9 Experiment design
Given everything we know about the source of the fMRI signal, what’s the right way to design an experiment?
The easiest way to answer that question is to work backward from our plans to analyze the data. When we analyze the data, we are going to assume that the neural events — the things that we want to discover — are not directly detectable with fMRI. Instead, they a hemodynamic response (blood flow and oxygenation changes) that we can detect with T2 or T2*-weighted imaging. And now we know a little bit about that hemodynamic response — it’s slower than the neuronal response, hits its maximum level 5 or 6 seconds after the neural response happens, and then slowly returns to baseline after about 16 seconds, maybe undershooting a little bit.
This slow, indirect response isn’t always what we want to use to understand the brain. Functional MRI has unparalleled spatial resolution, for separating structures inside the brain (without cutting it open!). But the temporal resolution and the specificity of the signal are not as good as some alternatives. So first, we’ll pause and consider the field of neuroimaging more broadly, to set the context for designing an fMRI experiment.
Both the Google Slides and a YouTube of me talking through them are linked here.
So when we design an experiment, we design around this sluggishness, and we’re careful to present our stimuli in a way that we know we’ll be able to separate them out from each other after they get blurred together by the hemodynamic response. To estimate the fMRI signal we’re going to get from an experiment we use convolution. Here’s a brief introduction to convolution, if you haven’t worked with it before.
And now, knowing what convolution is, we can start talking about experiment design. The basic, classic designs are block and event-related. Block-design experiments present one stimulus or task for an extended period of time — 10 or 20 seconds — and estimate the average signal during the time you’re doing that task or looking at that stimulus. Event-related designs, on the other hand, are structured so that brief events happen, and you mix up the types of events, estimating a response to each type. Event-related designs are more like real life, but they produce smaller signals and place more demands on the experimenter to get their stimulus timing under careful control.
After block and event-related designs, there are several other ways to organize an experiment. Rest fMRI lets subjects just rest (either with their eyes open or with their eyes closed — there are 2 styles) and uses correlations in signal fluctuations in different brain areas to discover “connectivity” in the brain. This type of analysis assumes that there is such a thing as rest, and places strong constraints on how you process the data to be sure you don’t introduce unwanted correlations and come to the wrong conclusions! And every year, there are a few creative studies published in which the experimenter didn’t define the task, but let the participants explore a (virtual) space on their own (watch a movie, solve a maze, search for an object … that kind of thing). Behavioral data (button presses every time they found an object, for example, or eye movements) provide the regressors necessary for data analysis.
Exercises
A Colab notebook that simulates an experiment is here. Make a copy of it for yourself so you can edit it, or open it in Playground mode (both of those options are in the File menu). And then play around with the numbers until you can answer these questions
- Why is the fMRI response less than 1 if the simulated amplitudes were 1 and 2? What do you need to change about the stimulus to get the simulated fMRI response to be larger?
- How far apart do the stimuli need to be so that the response to one stimulus doesn’t affect the response to another? (You can answer this by looking at the simulated BOLD responses *before* the noise is added.)
- Do you think you can analyze the data even if the responses overlap?
- If the simulation was realistic and the responses are really that small, compared to the noise, then why is it possible to run an fMRI experiment?