In situations with major class imbalance, ROC-AUC may not be a good metric to assess model concordance. Instead, as suggested in numerous places such as the scikit-learn documentation, the area under the precision-recall curve may be preferred. Functions already exist for PRC-AUC for the standard settings, but there is currently no function available in yardstick for the survival setting.
I've attached some code here that is an adaptation of the roc_auc_survival_vec function and the functions it depends on that, I believe, implements the survival version of PRC-AUC by using the principles of Vock et al., where they provide a general recipe for incorporating inverse probability of censoring weights to any model. The final step, after estimating the weights, is:
Apply an existing prediction method to a weighted version of the training set where each member i of the training set is weighted by a factor of $\omega_i$. In other words, if $\omega_i=3$ it is as if the observation appeared three times in the data set.
In situations with major class imbalance, ROC-AUC may not be a good metric to assess model concordance. Instead, as suggested in numerous places such as the scikit-learn documentation, the area under the precision-recall curve may be preferred. Functions already exist for PRC-AUC for the standard settings, but there is currently no function available in
yardstick
for the survival setting.I've attached some code here that is an adaptation of the roc_auc_survival_vec function and the functions it depends on that, I believe, implements the survival version of PRC-AUC by using the principles of Vock et al., where they provide a general recipe for incorporating inverse probability of censoring weights to any model. The final step, after estimating the weights, is: