Functional MRI is widely used to map human brain function, but anatomical variability across individuals can hinder traditional group analyses. Standard random-effects (RFX) group approaches often require very large samples to produce stable results because they do not explicitly account for inter-individual topographic differences. The authors evaluate an alternative analysis strategy, group-constrained, subject-specific (GcSS) analysis, which combines group-derived spatial constraints with subject-level fROI definitions to accommodate topographic variability while retaining spatial commonalities.
The study analyzed sixteen contrasts drawn from four Human Connectome Project tasks covering domains of language, social cognition, motor, and working memory. Sample sizes spanned N = 10 to 450, enabling assessment of how key GcSS outputs change across a wide range of dataset sizes. The source reports that analyses were performed on these contrasts to quantify effect-size differences and reliability metrics for group and subject-specific outputs.
The GcSS framework uses group-level information to define anatomical or functional parcels that act as constraints for defining subject-specific functional regions of interest (fROIs). In this two-stage approach, (1) group probabilistic maps identify spatial regions commonly engaged by a contrast, (2) group-level parcels are derived from those maps and used as anatomical constraints, and (3) subject-specific fROIs are defined within those parcels for each participant to estimate task-related responses. This design explicitly models inter-individual variability in the precise locations of functionally specific regions while leveraging group-level spatial similarity.
The authors focused on three GcSS outputs: (1) group probabilistic maps for each contrast, which indicate the probability that a given location is engaged across subjects; (2) group-level parcels, derived from probabilistic maps, that serve as anatomical constraints for defining subject-specific fROIs; and (3) the effect sizes of fROI responses to task conditions as estimated within individual subjects.
Reliability of the group probabilistic maps increased with sample size. For most contrasts, probabilistic maps reached 'good' reliability at N = 100 as measured by intraclass correlation (reported ICC = 0.75) and achieved 'excellent' reliability at N = 200 (ICC = 0.90). Group-level parcels showed substantial spatial agreement by N = 100, with Dice coefficients reported around DC = 0.80 for most contrasts. These metrics indicate that the spatial constraints needed for subject-specific fROI definitions stabilize once moderate group sample sizes are attained.
A central finding is that, after establishing robust group parcels, the subject-specific portion of the analysis can be carried out with much smaller samples. The authors report that with group parcels in hand, samples as small as N = 10 participants yield accurate effect-size estimates within subject-specific fROIs, detecting approximately 80% of practically meaningful effects defined in the study as effect sizes ≥ 0.2. Increasing the sample to N = 20 improves detection to about 90% of such effects. These results were observed across the contrasts examined.
The observed scaling patterns were not limited to cortical regions: the authors report that similar reliability and sample-size relationships held for analyses in the cerebellum. Thus, both cortical and cerebellar GcSS analyses followed comparable sample-size scaling properties in the contrasts and tasks studied.
The findings challenge the broad assertion that fMRI research universally requires very large samples for generalizable results. Instead, the study suggests a two-tiered strategy: use a moderately large sample (e.g., N ≥ 100) to define robust group-level parcels and probabilistic maps, and then conduct subject-specific fROI analyses with much smaller samples (e.g., N = 10–20) to estimate effect sizes that generalize. This approach leverages the stability of group-derived spatial constraints while preserving statistical efficiency in subject-level estimation.
The article is a preprint posted on bioRxiv and has not been certified by peer review. The abstract reports key numerical reliability and agreement metrics (ICC and Dice coefficients) and describes the set of contrasts and tasks analyzed, but additional methodological details and full results are contained in the full preprint. Funding sources declared include Georgia Institute of Technology and a Georgia Tech startup fund. The authors declared no competing interests.