Can I use imputed data for imputating another variable? I have a dataset with three columns A, B and C.
The column A has some missing values and the column B too. Column C is complete.
If I use a method like MICE for filling column A, can I then run the MICE algorithm once again using the imputed values from A to predict values for B?
 A: Yes you can. This will add some extra randomness, but if you use proper pooling rules afterwards, multiple imputation is able to handle multiple variables with missing. The R Mice package offers a lot of options to deal with such data. 
Do take care to think which predictors to use:


*

*If missing values in A are predicted by both B (& C), using A to predict missing values in B might cause a feedback loop where the association between A and B is unduly increased. MICE allows specifying which variables are used in which imputation model using the 'predictormatrix'. Avoid such loops, and check for any by plotting the imputation object after the procedure has run (plot(imp))

*It is good to think of additional variables which are specific for missingness in a variable with missing values, and add these to the imputation strategy. As overfit imputation models are not considered that much of a problem, having more additional variables is almost always better.

*When there are more variables with missing values, mice needs to know where to start. Default is to start with the left-most variable in your dataset. As imputation using mice is an iterative process, having enough iterations usually means you do not need to change this. However, some missing values needs some iterations to reach a 'probable parameter space'. So there's two options in mice to consider here: alter the 'visitsequence' to pick which order to complete missing values, and/or increase the 'maxit' (maximum iterations) from the default (5 I think) to, say, 30.

