10-12-2015 01:08 PM
I am going to perform MI to impute missing data in a dataset with 200 observations, 15 of which have some missing data; among those 15, 10 have monotone missing patterns. Assuming the data is continuous, multivariate normal it seems there are 2 choices: MCMC to fully impute the missing data, or MCMC with the monotone option, followed by monotone regression (ie, double-imputation). Are there any (statistical) reasons to favor one approach over the other? And secondly, if I were to use the double-imputation approach, I could do 100 imputations in the first MCMC step, followed by Nimpute=1 in the monotone regression step; or vice versa; or I could split them equally between the 2, say 10 in the MCMC step and 10 in the monotone regression step. Again, does this make any difference?