Event-related responses in pupil size and fMRI measurements
It’s essential to understand the signals you’re looking at. The standard approaches to fMRI analysis assume the haemodynamic impulse-response function has a specific shape. Although slight changes in its timing and shape can be accounted for, it is nonetheless important to inspect the shapes of the event-related responses in one’s data. One way to correctly fit these event-related responses is to use linear regression to estimate the finite impulse-response caused by specific types of events.
Finite impulse response fitting has some advantages over and above event-related averages. If the impulse-response your data were filtered by is slow, i.e. a low-pass filter, it may be the case that responses to multiple events overlap. This causes problems for epoch-based event-related averaging, since some parts of the data will end up in the event-related responses of several different event types. Provided enough events of all types are present in the experiment (aided even more when inter-event periods are randomly drawn from an exponential distribution), deconvolution can take into account this overlap and correctly attribute the signal to specific types of events.
We have created a python class package, fir, to do this. This package makes it very easy to perform quite sophisticated analyses, distilling event-related responses from timeseries signals. See below for a brief example, for more elaborate examples, see this IPython notebook on GitHub
The package can be installed by issuing
pip install fir --pre, where the
--pre allows you to install software that’s still in development – as fir is.
And this code might return the type of curves shown in the above figure. We encourage you to give this package a try when you have fMRI data from voxels, ROIs, or pupil size timeseries data.