The theta/alpha ratio, an index of electroencephalographic slowing, has been proposed as a quantitative EEG marker of cognitive impairment. The magnitude and statistical significance of reported associations may depend on numerous analytic choices made during preprocessing, spectral estimation, spatial summarization, and statistical modeling. This study used a multiverse analysis to evaluate how robust theta/alpha associations with cognitive impairment are across a wide range of plausible processing and analysis pipelines.
Two resting-state EEG datasets were analyzed. Dataset 1 comprised 49 control participants and 100 participants with Parkinson's disease spanning cognitive status from normal cognition to dementia. Dataset 2 comprised 29 controls, 36 people with Alzheimer's disease, and 23 people with frontotemporal dementia. The authors systematically varied eight analytic decisions to generate multiple analytic "universes." Across combinations of these choices, a total of 1,280 specifications were evaluated.
The eight decisions that were varied included: filtering choices, artifact correction approach, the use or not of independent component analysis (ICA) for artifact removal, referencing method, how spatial summaries of EEG were computed, the exact definition of the alpha band, the spectral method used to quantify power (including absolute fast Fourier transform (FFT) and Welch power), and whether covariates were included in statistical models. By combining these options the analysis explored the specification dependence of theta/alpha associations with cognitive measures.
Findings in Dataset 1 varied across the unrestricted multiverse, indicating sensitivity to analytic choices. However, a notable consistent pattern emerged within an exploratory family of specifications: combinations that applied ICA and used absolute FFT or Welch power produced robust results. In these universes, comparisons of controls and cognitively normal Parkinson's disease participants with Parkinson's disease dementia yielded p < 0.05. Moreover, theta/alpha ratio associations with Montreal Cognitive Assessment scores were significant in more than 95% of universes within these specifications. Across the wider multiverse, higher theta/alpha ratios were consistently associated with greater cognitive impairment, but the statistical significance and effect detection depended on the pipeline.
In Dataset 2, the choice of spectral power method substantially influenced results. Specifications using absolute FFT or Welch power supported control versus Alzheimer's disease differences in more than 95% of universes. Additionally, configurations that incorporated ICA combined with absolute power measures yielded significant associations with Mini-Mental State Examination scores in all specifications examined within that family. As in Dataset 1, the direction of association was consistent: higher theta/alpha ratios linked to greater cognitive impairment, but the degree to which this was detected statistically depended on analytic choices.
The multiverse analysis demonstrates that the theta/alpha ratio is a promising EEG marker of cognitive impairment, showing consistent directionality across datasets and many specifications. Crucially, the presence and strength of statistically significant associations are specification-dependent, with particular dependence on the use of ICA and on the method of spectral power quantification (absolute FFT or Welch power). These dependencies mean that reported results for theta/alpha may not generalize across studies that use different preprocessing or spectral analysis pipelines.
By mapping analytic conditions that produce stable theta/alpha findings, the study identifies key processing decisions that should be considered when comparing results across studies or when developing standardized pipelines. The findings support efforts to define reproducible practices for quantitative EEG markers of cognitive impairment and to report analytic choices transparently so that specification dependence can be assessed.
The report describes the datasets, the eight decision dimensions, and the results across 1,280 specifications as presented in the preprint. Details beyond those reported in the source (for example, effect sizes for each universe or full specification tables) were not provided in the summary text. The authors declared no competing interests in the preprint.