I'm new to the listserv and unfortunately don't have access to the
northwestern archive for this listserv as it appears to require login/pw, so
I apologize if this question has been asked before.

I'm performing multiple imputation on a rather large dataset (~450,000
obs) with about 9% overall missingness. I was getting "failure to converge"
errors initially. This resolved when I removed some variables. Oddly enough
though, I found a mistake in my coding of one variable and reworked the
entire dataset, and now no longer get the error at all.

I believe my original error was related to very unbalanced
binary/categorical variables (<2% in a category), partly because of the
reference white paper by Gullion, Chen, Meltesen (Applied aspects of using
multiple imputation in clinical data). It seemed that removal of these types
of variables eliminated the error message, but it was unclear why one
variable vs another that both had sparse categories (or the dummy variable
set equivalent).

Are there other plausible explanations for receiving this error in proc MI?
and are there common methods to troubleshoot such an error when it does
occur?

thank you for any advice you can provide,
- mike

--------------------------
Michael Liao, MD
Clinical Research Fellow
Department of Emergency Medicine
Denver Health Medical Center

Reply via email to