I am running a series of random forest models using proc hpforest. I am having a difficult time understanding NEGATIVE OOB gini coefficients from the variable importance output. Is one not to interpret any variables with negative OOB gini coefficients as meaningfully related to the target variable? Thus, should I interpret only those variables with positive OOB gini coefficients?
Also, to get to a final model should I be dropping variables from the model with negative (inbag) gini coefficients and/or negative (inbag) margin coefficients (as was suggested in previous posts)?
I can't find a good explanation to all this, so any help is appreciated! Thank you.
