|
|
Hi, I have a design with subjects from three different clinical
populations who have have been tested in an experiment consisting of
two within-subjects factors, one with eight levels, the other with
five. I've coded this in SPSS as one between-subjects factor with
three levels and 40 dependent variables, which I have coded into an
8x5 within-subjects structure using the GML repeated measures dialog.
In addition to the experimental data, I also have several pieces of
demographic, clinical, and performance information for each subject.
They also have been entered as separate columns/variables in the SPSS
data file.
I have two questions, and please pardon me if they are too naive.
The first question is about the assumptions that must hold to use the
ancova procedure appropriately. Apparently, the covariate, to be
useful in an ancova, must be correlated with the dependent variable.
From reading the list archives, it appears that glm separately
adjusts each one of the dependent variables that make up a within-
subjects factor so that the within-subjects factors can be part of
the ancova. Does this mean that the covariate should be strongly
correlated with each one? That is, all 40 in my 8x5 design? If not,
then how would one decide whether or not some particular covariate
could be used in an ancova? The average correlation coefficient?
Correlation to the overall mean of the 40 cells? Correlated
significantly in more than 50% of the cells? 75%? 90%?
The second question has to do with the applicability of the ancova
procedure and perhaps it isn't actually an spss question per se. If
there are two primary applications of the ancova (baseline adjustment
and eliminating nuisance variables), both aimed at increasing the
power of the anova to pick up significant differences, then it
follows that one should always be aiming for smaller mse's, bigger
F's, smaller p's, and bigger effect sizes as the result of an ancova
relative to a simple anova. Is this correct? Is it ever appropriate
to use a covariate to eliminate or reduce a significant anova effect
or interaction? It seems strange to me to do this, since you can wipe
out effects just by adding random noise, but perhaps I am missing
something.
Thanks in advance,
Greg Shenaut
|
Locked
