Re: st: Kenward-Roger method in Stata?

[Maarten buis <maartenbuis@yahoo.co.uk>]

--- On Fri, 4/6/10, Seed, Paul wrote:
> SAS's command PROC MIXED has an option " ddfm=KR " 
> which implements Kenward & Roger (1997), 
> improving on the asymptotic estimation of 
> SE, CI significance etc.  There is some evidence that
> it performs well with missing data and small samples 
> 
> Has anyone attempted to implement this in Stata, for 
> either -xtgee- or -xtmixed-?  

Not that I know, but I would start with trying to figure
out whether you would actually need it. One way of 
figureing out whether this is true for your dataset is
simulation:

1) estimate your model of interest, for the purpose of this
   simulation this will be the estimates in the "populaiton".
2) use -bsample- to draw a sample from your data
3) estimate your model of interest
4) test whether the parameter of interest is equal to the 
   "population"-parameter, and return the p-value.
5) repeat steps 2-4 many times (use the -simulate- command
   for that.)

If the standard standard errors work wel for your data, then
these p-value should follow a uniform distribution.  If I am
allowed some shameless self-promotion, I would say that I like
-hangroot-  for that. (see: -ssc d hangroot- and the 7th graph 
on <http://www.maartenbuis.nl/software/hangroot.html>). You can
also look at the coverage: if 5% of the samples have a p-value
less than .05. As the information necesary for computing this
coverage is contained in only 5% of the samples, you need a 
great many of those to get a reliable estimate, say 10,000, 
which means you would expect around 500 rejections.

Hope this helps,
Maarten

--------------------------
Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen
Germany

http://www.maartenbuis.nl
--------------------------


      

*
*   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/