Robust and replicable effects of ageing on resting-state brain electrophysiology measured with MEG
- Centre for Human Brain Health, School of Psychology, University of Birmingham, United Kingdom
- Oxford Centre for Human Brain Activity, Wellcome Centre for Integrative Neuroimaging, Department of Psychiatry, University of Oxford, United Kingdom
- Institute of Cognitive Neuroscience, University College London, United Kingdom
- Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty and University Hospital Düsseldorf, Heinrich Heine University, Germany
- Department of Experimental Psychology, University of Oxford, United Kingdom
- Wu Tsai Institute, Yale University, United States
- Department of Psychology, Yale University, United States
Figures
Figure 1
Effect of age on the relative magnitude spectrum across space and frequency.
(A) Model-predicted spectra (averaged across all sensors) for four equally spaced ages across the participant age range with inset histogram of participant ages within Cam-CAN. (B) Spectrum of t-values quantifying the age effect across space and frequency. Non-parametric permutations with maximum statistics to control for multiple comparisons across sensors and frequency bins. While this permutation testing was not cluster-based, contiguous clusters of significant sensors were computed post hoc for visualisation. The largest six spatially and spectrally contiguous areas of statistically significant effects are highlighted in frequency by black bands at the top of the spectrum and highlighted in space by sensors marked with a white circle in the adjoining topography. (C) A 2D frequency-by-space map of all statistically significant effects. Sensor-frequency combinations that do not reach statistical significance have a faded colour scale. Blue regions indicate decreasing spectral power with age and red regions indicate increases.
Figure 2
Cohen’s effect size for the linear effect of age.
(A) Cohen’s effect size for age across space and frequency. Each line is a sensor, with a colour matched to the sensors in the topography shown in the inset. The six topographies show spatial distributions across the sensor array at the peak frequency of the six significant effects identified in Figure 1B. (B) Contours showing the relationship between effect size and sample size for five different experimental power levels. As sample sizes get larger, there is sufficient power to reliably detect smaller effect sizes. (C) Effect size spectra with bootstrapped 95% confidence intervals at the peak sensor for the four largest effects identified in Figure 1B.
Figure 3
Quadratic effects of age on the neuronal frequency spectrum.
(A) The spectrum of t-values for the quadratic effect of age for all frequencies and channels. Positive values indicate a ‘U’ shaped effect and negative values indicate an ‘inverted-U’ shaped effect. Shaded regions and sensors indicated on the topomaps indicate statistical significance identified using non-parametric permutations with maximum-statistic correction for multiple comparisons. (B) A 2D frequency-by-space map of all statistically significant effects. Sensor-frequency combinations that do not reach statistical significance have a faded colour scale. Blue regions indicate inverted-U shaped effects of age and red regions indicate U shaped effects. (C) The spectrum of effect sizes for the quadratic effects of age.
Figure 4
Replicability of the parameter estimates, t-statistics, and effect sizes of ageing on the neuronal power spectrum.
(Ai) Spectrum of parameter estimates across frequency (averaged over all sensors) for each of the four datasets. (Aii) Correlation matrix indicating similarity of the sensor-averaged frequency profile of the age effect between datasets. (Bi and Bii) As (A) for null hypothesis test statistics. (Ci and Cii) As (A) for effect sizes.
Figure 5
The spatial and spectral profile of the effect of age on resting-state brain electrophysiology is replicable across datasets.
(A) GLM-Spectrum parameter estimates quantifying the age effect across space and frequency for the four datasets in the replicability analysis with distribution of participant ages shown in the inset figure. The four datasets share the key features highlighted in Figure 1. (B) As (A) for t-statistics testing the hypothesis that the parameter estimate of the age effect is different to zero. (C) As (B) but visualised as a 2D image in which each row indicates a single sensor with their y-axis position sorted by spatial location on the anterior-posterior axis. (D) As (A) for Cohen’s effect sizes for the age regressor in the GLM-Spectrum.
Figure 6
The age effect computed on the absolute magnitude of the power spectrum.
