One important thing I noticed in an earlier, partial analysis of ability scores and hit points (link) is that these improved high-level distributions are still bell-shaped. That is: You would not want to use a method like "roll d6's, minimum 3", because that causes a pileup of scores on the low end (e.g.: 3), which is to say, generates a right-skewed distribution. (For example, the "minimum score" method that Gygax uses for hit points in AD&D Unearthed Arcana, p. 74, is contra-indicated.)
What I came up for a simple dice-rolling method for such heightened, generally bell-shaped scores is to roll two dice of some type and add a fixed bonus, such that the maximum result is still 18. Consider these dice to be our toolset for this purpose: 3d6 (Expected value 10.5), 2d6+6 (E = 13), 2d4+10 (E = 15), and 2d3+12 (E = 16). Now we look at the average scores for our population generated by the Arena program (with an encounter roll modifier of -4), and try to find the best fit at each level:
I've batched up the levels into groups of about 3 at a time for convenience. At higher levels, the picture is pretty clear; scores in each of Strength, Dexterity, and Constitution trend upward as only the fittest characters survive to higher levels, as we would expect. In the 2nd-4th level tier, characters have an average of 13 across each of these abilities, arguing for a roll of 2d6+6 in three abilities of the player's choice. For levels 5th-7th, it looks like one ability is about 15 (2d4+10), and the two others 13 (2d6+6). For levels of 8th and above you have at least two scores at 15; in this case I'll take a bit of artistic license and use one 2d3+12, one 2d4+10, and two 2d6+6 scores. Not too bad.
But now let's focus on the slightly troublesome 1st level. Whereas the 0-level clearly depicts the straight 3d6-roll for newly generated normal men (average about 10.5), characters surviving to 1st level do evidence about a 1-point advantage across each of Strength, Dexterity and Constitution (average about 11.5). The problem is, using the dice "tools" above, none of the abilities quite reach halfway to the 2d6+6 roll method (mean of 10.5 and 13 = 11.75). But if we pool the overall improvement in all three scores, then we get a value a bit over 3 (3.2). So to reflect that basic trend, I'm going to give 1st-level characters in my games a boost of about 3 points by selecting one single ability -- likely their prime requisite -- and rolling 2d6+6. (This replaces my old house rule of rolling all 3d6 and then swapping two of the player's choice.)
In summary, here's my new evidence-based house rule for rolling ability scores:
- Level 0: Roll all abilities 3d6 in order.
- Level 1: Roll one selected ability 2d6+6, others 3d6.
- Levels 2-4: Roll three selected abilities 2d6+6, others 3d6.
- Levels 5-7: Roll one ability 2d4+10, two 2d6+6, others 3d6.
- Levels 8+: Roll one 2d3+12, one 2d4+10, two 2d6+6, and two 3d6.
Final Thoughts: While we've developed this system looking at the "NPC" population at the -4 encounter die level, the truth is, any other close modifier generates about the same ability scores and hit points upon surviving to higher levels. So I'm comfortable using this system for generating scores for both high-level PCs and NPCs. It's conceivable that future refinements of the Arena simulator program might change things here, but my expectation is that the method will remain pretty stable going into the future.