Matt,
Thanks for the useful ideas. By increasing the ridge prior I was able
to get Amelia to produce imputations (though I haven't analyzed them
sufficiently, yet, to decide if I believe them.) The preliminary
listwise deletion approach is probably not viable in my case because
there are only a small handful of states for which there is no missing
data. Ultimately, the Bayesian approach may prove the best choice,
but it'll take me some time to figure it all out.
Thanks again for the quick and helpful response.
Sincerely,
Ben
On Thu, Jan 26, 2012 at 8:39 AM, Matt Blackwell
<blackwel(a)fas.harvard.edu> wrote:
Hi Ben,
Generally, the model is not identified if there are more variables than
observations, which means imputations will be poorly behaved. You could run
this data through Amelia with a ridge prior (that is, setting the "empri"
option to a relatively high number). This would help identify the model by
shrinking the covariances toward zero.
Another option would be to run a preliminary factor analysis on the listwise
deleted data and find the subset of important variables (that is, those with
high factor loadings) and then use those variables in the imputation.
Last, you could eschew Amelia completely and implement a Bayesian analysis
with relatively strong priors to identify the model. In this option, you
could handle the factor analysis and the missing data together.
Hope that helps.
Cheers,
matt.
~~~~~~~~~~~
Matthew Blackwell
PhD Candidate
Institute for Quantitative Social Science
Department of Government
Harvard University
url:
http://www.mattblackwell.org
On Thursday, January 26, 2012 at 11:26 AM, Ben Highton wrote:
Hi Dan,
Thanks for your response. In this case, though, I don't think the
solution will work because the I don't have repeated measurements on
the same variables through time. The data is fundamentally
cross-sectional, not time-series cross-sectional.
Best,
Ben
On Wed, Jan 25, 2012 at 11:51 AM, Dan Matisoff <dmatisof(a)umail.iu.edu>
wrote:
You can longitudinalize the data... so for each year, you'll get 51 more
observations (which will also help generate improved imputations)
On Jan 25, 2012, at 2:30 PM, Ben Highton wrote:
Hi fellow Amelia users,
I have a dataset with 51 observations (one for each American state +
the District of Columbia) and nearly 150 variables. Is there a way to
do imputations when variables>observations? (One of my analysis goals
is to factor analyze the variables to attempt and identify what I
believe to be the 2-4 dimensions in the data.)
Because I primarily use Stata, I am using AmeliaView. An initial
attempt to impute missing values produced the following error:
"Amelia Error Code: 34 The number of observations in too low to
estimate the number of parameters. You can either remove some
variables, reduce the order of the time polynomial, or increase the
empirical prior."
If anyone has advice for me, I'd appreciate it. Thanks.
Sincerely,
Ben Highton
--
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
https://lists.gking.harvard.edu/mailman/listinfo/amelia
--
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
https://lists.gking.harvard.edu/mailman/listinfo/amelia
--
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
https://lists.gking.harvard.edu/mailman/listinfo/amelia