I am working on a classifier where some is be missing in the negative sample, but not in the positive one. The missing data is about gender (M/F) and age group (child/young/adult/old).
The data is MAR - missing at random (probability to be missing depends on other observed variables) and only in the negative samples (i.e. for the positive samples I always have all the data), thus deleting the incomplete rows would not be a good idea.
Running a classifier to predict the gender based on other features, I get a Matthew's correlation coefficient of more than 0.2, which means that the probability of a gender is influenced by those variables, but I can't perform a very good classification.
What would be the preferred way to perform imputation in this case? I tried some approaches (for instance, MICE imputation) but most of them give continuous results, while I am looking for categorical (binary, in case of gender) ones.