Hi Matthew,
The imputations you get will be based on the observed data you do have. So,
what Amelia would do is impute the earlier data using the relationships
between the always observed variables and the new DoJ items from the later
part of the data. This is perfectly valid as long as the always observed
variables are good predictors of the DoJ items in both time periods. It's
much less important that there are other variables that are also missing in
the earlier time period. This might increase the uncertainty of the
imputations, but this would be reflected in the variance of the imputed
values between imputations.
Here's a link to the vignette/manual:
http://cran.r-project.org/web/packages/Amelia/vignettes/amelia.pdf
You can also see the second and third pages of the 2001 APSR article:
http://gking.harvard.edu/files/gking/files/evil.pdf
Hope that helps!
Cheers,
Matt
On Thu, Nov 14, 2013 at 4:09 PM, DeMichele, Matthew <mdemichele(a)rti.org>wrote;wrote:
Matt:
I have a follow-up question. The situation is I’m pooling probation data
across states and years (1980 to 2011) in the early years there were only a
handful of questions, and in the mid-1990s the government (Department of
Justice) included several new items. So, for instance, they started to ask
about the number of people on probation for DWI. I’m thinking the answer is
no. By the way, I’m not seeing the vignettes – did you mean page 4 of the
Amelia user guide or one of the related publications? I checked both, but
must be missing it.
Thanks for your help.
Matthew
*From:* Matt Blackwell [mailto:m.blackwell@rochester.edu]
*Sent:* Thursday, November 14, 2013 3:55 PM
*To:* DeMichele, Matthew
*Cc:* amelia(a)lists.gking.harvard.edu
*Subject:* Re: [amelia] amelia help
Hi Matthew,
It is definitely acceptable to impute earlier questions as long as they
overlap with other questions that are asked throughout the entire dataset
and that those observed values predict the missing values. The key idea to
consider is whether or not that data is missing at random (MAR) given you
set of observed variables (see pg 4 of our vignette/documentation). If it
is, then you can and should impute those observations.
Hope that helps!
Cheers,
Matt
~~~~~~~~~~~
Matthew Blackwell
Assistant Professor of Political Science
University of Rochester
url:
http://www.mattblackwell.org[mattblackwell.org]<https://urldefense.proof…
On Thu, Nov 14, 2013 at 3:20 PM, DeMichele, Matthew <mdemichele(a)rti.org>
wrote:
I have a question about the appropriateness of using Amelia to impute some
data for items that were not asked in earlier years of a survey. I’m
working with a governmental survey collection that started in 1975 and
continues up to the present day. The data are in long form arranged by
state and year. My question is that certain items were not asked in the
earlier years that are asked now. Is appropriate to impute the data for
these earlier years? Initially, I planned to develop multiple datasets to
maximize years and states, but I started thinking this may be an imputation
issue as well.
Best regards,
Matthew
--
Amelia mailing list served by HUIT
[Un]Subscribe/View Archive:
https://urldefense.proofpoint.com/v1/url?u=http://lists.gking.harvard.edu/?…
More info about Amelia:
https://urldefense.proofpoint.com/v1/url?u=http://gking.harvard.edu/amelia&…
Amelia mailing list
Amelia(a)lists.gking.harvard.edu
To unsubscribe from this list or get other information:
https://urldefense.proofpoint.com/v1/url?u=https://lists.gking.harvard.edu/…