Daniel,

Please see inline responses below.

On Wed, Dec 11, 2013 at 10:12 AM, Daniel Weissman <[log in to unmask]> wrote:

Dear SPMers,

I'd like to run an event-related fMRI study to investigate the neural bases of sequential effects involving two trial types: A and B (each trial lasts 2 seconds and involves a stimulus for 300 ms, followed by the subject's response). To this end, I'd like to employ a first-order counterbalanced sequence to produce equal numbers of four sequential trial types.

1) A preceded by A

2) A preceded by B

3) B preceded by A

4) B preceded by B

Finally, I'd like to use a constant inter-trial-interval (ITI), because sequential effects in my task will likely change with the ITI. This is a bit different from my normal practice of jittering the ITI and so it raises a few questions in my mind.

1. Although the absence of jitter will make it nearly impossible to observe the common effect of all conditions vs. baseline, my understanding is that I will still be able to perform contrasts involving the 2 main trial types (e.g., A minus B) as well as the 4 sequential trial types (e.g., A preceded by A minus A preceded by B). This is because differences between beta values are relatively stable when no jitter is present, even though the individual beta values themselves are highly unstable. Is this correct?

The absence of jitter reduces the ability to determine the amplitude of the events because you can lead to an underdetermined model and unstable estimates. However, given the fact that you have 4 trials types, I don't think this will be a large concern, especially if you are assuming an hemodynamic response.

If you can't get stable estimates of the 4 trials types, then the subtractions will also be unstable.

One way to insert jitter into this design is to insert miniblocks --> ABABABABAB pause BABBABABAB pause. The pauses help insert the jitter. Another alternative is to use an ITI that is not matched to the TR. This way each trial starts at a different point in the TR leading to sampling the response at different timepoints throughout the response.

 

If so, then, at the nuts-and-bolts level, would I model the 4 sequential trial types above (i.e., A preceded by A, A preceded by B, B preceded by A, B preceded by B) and then form contrasts based on the resulting betas? If so, then would the A - B contrast be formed by contrasting 

AVERAGE (A preceded by A, A preceded by B) 

MINUS

AVERAGE (B preceded by A, B preceded by B)?

Yes. This is how to form the contrasts.

 

2. To be able to look at the common effect of all stimuli vs. baseline (e.g., to functionally define ROIs for subsequent orthogonal contrasts, I've been considering breaking up the run into alternating periods of task and rest as suggested by Rik Henson (2006, Chapter 15 from Human Brain Function, I think). As Rik states on page 209 of this chapter:

"Another problem with null events is that, if they are too

rare (e.g. less than approximately 33 per cent), they actually

become ‘true’ events in the sense that subjects may be

expecting an event at the next SOA and so be surprised

when it does not occur (the so-called ‘missing stimulus’

effect that is well-known in event-related potential (ERP)

research). One solution is to replace randomly intermixed

null events with periods of baseline between runs of

events (i.e. ‘block’ the baseline periods). This will increase

the efficiency for detecting the common effect versus

baseline, at a slight cost in efficiency for detecting differences

between the randomized event-types within each

block."

In this chapter, Rik doesn't refer to previous event-related studies that have employed this approach. Do you know of any?  

I am not aware of studies, but this is exactly what I suggested above. You will lose a few trials that don't have any preceding trial, but that shouldn't have a huge impact.

Also, if I employed Rik's suggested approach, does the following sound reasonable? I'd consider breaking up the 96-trial-long sequence in each run into four 24-trial-long sequences and then separating each of these 24-trial-long sequences with a fixation block. Since trial duration is about 2 seconds, this would amount to alternating between 48-second-long task blocks and (I was considering) 20-second-long fixation blocks. While this is far from the most efficient block design (i.e., 16 seconds on, 16 seconds off), I'm thinking it may still allow me to observe the common effect of all stimuli vs. baseline as well as the contrasts between different sequential trial types discussed earlier. Does this sound right?

I think that would be fine. Just remember to have fixation at the beginning and end of the run as well. I also think you need more than 96 trials. I would say you need at least 128 analyzable trials + the trials at the beginning of each block. If you only want to use correct trials, then you will need more than the 128 trials. 128 trials is 32*4.

 

3) Adding one more level of complexity to the design above, I may also wish to model with different regressors different trials from a given trial type (e.g., A preceded by B) depending on the speed with which a subject responds. For example, I might want to model the 20% fastest trials separately from the 20% slowest trials within each trial type. In general, event-related fMRI allows for such "post-hoc" creation of trial types, but I just wanted to check whether anyone sees a problem with this procedure in the designs I am asking about.

This isn't necessarily a problem; however, the post-hoc shorting will leave you vastly underpowered. At 128 trials (or 32 per event type), 20% would be 6-7 trials - which is not enough to be able to get a stable estimate of the BOLD response. Also, depending on the task, you might want to simply model RT as the duration of the trial. If you want to look at response speed, then you would need 640 trials to get enough power. This would be 32 trials in each trial type quintet. 

As an aside, using 0 second duration prohibits any PPI analysis on the data.

 

Finally, if you've read this far, THANK YOU!!!!!

Hope you find this useful.

 

Best regards,

Daniel

--
Daniel Weissman, PhD
Associate Professor
Department of Psychology
University of Michigan
1012 East Hall
530 Church Street
Ann Arbor, MI 48109