Print

Print
Helmut, Excellent points. After your in depth response, I simulated some data to see what impact trial averaging has on the accuracy of the estimates. 

The basic summary is, as before, the more variable the trial-to-trial response is the more trials you need. The more noise in the data, the more trials you need to get a stable response. After 25-36 trials, the variance from a qualitative inspection, looks pretty stable. This analysis was done assuming a HRF and used short, jittered ITIs based on optseq2.

Based on this recent simulation, a target of 30 trials seems reasonable.



Best Regards, 
Donald McLaren, PhD


On Fri, Aug 21, 2015 at 12:03 PM, H. Nebl <[log in to unmask]> wrote:
Dear Mike, dear Donald,

One comment on the Huettel & McCarthy paper, as this is quite important: They estimate the shape of the hemodynamic response, and they conclude that after >= 25 trials the estimate stabilizes. But also see the discussion:

The appropriate number of trials for an fMRI experiment
depends on its goals, whether activation detection or HDR
estimation [11]. Our results suggest that, if an experiment
is intended to detect whether there are active voxels within
an area, then a relatively small number of trials (<20) may
be sufficient. For example, detectible changes in fMRI
activation have been reported with single trials [12]. But, if
the experiment is intended to determine the spatial extent
of activation by detecting all or nearly all active voxels,
then many more trials (>100) should be averaged, far
more than typical for fMRI studies. Conversely, if the goal
is to estimate the HDR from a identified region, then
relatively fewer trials (25-36) are required.

As long as you're interested in group results only, based on some pre-defined hemodynamic response function like the one in SPM used for convolution, it should thus be okay to go with a small no. of trials (as their aspect on spatial extent should mainly be an issue for single-subject models, as you forward the beta estimates into second level statistics, ignoring the variance, and not the T maps).

However, also note that their interstimulus interval varied randomly between 14 and 18 s. As long as your predictors are reasonable this should also hold for shorter ITIs, but if the ITIs become too short then you cannot properly estimate activation levels for the conditions any longer (cond1 vs. baseline) as the BOLD response never turns back to baseline, while the differences between conditions can still be estimated (cond1 vs. cond2). Leaving aside that for very short ITIs succeeding trials might interact, and the resulting BOLD responses might not superpose linearly any longer, making your predictors suboptimal.

Another aspect to consider is error rate. Depending on task you should ensure that you have enough correct trials at the end of the day, it won't help if there are two correct responses and 28 incorrect ones that might have to be dropped. ;-)

Best

Helmut