dsm.var.gam when predicted values are 0

[dill]

If predictions are 0 (therefore on the linear predictor scale log(0)=-Inf), then getting derivatives for use in variance estimation is tricky. Numerically (grad()), it falls over.

This seems like it would be relatively rare (predictions of exact zeros seem unlikely due to the smooth nature of the model).

[dill]

This is a somewhat unlikely bug. It relies on a case where we have (numerically) zero predictions, which barely ever happens. It's actually pretty hard to reproduce.

[dill]

Throw a warning if there is a zero prediction?