Auxiliary variables excluded (loggedEvents, method = "pmm")

[NB-Bio]

Hi all,
I am working with an ecological dataset where I need to impute missing values for 11 variables of interest (species'traits). To support the imputation of NMAR traits, I included ~100 auxiliary variables, specifically phylogenetic eigenvectors generated using the PVR package (no missing data in these).

Following recommendations from Flexible Imputation of Missing Data by van Buuren, I reduced the number of predictors using quickpred():

pred.birds <- quickpred(imp.mod.birds, include = to.include, exclude = "scientificName", method = "spearman")
Where to.include contains the 11 variables I aim to impute.
This resulted in a predictor matrix with an average of ~25 predictors per variable, great.

Then I run mice

meth.birds <- make.method(imp.mod.birds)

imputation.test <- mice(imp.mod.birds, method = meth.birds, predictorMatrix = pred.birds, m = 1, maxit = 1)

The problem is that I get a list of logged events. In imputation.test$loggedEvents, I noticed that almost all the phylogenetic eigenvectors selected by quickpred() were excluded during the imputation process for most target variables (except 2 out of 11).

Here is an example from loggedEvents :

I understand that pmm refers to Predictive Mean Matching, but I don’t understand why these predictors are excluded.

Trying to investigate, I calculated the correlation between the excluded predictors and one of the affected target variables, for example, Population_density:

There was great variability in the correlation coefficient:
summary(cor_results$spearman_rho) Min. 1st Qu. Median Mean 3rd Qu. Max. -0.26880 -0.17143 -0.11164 -0.03304 0.13813 0.21181

and I am not sure if the cause of the logged events is the correlation between the target variable and the predictors in this case.

Then I checked the pairwise correlations among the excluded predictors themselves, with a histogram showing no indication of extreme multicollinearity (image below).

Why are these predictors excluded during imputation if they were selected by quickpred()?
Is the "pmm" method responsible for this exclusion? If so, under what criteria?
Are there hidden thresholds (e.g., for correlation, collinearity, or singularity) that trigger exclusion inside mice()?
Should I adjust my predictor matrix manually, or is there a better way to preserve useful predictors like phylogenetic eigenvectors?

Thanks in advance to anyone who can shed light on this

[stefvanbuuren]

It is likely that your variables are linearly dependent. The internal function remove.lindep() removes linear dependent variables. You can specify mice(..., eps = 0) to inactivate it, but that is likely to get you into trouble. See #225.

[NB-Bio]

Thank you so much for your insight! Very appreciated