On Tue, Nov 27, 2012 at 1:28 PM, Mahinda Yogarajah y.mahinda@gmail.com wrote:

Hi,

Thanks for the reference and explanation of the subtle statistical inferences. It was very useful.

You are right in what i am trying to show (c- controls, p1 to p4 patients with varying levels of endogenous factor) which is that when patients are subdivided into levels of endogenous factor there is correlation with the changes seen in dependent variables compared to controls. So far,

C v whole P gp - differences seen in plausible areas. C v P1 (greatest endo factor) subgp - no differences C v P4 subgp (lowest endo factor) - differences seen in plausible areas.

Given that it is not valid to compare c v p1 and c v p4 (and p1 v p4 shows nothing) is the following valid - perhaps it is similar to what you proposed:

...

...

...

"compare c v p1 and c v p4 (and p1 v p4 shows nothing)" These are all valid tests on their own. The issue waas that you were making qualitative interpretations of the results.

Consider c and p1 to p4 as one group of subjects (theoretically the only difference between subjects being level of endogenous factor) but assign each subject a variable which represents level of endoenous factor (c - 1 p1 - 2 p3 - 4 p4 - 5) - then run glm looking for correlation between endogenous factor and dependent variable.

Is this very different to what you suggested and any less valid ?

...

...

...

If you have you have entered 5 groups and want to test the linear change over the groups. You can use [2 1 0 -1 2] as the contrast. If you want the linear change over the patients, then you can use [0 1.5 .5 -.5 -1.5]. If you want to detect any difference in the patients, you could try an F-test of [0 1 -1 0 0; 0 0 1 -1 0; 0 0 0 1 -1].

Thanks.

M