This is all uniqueness bias talking, wanting to be reintroduced into your decisions when you’re trying to eliminate it. Don’t listen to it. Keep the process simple: Define the class broadly. Err on the side of inclusion. And adjust the average only when there are compelling reasons to do so, which means that data exist that support the adjustment. When in doubt, skip adjustment altogether. The class mean is the anchor, and the anchor is your forecast. That’s very simple, yes. But simple is good; it keeps out bias.
