Dear SPMers,
Thank you in advance for any advice you can offer. Briefly, my question
concerns how to properly handle task parameters that are correlated with,
but theoretically extraneous, to a parameter of interest. In our
blocked-design, each block gets a value based on the content (level of
abstraction) of the participant's responses. However, the length (word
count) and latency of their responses are correlated with the content. Our
goal is to conclude that these extraneous parameters do not explain the
effects we observe when examining regions that are modulated by our
parameter of interest.
My intuition was to simply model these extraneous parameters (length and
latency) in the same way that I modeled the parameter of interest. This
yields the following simple characterization of our model:
Task + POI + EP1 + EP2
Where POI is the parameter of interest and EP1 and EP2 are the extraneous
parameters. When I implemented this, the parameter for EP2 could not be
estimated (black bar on the design matrix; says "beta not uniquely
specified"). I centered all parameters; the correlation between EP2 and EP1
is low (.13), while the correlation between EP2 and the POI is quite high
(-.88).
Which yields the following questions:
1. Is the above procedure correct? If so, why can't the beta for EP2 be
estimated?
2. If the above is incorrect, or if there is a better procedure, then what
is it?
Thank you in advance for your help!
Sincerely,
Bob Spunt
|