Hi Mark,
You definitely want to use 1.7.1 for work with priors. One slight laborious
approach to detecting issues is to run Amelia with a small subet of
covariates, then increase that subet slowly to find the culprit.
Having said that, a correlation of 0.92 is quite high. You might want to
try to omit one of those variables first.
Cheers,
Matt.
On Friday, April 5, 2013, Mark Seeto wrote:
Thanks for your reply Matt (and thank you to you and
your colleagues
for providing this great software!).
If I ignore my two nominal variables, the highest correlation is 0.92
(and I'm confident that there's no correlation problem with the
nominal variables). I tried using findLinearCombos, but it appears not
to work with data having missing values.
I'm happy to just use version 1.6.1, but I'm looking to use the
"priors" argument for the first time. I note that the package news
page says that in version 1.7 a bug with priors was fixed - how
serious was that bug, and does it mean that using priors with version
1.6.1 would be a bad idea?
Thanks,
Mark
On Fri, Apr 5, 2013 at 11:26 PM, Matt Blackwell
<m.blackwell(a)rochester.edu <javascript:;>> wrote:
Hi Mark,
We changed the infrastructure of Amelia in 1.7 to improve the speed
considerably, but there may be consequences in terms of tolerance for
highly
correlated data. You might want to check to see
if there are variables in
the data with correlation above .95 or use the "findLinearCombos"
function
from the "caret" package to see if
there are linear combinations hidden
somewhere in the data. Unfortunately, a lot has changed from 1.6.1 to
1.7.1,
so it's hard to exactly pin down your
issues, but it seems very likely
due
to the structure of the data.
Hope that helps!
Cheers,
matt.
~~~~~~~~~~~
Matthew Blackwell
Assistant Professor of Political Science
University of Rochester
url:
http://www.mattblackwell.org
On Fri, Apr 5, 2013 at 12:27 AM, Mark Seeto
<markseeto@gmail.com<javascript:;>>
wrote:
Dear Amelia mailing list,
I have some code that runs in Amelia 1.6.1 (with R 2.15.0) without any
problems. When I run that same code in Amelia 1.7.1 (with R 3.0.0), a
few imputations might finish without any error messages, but many
finish with the message "The resulting variance matrix was not
invertible. Please check your data for highly collinear variables".
I understand that this suggests that there may be a problem with the
data, but I don't understand why there would be a problem in version
1.7.1 but not in 1.6.1.
Also, when I'm using Amelia 1.7.1 in R 3.0.0, R will sometimes crash
during the imputation process. I'm running R in emacs in Windows, and
I'll sometimes get a pop-up window saying "R for Windows terminal
front-end has stopped working", with the following message appearing
in emacs:
"error: chol(): failed to converge
This application has requested the Runtime to terminate it in an unusual
way.
Please contact the application's support team for more information.
terminate called after throwing an instance of 'std::runtime_error'
what():
Process R exited abnormally with code 3 at Fri Apr 05 15:04:58 2013"
Amelia 1.7.1 worked without any problems when I tried it on some small
made-up data, but not on my real data. I understand that it's almost
impossible to diagnose problems without a reproducible example, but I
was wondering whether others have had similar problems.
Thanks for any comments you can give.
Mark Seeto
--
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 <javascript:;>
To unsubscribe from this list or get other information:
https://lists.gking.harvard.edu/mailman/listinfo/amelia