Hi Nadine,
A few points.
1) You should expect that the convergence properties will change when adding priors since these change the optimization problem. Are the imputations converging to different places in the plot? That would be more of a problem.
2) The sensitivity is strange, but I wonder if it is due to the number of parameters the imputation model is trying to fit. Note that you have 133 variables plus a fixed effect for each country (as specified by your call to amelia), which means you are estimating roughly 6000 parameters in the model. You have added a ridge prior, which will help, but it might not be enough. The imputations could still be poorly behaved with this large of a dataset.
3) It's hard to say anything else without seeing your data at all. If you would like me to take a look at it, feel free to send me mail off-list, so that I can track down the problem.
Hope that helps.
Cheers,
matt.
On Wed, Aug 10, 2011 at 5:56 AM, Nadine Reibling
<nreiblin@mail.uni-mannheim.de> wrote:
Hello everyone,
I encountered a problem trying to introduce observation-level priors. I expected that the imputation for the specifed observations would be better (closer to my expectations, smaller variation). What indeed happens is three things:
1) The convergence process completely changes. The number of iterations goes down by at least 50% compared to not using the prior. I tried different priors specifiying information from 3 to 50 observations; the same thing always occurred. The imputation without the priors works fine and diagnostics indicate a good fit of the imputation model.
2) The imputations for a lot of other variables and observations not specified by the prior changed tremendously. Imputation diagnostics look a lot worse for many variables even if I specify only three variables.
3) Imputations for the observations for which priors are introduced change only slightly.
My code:
pr3 <- matrix(c(75, 76, 77, 25, 25, 25, 1, 1, 1, 2.29, 2.29, 2.28, 0.95, 0.95, 0.95), nrow=3, ncol=5)
pr3
a.out.new3 <- amelia(macimp, m=5, ts = "_j", cs = "countries", polytime=0, intercs=TRUE, p2s=2, empri=0.01*nrow(macimp), priors = pr3)
My data encompass 1127 observations and 133 variables (of which 48 are year dummies).
A non-related problem occurs when I use the imputation without priors and try to create the overdispersion plot with 5 imputations.
I get an error message:
A vector of this size cannot be allocated [translated from German].
When I use only 3 imputations, I get an overdispersion graph that looks fine.
Thank you for any suggestions how to solve these problems!
Nadine
--
Nadine Reibling
PhD candidate
Mannheim Centre for European Social Research
A5,6
68131 Mannheim
0049-621-181-3302
nreiblin@mail.uni-mannheim.de
Website: http://cdss.uni-mannheim.de/index.php?id=346