Statistical Inference Methods in Functional Neuroimaging

Functional neuroimaging relies on statistical inference to distinguish meaningful patterns of brain activity from high-dimensional, noisy data. Central to most analyses is the mass-univariate general linear model (GLM), which estimates voxelwise responses by modelling signal intensity as a combination of experimental factors and covariates. To address the challenge of multiple comparisons across tens of thousands of voxels, parametric approaches such as random field theory control the family-wise error rate (FWE) by modelling the spatial smoothness of statistical maps, while false discovery rate (FDR) procedures adaptively threshold tests to balance sensitivity with error control. Cluster-extent and peak-height inference exploit spatial coherence to enhance detection power under localised activations. Non-parametric methods—particularly permutation tests—offer exact control of false positives with minimal distributional assumptions, albeit with greater computational cost. Recent methodological advances have focused on improving localisation accuracy, boosting statistical power through multivariate and model-based prediction techniques, standardising analysis pipelines, and promoting reproducibility. These developments underpin a wide array of applications—from mapping cognitive processes to guiding clinical interventions—by enabling robust identification of functional networks and reliable brain–behaviour correlations on a global scale.

Research from Nature Portfolio

Recent studies have highlighted the critical role of sample size in the reliability of task-based functional MRI findings. By evaluating multiple cognitive paradigms across large cohorts, researchers demonstrated that commonly used sample sizes (n = 20–30) yield only modest replicability and high variability in effect localisation. The work introduced intuitive metrics for quantifying replicability at both voxel and cluster levels, revealing that substantially larger cohorts—often exceeding n = 100—are needed to stabilise effect estimates. These insights have driven a shift towards pre-registered analysis plans and the aggregation of data from multisite consortia, fostering more robust and confirmatory research designs.

Research from all publishers

Innovations in cluster-based inference have extended traditional spatial extent thresholding by embedding it within closed testing procedures. This approach enables the detection of activation presence in regions of interest and the quantification of active voxel proportions, all under strict FWE control without additional alpha adjustments. Complementing this, foundational work on permutation inference for the GLM has introduced a generic framework that accommodates complex experimental designs and nuisance effects. These permutation methods guarantee exact error-rate control and permit the use of bespoke test statistics, and have been packaged into open-source tools to streamline adoption and ensure analytical rigour.

Statistical Inference Methods in Functional Neuroimaging publication trend

The graph below shows the total number of articles in statistical inference methods in functional neuroimaging across all publications each year (not limited to Nature Index journals).

Technical terms

General linear model (GLM): A statistical framework that models each voxel’s signal as a linear combination of explanatory variables plus noise.

Family-wise error rate (FWE): The probability of making one or more false-positive inferences when conducting multiple statistical tests.

False discovery rate (FDR): The expected proportion of false positives among all rejected hypotheses in a multiple-testing scenario.

Cluster-extent inference: A method that assesses the significance of spatially contiguous voxel clusters rather than individual voxels.

Permutation test: A non-parametric technique that determines statistical significance by repeatedly shuffling data labels to build an empirical null distribution.

References

1. Cluster extent inference revisited: quantification and localisation of brain activity. Journal of the Royal Statistical Society Series B Statistical Methodology (2023).
2. Thresholding of Statistical Maps in Functional Neuroimaging Using the False Discovery Rate. NeuroImage (2002).
3. Permutation inference for the general linear model. NeuroImage (2014).
4. Small sample sizes reduce the replicability of task-based fMRI studies. Communications Biology (2018).
5. Analysis of family‐wise error rates in statistical parametric mapping using random field theory. Human Brain Mapping (2017).
6. Non‐parametric combination and related permutation tests for neuroimaging. Human Brain Mapping (2016).