|
|
Re: Correct way to normalize an rmsd-based distance metric used in
repeated trials of pairs
Posted:
Apr 11, 2012 12:30 PM
|
|
On Apr 11, 5:00 am, djh <halitsk...@att.net> wrote:
> Since we're talking about the simple yield counts, "19*36 = 684"
> in the above post should have been "19+36 = 55" t-tests.
>
> So these two paragraphs:
>
> "So, for a given (L,p,q) triple, we can perform 19*36 = 684 t-tests,
> and then see what percent P1 of these 684 t-tests showed a (j-i<=1)
> yield to be signficantly greater than a (j-i>1) yield.
>
> And since we can do these 684 t-tests 1050 times (once for each
> (L,p,q) triple), we can determine what percent P2 of these 1050
> trials resulted in a P1 > 95%. "
>
> should have read:
>
> "So, for a given (L,p,q) triple, we can perform 19+36 = 55 t-tests,
> and then see what percent P1 of these 55 t-tests showed a (j-i<=1)
> yield to be signficantly greater than a (j-i>1) yield.
>
> And since we can do these 55 t-tests 1050 times (once for each
> (L,p,q) triple), we can determine what percent P2 of these 1050
> trials resulted in a P1 > 95%."
>
> In other words, we test the mean of the yields for the 36 against
> each of the yields for the 19, and test the mean of the yields for
> the 19 against each of the yields for the 36, for 55 tests in all.
>
> When I typed 19*36=684, I was getting ahead of myself and thinking
> about t-testing the values for the "span" variable I mentioned
> earlier, not the simple yield variable itself.
Neither the original nor the revised analyses would ever be
statistically acceptable. Summary proportions such as those might
appear in non-technical writing, but they would never be considered
proper evidence in a rigorous argument. You really need to find a
model that fits the data acceptably, ignores no factors that are
generally thought to matter, and has parameter estimates that lead
to the desired conclusion. (In the logistic regressions done so far,
you would be looking for the X2 coefficient to be significant and
negative.)
You said (in a private email) that the unit of measure of U (the
continuous variable that underlies the discrete V) is kcals/mols.
I don't know the relevant physics, but if what actually matters
is the ratio U1/U2, as opposed to the difference U1-U2, then the
boundaries of the V-categories should have constant ratios instead
of constant differences.
|
|
