Posted by B. Douglas Ward on November 19, 2002 at 10:35:40:

In Reply to: multicollinearity posted by Hackjin Kim on November 19, 2002 at 02:07:25:


Hello Hackjin:

With a block design, the estimated IRF is very noisy. Furthermore, the
consecutive IRF coefficients tend to be negatively correlated. Therefore,
I recommend summing the IRF coefficients, which approximates the "area under
the IRF curve", or AUC. The physical significance of the AUC is this:
The AUC represents the steady state response to a step function (assuming,
as usual, that the system is linear and time invariant). The IRF coefficients,
for each input stimulus function, can be summed using the -glt option.
Also, the contrast between the responses to different stimuli can be similarly
calculated as the difference between the sums of their respective IRF's.

For accurate estimation of individual IRF coefficients, I recommend using
a random binary sequence, or some other very randomized design.

Doug Ward