My general concept for judging the robustness of threads amounts to measuring how square they are. That is, a robust thread has lots of posts to the original root post, and those replies in turn have lots of replies. I think that would lead to a pretty simple metric: The degree to which the number of vertical branches and the depth of replies are equal. If you define it as height / width, then 1.0 is optimum; > 1 indicates a root post that got lots of initial attention but generated less ongoing discussion; < 1 indicates a small number of people having a lengthy conversation that didn't generate widespread interest.

The formula could have an additional factor to gauge popularity: The ratio of the number of posts / number of people. Larger numbers mean fewer people are involved. This would compensate for the possibility that just two people could still have a robust thread, if both have the powers of concentration to keep track of a conversation that branches in different directions. I arranged it as posts/people so that as a multiplier it would exaggerate the deviation from 1.0.

( height / width ) * ( posts / people )

Hmmm...probably still a bit too simplistic. Height and width can be skewed if, for example, the root posts gets lots of replies, but only a few of those replies in turn generate follow-on discussion. It probably needs a way of counting the number of posts at each level down from the top, weighting each post by its number of levels. Branches at each level can count toward the height value, with inverse weighting: Branches at the top count more than branches at the bottom. Something like that--I can see it needs further thought. But it's off to an interesting start, anyway.