Competing risk and variable selection

[turgeonmaxime]

It would be very useful to be able to perform variable selection with competing risks. It would seem that it should be straightforward, since glmnet has a multinomial family. But as we've discovered, they use a different parametrization for the multinomial model which implies that constant offsets are dropped.

I see at least two solutions:

• Find a way to adjust the glmnet fit to account for the missing offset term.

• Implement our own version of penalized multinomial regression.

For obvious reasons, the first solution (if possible) would be preferable.

[sahirbhatnagar]

The CVXR package and paper might also be an option, but would require some thought, as I didn't immediately find a multinomial regression example. But they do say it's possible. What's nice about this package is that it allows you to define your own objective function and arbitrary constraints, so we know exactly what's going on behind the scenes.

As for the offset, I believe this is just another constraint achieved by setting the beta for the offset term to be 1.

[turgeonmaxime]

Thanks for sharing, that looks like a great option to look into!