Matthew Frehlich1, Luis Garcia Dominguez2, Sravya Atluri1, Natasha Radhu1, Yinming Sun1, Zafiris J Daskalakis1, Faranak Farzan3. 1. Temerty Centre for Therapeutic Brain Intervention, Centre for Addiction and Mental Health, Toronto, Ontario, Canada; University of Toronto, Toronto, Ontario, Canada. 2. Temerty Centre for Therapeutic Brain Intervention, Centre for Addiction and Mental Health, Toronto, Ontario, Canada. 3. Temerty Centre for Therapeutic Brain Intervention, Centre for Addiction and Mental Health, Toronto, Ontario, Canada; University of Toronto, Toronto, Ontario, Canada. Electronic address: faranak.farzan@camh.ca.
Abstract
BACKGROUND: Recent increase in the size and complexity of electrophysiological data from multidimensional electroencephalography (EEG) and magnetoencephalography (MEG) studies has prompted the development of sophisticated statistical frameworks for data analysis. One of the main challenges for such frameworks is the multiple comparisons problem, where the large number of statistical tests performed within a high-dimensional dataset lead to an increased risk of Type I errors (false positives). A solution to this problem, cluster analysis, applies the biologically-motivated knowledge of correlation between adjacent voxels in one or more dimensions of the dataset to correct for the multiple comparisons problem and detect true neurophysiological effects. Cluster-based methods provide increased sensitivity towards detecting neurophysiological events compared to conservative methods such as Bonferroni correction, but are limited by their dependency on an initial cluster-forming statistical threshold (e.g. t-score, alpha) obstructing precise comparisons of results across studies. NEW METHOD: Rather than selecting a single threshold value, unbiased cluster estimation (UCE) computes a significance distribution across all possible threshold values to provide an unbiased overall significance value. COMPARISON TO EXISTING METHODS: UCE functions as a novel extension to existing cluster analysis methods. RESULTS: Using data from EEG combined with brain stimulation study, we showed the impact of statistical threshold on outcome measures and introduction of bias. We showed the application of UCE for different study designs (e.g., within-group, between-group comparisons). CONCLUSION: We propose that researchers consider employing UCE for multidimensional EEG/MEG datasets toward an unbiased comparison of results between subjects, groups, and studies.
BACKGROUND: Recent increase in the size and complexity of electrophysiological data from multidimensional electroencephalography (EEG) and magnetoencephalography (MEG) studies has prompted the development of sophisticated statistical frameworks for data analysis. One of the main challenges for such frameworks is the multiple comparisons problem, where the large number of statistical tests performed within a high-dimensional dataset lead to an increased risk of Type I errors (false positives). A solution to this problem, cluster analysis, applies the biologically-motivated knowledge of correlation between adjacent voxels in one or more dimensions of the dataset to correct for the multiple comparisons problem and detect true neurophysiological effects. Cluster-based methods provide increased sensitivity towards detecting neurophysiological events compared to conservative methods such as Bonferroni correction, but are limited by their dependency on an initial cluster-forming statistical threshold (e.g. t-score, alpha) obstructing precise comparisons of results across studies. NEW METHOD: Rather than selecting a single threshold value, unbiased cluster estimation (UCE) computes a significance distribution across all possible threshold values to provide an unbiased overall significance value. COMPARISON TO EXISTING METHODS: UCE functions as a novel extension to existing cluster analysis methods. RESULTS: Using data from EEG combined with brain stimulation study, we showed the impact of statistical threshold on outcome measures and introduction of bias. We showed the application of UCE for different study designs (e.g., within-group, between-group comparisons). CONCLUSION: We propose that researchers consider employing UCE for multidimensional EEG/MEG datasets toward an unbiased comparison of results between subjects, groups, and studies.
Authors: Faranak Farzan; Sravya Atluri; Matthew Frehlich; Prabhjot Dhami; Killian Kleffner; Rae Price; Raymond W Lam; Benicio N Frey; Roumen Milev; Arun Ravindran; Mary Pat McAndrews; Willy Wong; Daniel Blumberger; Zafiris J Daskalakis; Fidel Vila-Rodriguez; Esther Alonso; Colleen A Brenner; Mario Liotti; Moyez Dharsee; Stephen R Arnott; Kenneth R Evans; Susan Rotzinger; Sidney H Kennedy Journal: Sci Rep Date: 2017-08-07 Impact factor: 4.379