Hi Tiffany,
Ideally, an experiment with sequentially dependent stimuli should include catch trials, with only stimulus A. These would reduce the covariance between the regressors for stimuli A and B and improve estimation of contrast A-B. Assuming your design does not include catch trials, you could try:
1) Contrasting A and B using a parametric modulator
Enter both stimuli in a single regressor, and include a modulator that is +1 for A and -1 for B. The resulting design matrix will then comprise two orthogonal regressors, with the main regressor testing for common activity for A and B and the modulator testing for differences between A and B after accounting for common activity.
2) Use a FIR model
In this case, responses are modelled over a fixed time window, relative to a set point in time within each trial. You can then run tests between the time bins in that window. For example, you could define your window as starting with the first stimulus, and having a width spanning both stimuli, and then test for differences between stimuli within that window. In your case, the delay between A and B is variable, so implementing a FIR model would be tricky. If you had, say, three different intervals between A and B, you would have to model each interval separately.
Having said all this, if SPM estimates your model, and the results make sense, you have a win. If I understand the posts made by people knowing more than me on the list correctly, the pseudoinverse correction for correlated regressors will impair estimation of contrasts of each regressor alone (e.g. [1 0] or [0 1]) but not against each other [1 -1], which is what you want to test.
Hope this helps,
Paul Wright
--
Paul Wright PhD
Centre for Speech, Language, and the Brain
Dept. of Experimental Psychology Tel: +44 (0)1223 766559
University of Cambridge Fax: +44 (0)1223 766452
Downing Street email:[log in to unmask]
Cambridge http://csl.psychol.cam.ac.uk
CB3 2EB
On Mon, 16 Aug 2010 21:11:59 -0400, Tiffany Huang <[log in to unmask]> wrote:
>Dear SPMers,
>
>In my experiment, I have two stimuli presented before subjects making a
>response, stimulus A and B. Stimulus A always appears before stimulus B by a
>variable period, say a couple seconds. Because I am interested in the
>difference between the two stimuli, I entered a column of event time for
>stimulus A and and another for stimulus B in the same design matrix. Before
>defining a contrast, I calculated the correlation coefficient between the
>two regressors (stimulus A and B). It was highly correlated (r~=.6 and
>p<.000). Here are my questions:
>
>1. Is this design still making sense for the contrast that I am interested
>in (i.e., A-B) given such a high correlation between the two regressors?
>
>2. To determine whether two regressros are dependent or independent, do I
>just use the p value or is there any common values in the coefficient that I
>should look for?
>
>3.In general, how do dependent regressors affect the model and results?
>Should I get rid of dependent regressors even though SPM was able to
>estimate the beta?
>
>Any help on any of these questions would help. Thanks in advance.
>
>Tiffany
>
|