I’m running an amplitude modulated deconvolution using 3dDeconvolve, with TENT as the basis function. I’ve specified a 20 second window with 1 second radius (my TRs are 1 sec). The relevant part of the script for this looks like:
-stim_times_AM2 1 timing_file.1D ‘TENT(0,20,20)’ -stim_label 1 variable_name \
In my timing file, each stimulus is listed like this:
start_time_of_stimulusmodulator1,modulator2,modulator3:duration_of_stimulus (e.g., 52.1,4.3,3.8:1.2)
The output for each variable contains 100 beta weights. I’m seeking confirmation on my interpretation of these coefficients please. Here’s how I’m interpreting them based on the documentation:
0 -19: unmodulated (mean) response that depends upon the stimulus duration, at each of the TENT functions
20 - 39: response modulated by modulator1, at each of the TENTs
40 - 59: response modulated by modulator2, at each of the TENTs
60 - 79: response modulated by modulator3, at each of the TENTs
80 - 99: This is the one I’m unsure about. I thought it might be the response modulated by duration, but my understanding is that the first set of coefficients incorporates the duration. Can you please offer guidance about this last set of beta weights, or whether the first set does not incorporate duration? Thank you.
Your interpretation seems correct.
Note that TENT(0,20,20) will not be TR-locked. You probably want TENT(0,20,21), and Gang would suggest using TENT(1,19,19), to omit time points for which you expect the betas to be zero.
The 20-second interval from 0 20 needs 21 betas to include the end points, measured at times 0…20 (so 21).
As you thought, the duration modulation part of the stimulus is indeed being treated as amplitude modulation, because it is not possible to use DM with TENTs. Doing so would mean either spacing the 21 knots at varying time points or using a variable number of knots, either of which would alter the interpretation of the betas across events. The program does not give a warning or error, because it has a liberal interpretation of the punctuation, applying as “seems appropriate”.
Thank you for your answer; it was helpful. It sounds like unless I have hypotheses that the duration of my stimuli would impact the amplitude of the BOLD response, it may be better to remove duration from the model, as the deconvolution TENT functions will account for narrower/wider HRF functions by nature of being data-driven (as opposed to a convolution, where I might want to include the duration even if I don’t expect it to modulate amplitude).
Regarding the first 20 beta weights (0 - 19 in the initial question), can you please confirm that the unmodulated response is the overall response (aka the contribution from the modulators is included)? Or does the unmodulated response subtract the contributions from the modulators and provide the remaining mean response?
Yes, the first set of 20 betas (the unmodulated ones) should basically be the main effect. Note that any modulation values are first demeaned before being used to create additional regressors. So those regressors will be approximately orthogonal to the main ones, leaving the main ones to reflect the consistent, overall effect.
And to be clear, your hypothesis is that duration will affect amplitude above and beyond the amplitude of the convolution? The convolution amplitude effect has a unit height regressor for brief events (e.g. around 1s) convolving up to a height of more than 5 for long events (e.g. 12-15s+). It is not just a shape change. Just to be sure…