Multiple Imputation rates

A basic identi ability requirement for the imputa-
tion model is that for each variable in the imputation the number of
observations should be at least as many as the number of variables in
the imputation model. Suppose that there are 50 variables in the im-
putation model and that there are 1000 observations in the data set.
Suppose that the rst variable has 960 missing values and only 40 ob-
served values. Since 40 < 50 the imputation model is not identi ed. That variable should be removed from the imputation or it should be imputed from a smaller data sets, for example with 30 variables. If for example there are 60 observed values and 940 missing values the model will be identi ed but essentially 51 parameters(1 mean parameter, 1 residual variance parameter and 49 regression coe cient parameters) in the imputation model are identi ed with only 60 observations. This may work, but would likely lead to slow convergence and poor mixing. So even though the model is identi ed, this variable would cause slow convergence and if the variable is not important one should consider dropping that variable from the imputation

From Asparouhov, T., & Muthen, B. (2010) Multiple Imputation with Mplus. http://www.statmodel.com/download/Imputations7.pdf