¶ Terminology
¶ General MRI terminology
For a good and brief introduction on the terminology used in fMRI, please see the book by Huettel et al on Functional magnetic resonance imaging Chapter 8 box 1.
¶ Subjects, participants, patients or volunteers?
There is some debate on whether subjects, participants or volunteers are the proper terminology. There is a similar discussion on when and when not to use the term patients.
Subject is the traditional term but has gone out of fashion because the term emphasises the passivity of the individual who is taking part in research.
Participant (or volunteer) on the other hand emphasises the autonomy of the individual and highlights the voluntary and active nature of their participation.
In our line of research, all individuals volunteer and actively give their permission to use their data. Thus, the term participants is more appropriate.
Note that permission is not obtained in all research, for example historical data, animal data, or when harvesting data on human behaviour on the internet.
For a helpful discussion on the matter see: What's in a name? Research"participant" versus research "subject" and Why We Need to Keep the Term "Research Subject" in Our Research Ethics Vocabulary.
¶ Participant design considerations
¶ Number of participants
We aim to get our effect significant in every single participant. This is in contrast with most neuroimaging studies where effects only reach significance when averaging across participants.
Therefore, the number of participants is needed to build confidence that your effect is reproducible across participants, i.e. participants are replication units not measurement units.
For discussion in defence of small sample sizes (provided strong measurements) see Smith and Little, Psychonomic Bulletin & Review, 2018 and Normand Front. Psychol. 2016. Another way, to put it is that we conduct focussed studies that maximise signal and minimise noise to inform us about effects in individuals, which is ultimately required to guide clinical care (Gratton et al. Neuron 2022).
To get significant results in single participants, we may scan the participant more than once. Thus we focus on sufficient trials per participant rather than sample size.
For a discussion on the relationship and influence on statistical power between trials and sample size see Baker et al. Psychological Methods, 2021., basically statistical power is a tradeoff between number of trials per participant and number of participants. So studies with fewer participants but more trials can have the same statistical power as more participants with fewer trials.
Typically, we aim for about 10 participants. This is in line with many previous studies that use comparable experimental (model-based) approaches, for example Kay et al, Nature, 2008 (2 participants), Dumoulin and Wandell, Neuroimage, 2008 (6 participants), Huth et al, Nature, 2016 (7 participants), Harvey et al, Science, 2013 (8 participants).
Furthermore, our experiments have build in replications. For example, the notion that we have visual field maps and pRFs have been replicated literally over a 100 times. Most of our studies build upon other studies. In addition, we have build in replications, e.g., Harvey et al, Science, 2013 replicated the numerosity maps in the same participants on 5 separate days with 5 different stimuli configurations.
¶ Yourself or colleagues as participants
There are several reasons why you (or your colleague) would not participate in your own experiment:
- Concerns about the validity of the work. Basically, are you cheating to get the result you want (consciously or unconsciously)?
- Concerns about undue pressure. Basically, are you pressured by your peers or supervisor to participate?
- Concerns about disclosure of personal health data. Basically, you may not want to participate because it reveals health conditions that you do not necessarily want all your colleagues to know.
The first concern is (mostly) solved by proper experimental design. For example, the computer randomises conditions and brain responses cannot be faked (e.g. you couldn't make a numerosity map in your brain even if you really wanted too). But, in more complex design, for instance deception tasks, this is an obviously a problem. You can also deal with this issue by showing individual participant results. Naive and non-naive participants should not differ.
The second and third reasons are valid however. To prevent undue pressure, colleagues who want to participate in each others experiments need to register in the Sona system.
There are several reasons why you (or your colleague) would participate in your own experiment:
- Educational. You only really understand your experiment if you participate in it yourself. This is especially true for beginning MSc and PhD students. Participating in an experiment is absolutely necessary if you want to design your own experiment and explain it to other volunteers.
- Design. Participating in your own experiments is the best way to catch design or stimulus flaws.
- Motivation. Basically, authors and colleagues and most attentive and disciplined. They completely understand the necessity of good data. They typically produce the best behavioural and fMRI data. There are notable exceptions in both directions however.
There are many papers where the co-authors are participants.
- Most of our papers are, as are most conventional psychophysics papers and fMRI papers who follow similar methods. In the old days, participant initials were used, which makes detection of author participation clear. Often it is explicitly mentioned. Due to privacy reasons this is not allowed any more.
- The most famous example that involved the author is Nobel laureate Barry Marshall for the demonstration that the bacteria Helicobacter pylori was the cause of stomach ulcers. See review paper by Ahmed, Ann Clin Microbiol Antimicrob. 2005: "Marshall underwent gastric biopsy to put evidence that he didn't carry the bacterium, then deliberately infected himself to show that it in fact caused acute gastric illness."
¶ Stimulus design considerations
¶ Keep it simple
We aim to keep most of our experiments as simple and as robust as possible (following the KISS principle). For example, many (if not most of our) experiments have been successfully executed using bar-shaped stimuli (introduced in Dumoulin and Wandell, Neuroimage, 2008) and dot-stimuli (introduced in Harvey, Klein, Petridou, Dumoulin, Science, 2013), with the tasks involving simple red-green or black-white judgments.
In most of these experiments, the effects sizes are large enough to be observed by eye from the raw fMRI data in individual subjects. If you need statistics to prove a difference, the difference is likely too subtle to be meaningful.
¶ Start the stimulus before data acquisition
The hemodynamic response function lasts about 20 to 30 seconds. Therefore, we start the stimulus before the fMRI data acquisition in order to avoid startup transients due to the hemodynamic response, stimulus and task. This way you'll record a steady state response. This is critical for traveling wave or phase-encoded designs but also for other designs. To achieve this, you can also start them both simultaneously and later remove the first volumes from your fMRI data-set.
¶ MRI design considerations
¶ Length of fMRI session
We found that 1.5 hour is a good maximum length of an (f)MRI session for healthy participants. Any longer starts to get uncomfortable and the ability to lie still will decrease. If the participant is uncomfortable, this will result in more motion artefacts, poorer performance and ultimately data loss. Also, the subject is less likely to volunteer again.
Shorter is better especially for naive subjects (if you want them to return) and patients. In this case, we recommend keeping the session length under one hour.
It is not unusual that delays at the start will push your fMRI session over the recommended length. Be prepared to decrease the amount of fMRI scans, in order to keep the length of the fMRI session 1 to 1.5 hours. Therefore, it's important to consider which fMRI scans you deem most important and which less important that you can potentially skip.
¶ Length of fMRI scan
Every fMRI scan is typically 4 to 6 minutes. Several shorter scans are preferred over fewer longer scans.
It's not unusual to lose scans within a session, for example due to subject fatigue or subject motion. The most conservative approach is to remove those scans from the analysis. Using several shorter scan is more robust to data loss than fewer longer scans.
Especially, when using a difficult task, we found that 4 to 6 minutes is doable for the subject. For longer scans it's difficult to maintain the same level of task performance throughout.
Having several scans also allows for cross-validation and computation of noise-ceilings. These are becoming more popular to characterise the goodness-of-fit of the models and characterisation of the (maximal) signal-to-noise.