Date: Dec 7, 2012 8:57 AM Author: Halitsky Subject: Thanks for the guidance on how to evaluate the contribution of u^2 in<br> the second model. Thanks for the guidance on how to evaluate the contribution of u^2 in

the second model.

I can easily clone some existing code to compute t and its associated

p for u^2 for each cell of the model (but note that I will only do so

when the cells are non-empty for all four combinations of subset x

MoSS | each (fold, set, length.)

But before doing so, I have one minor question about your formula:

estimated coefficient of the extra predictor in the second model

t = ----------------------------------------------------------------.

estimated standard error of that estimated coefficient

By analogy to the formula for t that you gave me for doing the 2-way

t?s:

(coeff1-coeff2) + (coeff3-coeff4)

t = --------------------------------------------

var1 + var2 + var3 + var4

the denominator term ?estimated standard error of that estimated

coefficient? should simply mean the variance of the coefficient in

question. Is that the case? If not, what do you mean by ?estimated

standard error? and why isn?t it simply the variance in this case?

Also, can you answer the following questions.

I.

You wrote:

?A totally equivalent way is to get F = (RSS1/RSS2 - 1)*(n-k), where

RSS1 and RSS2 are the Residual Sums of Squares from fitting the two

models. Refer F to the F-distribution with df = (1,n-k). F = t^2.

A. If I understand what you wrote here, then I really don?t have to

fit the model to get RSS1,RSS2 and then compute F from them. Rather,

I can just square t. Is that the case?

B. Even if I can just square t to get F, is getting RSS1 and RSS2 by

?fitting the two models? something I can do programatically myself

with instructions from you? Or do I have to look for an appropriate

GSL module akin to Ivo?s regression module, one that I can tuck inside

a PERL program? (Remember ? I am R-less and Mathematica-less.)

II.

You wrote:

?No matter how small the p is, it doesn't mean SEP1/SEP2 is big enough

to get excited about, that using the first model instead of the second

would increase the SEP by enough to worry about.?

Since there is no equivalent of u^2 in the regression c on (u,e,u*e),

I don?t understand where the other SEP is coming from.

By SEP1 and SEP2, do you mean the SEP?s of the two regressions

themselves ( c on (u,e,u*e) and c on (u,e,u*e,u^2) ), rather than the

SEP?s of any particular coefficients of these two regressions?

If so, can you tell me how to compute the SEP for a regression

itself? So far, I don't think I've done this yet, and Ivo's module

only reports out the SE's of coefficients, not an SE for a regresssion

itself.