A lotta work for nothing

Alert readers will recall that a week ago I spent the better part of two days sorting through derivations to find a term that almost always evaluated to zero unless you blended the estimators - then it was not anything like zero.

Well, it's back, but in reverse. Now, since I'm not blending estimators in my most general case, I do want the term. It still comes out to pretty much nothing, but I don't want reviewers pinging me for dropping terms just because I don't think they matter. I still do drop terms, but I'm being very careful about making sure they really are insignificant. I've wasted too much time chasing down bad assumptions.

FWIW, here's the term:




I do drop a couple terms in the third step because the hit rate and mean observation aren't supposed to change from one partition to the next so their variance should be very close to zero, especially with large row counts. That's an assumption and I could get called on it, but since I call it out myself in the paper, I'm OK with it.