imputation - Is this a logical/safe way to deal with multicollinearity warnings in MICE?

Question

Patients who experienced a stroke completed a questionnaire regarding their language history, i.e., when they acquired their second language (L2), what percentage of the day do they use their?first language (L1) vs. their second (L2).

Various values were missing among the participants along the variables of interest, so I imputed the missing values using MICE in R.

The problem is that MICE is warning against imputing columns that present with suspected multicollinearity. This is common in this questionnaire as certain pairs of responses when converted to percentages must add up to 1.00. For example, when added together, Daily Use in L1 and Daily Use in L2 must add up to 100%, since anything greater or less does not make sense, so there is a dependency between those variables.

In this case, MICE was detecting that L1 and L2 post-stroke daily usage (L1PostUse and L2PostUse), when combined, were adding up to 1.00, so it did not impute L2PostUse.

For the?first imputation: I forced MICE to impute the L2PostUse column in spite of multicollinearity.

Forced imputation:

iLUQ <- mice(MICELUQ,m=5,maxit=50,meth='pmm',seed=500, remove.collinear=FALSE)

For the second PCA: I allowed MICE to?not?impute L2PostUse, and then I simply subtracted?L1PostUse from 1.00, i.e., 1-L1PostUse, given that L1 and L2 Daily Use when combined must equal 1.

The results for all?imputations were the same, even including the manual subtractions for the L1/L2PostUse columns, except for one participant. The results with the manual subtractions made more sense/fit the trend of the data -- but we are just wondering if that is logically/statistically sound practice?

Already tried messing with ridge!

Answer 1

等待大神答复