st: RE: -nlcom- issues

["Feiveson, Alan H. (JSC-SK311)" <alan.h.feiveson@nasa.gov>]

Hi Stas - The previous time issue you were referencing is in 

http://www.stata.com/gsearch.php?q=increasing+time+burden&site=stata&client=stata&proxystylesheet=stata&output=xml_no_dtd


Numeric issue - Just curious - can you make this happen with nl testing after a "standard" estimation command/ I would like to see an example.

AL


-----Original Message-----
From: owner-statalist@hsphsun2.harvard.edu [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Stas Kolenikov
Sent: Monday, May 17, 2010 12:34 PM
To: statalist@hsphsun2.harvard.edu
Subject: st: -nlcom- issues

I have a couple of issues with -nlcom- that seem odd. To be specific
(but I don't think that matters), I am estimating percentiles of an
income distribution in two years of data (by a custom eclass command),
and wrap them into a growth model (by -nlcom- -- I need a ratio of say
25th percentiles in two waves of data, pro-rated to be on the per
annum scale).

1. Time issue: each next -nlcom- takes longer to run than the previous
one, although at the face of it they should run at the same speed. I
think the issue came up a few months ago on the list, and was traced
down to how -testnl- (on which -nlcom- relies in computation of the
standard errors) handles memory. In the original post, it even crashed
Stata in some simulations when it used all available handles to
temporary objects (or something of that nature). I was under
impression that this oddity was fixed, but apparently it was not.

2. Numeric issue: the standard errors reported by -nlcom- tend to go
down with the number of parameters. That is, I estimate say
percentiles in numlist 10(10)90, and percentiles in numlist 2(2)98,
and I form my growth in the percentile estimates from these two sets
of estimates. While the estimates and their standard errors are
identical between the two lists when they intersect (i.e., for the
percentiles in the first list), the -nlcom- results match only in
point estimates. The standard errors for these identical point
estimates are smaller in the second case. I would say, disturbingly
smaller, by about 20-25%.

I'd be happy to share the code (and probably the data, too) with
tech-support, if needed.

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/