Hi Matt
Just to let you and the others know: we now did a correlation matrix and
excluded variables involved in high correlations. Still we got the same
error. So maybe the threshold of 0.95 was too high. Anyway: we've now
also set the prior to 1%. Now the imputations seem to run very good and
at the same speed of chain increments (single imputations on multiple
processors).
Thanks again for your help!
Best,
Philippe
On 24.09.12 15:00, Philippe Sulger wrote:
Hi Matt
Thanks for this. Yes, that's what I've also been thinking of to do
(both corr-analysis and ridger prior).
Now, as mentioned, some imputations work fine. What do you think about
the "stability"/"instability" of the chain lengths across these
(successful imputations)? They vary from 250-550. Every single
imputation reports me "normal EM convergence". Would you just run
imputations until having enough imputed datasets, that is, ignore the
failures and just reset them up? Or do we have to be really worried
about these collinearity (and stability) problems a priori? I ask that
because a ridger prior could also involve some "costs" by adding some
bias.
Thanks for your help!
Best,
Philippe
On 24.09.12 14:31, Matt Blackwell wrote:
Hi Philippe,
This is almost certainly the same issue with collinearity causing
problems. Most of the time it is not nominals, but rather a series of
variables that are linear transformations of each other (or quite
close to being so). I would take a look at a correlation matrix of
your dataset and see if there are any extremely high correlations
(0.95 or greater) and see if removing those helps solve the problem.
You might also try to add a ridge prior to the data by setting the
=empri= argument in the =amelia()= function (see section 4.6.1 in the
Amelia manual).
Cheers,
matt.
~~~~~~~~~~~
Matthew Blackwell
Assistant Professor of Political Science
University of Rochester
url:
http://www.mattblackwell.org
On Fri, Sep 21, 2012 at 5:46 AM, Philippe Sulger
<philippe.sulger(a)econ.uzh.ch <mailto:philippe.sulger@econ.uzh.ch>> wrote:
Dear Amelia users
We are running impuations. In the model we are currently running
we intentionally avoid to use ordinal-specifications. We have
some specifications for nominal variables. We have successfully
done multiple imputations with the same dataset specifying many
ordinal variables and some more nominal variables.
However, now we are running into the problem that for some reason
we get for many single-imputation runs the error mentioned in the
header of this email. Some runs seem to work fine (up to now). From
https://lists.gking.harvard.edu/pipermail/amelia/2011-February/000547.html
I see that this problem might be related to the specification of
the nominals and collinearity among them in their "decomposed"
nature. Note that we have "task 1 failed" and not "La.svd(x, nu,
nv)" (followed by "error code 1..."). If this is still related to
near-singularity problems, I have difficulties to understand why
this happens now in a model where we have no ordinals and fewer
nominals.
Do you have a suggestion?
Best,
Philippe
--
===============================================================
Philippe Sulger
Visiting Scholar
Institute of Criminology
Sidgwick Avenue
University of Cambridge
CB3 9DA Cambridge, UK
===============================================================
Email:philippe.sulger@econ.uzh.ch <mailto:philippe.sulger@econ.uzh.ch>
ps596(a)cam.ac.uk <mailto:ps596@cam.ac.uk>
Web:http://www.econ.uzh.ch/faculty/sulger.html
http://www.crim.cam.ac.uk/people/visitors/philippe_sulger
===============================================================
--
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
<mailto:Amelia@lists.gking.harvard.edu>
https://lists.gking.harvard.edu/mailman/listinfo/amelia
--
===============================================================
Philippe Sulger
Visiting Scholar
Institute of Criminology
Sidgwick Avenue
University of Cambridge
CB3 9DA Cambridge, UK
===============================================================
Email:philippe.sulger@econ.uzh.ch
ps596(a)cam.ac.uk
Web:http://www.econ.uzh.ch/faculty/sulger.html
http://www.crim.cam.ac.uk/people/visitors/philippe_sulger
===============================================================
One has a right to take an action if and only if one does not have a duty to refrain from
it.
--- Ken Binmore
--
===============================================================
Philippe Sulger
Visiting Scholar
Institute of Criminology
Sidgwick Avenue
University of Cambridge
CB3 9DA Cambridge, UK
===============================================================
Email: philippe.sulger(a)econ.uzh.ch
ps596(a)cam.ac.uk
Web:
===============================================================
One has a right to take an action if and only if one does not have a duty to refrain from
it.
--- Ken Binmore