithm (Broyden, Fletcher, Goldfarb, and Shanno)
by default (with or without the -d1- option). It appears that -nlogit- is
having difficulty with BFGS and Bo's data.
Bo can use the -technique(nr)- option to get -nlogit- to use the
Newton-Raphson algorithm instead of BFGS.
P.S. The -d1- option of -nlogit- has no effect if your Stata is up-to-date,
it has be superseded by the new -technique()- option. It's only effective
under version control.
* For searches and help try: