Frequently Asked Questions - Jitter
1. What is jittering?
It's the practice of varying the timing of your TR relative to your stimulus presentation. It's also often connected to, or even identified as, the practice of varying your inter-trial interval. The idea in both of these practices is the same. If your TR is 2 seconds, and your stimulus is always presented exactly at the beginning of a TR and always 10 seconds long, then you'll sample the same point in your subject's BOLD response many times - but you might miss points in between those sampling points. Those in-between points might be the peak of your HRF, or an inflection point, or simply another point that will help you characterize the shape of your HRF. If you made your TR 2.5 seconds, you'd automatically get to sample several other points in your response, at the expense of sampling each of them fewer times. That's "jittering" your TR. Alternatively, you might keep your TR at two seconds, and make the time between your 10-second trials (your inter-stimulus interval, or ISI) vary at random between 0 and 4 seconds. You'd accomplish the same effect - sampling many more points of your HRF than you would with a fixed ISI. You'd also get an added benefit - you'd "uncover" a whole chunk of your HRF (the chunk between 10 seconds and 14 seconds) that you wouldn't sample at all with a fixed ISI. That lower portion can help you better determine the shape of your whole HRF and find a good baseline from which to evaluate your peaks. This added benefit is why most people go the second route in trying to "jitter" their experiment - varying your ISI gets you all the benefits of an offset TR, plus more.
2. Why would I want to do this in my experiment?
If all you care about is the amplitude of your response, you probably wouldn't. In this case, you're assuming a certain shape to the hemodynamic response, and all you care about is how "high" the peak of the HRF was at each voxel for each condition. You'd want a design with very high statistical power - the ability to detect amplitude. On the other hand, you might not want to assume that every voxel had the same HRF shape for every condition. If you'd like to know more about the shape, without assuming anything (or less than everything, at least), you need a design with high statistical efficiency - the ability to accurately estimate shape parameters, without assuming a shape. See Liu and Dale (JitterPapers) for more on the tradeoff between power and efficiency.
Varying your ISI is a strategy to increase the efficiency of your estimates at the expense of your power. Clearly, because you'll be sampling each point of your HRF fewer times, you'll necessarily have less confidence in the accuracy of any given estimate. But because you'll have so many more points to sample, you'll have much more confidence about the true shape of your HRF for that condition. This is critical when you believe your experiment may induce HRFs of different shape in different region, or if knowing the shape (lag, onset time, offset time, etc.) of your response is important (as it is in mental chronometry). With a variable-ISI design, you can run a rapid event-related design and pack many more trials into a given experimental time than you would for a fixed-ISI design that sampled the whole HRF, or you can sample much more of the HRF than a fixed-ISI design could with the same number of trials.
3. What are the pros and cons of jittering / variable-ISI experiments? When is it a good/bad idea?
Variable-ISI experiments are a way of making the tradeoff between power and efficiency in an experiment. Fixed-ISI designs are extremely limited in their potential efficiency. They can have high power by clustering the stimuli together: this is a standard block design. But in order to get decent efficiency in an experiment, you need to sample many points of the HRF, and that means variable ISI. Any experiment that needs high efficiency - say, a mental chronometry experiment, or one where you're explicitly looking for differences in HRF shape between regions - necessarily should be using a variable-ISI design. By contrast, if you're using a brand new paradigm and aren't even sure if you can get any activation at all with it, you're probably better off using a block design and a fixed ISI to maximize your detection power.
From a psychological standpoint as well, the big advantage of variable-ISI designs is that they seem far more "random" to subjects. With a fixed ISI, anticipation effects can become quite substantial in subjects just before a stimulus appears, as they catch on to the timing of the experiment. Variable ISIs can decrease this anticipation effect to a greater or lesser degree, depending on how variable they are. Liu (JitterPapers) explores the conept of "conditional entropy" as a measure of randomness in how an experiment "seems," and it is, predictably, intertwined with the power/efficiency tradeoff.
4. How do I decide how much to jitter, or what my mean ISI should be?
Great question. Depends a lot on what your experimental paradigm is - how long your trials are, what psychological factors you'd like to control - as well as what type of effect you're looking for. Dale (JitterPapers) lays out some fairly intelligible math for calculating the potential efficiency of your experiment. Probably even easier, though, is to use something like Tom Liu's experimental design toolbox (JitterLinks) or, even better, AFNI's 3dDeconvolve -noinput option, which will take a given experimental paradigm and calculate how good it is in terms of power and efficiency. Once you're within the ballpark for the type of paradigm you like, these tools can be an invaluable way to optimize your design's jitter / ISI variation, and are highly recommended for use.
5. But how do I get better temporal resolution than my TR?
Simple: don't always sample the sample points of your response. If you always sample the BOLD response 2 seconds and 4 seconds and 6 seconds after your stimuli are presented, for your whole experiment, you'll have a very impoverished picture of the shape of your HRF. But if, for example, you sampled 2 sec. and 4 sec. and 6 sec. post-stimulus for half the experiment, then cut one second between trials and sampled 1 sec. and 3 sec. and 5 sec. for the rest of your experiment - why, then, you'd have a better picture. The cost, of course, is reduced power and expanded confidence intervals at the points you've sampled.
With a good picture of the shape of your HRF, though, you could then compare HRFs from two different regions and see which one had started first, or which one had reached its peak first. If HRF timing is connected in some reliable way to neuronal activation, you then don't need to sample the whole experiment at a super-fast rate - you could infer from only a limited-sample picture of one part of the HRF where neuronal activity had started first and where it had started second, which allows you to rule out certain flows of information.