Ok, that helps. I think you’re saying the issue arises when the set of constraints limit the observed events to a number too small to draw appropriate conclusions from.

I’m hesitant to shy away from “bizzare” constraints. If there are enough data points for that scenario to draw some statistical correlation… then that just is the reality even if we can’t explain it (yet).

If the coaches wife sits in a different section for 20% of the games, and they disproportionately lose when she sits there… that’s the correlation.

Could be she sits in a further away section if she’s pissed after a fight with her husband the night before, which is a signal the coach also had a bad night, and is fatigued and unfocused during the game now.

But yeah, you need enough observed instances.