(A) Model-predicted spectra (averaged across all sensors) for four equally spaced ages across the participant age range with inset histogram of participant ages within Cam-CAN. (B) The effect size for the age regressor computed on the absolute magnitude of the power spectrum. (C) Top: Spectrum of t-values quantifying the age effect across space and frequency. Non-parametric permutations with maximum statistics to control for multiple comparisons across sensors and frequency bins. While this permutation testing was not cluster-based, contiguous clusters of significant sensors were computed post hoc for visualisation. The largest six spatially and spectrally contiguous areas of statistically significant effects are highlighted in frequency by black bands at the top of the spectrum and highlighted in space by sensors marked with a white circle in the adjoining topography. (C) Bottom: A 2D frequency by space map of all statistically significant effects. Sensor-frequency combinations that do not reach statistical significance have a faded colour scale. Blue regions indicate decreasing spectral power with age and red regions indicate increases.
Figure 7
Covariate effect sizes and their impact on the ageing effect.
(A) Box and whisker plot with paired kernel density plot for the distributions of Cohen’s effect size for each alternative covariate, when the alternative covariate is included as the only regressor in a GLM, along with an intercept term. The Cohen’s distributions are collected over every single sensor-frequency pair from the sensor-space dataset (i.e. over 102 sensors and 189 frequency bins). Full GLM-Spectrum visualisations of the results are included in Appendix 4. For comparison, the same is shown for the age using a GLM that contains regressors for age and an intercept. (B) Change in the Cohen’s effect size of age between (1) a GLM that includes an age regressor plus intercept, and (2) a model that includes an age regressor, a single alternative covariate regressor, and intercept. This is shown for each alternative covariate in turn. Age is excluded from this panel. (C) Pearson’s correlation coefficient quantifying the univariate linear relationship between age and each covariate. Age and sex are excluded from this panel.
Figure 8
The age effect is reduced heterogeneously across space and frequency by including global grey matter volume (GGMV) in the model.
(A) The Cohen’s spectrum for age using a GLM that contains regressors for age and an intercept. Replicated from Figure 3. (B) As (A) when a grey matter volume regressor is added to the GLM. Accounting for grey matter volume broadly reduces the effect size of age across the spectra, but some moderate effects remain. (C) Spectrum of t-values quantifying the age effect across space and frequency replicated from Figure 1. Non-parametric permutations with maximum statistics to control for multiple comparisons across sensors and frequency bins. While this permutation testing was not cluster-based, contiguous clusters of significant sensors were computed post hoc for visualisation. The largest six spatially and spectrally contiguous areas of statistically significant effects are highlighted in frequency by black bands at the top of the spectrum and highlighted in space by sensors marked with a white circle in the adjoining topography. (D) As (C) for the age effect in a model that includes a GGMV covariate. Most effects are reduced in size with smaller statistically significant regions surviving the post hoc clustering procedure.
Appendix 1—figure 1
Model projected age effect retains non-linear properties of decrease in alpha peak frequency and magnitude.
(A) Model-predicted spectra (averaged across all sensors) for the oldest and youngest participants in Cam-CAN with the alpha peak for intermediate ages overlaid in the thick line. (B) The parameter estimates of age that combine to describe the peak frequency shift seen in (A). (C) Scatter plot of individual alpha peak frequencies against age. A linear regression line fitted directly to the individual alpha peaks is shown in red, and the GLM-Spectrum-derived alpha peaks from the model projected spectra is shown in blue. (D) As (C) for individual alpha peak magnitude.
Appendix 2—figure 1
Effect of age on the relative magnitude spectrum across space and frequency in linearly constrained minimum variance (LCMV) source reconstructed and parcellated data.
(A) Model-predicted spectra (averaged across all sensors) for four equally spaced ages across the participant age range. (B) Spectrum of t-values quantifying the age effect across space and frequency. Source topographies are shown in the frequency bands with significant effects at sensor space. The permutation statistics are not repeated at source space. (C) A 2D frequency-by-space map of all statistically significant effects. Sensor-frequency combinations that do not reach statistical significance have a faded colour scale. Blue regions indicate decreasing spectral power with age, and red regions indicate increases.
Appendix 3—figure 1
Effect of age with and without independent component analysis (ICA) denoising is largely consistent.
