I read in your paper that "URIEL+ prevents distance computation between language pairs that lack shared features.". I thought you might be interested to know that within the Grambank project we generally use the cut-off of <25% missing values for proceeding with analysis. That's what we did in the release paper (Skirgård et al 2023) and what we have as the default for the R function crop_missing_data in rgrambank. Only after that cropping do we do some imputation when necessary, and we generally do it with random forests. Doing imputation with phylogenies would be problematic for us, as we later want to do things like calculate the phylogenetic effect on variation etc.
I read in your paper that "URIEL+ prevents distance computation between language pairs that lack shared features.". I thought you might be interested to know that within the Grambank project we generally use the cut-off of <25% missing values for proceeding with analysis. That's what we did in the release paper (Skirgård et al 2023) and what we have as the default for the
Rfunction crop_missing_data in rgrambank. Only after that cropping do we do some imputation when necessary, and we generally do it with random forests. Doing imputation with phylogenies would be problematic for us, as we later want to do things like calculate the phylogenetic effect on variation etc.