Hi Suveyda,
Unfortunately, comparing the density of the imputed and observed values isn't a surefire way to judge the quality of the imputations. In fact, it can be a good sign---the observations with missing data might be very different than those that are observed and Amelia is correcting for this problem. There is no need for the imputed values to follow the same distribution as the observed data. The compare density plots can be useful for detecting large problems, such as imputations that are radically outside the range of the observed data.
For overimputation, you should definitely check section 4.7.2 of the documentation, found here:
In overimputation plots, we are checking how the imputation model does for observations that are actually observed (thus, it is similar to cross-validation). The x-axis is the original, observed value and the y-axis are the imputed values when we treat the observed value as temporarily missing and impute it from the model. The vertical lines show the the 90% confidence intervals for those imputations. The colors of those lines indicate what percentage of the other variables are missing for that observation (with red being more missing and blue being less missing). Again, see the documentation for an example of this.
There's no need to perform one-by-one regressions on each of the imputed datasets. The combination rules will put them together and take care of the added uncertainty due to the imputations.
Hope that helps!
~~~~~~~~~~~
Matthew Blackwell
Institute for Quantitative Social Science
Department of Government
Harvard University
url: http://www.mattblackwell.org
On Tuesday, June 19, 2012 at 2:20 AM, Karakaya, Suveyda wrote:
Hi,
I have a question about imputation disgnodtics, the imputed and observed densities of polity 2 are quite different, first one with central density, second one with two hills. Also, I could not figure out how to interpret overimputation graph. There are vertical lines all around with different colors and I dont know which one denotes observed or imputed values. For 90 percent comfidence , should these vertical lines concentrate on the leftside of the graph? My vertical lines is scattered all around, is this a sign of trouble? One last question, since there are five imputations, should I try all of them in stata for robustness check or would using just one of them is fine? Thank you very much
Suveyda Karakaya
PhD Candidate
The University of Tennessee
--
Amelia mailing list served by HUIT
Amelia mailing list