Cluster-Level Parametric Inference is Valid for Electrophysiological Data in 1D and 2D
Poster No:
1323
Submission Type:
Abstract Submission
Authors:
Suvadeep Maiti1, Tim Tierney2, Vladimir Litvak3
Institutions:
1University College London, London, London, 2University College London, London, United Kingdom, 3UCL Queen Square Institute of Neurology, London, London
First Author:
Co-Author(s):
Introduction:
Parametric statistical inference using the General Linear Model (GLM) with multiple comparisons correction based on Random Field Theory (RFT) is a powerful and versatile framework, widely applied for various neuroimaging modalities. Cluster-level inference within this framework is particularly advantageous for analysing evoked and induced electrophysiological responses, as true effects typically extend across space, time and frequency.
In their seminal 2016 study, Eklund et al. (Eklund et al., 2016) demonstrated that for fMRI 3D images, cluster inference can fail to control the type I error rate unless the cluster-forming threshold is stringent (e.g., p < 0.001, the standard recommendation). However, applying this heuristic to electrophysiological data (e.g., 1D time series, 2D scalp topographies, and time-frequency images) may be overly conservative, potentially missing most of true effects. While non-parametric cluster inference is valid regardless of threshold selection, its applicability is severely constrained by experimental design limitations.
As the Euler characteristic density, which underpins RFT-based p-value approximations, is highly sensitive to dimensionality, it remains uncertain whether Eklund et al.'s observations generalise to 1D and 2D data. We hypothesised that more lenient thresholds could maintain validity in these lower-dimensional contexts. Here, we present preliminary evidence supporting this hypothesis, derived from analyses of real data with permuted condition labels.
In their seminal 2016 study, Eklund et al. (Eklund et al., 2016) demonstrated that for fMRI 3D images, cluster inference can fail to control the type I error rate unless the cluster-forming threshold is stringent (e.g., p < 0.001, the standard recommendation). However, applying this heuristic to electrophysiological data (e.g., 1D time series, 2D scalp topographies, and time-frequency images) may be overly conservative, potentially missing most of true effects. While non-parametric cluster inference is valid regardless of threshold selection, its applicability is severely constrained by experimental design limitations.
As the Euler characteristic density, which underpins RFT-based p-value approximations, is highly sensitive to dimensionality, it remains uncertain whether Eklund et al.'s observations generalise to 1D and 2D data. We hypothesised that more lenient thresholds could maintain validity in these lower-dimensional contexts. Here, we present preliminary evidence supporting this hypothesis, derived from analyses of real data with permuted condition labels.
Methods:
We used single-channel EEG data capturing auditory oddball responses from one subject, provided in the Statistical Parametric Mapping (SPM) toolbox example dataset (https://www.fil.ion.ucl.ac.uk/spm/data/eeg_mmn/). The data were pre-processed following the standard SPM tutorial pipeline. After epoching, the data underwent time-frequency analysis, generating two input datasets for statistical analysis: a 1D time-domain epoched dataset and a 2D time-frequency epoched dataset.
Condition labels were randomly reassigned 1000 times, drawing 480 'standard' and 120 'oddball' trials from the total trial pool. Two-sample t-tests with cluster-level correction were performed and their significance was assessed with one-sided t-tests and F-tests. The percentage of iterations producing significant clusters by chance was quantified across varying cluster-forming thresholds (p< 0.001, 0.005, 0.01, 0.05, 0.1). The cluster-level threshold was set at p < 0.05; accordingly, we expected the percentage of false positives to remain below 5% if the type I error rate was properly controlled.
Condition labels were randomly reassigned 1000 times, drawing 480 'standard' and 120 'oddball' trials from the total trial pool. Two-sample t-tests with cluster-level correction were performed and their significance was assessed with one-sided t-tests and F-tests. The percentage of iterations producing significant clusters by chance was quantified across varying cluster-forming thresholds (p< 0.001, 0.005, 0.01, 0.05, 0.1). The cluster-level threshold was set at p < 0.05; accordingly, we expected the percentage of false positives to remain below 5% if the type I error rate was properly controlled.
Results:
The percentage of false positives remained below 5% across all analyses for cluster-forming thresholds of p<0.01 and below. For p<0.05 the percentage was close to 5% for 2D but exceeded this level for 1D. For p<0.1, the family-wise error rate was not well-controlled and was between 9 and 18% for the different analyses (see Figure 1 for details).
Conclusions:
These findings support the validity of cluster-level inference for low-dimensional data with cluster-forming thresholds of up to p<0.01 and, for 2D data, possibly even p<0.05. Future work will expand these analyses by incorporating simulated data with varying degrees of smoothness to refine the conditions under which lenient cluster-forming thresholds may be safely employed.
Modeling and Analysis Methods:
EEG/MEG Modeling and Analysis 1
Methods Development 2
Univariate Modeling
Keywords:
Electroencephaolography (EEG)
ELECTROPHYSIOLOGY
Experimental Design
MEG
Open Data
Open-Source Code
Open-Source Software
Statistical Methods
Other - General Linear model, GLM, Random Field Theory, RFT, cluster inference, multiple comparisons
1|2Indicates the priority used for review