Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: RE: Goodness of fit and model comparison in xtgee


From   "Nick Cox" <n.j.cox@durham.ac.uk>
To   <statalist@hsphsun2.harvard.edu>
Subject   st: RE: Goodness of fit and model comparison in xtgee
Date   Thu, 20 Sep 2007 19:28:23 +0100

This raises questions on two levels. 

1. Residuals from -xtgee- 
=========================

Seyi is referring to an email 

http://www.stata.com/statalist/archive/2005-12/msg00505.html

This chunk still seems to be true: 

"It seems that the general comments on -predict- 
after -xtgee- are an exaggeration in this case, 
as no specific option is available that 
produces any flavour of residuals. 

No matter, you can still use -predict- 
to generate predictions and (e.g.) plot those
against observed. Use the GIC: the 
graphical investigation criterion. Or 
calculate your own residuals from observed 
and predicted. 

That might tell you more than any other ?IC
purporting to encapsulate all information on
badness of fit in a portmanteau statistic. 

This is all supposing that there isn't 
some well-known argument showing that
with -xtgee- residuals are a snare, 
a delusion and generally poor citizens."

-xtgee- is a very general and complex command: perhaps 
the programmers got distracted or too close
to their deadline to include options for
various flavours of residuals. It happens. 

No matter; you can calculate your own as 
suggested above. 

2. Categorisation of continuous predictor
=========================================

(I dislike the terms dependent and independent 
variables intensely.) 

There is quite a large literature arguing 
that this is a very bad idea even given 
some subject-matter justification. This 
has been much discussed in medical statistics. 
This argument was one of the main justifications 
for fractional polynomials. 

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

Soremekun, Seyi
 
> I have panel data with a normal continuous dependant variable, and I'm
> using the xtgee command to fit it. I'm trying to categorise one of my
> independent variables (which is also continuous) partly 
> because of lack
> of data points but also because it may be biologically sensible to do
> so. At the moment, I'm varying the cut-off points for my 
> categories and
> then running the models. 
> 
> I would like to find which cut-off points give me the best fitting
> model, but I'm having trouble finding goodness-of-fit tests that are
> compatible with xtgee. Are there any that stata can run? I 
> know there is
> a stata book available on GEE which might mention it but I don't have
> access to it.
> 
> I saw a similar query on this a few years ago which Nick Cox answered
> ("residuals for gee") and I wondered if there had been any 
> advance since
> then. At the moment I'm just plotting the model residuals and also
> looking at their sizes to compare them.
 

*
*   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/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index