"Comparing covariances of estimates is even trickier than comparing standard errors, but if you have nothing better to do then fine, look at their regressions on length. Plots will probably be more informative than regression coefficients etc."
Thanks for the go-ahead but before I drop the 2-way approach per fold and length using the new average slopes, would you evaluate the following table with an eye toward deciding:
a) whether we have enough method/subset 2-ways per fold and length to justify continuing to work with the average slopes of the current set of regressions:
Ruq = c on (u,u^2) Rub = c on (u, e, u*e) Rubq = c on (e, u, u*e, u^2)
Auq = average slope for Ruq Aubu = average slope for u coefficient of Rub Aube = average slope for e coefficient of Rub Aubqe = average slope for e coefficient of Rubq Aubqu = average slope for u coefficient of Rubq
b) whether enough of these 2-ways have p's < .05 to justify continuing to work with the current set of regressions.
If we don't have enough average slope 2-ways, or enough with p's < . 05, I will repeat the same analysis using the special covs for Rub and Rubq.
Finally, if we do have enough p's < .05, can you suggest the correct Bonferroni framework for "correcting them"?
Thanks as always for considering this matter.
Table: # of 2-ways Method Avg per # with (N/R) Fold Slope len p < .05