Multimodal neuroimaging research frequently compares spatial brain maps from distinct imaging modalities to evaluate their spatial correspondence and infer biological coupling. Analyses that quantify intermodal coupling are increasingly common in studies of development, aging, and disease, where participant heterogeneity is substantial. Reliable statistical inference in these analyses requires careful attention to assumptions underlying the chosen tests, because apparent correspondence between modalities can arise from systematic differences across subgroups rather than genuine within-subject relationships.
One widely used approach to assess intermodal correspondence is the simple permutation-based inter-modal correspondence (SPICE) test. SPICE evaluates whether within-subject correspondence between two modalities exceeds what would be expected by chance by constructing a null distribution via permutation of subject labels for one modality. The test is appealing because it is straightforward to implement and makes relatively few spatial assumptions.
However, permutation inference relies on the critical assumption that subjects are exchangeable under the null. When subject-level covariates—such as age, sex, or disease status—systematically shift the distribution of brain maps, this exchangeability assumption may be violated. Such violations can produce misleading or spurious results. The authors highlight that these effects can resemble Simpson's paradox, where population-level patterns arise from differences between subgroups rather than true within-subject coupling. Thus, relying on simple label permutations without considering covariate structure risks conflating between-group differences with within-subject associations.
To address this limitation, the authors propose a formal test built on a U-statistic framework to detect covariate effects that would invalidate permutation-based coupling inference. The test is presented as serving two complementary purposes: first, as a diagnostic to determine whether the exchangeability assumption is violated in a given dataset; and second, as a scientific tool to identify and quantify covariate-driven effects on intermodal coupling.
The U-statistic approach provides a formal hypothesis test for covariate confounding in the coupling of multimodal imaging data. The abstract reports that the framework allows both evaluation of assumption violations and discovery of covariate effects, improving interpretability of intermodal analyses. Specific mathematical formulation, test statistics, and implementation details are provided in the full preprint but are not described in detail in the abstract.
The authors validated their proposed method using synthetic and semi-synthetic neuroimaging datasets. According to the abstract, these experiments show that the U-statistic-based test maintains well-controlled Type I error and achieves high statistical power under the scenarios examined. Exact numeric performance metrics, simulation parameters, and full experimental settings were not reported in the abstract and are available in the main text and supplementary materials of the preprint.
To illustrate real-world utility, the method was applied to assess potential confounding by age and sex in coupling analyses using two pairs of imaging modalities from the Philadelphia Neurodevelopmental Cohort. The abstract indicates that these applications demonstrate the method's ability to probe for covariate-driven confounding in developmental neuroimaging data. The abstract does not report the detailed outcomes of these cohort analyses; readers should consult the full preprint for the specific results and interpretations.
The proposed testing framework is intended to increase the rigor and interpretability of intermodal coupling analyses in heterogeneous populations. By formally testing for covariate effects that violate permutation exchangeability, investigators can better distinguish true within-subject coupling from population- or subgroup-driven correspondence. This is particularly relevant for studies of development, aging, and disease, where systematic differences across subjects are common and may otherwise lead to erroneous conclusions.
This work is reported as a preprint on bioRxiv and has not been peer reviewed. The authors declare no competing interests. Funding sources listed in the abstract include the National Institutes of Health (R01MH112847, R01MH123550) and support from The Research Institute of the Children's Hospital of Philadelphia. For full methodological details, simulation results, and cohort analysis outcomes, consult the full preprint and supplementary material linked in the source.