On Sep 23, 10:13 am, djh <halitsk...@att.net> wrote: > My apologies, Ray: I was incorrectly computing LL and HH for use in > the formula: > > x -> P(x) = A + (x-LL)*(B-A)/(HH-LL) > > The error was two-fold in nature: > > a) I was assuming that LL would be the lower limit for the CI with the > lowest mean and that HH would be the upper limit for the CI with the > highest mean; these two assumptions are obviously nonsensical. > > b) I was computing LL and HH PER FOLD PER COEFFICENT (and therefore > across only 12 values), when I should have been computing LL and HH > across all 72 values for all folds and all coefficients (otherwise, > it's impossible to compare CI's ACROSS FOLDS).
Not across all coefficients. For each coefficient there are 72 CI's: 6 folds x 3 dicodon sets x 2 subsets (core,complement) x 2 u-levels. LL is the lowest L, and HH the highest H, over those 72 CI's, which appear as the rows of a single plot. CI's are not comparable between coefficients. Each coefficient gets a separate 72-row plot.
> > I will recompute and repost the results. But in the meantime, I think > the questions I asked in my previous post are still valid, even though > asked relative to erroneous results. (That is, I think that if one > assumes the posted CI results are correct, then valid answers can be > given to the questions.) > > So if and when you have a chance to answer these questions, your > answers would be much appreciated. > > Again, my apologies for the erroneous computation.