family argument in fit_hal is funky about hal9001 HOT 7 CLOSED

family is passed directly from fit_hal to cv.glmnet. Could you clarify what g0_linear and Q0_trig1 are in your example above? (I can't reproduce the reported error without them.)

from hal9001.

The issue is that predict is currently hard coded to x*coef, so predictions when family="binomial" are on the logit scale. Side note: halclassic ignores its family argument https://github.com/benkeser/halplus/blob/master/R/hal.R#L190 so comparisons should be made for gaussian only

from hal9001.

Also, including an argument family may be inappropriate for theoretical reasons. See the comments in this issue (in the halplus repo) for details.

from hal9001.

Good to know. So what should we do going forward? Disable support for passing family to cv.glmnet in hal9001 (and maybe implement bounding instead), or support it correctly?

from hal9001.

Also, @jlstiles to @nhejazi's point earlier, it would be helpful if you could try to make your examples minimum and reproducible. Here's a discussion on the topic: https://stackoverflow.com/a/5963610/3243525

from hal9001.

I think we're good to keep the family argument in. I was too lazy to think it through previously, but I think Jonathan has convinced me that it's kosher.

from hal9001.

Resolved as part of #24

from hal9001.