 |
 |
Hello all... In reviewing the past posts regarding parametric analysis techniques, there is a lot of information on how to conceptualize and implement an fMRI parametric analysis using SPM. However, I'm wondering if anyone can shed some light into basic data considerations for how SPM treats the modulator variable within the GLM parameters. In other words, what constraints exist when setting up the parametric modulator? I'm asking because I am dealing with a specific situation: I recently received a manuscript review that questioned whether my parametric term was appropriate under basic GLM assumptions. My event- related model tested whether there was the linear effect of time interval between successive occurrences of an event class, suggesting greater amplitude of hemodynamic response the longer one had to wait before the next event in a class. For this task, in each of two sessions there were 24 target stimuli and 24 nontarget stimuli, presented pseudo-randomly over the course of ~8 minutes. I admit, the paradigm was not originally designed for this purpose. So when I worked out the model, I got approximately 15 unique 'bins' of interval-lengths for each of the two event classes. The intervals ranged from about 8 seconds to just under 60 seconds. Clearly, as you can guess by this math, only a few 'events' comprised each bin, with some bins having many more than others, but some having having only 1-2 events at each level. So my primary concern is whether a minimum number of events is needed at EACH level of the parametric modulation term during first-level (individual subject) analysis? I confess that this did not initially occur to me, perhaps because I had assumed that parametric analysis essentially fitted a linear (or quadratic, etc.) term through ALL the data, which would be more insensitive to levels with few observations. Conceptually, I envisioned this as similar to estimating the best fitting line through a big scatterplot of data. However, in retrospect I wonder if I was missing something. I can add, however, that we approached this having a rather large sample, which at least provides a measure of control at the between subjects level. Secondarily, I'm wondering what other GLM assumptions might apply to the parametric term at the first-level model, and how to test for them (i.e., homogeneity of variance, etc.) Any statistical insight or guidance would be greatly appreciated! Mike