RE: st: Re: Change in QIC output

["Nick Cox" <n.j.cox@durham.ac.uk>]

This thread started a week ago. I didn't see a response from the author
of -qic-, Jisheng Cui, who may not be a member of the list. I would
contact him directly. 

The code for -qic- carries notes on its revisions: 

*! Version 1.1.9, 19 September 2008
*! Author: Jisheng Cui, Deakin University
*! Calculate QIC and QIC_u proposed by Pan (2001)
*! Update 30Sep06: include if, in and weight
*! Update 10Oct06: include rclass
*! Update 18Jan07: change to scalar
*! Update 25Apr 2007: change scalar p to par
*! Update 27Nov 2007: extend to general negative binomial
*! Update 04Jan 2008: Parameter in negative binomial consistent with
Stata
*! Update 19Mar 2008: Restore to version 9.0 and add a trick for
nodisplay
*! Update 19Sep 2008: Remove constant terms for Poisson and negative
binomial

Nick 
n.j.cox@durham.ac.uk 

Pamela Phillips

Thanks for your prompt and helpful reply.

Looking back, I think it is likely that I downloaded the previous
version
(2006) of -qic- for my analyses in 2008.

Does anyone know of any settings that have changed between the 2006 and
the 2008 versions of -qic-? Is the previous 2006 version still available
anywhere?

Nick Cox 

> First off, there is (strictly) no -qic- command in Stata. -qic- is a
> user-written program from SSC.
>
> What's changed, presumably, is not -qic-: although your information
does
> not absolutely exclude the possibility that you used different
versions
> of -qic-, the version of -qic- on SSC is 1.1.9 9 Sept 2008 and is
likely
> to be the same.
>
> What's changed is the internal engine that Stata uses for
optimisation.
> You may need to tweak the fitting using the optimize options flagged
in
> the help.
>
> I guess you should be worried that your QIC values are, or appear to
be,
> not well determined, but I can't advise further.
>
> The more-or-less peaceful coexistence of various *IC and of so many
> other single-number figures of model merit reminds me of the old joke
> "The great thing about standards is that there are so many to choose
> from".

Pamela Phillips

> In October 2008 I fitted some Poisson generalised linear models using
> generalised estimating equations. I calculated the QIC
(quasilikelihood
> information criterion) for each model, using the qic command in STATA.
>
> On revisiting the same datasets and the same code in April 2010, I
find
> that the QIC values have changed by a factor of 1.5. The command
output
> is
> the same, except for the first line, which has changed from 'Iteration
> 1:
> tolerance = 1.166e-11' to 'Iteration 1: tolerance = 2.528e-12'.
>
> Can anyone tell me what has happened? Have any default settings been
> changed for the qic command? I am using STATA 10.0.

*
*   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/