To Lyte, and other people deeply entrenched in matchmaking: (not much of this affects me, but I've got a hungering curiosity, honestly don't have my apples in any particular basket... if even one does keep apples in baskets.)
I'm going to keep this conversational, but if you'd prefer more mathematical specifics I can oblige (though it will take a bit more time to do formal writing as I'm sure you're aware), though I'm purposely avoiding LaTeX and an attached document (since I assume they get less reads than posts).
I don't know if you're able to elaboate too much, but I would like to understand the reasoning behind why even having a 'newbie island'.
From a mathematical and historical standpoint (of ELO chess, and the assumption that the modified version that you use in League is somewhat similar) we would see something of the logistic distribution for an average population. Whether or not you have maintained the idea of zero-sum for a single game or overall (about the average value or wherever it lies), we have two mechanisms for lowering rating, decay and dodging, and none for increasing it. This means that we can only see a steady decrease in the average rating of the general populace.
This means, that our newbie island is a separate distribution from that of our general population. (Which of course, is the intent.) But the repercussion is that our average newbie island distribution is centered about the specified a priori value while our general population's value is monotonically diverging from this specified value.
It can also be argued that the more games people play, the higher their average skill-elo correlation. (Of course I have only anecdotal evidence to prove this one, which as we know, is worth next to nothing. However, you do have the data to show this correlation, should it exist.) However, seeing how this is a game whose interactions are varied among a number of champions nearing a hundred, and most interactions are something that which needs to be learned in a mechanical sense, it might be reasonable to assume that there is a number of games before which this skill-rating correlation might plateau. I will assume that it does exist for the aforementioned reason, but if you wish to refute this one I would love to hear as such.
This means, that independently of the general populace's rating slowly decreasing, our newbie island is, on average, going to be below the average of general rated players.
By generating two populations like this, we do isolate new players, but without the appropriate later merging (which is impossible without considering the appropriate correlation between the two populations) we obtain more conflicts with this newbie island than had we simply introduced players directly into the single population.
These analyses aside, I'm wondering what benefit is seen from extending this newbie island (and increasing the divergence in these populations).
The idea of using normal games to seed into ranked, honestly, seems to me to be the best choice. Of course, it would be required to make an analysis of the standard deviation of a player's normal rating as compared to rated players, since anecdotal evidence would suggest that people play 'looser' with their normal games, but again, without a real analysis of data, who knows?
Ravencoff - Adding Search Key