Thinking about it, in the instance of important but esoteric metrics, do you think that the most frequent case has to do with the [expected] ability to impact or exploit it meaningfully? I don't do a lot of work with numbers, but it seems to me that in terms of affectation, those dumb statistics are the low hanging fruit with low investment cost in terms of manipulation, thus yielding "better" results and isolating them to the use case while alienating the complex metrics. It's a lot easier to take a metric that isn't contextually affected than the obverse, and up/downregulate it.
I think my example would be the quasi-metric of the single-sample workup you'd get at a hospital. You get a single datapoint. Compare that to a longer term study where samples are taken at 4 hour intervals. The latter brings up a lot of subsequent questions and has a lot of interconnections with the world at large and the individual's biological characteristics. E.g. getting caught by a train 2/5 days, boosting cortisol and norepinephrine, or low receptor affinity for X, thus producing the illusion of high X levels. As you modify you also complicate with an increasingly large list of rules. Or you just write a script for statins and pat Tom on the ass, noting next visit that serum cholesterol is down.