|
|
Re: I'm sending the a1 CI table off-line, since the format didn't
hold here.
Posted:
May 7, 2012 5:44 PM
|
|
I have begun analyzing the annotations you sent offline for the CI
tables, i.e. your annotations indicating which of the predictors in
our current 7-predictor model are significant and/or required for the
study groups and control groups in our six folds.
A full analysis of your annotations logically requires considering
them in relation to the sets of slopes and intercepts for:
a) the underlying correlations of the form loge-logL on logc-logL
which generate the residuals whose dichotomizaton determines the
values x1=0 and x1=1
b) the underlying correlations of the form loge-logL on logc-logL
which generate the residuals whose dichotomizaton determines the
values x2=0 and x2=1
(Recall that there is a slope and intercept for each instance of these
correlations for each length interval for each fold for each group.)
But before I continue with this analysis, your annotations reveal a
possible result which may be highly desirable, and which I have to ask
you about first.
In particular, from your annotations as to which predictors differ
significantly between study and control group, this table can be
constructed:
Table I
x1 lnLx1 x2 lnLx2
a1 1 1 0 0
a3 1 0 0 0
b1 1 1 1 1
b47 0 0 0 0
c1 1 0 1 1
c2 1 1 0 0
("1" means the predictor differs significantly beween study and
control group; "0" means it does not.)
So from my ignorant perspective, the first question is whether there
are enough cells in this matrix for you to be able to reliably
determine whether the following is true:
"the predictors x1 and lnLx1 play a SIGNIFICANTLY greater role in the
difference between the study and control groups than the predictors x2
and lnLx2".
If not, how many more rows would have to be added to this matrix, i.e.
how many new folds would have to be analyzed in order for you to be
able to reliably evaluate the results?
And my second question is the following.
Suppose that:
i) the entire experiment were repeated using the new sets of study and
control dicodons that I've mentioned in offline emails on which you
were cc'd;
ii)this second experiment resulted in the table:
Table II
x1 lnLx1 x2 lnLx2
a1 1 1 0 0
a3 1 1 0 0
b1 1 1 0 0
b47 1 1 0 0
c1 1 1 0 0
c2 1 1 0 0
Then given the sharpness of the bifurcation in this table, would you
be able to reliably determine whether x1 and lnLx1 differ play a
SIGNIFICANTLY greater role in the difference between the study and
control groups than the predictors x2 and lnLx2".
As always, Ray, thanks very much for considering these two questions.
I feel justified in asking them because as you have noted above, the
two chi-square tests have already established a statistically
significant difference between the two groups. if it were not for that
fact, I wouldn't even bother asking you these two questions.
Also, Please note that in deference to your role at this point, I have
NOT communicated anything to the team regarding the possibkeresult
shown in Table I, nor the questions I'm asking you regarding this
possible result.
Finally, please note that by bringing up the topic of a second
repetition of the experiment, I am not trying to "change horses in mid-
stream". JRF, AML, and MS all agre from discussion at least a month
ago that the "energetic" distinction between our current study and
control groups is unnecessarily LESS sharp than it could be because
these groups do not take into account the equivalence of certain
energetic values under an operation called "Watson-Crick canonical
reverse complementation." And the new study and control group merely
sharpen the "energetic" distinction between study and control by
taking this factor into account.
|
|
