If your covariates don't predict well, the key is that the imputations
have enough variability to represent that fact. Amelia will normally do
this right. It may be in your application that you wind up with very
large standard errors and confidence intervals, but they would in that
case be accurate. If that's the case, you could try finding better
variables (perhaps coded as a function of the existing ones) or finding
better data. Not much else you can do...
Gary
p.s. the list address is Amelia Listserv <amelia(a)lists.gking.harvard.edu>
not owner-amelia_at_lists_gking_harvard_edu(a)mail.hmdc.harvard.edu
On Fri, 9 Feb 2007, Christopher Parker wrote:
To whom it may concern,
I wrote a while back to ask about using MI, much in the same way that
Scheve (2006) did in recent SAPD paper in which he imputed data for a
variable that was not observed in the dataset of interest.
Here's the question. I'm attempting to expmine the effects of mental
health on politics. It's a revisitation of some of Lasswell's work. As you
know, there's nothing in the way of data on this. So, I'm trying to use MI
as a temporary solution. I plan to use data from a national survey on
mental health, the National Comorbidity Study, to fill in the
"missingness" within the NES for depression, anxiety, and PTSD.
The problem is that there are only 30 variables, all of which are
demographic in some way, that overlap between the two datasets and can be
used as information. Moreover, when I tried to predict depression,
anxiety, and PTSD in the NCS, the r-squares average .05. Perhaps this has
something to do with the fact that the NCS has 9K observations and over 1K
variables. That said, I'm concerned that once imputed into the NES the
mental health items will be too noisy to predict anything. Is this
something about which I need to be concerned, i.e., the low explained
variances?
Any advice is welcomed.
Best,
Christopher
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive:
http://lists.gking.harvard.edu/?info=amelia