It’s just really hard to eliminate adverse selection in population studies.

If people only choose to put on their helmets when doing dangerous things, but don’t bother when doing relatively safe things, then it’s going to look like helmet use is correlated with injuries. Even if, on a per-person basis, the helmet actually reduces injury severity.

And it’s difficult to test helmet effectiveness directly because it is widely considered unethical to randomly subject people to blows to the head.