Group-level inference of fMRI data through non-parametric combination.
Poster No:
1943
Submission Type:
Abstract Submission
Authors:
Matteo Lionello1, Luca Cecchetti1
Institutions:
1IMT School for Advanced Studies Lucca, Lucca, Italy
First Author:
Co-Author:
Introduction:
A large extent of neuroimaging findings relies on group-level inference, which consists in computing summary statistics and Neyman-Pearson significance from the output of first-level analyses. In this regard, several studies pointed to the importance of using permutation-based null distributions in estimating group-level significance[1,2]. Recombination strategies are available to probe most study designs[3], such as permuting group labels in two-sample tests or sign-flipping. Regarding the latter, in alternative permutation strategies, if the null hypothesis is true, the test variable should have values equally dispersed around the central value of the H0 distribution. This assumption does not hold for unsigned testing variables, such as the R2 in encoding studies. Thus, how can the group-level significance of a multi-column encoding model be assessed? Here, non-parametric combination (NPC), a method conceived to combine multiple neuroimaging modalities within a subject[4], shows to reliably detect group-level effects in encoding studies while controlling for false positives. The code to implement group-level NPC is publicly available at github.com/mlionello/NaPuCco.
Methods:
900 volumes are generated with 27 effected voxels (R2~U[0,1]) while 120 volumes are generated with 8000 non-effected voxels. A synchronised timepoint-shuffling schema is created for 2000 permutations across participants and voxels. Single participant data are then created on effected and non-effected volumes. Parametric p-values are estimated for each voxel and combined into a joint statistic across a variable number of subjects (from 10 to 80) via the Fisher method. In power analysis, 150 experiments are simulated varying the percentage of participants carrying effects. In the clusterbased correction (CFT=0.001), effected volumes are padded by non-effected voxels (padding width=8). Voxel- and cluster-level family-wise false positive rate (FPR) are estimated by simulating 500 and 1000 experiments, respectively. Lastly, in real data, voxelwise FPR are measured with 198 resting-state acquisitions from the Cambridge Buckner dataset[5], a 10-column random noise encoding model, and simulating 1000 experiments. Participants were combined in groups from 5 to 60. All results are reported at pFWC<0.05.
Results:
As expected, to maintain statistical power constant, the overall sample size must increase proportionally to the decrease in the number of participants carrying an effect (Fig 1 right). In cluster-based correction, with R2~N(0.074, 0.17), the slope of the power curve at 0.80 increases by 0.06 per participant from 100% down to 70% of effected subjects included, and by 0.05 per participant when considering 60% of individuals carrying an effect (Fig 1 top-centre). Similarly, in the voxelwise correction and R2~N(0.074, 0.17), the slope of the power curve increases by 0.30, 0.15, 0.06, 0.03, 0.01 per participant when 100%, 90%, 80%, 70%, 60% of subject carries an effect(Fig 1 bottom-centre).
FPR confidence intervals are in line with the nominal family-wise corrected alpha level (Fig 2). When replicating the same analysis using real data, voxel-level FPR ranges from 4.9% to 6.9% with no association between the number of combined subjects and the number of false positives (p=0.317).
FPR confidence intervals are in line with the nominal family-wise corrected alpha level (Fig 2). When replicating the same analysis using real data, voxel-level FPR ranges from 4.9% to 6.9% with no association between the number of combined subjects and the number of false positives (p=0.317).
Conclusions:
This project systematically analyses and validates the use of NPC for group-level inference with unsigned statistics. In line with statistical theory, given a certain power, increasing the number of participants carrying an effect reduces the samples size needed to detect a group-level effect. The FPR aligns with nominal data and with findings in the literature for other permutation methods. In conclusion, the findings suggest that NPC is suitable for establishing group-level significance in encoding studies when the coefficient of determination represents the testing variable. This is crucial, especially when idiosyncrasies exist in how brain activity relates to individual features of an encoding model.
Modeling and Analysis Methods:
Methods Development 1
Univariate Modeling
Other Methods 2
Keywords:
FUNCTIONAL MRI
Other - non-parametric combination NPC; power analysis; false positive rate; significance; group-level inference; unsigned statistics; encoding
1|2Indicates the priority used for review
Provide references using author date format
[2] Eklund, A., Nichols, T. E., and Knutsson, H. (2016), 'Cluster failure: Why fmri inferences for spatial extent have inflated false-positive rates', National Academy of Sciences of the United States of America, vol. 113, no. 28, pp. 7900–7905.
[3] Winkler, A. M., Ridgway, G. R., Webster, M. A., Smith, S. M., & Nichols, T. E. (2014), 'Permutation inference for the general linear model', NeuroImage, vol. 92, no. 100, pp.381–397
[4] Winkler, A. M., Webster, M. A., Brooks, J. C., Tracey, I., Smith, S. M., and Nichols, T. E. (2016), 'Non-parametric combination and related permutation tests for neuroimaging', Human brain mapping, vol. 37, no. 4, pp. 1486–1511.
[5] Biswal, B., Mennes, M., Zuo, X.-N., Gohel, S., Kelly, C., Smith, S., Beckmann, C., Adelstein, J., Buckner, R., Colcombe, S., Dogonowski, A.-M., Ernst, M., Fair, D., Hampson, M., Hoptman, M., Hyde, J., Kiviniemi, V., K ̈otter, R.,
Li, S.-J., and Milham, M. (2010), 'Toward discovery science of human brain function', Proceedings of the National Academy of Sciences of the United States of America, vol. 107, pp. 4734–9.