Mike Lacy <[email protected]> asks about convergence problems in
-ologit- when bootstrapping its estimation results:
> I am doing some bootstrapping experiments using OLOGIT . I have
> deliberately chosen a skewed response variable (P = 0.1, 0.1, 0.1, 0.1,
> 0.1, 0.6) and 6 covariates. The covariates are all scaled to the same
> order of magnitude. Perhaps not surprisingly, at relatively smaller sample
> sizes, (e.g., N = 100) , OLOGIT occasionally (say 1/100 reps) fails to
> converge at any reasonable tolerance setting, and halts BSTRAP. It would be
> nice if I could prevent this by tweaking the estimation process, but it
> seems that the tolerance is the only thing I can change on the maximization
> routine underlying OLOGIT. I presume but don't know that the routine is
> simply reacting to a really degengerate distribution on the response
> variable because leaving out some of the covariates doesn't seem to help.
> If anyone has ideas for promoting convergence, that would be good, but I'm
> not hopeful.
> Presuming the convergence problem is intractable:
> I would like to find a way for BSTRAP to continue, simply ignore the
> occasional failure to converge, and to accumulate the results from the
> reps that do converge. Is there a way this can be done?
It appears that Mike is using Stata 7 or earlier, based on his comment that
-bstrap- stops when -ologit- fails to converge. If Mike is using Stata 7, I
would suggest using the -bs- command; it will post missing values if -ologit-
stops with an error. Besides that, it has a better syntax and doesn't require
the user to write a program. For example, suppose I wanted to bootstrap the
coefficients from an -ologit- fit, then my do-file would look something like
the following:
. version 7
. use auto, clear /* assuming auto.dta is in the current dir */
. set seed 1234
. bs "ologit rep mpg weight" "_b[mpg] _b[weight]", reps(100)
In Stata 8, -bs- and -bstrap- are synonyms for the new -bootstrap- command.
-bootstrap- has an even more flexible syntax (similar to the -jknife- command)
along with other new options. In Stata 8, my do-file would look like:
. version 8
. sysuse auto, clear
. set seed 1234
. bs "ologit rep mpg weight" _b, reps(100)
Where by -_b-, -bootstrap- (and thus -bs-) understand that to mean all the
elements of the coefficient vector -e(b)-.
If Mike is performing intermediate calculations after the -ologit- fit, he can
modify the program he wrote for the (Stata 7) -bstrap- command so that it
returns the results in -r()- (that is, Mike can turn his program into an
-rclass-). Then he can refer to these results in the list of statistics for
the -bs- command. Suppose I have a program called -mycalc-, and -mycalc-
places its result in -r(answer)-. Then my do-file for bootstrapping the
results of -mycalc- would look something like
. version 7
. use auto, clear
. set seed 1234
. bs "mycalc rep mpg weight" "r(answer)", reps(100)
or something similar for Stata 8.
For more information on -rclass- programs, see [P] return.
Also see [R] bstrap in the Stata 7 manual set, [R] bootstrap in the Stata 8
manual set.
--Jeff
[email protected]
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
