mstop = 1 should be possible for non-cyclical fitting

[hofnerb]

glmboostLSS(y ~ ., families = NBinomialLSS(), data = dat,
                     control = boost_control(mstop = 1), method = "outer")

currently gives the errror

Error in mboostLSS_fit(formula = formula, data = data, families = families,  : 
  ‘mstop’ has to be an integer larger than 2

and if one fits a larger model and reduces it to 1, one gets

Warning message:
In attr(x, "subset")(i) :
  Minimal number of iterations: 2 (at least one iteration for each distribution parameter)

Does this really make sense for non-cyclical models?

[hofnerb]

Why do we force update all boosting parameters in the first iteration? Is this really necessary? I know that we need to build up all mboost objects but couldn't we (and don't we have to) just use the best update already in this first step? (i.e. alway refit already the first step if necessary???)

[ja-thomas]

In the current implementation we have to initialize all models before the noncyclical model can run.
This was something that isn't ideal when I implemented it initially.

But in general you are completely right that it should work like you suggested. I'm not sure how easy it would be to change it like that. This is something that I would like to solve together with #24.

[ja-thomas]

@hofnerb is there a way to create a mboost object with 0 iterations?

So basically a intercept only model, with mstop = 0

[hofnerb]

No, this is currently not possible. We should have a look at mboost to fix this. Shouldn't we?

[ja-thomas]

That's imo the smartest way to fix this, I'll open an issue there

[hofnerb]

Acutally this might be pretty easy. I just have a look at it. We simply might skip this line. Just need to make sure that no unwanted side effects are induced.

[hofnerb]

Should be working now. Both, subsetting (mstop(mod) <- 0) and setting mstop = 0 initially.

Please go ahead (and let me now if you spot an open issue regarding mstop = 0)

[ja-thomas]

Awesome! Thanks a lot

[hofnerb]

@ja-thomas I hope you always use risk(mod) and never mod$risk. See boost-R/mboost#66 for possible future problems.

I think we should also allow mstop = 0 in gamboostLSS and hence also cvrisk starting from 0.

[hofnerb]

I modified risk(mod) such that it now returns a vector of length mstop + 1 as it includes the risk for the offset model.