Dear all,
These are a couple of beginner questions regarding how to best model
experimental variance (and decrease error) when only interested in a subset
of conditions that were presented to participants. Any advice is highly
appreciated!
1) The experiment I am currently analyzing was originally designed as a
2(temperature) x 3(instruction) repeated-measures, but for my current
purposes, I only wish to analyze two of these instructions within only one
of these temperatures (i.e., two conditions, two simple subtractions). I
understand that one wants to explain as much variance as possible in the
individual level analyzes, so all known sources of variance should be
modeled there. At the group/higher level analyzes, however, is where my
question resides- and sorry if this is a naive question: if I am only
interested in these two contrasts mentioned above, is there any reason to
include/model the other conditions which won't be used *also* at the higher
level analyzes? Will it make a difference for the degrees of freedom of the
group tests? From my reading on fsl documentation, it seems that the group
level analyzes use variance explained/ error variance and degrees of freedom
from the *individual/lower level contrasts* for each subject, so . . . it
seems like it would not matter whether or not I include the conditions I am
not interested in the higher level stats. Is that indeed the case, where
including those is neither beneficial nor detrimental to the
sensitivity/power of the higher level analyzes?
2) This is another stats question regarding the most appropriate way to
reduce error and model the conditions of interest: up until this point, what
I have been doing when modeling the aforementioned conditions above is to
use one sample t-tests (for example, a subtraction of two instructions at a
given temperature), at both individual and group level analyzes. For these
specific analyzes, I am not interested in the interaction between
temperature x instruction, nor am I interested in one of the instructions
(enhance). That being said, is there any statistical reason (sensitivity,
power) why I should still be modeling this repeated-measures experiment as 2
x 3 factorial instead, if only interested in the aforementioned simple
comparisons?
Thank you very much for any feedback!
Regina
|