Clarification on the Role of predictorMatrix in the mice Package for Imputation

[JASMINE0926]

Hi,
I'm using the mice package for imputations and wanted to speed up the process by using quickpred. I understand that in the default case (pred <- quickpred(data, minpuc = 0, mincor = 0.1)), each row's variable is predicted using columns marked with 1.

> print(pred) OSESmean.income PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10 sex age TIV total_WhiteSurfArea_area total_MeanThickness_thickness edu OSESmean.income 0 1 1 0 0 0 0 0 0 0 0 0 1 0 0 0 1 PC1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 sex 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 age 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 TIV 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 total_WhiteSurfArea_area 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 total_MeanThickness_thickness 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 edu 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 0 height 0 1 0 0 0 0 0 0 0 0 0 1 0 1 1 0 1 weightMRI 0 1 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 urban_score15 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 urban_score18 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 lh_bankssts_area 0 0 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 lh_caudalanteriorcingulate_area 0 0 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 lh_caudalmiddlefrontal_area 0 0 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 lh_cuneus_area 0 0 0 0 0 0 0 0 0 0 0 1 0 1 1 0 0 height weightMRI urban_score15 urban_score18 lh_bankssts_area lh_caudalanteriorcingulate_area lh_caudalmiddlefrontal_area OSESmean.income 0 0 1 1 0 0 0 PC1 0 0 0 0 0 0 0 PC2 0 0 0 0 0 0 0 PC3 0 0 0 0 0 0 0 PC4 0 0 0 0 0 0 0 PC5 0 0 0 0 0 0 0 PC6 0 0 0 0 0 0 0 PC7 0 0 0 0 0 0 0 PC8 0 0 0 0 0 0 0 PC9 0 0 0 0 0 0 0 PC10 0 0 0 0 0 0 0 sex 0 0 0 0 0 0 0 age 0 0 0 0 0 0 0 TIV 1 1 0 0 1 1 1 total_WhiteSurfArea_area 1 1 0 0 1 1 1 total_MeanThickness_thickness 0 0 0 0 0 0 0 edu 1 0 1 1 0 0 0 height 0 1 1 1 1 1 1 weightMRI 1 0 1 1 1 1 1 urban_score15 1 1 0 1 0 0 0 urban_score18 1 1 1 0 0 0 0 lh_bankssts_area 1 1 0 0 0 1 1 lh_caudalanteriorcingulate_area 1 1 0 0 1 0 1 lh_caudalmiddlefrontal_area 1 1 0 0 1 1 0 lh_cuneus_area 1 1 0 0 1 1 1 lh_cuneus_area OSESmean.income 0 PC1 0 PC2 0 PC3 0 PC4 0 PC5 0 PC6 0 PC7 0 PC8 0 PC9 0 PC10 0 sex 0 age 0 TIV 1 total_WhiteSurfArea_area 1 total_MeanThickness_thickness 0 edu 0 height 1 weightMRI 1 urban_score15 0 urban_score18 0 lh_bankssts_area 1 lh_caudalanteriorcingulate_area 1 lh_caudalmiddlefrontal_area 1 lh_cuneus_area 0

However, when I set minpuc = 0.6 and mincor = 0.6, some rows (i.e. OSESmean.income or lh_bankssts_area, including those with missing data) have all zeros. How is the target variable imputed when its row in predictorMatrix has no predictors (all zeros)?

> print(pred) OSESmean.income PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10 sex age TIV total_WhiteSurfArea_area total_MeanThickness_thickness edu OSESmean.income 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC5 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC6 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC7 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC9 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 PC10 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 sex 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 age 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 TIV 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 total_WhiteSurfArea_area 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 total_MeanThickness_thickness 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 edu 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 height 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 weightMRI 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 urban_score15 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 urban_score18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 lh_bankssts_area 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 lh_caudalanteriorcingulate_area 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 lh_caudalmiddlefrontal_area 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 lh_cuneus_area 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 height weightMRI urban_score15 urban_score18 lh_bankssts_area lh_caudalanteriorcingulate_area lh_caudalmiddlefrontal_area OSESmean.income 0 0 0 0 0 0 0 PC1 0 0 0 0 0 0 0 PC2 0 0 0 0 0 0 0 PC3 0 0 0 0 0 0 0 PC4 0 0 0 0 0 0 0 PC5 0 0 0 0 0 0 0 PC6 0 0 0 0 0 0 0 PC7 0 0 0 0 0 0 0 PC8 0 0 0 0 0 0 0 PC9 0 0 0 0 0 0 0 PC10 0 0 0 0 0 0 0 sex 0 0 0 0 0 0 0 age 0 0 0 0 0 0 0 TIV 0 0 0 0 0 0 0 total_WhiteSurfArea_area 0 0 0 0 0 0 0 total_MeanThickness_thickness 0 0 0 0 0 0 0 edu 0 0 0 0 0 0 0 height 0 0 0 0 0 0 0 weightMRI 0 0 0 0 0 0 0 urban_score15 0 0 0 0 0 0 0 urban_score18 0 0 0 0 0 0 0 lh_bankssts_area 0 0 0 0 0 0 0 lh_caudalanteriorcingulate_area 0 0 0 0 0 0 0 lh_caudalmiddlefrontal_area 0 0 0 0 0 0 0 lh_cuneus_area 0 0 0 0 0 0 0 lh_cuneus_area OSESmean.income 0 PC1 0 PC2 0 PC3 0 PC4 0 PC5 0 PC6 0 PC7 0 PC8 0 PC9 0 PC10 0 sex 0 age 0 TIV 0 total_WhiteSurfArea_area 0 total_MeanThickness_thickness 0 edu 0 height 0 weightMRI 0 urban_score15 0 urban_score18 0 lh_bankssts_area 0 lh_caudalanteriorcingulate_area 0 lh_caudalmiddlefrontal_area 0 lh_cuneus_area 0

Thanks for clarifying this behavior!
Bests,
Jasmine

[stefvanbuuren]

If there are no predictors, mice uses the intercept-only model and imputes the mean.

[dhy1202]

Hi, I have the following three questions to ask you when using the mice package for imputation.
1.When using the default predictorMatrix, the error 'Error in solve.default(xtx + diag(pen)): system is computationally singular: reciprocal condition number = 2.23128e-17' occurs. Therefore, I used the quickpred function to filter the predictor matrix, and the error can only be avoided when mincor=0.6. However, under this parameter, some variables end up without any predictors. How can I solve this problem?
2.Is it appropriate to use an intercept-only model to impute target variables that have no corresponding predictor variables? If it is not appropriate, what solution should I adopt?
3.Is there an objective standard for setting mincor and minpuc in the quickpred function?

[stefvanbuuren]

minor = 0.6 is very high, unusual in most areas I am familiar with.

It sounds as if you have duplicate or linearly dependent variables that cause the multicollinearity error. If so, you can remove some. The error can occur if there are sparse categorical variables, so check for that.

[dhy1202]

Thank you very much for your patient response. I would like to ask a few more questions:

1. Do you have any recommended methods for objectively removing collinear variables?
2. If we want to retain collinear variables, can we use methods like “cart” or “rf” instead of the “pmm” imputation method?
3. When imputing the independent variables, should we also include the dependent and covariate variables required for subsequent analysis in the matrix, even if these variables do not have missing values?

[stefvanbuuren]

These things are also discussed elsewhere, so I suggest you do some additional searching.