If I understand your question, many people use either 1 or 2 STD. This is common in determining settling times of variable that have some inherent noise. Just eyeballing the data you should be able to say that this range captures the values you would pick if you were doing it by visual inspection.

It depends on the quality of your data. If you have a lot of normal adult subjects you can use the strictest thresholds to reach clean data for statistical analysis. 

If you have pediatric or patient data set you have to achieve a balance between data clearing and saving enough data for statistical analysis. That's why we don't have definite recommendations. Just look over your data and choose the strictest threshold allowed by the data. 

I agree with Roza's answer. It depends on the data and what inherent amount of noise you would expect from your population. Adults can be really still; children move a lot and forget not to wiggle in the scanner. 

Another important consideration is that movement and noise are correlated. Movement inherently produces changes in the BOLD signal. So, typically, scans that show movement are also the same ones that show dramatic changes in global signal. When considering movement, it is important to keep in mind that the movement that really counts in the scan-to-scan delta, rather than the global movement from the initial scan. The global change is often caused by normal and expected "drifts", like the slow vertical sinking of a subject's head into the padding or drift in the scanner's reference frame. Often, even very large cumulative drifts pose little or no problem, because the movement from scan to scan is minimal. As a rule of thumb, scan-to-scan movement > 1mm in any direction already produces visible artifacts in the corresponding image. Adult subjects are typically very good at moving less than this; most of my subjects move lass than 0.3mm. Remember, this is the delta movement from scan to scan, which you can visualize on ART by checking the appropriate checkbox. You might also want to use a composite movement index, rather that the six displacement indices (x, y, z, roll, pitch, yaw) separate. Once the outlier movement scans are identified, typically you have also identified most of the outlier global intensity scans. In general, I see no point in using a global threshold of 2 STD or less---these thresholds are already very strict. I have always used, and recommended to my students, to use 3 or more.  

Amazing! Thanks a lot, it really helped! We have adults with ADHD, so they do move a lot. We now used 3std and 1mm movement which seems fine for most subjects (although most have between 5-10 outliers) and two subjects have to many that its seems impossible to extract meaningful regressors, we are considering to scrap them.

Dear all, I found your responses very helpful, thanks!

I would like to ask you further clarifications by considering your expertise. 1) do you confirm me that I can enter the smoothed images?

2) As not composite motion measure I would like to use your suggested 1mm with a rotation of 0.02. I have groups of different psychiatric patients. Is 0.02 a suitable threshold? (I do not use composite motion by considering that 1 mm is a stricter threshold probably due to its calculation (Euclidean distance).

3) It is not totally clear to me if I have to flag the 'Use diff. global". Do you have any suggestions? Results deeply change by considering or not this parameter. I would not use it.

4) Finally, what about removing subjects which have more than 20% of outlier volumes? Is it a good strategy?

Thank for your help!