Dear Ivan,
> I got a problem with the relative intensities of some regressors in my
> design matrix. I"m modelling an event-related study with events of
> different length in spm99. I'm using an hrf+dt basis set for a discrete
> event, and an hrf alone for an event of variable length (specified by
> using the "variable durations" option). The model looks fine, apart for
> one feature. The scale of the basis functions ranges from -2 to 2.5 for
> the discrete event (as seen in the "regressors" plot visualized by
> "Explore design"), and from -20 to 80 for the event of variable
> length. However, the range displayed in the "basis set and
> peristimulus sampling" plot is approximately the same for both types of
> events (i.e., from -2 to 2.5 for the discrete event and from 0 to 2.5
> for the event of variable length). The same difference is present in
> the actual design matrix, of course. This scaling discrepancy is then
> reflected, I guess, in the plotting of the estimated responses. When I
> plot both event types together, I got a flat line for the event of
> variable length, and a pronounced response for the discrete event, even
> if the voxel is arising from a t-contrast asking for differences
> between the former event and the latter one.
The reason why the regressors for the longer duration trials are bigger
then those for events is that the longer trials are modelled by a train
of events at every time bin (default = TR/16 seconds). This means that
a short epoch of 1 TR will have regressors that are [roughly] 16 times
the height of the corresponding regressors for the true event. The
parameter you are estimating, and making inferences about, is the
estimated response to events per time bin (i.e. event density). When
you come to plot event/epoch related responses these are for one event
only. For example if a trial that lasted for 1 TR evoked roughly the
same response as the true event the estimated response per event would
be 16 times smaller (because the epoch was 16 time longer). I am
surprised that you get voxels that are significant for epochs > events
because nonlinear interactions (neuronal and hemodynamic) will
compromise the response to longer events. I suggest that if you use
the variable duration option avoid mixing true events and small epochs
but enter a sensible duration for all trials.
> I was a bit surprised to find such a difference in the average values
> of the regressors of the design matrix, since in the past these were
> detrended (at least some partitions of it). However, I couldn't find
> trace of any such operation in the new code, but maybe I missed it.
> My questions are: i) shouldn't different columns of the design matrix
> have a comparable mean, so that they are modelling only different
> slopes and not different intercepts, too ?
The regressors do not need to be explicitly centered because the mean
is modeled as a confound.
> ii) am I right in assuming that the different relative intensities of
> the responses, as assessed in the plot section, are due to the
> different relative intensities of the regressors, in the design matrix
Yes and no. The estimated responses do not care about the scaling of
the regressors (the bigger the regressor the samller the parameter
estimate). The critical thing here is that you are plotting responses
to one event in a train of events, when using the event/epoch-related
option. Plotting the fitted responses per se will show the response to
be quite sensible.
With very best wishes - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|