|
This is just a guess, but it sounds to me like you might being using the draws from the Gibbs sampler for reported observations as well as missing observations. Assuming you have 5 multiple imputations, you want to simply make 5 copies of the reported data for the variable on a responding case. With this procedure, the post-imputation variance estimate will equal the complete-data variance estimate and so the fraction of missing information will be 0.
--Dave Judkins
-----Original Message-----
Dear imputers I'm a Ph.D student and working on missing data, multiple imputation...I'm a little bit confused and have some questions :
1. When I estimate the fraction of missing information from a multiple imputation with Gibbs sampling , I obtained always higher the fraction of missing information than multiple imputation with stochastic EM. Is this normal? Do you have similar results?
2. Let's assume that I have data contains some complete and some incomplete variables, and I want to estimate the fraction of missing information. I expect that the fraction of missing information for complete variables should be 0. Is this idea wrong? Especially, if we impute data with Gibbs, they are not equal to zero..
3. Let's say I want to estimate the fraction of missing information. I have two options to impute data: First, I can estimate parameters from Gibbs complete data(i.e. after each draw of Ymissing, I have a complete data) In this case, I obtained 0 fraction of missing information for variables that I have complete data. (Is this improper multiple imputations?). Second, I can use parameter draws from Gibbs(of course, after convergence) and estimate the fraction of missing information(I think this os proper imputation??). In this situation, I don't have 0 fraction of missing information for variables which are complete. Which method is correct?
4. I have two different designs(missing by design) for the same data set and I want to compare these two different designs(i.e. different missing data patterns) using the fraction of missing information of parameters. Does the fraction of missing information show only missing information after imputation?Let's say if the imputation works very well for both designs, then shall we expect the fraction of missing information be the same amount for both designs? Do you suggest me any other methods(statistics) to show which designs contain more information before imputation?
I hope these are not stupid questions and I can get some reply. Thanks in advance for any help. Feray Adiguzel |
_______________________________________________ Impute mailing list [EMAIL PROTECTED] http://lists.utsouthwestern.edu/mailman/listinfo/impute
