Roll 6d6-drop-lowest-3 once, 5d6-drop-lowest-2 once, and then 4d6-drop-lowest for the other four stats.
(My understanding is that on average adding a dice is reasonably equivalent to a +1 bonus, so that's the equivalent of a +2 and a +1. Obviously, that rule of thumb doesn't work if you keep adding more and more dice.)
I haven't worked out the odds of getting an 18 in that system. I don't think they're terrible, but I suspect a majority of characters would still miss out.
Yeah, that's about right. Chance of 18 is 6.25% on 6d6k3, 3.55% on 5d6k3, and 1.62% on 4d6k3. Overall chance of getting at least one 18 would be 1-(1-0.0625)(1-0.0355)(1-0.0162)^4 = 0.153, or just a hair under 15.3%, about one in six. For comparison, the all-4d6 method is 1-(1-0.0162)^6 = 0.0933, or about 9.33%, about one in eleven. Certainly better odds, but one shouldn't be surprised to have no one in the group roll 18.
It's not bad if you're wanting to keep things simple while ditching racial ability bonuses, but I can definitely see why folks would pass on it. You still have a ~50% chance of not even getting a 15 on the 6d6k3, and just shy of that of not even getting 14 on the 5d6. The average
does move up the way you've described, however. Average result for 6d6k3 is 14.27, for 5d6k3 it's 13.43, and for 4d6k3 it's 12.24. Impressively, that's exactly as far as the rule of thumb goes; if you went up to seven dice, the average would be only 14.90, barely more than half what you'd want. But (if it ever mattered) 8d6k3 and 11d6k3 are close enough to being "+3" and "+4" in this scheme.
Alternatively, you could start replacing dice with d8s, but treating 7 and 8 as though they were 6. That is, just rolling 4d8k3, but treating 7/8 as 6, is almost identical to 6d6k3, but a
much better chance of getting 18 on any given roll. Might be worth thinking about, to get people to try that method. The replace-with-truncated-d8 trick doesn't work for the approximated +1, however. But having a 15%
base chance to get an 18 (about 23.3% overall chance) could be tempting enough to get folk to try rolling.
Or, for a wholly different alternative, switch to 9d4k5-2 instead of 6d6k3. The -2 keeps it in the same 3-18 range, but switching to larger numbers of d4 means a more "peaked" distribution. The chance of getting an 18 isn't much higher, but more than 50% of the weight lies in the 15-18 range. Not quite as simple to
roll, I admit, unless you use a computer to do the rolling for you. But it has near-identical average results while still being much "safer" than the d6 method.
Here's
an Anydice program comparing the three, if you're curious.