Date: Nov 30, 2012 11:09 PM
Author: Ray Koopman
Subject: Re: En passant question: What if a plot of slope CI<br>	’s is lousy, but splits the “m’s” perfectly?

On Nov 30, 7:54 am, djh <> wrote:
> Consider the following plot of slope CI?s (paste into fixed font
> document if it wraps):
> ----------m----------
> a1,1,C
> --------------------m--------------------
> b1,1,C
> ---------m---------
> c2,1,C
> -----m-----
> b47,1,C
> -----m----
> c1,1,C
> -----m-----
> a3,1,C
> -----m----
> c2,1,S
> -----m------
> c1,1,S
> ------m-----
> a1,1,S
> -----m-----
> b47,1,S
> ----------m----------
> a3,1,S
> ------------------------m------------------------
> b1,1,S
> Between ?C? and ?S? slopes, there is obviously no CI split at all.
> But there is a perfect split of ?m?s?.
> Can?t anything be concluded from the perfect split of ?m?s??
> (I think I know what your answer will be but wanted to check
> nonetheless.)

If all the true slopes were identical and the sampling distributions
of the sample slopes were symmetric then the probability of having
all C < all S, or vice versa, would be 2*6!^2/12! = 1/462 = .0022,
so it certainly looks like there's a subset effect. It might be more
to the point to do S vs C t-tests for each fold separately. And, in
the same spirit, for each cell in the Fold x LengthInterval design.
(Those are the same heteroscedastic t's you've been doing.
Don't give the data to Excel and tell it to do a t-test.)