A Potential Source of Bias in Group-Level EEG Microstate Analysis

Abstract

Microstate analysis is a promising technique for analyzing high-density electroencephalographic data, but there are multiple questions about methodological best practices. Between and within individuals, microstates can differ both in terms of characteristic topographies and temporal dynamics, which leads to analytic challenges as the measurement of microstate dynamics is dependent on assumptions about their topographies. Here we focus on the analysis of group differences, using simulations seeded on real data from healthy control subjects to compare approaches that derive separate sets of maps within subgroups versus a single set of maps applied uniformly to the entire dataset. In the absence of true group differences in either microstate maps or temporal metrics, we found that using separate subgroup maps resulted in substantially inflated type I error rates. On the other hand, when groups truly differed in their microstate maps, analyses based on a single set of maps confounded topographic effects with differences in other derived metrics. We propose an approach to alleviate both classes of bias, based on a paired analysis of all subgroup maps. We illustrate the qualitative and quantitative impact of these issues in real data by comparing waking versus non-rapid eye movement sleep microstates. Overall, our results suggest that even subtle chance differences in microstate topography can have profound effects on derived microstate metrics and that future studies using microstate analysis should take steps to mitigate this large source of error.

Keywords: EEG; Error; Microstates.

Figure 1.

Schematic of the study. Red violin plots are “cases” and black plots are “controls”.

Figure 2.

Subgroup maps show consistently higher type I error rates than combined maps. A) Proportion of permutations with false positives for each microstate and parameter. Gray bars are combined maps and black bars are subgroup derived maps. B) Histogram of p-values for all comparisons in all permutations. The dotted line indicates p = 0.05. There is a large peak of p values less than 0.05 In the subgroup derived maps.

Figure 3.

An example of elevated type I errors when using subgroup maps for a single permutation. A) Subgroup derived case and control maps and the difference between them. B) Correlations between the coverage metrics calculated with subgroup maps for cases (red) and controls (black) for each microstate. C) Coverage, duration, global explained variance (GEV), occurrence, and spatial correlation (SPC) for cases and controls with subgroup maps (left column) and combined maps (right column). * p < 0.05.

Figure 4.

Impact of true group differences in microstate maps when using a combined/single set of maps. In a proof-of-principle single simulation based on the original $N=30$ dataset, we created a second dataset which was identical except for having partially shuffled EEG channel labels. As the standard microstate segmentation procedure does not consider channel location, these two datasets are effectively identical through the lens of microstate analysis (except for the different microstate topographies). A) Maps from a $K=4$ segmentation of the original (upper row) and channel-labeled shuffled datasets (lower) along with the spatial correlations between corresponding maps. B) Correlations between microstate metrics in the original (O) and shuffled (S) datasets (which should in principle be 1.0); using a single/combined set of maps leads to lower correlations. C) Unsigned -log10(p) significance values between metrics from the original and channel-shuffled dataset (showing p<0.05 results only). Single/combined maps show significant differences in microstate metrics (due to unmodelled differences in microstate maps rather than true differences in microstate dynamics).

Figure 5.

Application to real data comparing waking (quiet rest) and stage N2 sleep. A) $K=4$ maps for wake and stage N2 sleep in those same individuals, yielding qualitatively different maps in this particular set of $N=30$ healthy controls. B) $K=6$ maps for wake (top row), stage N2 sleep (middle row) or from a segmentation on data combined across states (bottom row). Spatial correlations reflect the similarity of wake and sleep maps, which are generally high except for microstate D. C) Between group (wake vs sleep) analyses based on the K=6 segmentation; for nominally significant (p < 0.05) tests, values indicate the signed −log10(p) from a paired-sample t-test, with positive (green) values indicating higher values during sleep compared to wake. COV = coverage, DUR = duration, GEV = global variance explained, OCC = occurrence. Results for spatial correlations not shown as all methods showed similarly significant results for all microstates (higher values in sleep).