Adding a random number really does nothing. Best to remove the pmod if its truly constant. The pmods are estimating the slope of change in response per unit change in pmod.

It would be best to concatenate the runs, but you have to make sure that each run as separate constants, filters, and auto-regressive terms. You can look at the gPPI code for how concatenation can be done across runs.

The third approach is probably the easiest.

Best Regards,

Donald McLaren, PhD

On Fri, Feb 10, 2017 at 5:29 PM, Jerry Zhu <[log in to unmask]> wrote:

I have seen people adding a tiny random number to pmod if pmod remains constant, thereby "bypassing" the restriction. But I am not sure how accuracy this would be.

Also, if there are several runs for that task, if pmod is constant in one run, but not in the other runs. I am wondering if researchers can combine/concat the runs so that overall you can have varied pmod.

A third alternative might be: treat different levels of pmod (like ratings 1,2,3) as different regressor. For example Emotion1, Emotion2, Emotion3, Neutral1, Neutral2, Neutral3.

On Wed, Nov 23, 2016 at 10:28 AM, Mike <[log in to unmask]> wrote:
Dear Mclaren,

Thank you for your response.

By the way, I wonder why SPM uses an orthogonalization strategy to deal with parametric modulator (pm), I mean, the first pm explains the rest variance unexplained by the main regressor, and then the second pm explains the rest variance unexplained by both the main regressor and the first regressor (i.e., main regressor, first pm, and second pm are all orthogonal to each other, right?). Why not just use a multiple linear regression model, like the one in the multiple regression in the second-level model in SPM?

Mike