Re: st: combining chi square and p values of multiple imputed data

[Stas Kolenikov <skolenik@gmail.com>]

On Fri, Jun 18, 2010 at 4:29 PM,  <jl591164@albany.edu> wrote:
> My data is multiple imputed and I need to test whether two dummy variables
> are associated. I used -mim: tabulate y x-. The results give chi square,
> and p value of each imputed dataset. How I can obtain the combined test of
> the significace across imputed datasets by stata command or by calculating
> if i have the formula? I read that it is not o.k. to average the test
> statistics across imputed dataset. Thanks in advance.

That is one of those areas where you think there is an obvious answer,
and surprised not to find it in Stata. If there were, it would've been
implemented; and the reason it is not implemented is that it is not
possible to come up with a meaningful way to combine them. You can
combine the data sets for point and variance estimation purposes, but
that's all. If something like -testnl- is supported after -mim-, then
you can combine the cell probabilities into a Wald test of
independence. That would be quite an exercise of manipulating the
parameter names though. In a 2x2 example with cell probabilities p11,
p12, p21 and p22, the margins are (p11+p12) and (p21+p22) for one
dimension, and (p11+p21) and (p12+p22) for the other. So the test for
independence will be: H_0: p11 = (p11+p12)*(p11+p21). You can
substitute any other cell, I believe they will all be equivalent, as
it would result in a 1df test whichever way. In larger tables, you
will have longer expressions for the margins, and more of these
equation you could test.

It would be easier for you to run

mim: logit dummy1 dummy2

and get your p-value from Wald test there.

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   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/