Simple, I added +1 to each result, so the 3s became the 4s and so on up to 18, which is the total of 17 and 18 together. I did this for 2 reasons:
1) Many months ago, I was trying to decide if octs were worth it, so I collected all my tri/hex rolls and those of other players. In the end, I determined that only about half had a stat of 9 or less and would be rerolled. Oddly, the 10.5 percent works into this rather nicely.
2) During that same time, I was rolling a fair number of 18s. They were usually followed immediately by garbage, and I don't think I was pulling in quite 1 in 4.. but the frequency wasn't horrid. This also follows on that line. I'd have to check out my reroll file to get the exact percentage though.
What I don't like about it is that it figures using these numbers and your equation that a 3x18 happens about 1 in 90 rolls.. and I rolled over 200 and never got one. Then again, I know folks that rolled like 10 times and got stats better than I've ever dreamed of. Go figure... law of averages and the universe hates me.
Another thing, while I've RARELY heard of someone rolling a 4, I think the only 3 I've seen was wearing eq. This also fits with that.
It's a guess. It could very well be wrong, but so could everything else that's been posted. It fits better with my experience than the others, so at the moment, it's my favorite theory. If you have reason to believe it's wrong, I'd love to hear it