Hi Alex,
The main reason why Amelia was able to converge with this setup is the
ridge prior set by the "empri" argument. Essentially, this will smooth the
estimates from Amelia toward a multivariate normal with 0 covariance
between the variables. Thus, in your case, the model is identified from the
prior exclusively. Thus, this type of imputation is probably unreliable. It
makes more sense to remove the indicator variables from the imputation
model by including them in the "idvars" argument.
Hope that helps!
Cheers,
Matt
~~~~~~~~~~~
Matthew Blackwell
Assistant Professor of Political Science
University of Rochester
url:
http://www.mattblackwell.org
On Tue, Sep 24, 2013 at 4:14 PM, Alex Sutherland <as2140(a)cam.ac.uk> wrote:
Dear all,
I have a couple of follow-up questions concerning the treatment of
(obviously) collinear variables in Amelia II. Consider a dummy variable
with missing values (y), an indicator variable indicating the missingness
of y (also a dummy, say, i), and another indicator indicating the
missingness of y but with reversed coding to i, say j. So whenever y is
missing i=0 and j=1, when not, i=1 and j=0.
A dataset with such a setting (and many other variables) has been
successfully imputed with Amelia (m=21). Chain lengths vary from about 240
to 275, convergence has been reported "normal"; with empri =
.03*nrow(data) (where "data" is the raw data set).
From the various Amelia II discussion list responses I've seen (and
responses I've got from there since asking about this), I am surprised that
we did not get errors (singularity or similar). Given that we were
successful with the imputation, my questions are:
(1) are the results "reliable"?
(2) Do we need to be concerned about the quality of the imputation?
Regards,
--------------------------------------------
Dr. Alex Sutherland
Social Sciences' Research Methods Centre / Institute of Criminology,
University of Cambridge
as2140(a)cam.ac.uk
http://www.crim.cam.ac.uk/people/academic_research/alex_sutherland/
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive:
http://lists.gking.harvard.edu/?info=amelia
More info about Amelia:
http://gking.harvard.edu/amelia
Amelia mailing list
Amelia(a)lists.gking.harvard.edu
To unsubscribe from this list or get other information:
https://lists.gking.harvard.edu/mailman/listinfo/amelia