David Allison

So those are things where the crossover has a limit.

And the question is,

Does that mean you try to do what David Ludwig is, I think, arguing, again, I hope I'm not inappropriately putting words in his mouth, is saying, they're just invalid, just either don't use them at all, or you can only use them in this way, and if you get what looks like it might be a period by treatment interaction, then discard the study, it's invalid, so on and so forth.

Or do you say, which I think Kevin would say, and I would side with, is accept the limitations of the study.

It doesn't mean it's a flawed or incorrect study.

It means it's a limited study.

And as long as you point those limitations out, you may or may not want to accept it.

In the same sense, if you have an observational epidemiologic study,

I think most of us, again, there are exceptions, but most of us would not say never do one again.

They have zero value.

What we'd say is they're not invalid studies in and of themselves.

What they do is they leave open alternative explanations for findings other than causation.

Could be some bias of measurement error.

It could be some bias of sampling.

It could be some reporting bias.

It could be some confounding, et cetera, et cetera.

And we say, as long as you acknowledge those, then the study shows what it shows.

It's weak inference, but it's not nothing.

And I think the same thing is with crossover.

And we may have to accept that with lots of stuff.