F-tests begone!

May 18, 2009 in Math

Andrew Gelman just wrote a blog post regarding interaction terms in multiple regression and concludes:

You never have to do an F test. Just forget about that stuff!

I find it incredibly refreshing when someone (a professor, no less!) is willing to cut through the math and get down to common sense. And I particularly hate F-tests in regressions, since they are so frequently unhelpful. Yes, I get it - with many variables you increase the probability of a type I error and the F-test overcomes that. But I have yet to come across a vetted, sensible and non-hypothetical model in which F-test failed to reject the null but at least one of the marginal t-tests did. I know, never say never, etc.

Of course, I realize as I write this that you do have to follow quite a bit of math in this case, so we're not really cutting through the math as much as "intelligently disregarding it", but that's why we're nerds, isn't it?

Anyone?

Just me?

Leave a Comment

{ 1 trackback }

Previous post:

Next post: