Thanks Junpeng – helpful, as always!
But this would mean you can’t implement a hierarchical structure for GLMs, wouldn’t it? I’m quite puzzled by that implication.
I can actually test this hypothesis: the double normalization is there to take care of non-available categories. So, if I restrict the data to cases where all categories are available, then I only need one normalization and I’ll see if PPC improve – I’ll keep you posted on this experiment! 