It's a pharmacological intervention wherein we are trying to test an a priori hypothesis about decreased functional connectivity in a subset of networks involving the hippocampus in a 9.4T scanner.
For an experimental study to be considered parametric, and therefore have statistical power, you need to have at least 21 subjects. http://vassarstats.net/textbook/parametric.html
Sample size is determined mainly by the effect size and power to detect the change. The effect size is also determined by the minimum possible change that one can detect using the intervention/experiment you are doing. In many cases it is also determined by the sample variance. For fMRI studies, usually 20 or more subjects are reported. However, in many cases involving higher fields (> 3T) usually the significance level may be reached by lower numbers, although due to big data cross-site collaborations and improved statistical techniques, there are many studies that report more than 60 subjects per group. There are many studies with 400+ subjects. You could look at the recently published papers in the field pertaining to the question you are addressing to get an idea of the latest trends.
Thank you for the responses! I am sorry I inadvertently left the critical part of the question out -- I record resting-state activity in rats in a 9.4T scanner.
How does the number of subjects translate for this rodent study for at least a medium effect size (d > 0.5)?
I know of a study where they used 3 test v. 18 control monkeys in a split brain study (O'Reilly et al. 2013, http://www.pnas.org/content/110/34/13982.full#sec-8). I realize non-human primate studies are expensive but isn't 3 still too less to establish statistical significance?
I apologize for the inconvenience this may have caused.
Based on that example, I think it would very much depend on the statistics that were used. You could still get significance but standard parametric analyses might not be appropriate. I have quickly scanned the paper you linked to and they appear to have used Fisher's z test, which seems like a fine alternative. One other option in an uneven 'treatment'-control situation might be to use Crawford's modified t-test and treat each 'test' subject as an individual case and compare each against your normative control sample (http://homepages.abdn.ac.uk/j.crawford/pages/dept/pdfs/PREprints/Neuropsychology_dissociations_tests.pdf). I'm not sure if it is appropriate for fMRI. Alternatively, you might be able to use the reliable change index (RCI) often used in medicine, in pre-post intervention type studies (http://www.ncbi.nlm.nih.gov/pubmed/12206571). Again, it might be inappropriate for fMRI
- Badges
- Science topic
- Similar topics
- Medicine
- Disease
- Disease Models
- Animal Models
- Rodent Models