Re: st: Re: micombine and overall model fit statistics

["Rodrigo A. Alfaro" <raalfaroa@gmail.com>]

> --- "Rodrigo A. Alfaro" <raalfaroa@gmail.com> wrote:
> > Usually, when the method works the R2's for each imputation are not
> > too much different. I suggest you to report the average writing a
> > small note about the dispersion of the all R2's that you get... maybe
> > something like that in the table "R2  = 0.45" and in the note 
> > "Average R2 obtained from 5 imputed datasets. The min R2 obtained was
> > 0.42 and the max 0.47". 
> "Maarten buis" <maartenbuis@yahoo.co.uk> wrote
> That's fine, even better would be if you computed the log of each R2,
> than compute the mean, and than transform the mean back (with the -exp-
> function) to the R2 metric. You can find the rational for this in:
> http://www.stata.com/statalist/archive/2007-04/msg00900.html .
I read the Maarten's link to Rubin reply about the topic. It is a very 
interesting solution to compute the geometric average of R2 instead 
of arithmetic one. However, this should be used paying attention on
was it behind whatever average you take

For example with m=5 and R2 = {0.10 0.10 0.90 0.90 and 0.90} 
the averages are 0.58 (arithmetic) and 0.374 (geometric), but for 
R2= {0.35 0.36 0.38 0.38 and 0.4} the averages are 0.374
(arithmetic and geometric). In terms of geometric averages both
experiments have the same measure, but putting more attention it
seems that something weird is behind the imputed datasets!! It will
be clear for the reader and yourself f you report that min/max (in
this extreme case you may consider to rewrite the model). However
reporting the geometric average only will show that both experiments
have similar fit, even more if you put all these into a automatic code 
probably you will never see this!!

Rodrigo.




*
*   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/