Hi Claire
With a breath hold protocol (eg events ~20 secs), I think that you don't have the choice & that using boxcar is the more straightforward solution. Fitting your own model might be very difficult because your hypercapnic events (HC) are very short. If you plot your raw data, you will see that the HC is extremely variable, much more than a neurally triggered BOLD response.
An additional problem with boxcars when you model HC responses (from onset to offset of breath hold) is that the uptake and washout periods are badly modelled.
Uptake and washout periods can be very long (>30secs if your subject can hold his/her breath for 2 mn ;-) and will also vary a lot from one trial to the other depending on factors that are difficult to control (eg because of saturation of the HC response in time, quality of breath-hold, duration of breath-hold ...).
I therefore strongly suggest to regress out at least the washout period from the baseline using dedicated regressors, or simply by throwing volumes.
I would also add a parametric modulator for time. Again if you plot your data and compare the first HC event with the last HC event of your protocol, you should see that the former is weaker than the latter.
To minimize all these problems we use long blocks of ~3 to 5mn here.
Subjects
wear a loose fitting mask which delivers ambient air for 1mn30, then a
mix of air & 5% CO2 (2mn), then ambient air again for 1mn30.
We
model the uptake & washout periods using dummy regressors (1
regressor per time point) so that they don't fall in the baseline (the
unmodeled part of the GLM). Having a tidied baseline is essential if you want to have a good contrast with your HC response.