Re: [S] discrepencies found using GEE between Splus and STATA

Prof Brian Ripley (ripley@stats.ox.ac.uk)
Tue, 7 Apr 1998 07:42:45 +0100 (BST)


I am posting a reply, as this is yet another enquiry about a third-party
package posted to S-news rather than referred to the author. Please
note that this list's introduction asks that such things be raised
first with the package's author. In particular, this is one of several
packages for which the version on statlib is not current, so the author
may be able to tell you about existing changes and enhancements.

Stefan Ma wrote:
>
> Dear S+ friends,
>
> I am trying to use GEE in Splus for my data, however,
> I firstly use the sample data (testgee) to practise
> the programming before implmenting in my real data.

There is no `GEE in Splus'. I think you are referring to a library
written by Vincent Carey, an obselete version (with known errors)
of which is available from statlib. You need to get a current
version from his web site http://biosun1.harvard.edu/~carey/gee.html.
This was version 4.13 last time I looked.

> Eventually, I also use the sample data in STATA software,
> but I found there are discrpancies of estimates and both
> naive and robust SE from the results between
> two software even quite they are close up to 6 decmial places.

It is not clear to me that GEE is a sufficiently well-defined
procedure that I would expect two packages to give exactly the same answer,
but if they are giving similar answers this may just be the effect of
stopping numerical computations too early for complete agreement
(but late enough for statistical purposes). For example, it is
well-known that parameter estimates given by glm() can be quite a long
way off the mles unless the convergence tolerance is tightened.

How many `decmial places' of agreement is usually irrelevant; what is
important is the number of significant figures (for SEs) and proportion
of an SE for parameter estimates.

> In addition, I further check the GEE in Splus, I suppose
> using " corstr="independence" " in GEE, the Naive S.E. in
> GEE output are same as the S.E. in GLM output, unfortunately
> they are not same in Splus output, but they agreeed in STATA.

Which is `right'? Please check out the definitions used, which may differ.

> PS: I am using 1) GEE ver 4.3 and S-Plus ver 3.4 in Solaris 2.5.1

That was useful information.

> 2) STATA in win95

so I presume STATA has no version number?

-- 
Brian D. Ripley,                  ripley@stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272860 (secr)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595
-----------------------------------------------------------------------
This message was distributed by s-news@wubios.wustl.edu.  To unsubscribe
send e-mail to s-news-request@wubios.wustl.edu with the BODY of the
message:  unsubscribe s-news