Hello,
I am new to using Amelia. I have two questions. First, when I run
the overdispersed start values diagnostic, it converges to two
different modes. Adding priors didn't resolve the problem, so I
deleted some variables that were highly correlated and it converged to
a single mode. I was hesitant to remove the variables since I want to
include as many variables in the imputation model as possible. Is
there another way of managing this kind of situation? Unfortunately,
I have a high rate of missingness for many variables and a small
sample size of 208 countries. Originally, I had 31 variables, but
after deleting 6 highly correlated variables, I now have 25 variables,
3 of which are completely observed. I realize that this means I have
more parameters than observations, but I don't know what else to do
when I can't add any more countries to the sample. With a ridge prior
of 1, the imputations seemed to run well, however.
Second, with my smaller data set that appears to run well, as soon as
I begin running the overdispersed starting values diagnostic, the
first imputed data set (outdata1.csv) disappears from the folder
leaving only data sets 2-10. Is this an error or should it be deleted
for purposes of a burn-in? If this is not an error, should I run 11
imputations to get 10, not using the first data set? Alternatively,
should I make a copy of the data sets before running the diagnostics
and use the first imputed data set? This has not happened in previous
imputation models that I have run, so I have re-run the same model
three times and the first dataset appears after running the
imputations, but is disappears once I begin running diagnostics. The
overdipersed start values diagnostic graph shows that the values do
not converge until around 1,000 imputations. I increased the ridge
prior to 2 and the same thing happened.
I appreciate your response.
Erin Saunders
Portland State University
-
Amelia mailing list served by Harvard-MIT Data Center
[Un]Subscribe/View Archive:
http://lists.gking.harvard.edu/?info=amelia