Tom Wenseleers

Well but the reference category probability is always 1-the sum of the probabilities for the other categories, by definition, no matter whether you have 2 or >2 outcome levels. And...

And yes sorry I just temporarily stored those quick tests in that sandbox to show to you that they work. Do you need additional tests actually on top of the...

One reason I would prefer to have all outcome levels returned is that transforming between the latent & response scale is a hell easier if they all come out &...

Yes I ran inst/tinytest/test-pkg-net.R - but as I mentioned I could only run it if I added ".//inst//tinytest//" in front of all the testing_path paths - not sure why that...

Ha OK I thought that for a package one always assumed the project root directory to be the working directory... If we would like the reference level to come out...

Yes indeed, because then one can just take the softmax transform of the predicted values (by rowid) to get back to the response scale. emmeans doesn't have type=link, it only...

For type=probs I can maybe do something similar - use the original multinom predict method, but add the probability for the reference level if it's dropped? Then it would behave...

OK so what do you prefer - that I add a zero column for type link or not? I feel it's more elegant to do it at that stage than...

To give you a bit of an update on this: so in the terminology that they use in compositional data analysis, a multinomial model is fitted on the additive logratio...

Well it's just the case that other packages do it in different ways - the Effects package e.g. constructs confidence intervals on the logit scale & then backtransforms for multinomial...