Dear Sylvia et al.,
I did a simulation to look at how the inclusion of varying numbers of
error trials would affect the variance of estimates of the real effect.
This was done over a simulated 400 second scan with 20 real events and
then a varying number of error events (using the canonical HRF). What I
did was create a model and then test it on noise data (with a mean of
1000) in order to see what the variance looks like. What you see below
is the variance of three model parameters (the task regressor, the error
regressor, and the mean) along with their SD in parens, across different
numbers of error trials. What you will see is that varying the number
of error trials has an obvious effect on the variance of the error trial
effect, but it has virtually no effect on the variance of the estimates
for the task or mean effects, even with only two errors trials over a
400 second scan. This is of course just one limited exploration, but it
does suggest that including a regressor with just a few trials should
not have much of an effect on the estimates for your conditions of
interest.
here are the results:
> nerr mean(sd)
> 0 -0.015 (0.826) -0.000 (0.000) 1000.000 (0.162)
> 2 0.006 (0.829) -0.031 (2.498) 999.997 (0.164)
> 4 -0.004 (0.824) 0.010 (1.768) 1000.000 (0.166)
> 6 -0.001 (0.834) 0.010 (1.433) 1000.000 (0.168)
> 8 0.001 (0.824) -0.006 (1.264) 1000.001 (0.168)
> 10 -0.022 (0.831) 0.009 (1.140) 1000.002 (0.170)
if you are intersted in playing with the simulation, you can get it
from:
http://www.nmr.mgh.harvard.edu/~poldrack/spm/error_model_sim.m
I'll be interested to hear what others have to say on this issue.
cheers,
russ
--
Russell A. Poldrack, Ph. D.
Assistant Professor of Radiology, Harvard Medical School
MGH-NMR Center
Building 149, 13th St.
Charlestown, MA 02129
Phone: 617-726-4060
FAX: 617-726-7422
Email: [log in to unmask]
Web Page: http://www.poldracklab.org
|