FCS imputation specify different ignored rows for different imputation models

[tinneuro]

Hi,

I want to impute missing data for two variables of my dataset with 10 colums/variables. For these two variables I use different imputation models by specifying the formula used for each of the variables. Thus for the two variables x9 andx10 I have individual differing formulas for the imputation. i.e. formula for x9~ x1 + x2 + x3 and x10~ x 1+ x4 + x5 where we can see that x9 is not a predictor of x10 and x10 is not used as a predictor for x9.
In addition to the different formulas I also want to train the two imputation models on different rows of the dataset where I would use the ignore functionality of mice. But with the ignore I can only specify one vector of TRUE and FALSE which would then be the same for both models (i.e. x9 and x10) but I want the models to be trained on differing rows of the data. How can I achieve that? Maybe I could solve that with blocks but I did not find more guidance about how to solve this. Thanks a lot!

[stefvanbuuren]

Did you consider splitting the data into two sets: {x1, x2, x3, x9} and {x1, x4, x5, x10}, and impute these separately with their own ignore vector?

[tinneuro]

The other option I saw when having two seperate imputed datasets was then to join the obtained imputed values for x10 to the first imputed dataset. But for this I probably need to transform with complete and then again as.mids(). Is this an option which could work?

[stefvanbuuren]

mice::cbind.mids() combines the two mids objects into one. Give that a try.

[tinneuro]

Thanks! The problem is that I actually need to join the two imputed datasets by an Id. Thus I am now considering to rather use mice::rbind.mids() and have a long data format and then postprocess with complete and pivot_wider() to a wide format and transform back to a mids object. Unfortunately this is tedious...

[stefvanbuuren]

Yes. I see. cbind.mids() requires the same number of rows in both datasets. Otherwise, you're in the woods, indeed. Good luck.

[tinneuro]

Haha. Yes I also realised that. But still thanks a lot for your feedback it helped me!