benkeser
Implementing new SuperLearner method
One idea to think about that would be a really valuable software addition (i.e., it's an idea that's been around for a long time, but as far as I know has never been implemented): it would be great to write our own super learner method for the hazard-based estimator that uses a(n IPCW?) loss function based on the conditional cumulative incidence function rather than the hazard-based log-likelihood loss function. This would probably take real work to implement and would likely be enough to constitute a short paper to IJB studying it's practical performance. It is, at the very least, theoretically appealing in that our super learner is truly targeted towards our goal, which is estimating cumulative incidence, and oracle inequalities would imply that we're getting as close to the true conditional cumulative incidence as possible.