(A) Spectrum of t-values quantifying the age effect across space and frequency. Non-parametric permutations with maximum statistics to control for multiple comparisons across sensors and frequency bins. While this permutation testing was not cluster-based, contiguous clusters of significant sensors were computed post hoc for visualisation. The largest six spatially and spectrally contiguous areas of statistically significant effects are highlighted in frequency by black bands at the top of the spectrum and highlighted in space by sensors marked with a white circle in the adjoining topography. (B) A 2D frequency-by-space map of all statistically significant effects. Sensor-frequency combinations that do not reach statistical significance have a faded colour scale. Blue regions indicate decreasing spectral power with age and red regions indicate increases. (C) Spectrum of the difference in estimated effect size of age in data that has been cleaned with ICA and data that has not been cleaned with ICA. Positive values indicate that the effect was larger with ICA denoising and negative values indicate that the effect was smaller with ICA.
Appendix 4—figure 1
GLM-Spectrum of parameter estimates for all covariates.
Appendix 4—figure 2
GLM-Spectrum of t-statistics for all covariates.
Appendix 4—figure 3
GLM-Spectrum of effect sizes for all covariates.
Appendix 4—figure 4
GLM-Spectrum of changes in estimated age effect sizes when including a covariate in the model.
Age itself is excluded from this figure.
Appendix 5—figure 1
Correlations between head position and age for the Cam-CAN dataset.
Appendix 5—figure 2
GLM-Spectrum results showing the effect of age estimated for two different sensor normalisations and three different Signal Space Separation (SSS) head position correction types.
(A) The age effect for the absolute magnitude with SSS applied without head position correction. (B) The age effect for the relative magnitude with SSS applied without head position correction. (C) The age effect for the absolute magnitude with SSS applied with head position correction applied within each dataset (−movecomp in maxfilter software). (D) The age effect for the relative magnitude with SSS applied with head position correction applied within each dataset (−movecomp in maxfilter software). (E) The age effect for the absolute magnitude with SSS applied with head position correction applied within each dataset (−movecomp in maxfilter software) and head position alignment to a reference datafile (−trans in maxfilter software). (F) The age effect for the relative magnitude with SSS applied with head position correction applied within each dataset (−movecomp in maxfilter software) and head position alignment to a reference datafile (−trans in maxfilter software). (G) The spectral generalisation of the topography of the absolute magnitude age effect. (H) The spectral generalisation of the topography of the relative magnitude age effect.
Appendix 6—figure 1
Model residuals and coefficient of determination as a function of frequency for each channel show that the group model performs well.
(Left) The sum-square residuals of the first level models with each individual data recording (grey lines) and the overall group average (black line). (Right) The R2 coefficient of determination for the group-level model showing that the model is able to describe a high amount of variability across both channels and frequencies.
Appendix 6—figure 2
Model validation statistics and diagnostics for the group-level GLM containing an intercept, a z-transformed age regressor, and a z-transformed global grey matter volume (GGMV) regressor.
(Left) The correlation matrix between the three regressors shows a large negative correlation between age and GGMV. (Centre) The singular values of the design matrix stay well above zero for all components, indicating that the collinearity between regressors does not cause the model fit to critically fail. (Right) The variance inflation factors for all three regressors shows peak values of 2.35, which indicates inflated variance (VIF>1). Specifically, this corresponds to an increase in standard error by a factor of 1.533.
Tables
Table 1
Summary of significant spatio-spectral regions containing an age effect in Cam-CAN.
The estimated sample sizes are the range of sample sizes needed to have an 80% chance of detecting the effect in a future sample, assuming that both the effects here are true and that the effect sizes are well estimated. Sample size ranges are computed from bootstrapped 95% confidence intervals for effect size estimate. The lower bound of the 95% confidence intervals can be taken as a smallest effect size of interest. The sample size forecasts are not informative about the present results and serve only as a guide for future study planning.
| Descriptive label | Peak freq (range) | Peak t-value | Peak region | Peak Cohen’s F2 (95% CI) | Estimated future sample for 80% power |
|---|---|---|---|---|---|
| Linear effects | |||||
| Low freq | 5 Hz (1–7) | –12.11 | Frontal | 0.264 (0.187–0.375) | 26–51 |
| Low alpha | 8 Hz (7–8.5) | +6.44 | Occipital | 0.074 (0.044–0.122) | 79–218 |
| High alpha | 10.5 Hz (9.5–12.5) | –12.43 | Occipito-parietal | 0.278 (0.205–0.387) | 25–47 |
| Beta | 15 Hz (12.5–26) | +11.22 | Central | 0.227 (0.162–0.320) | 30–60 |
| Low gamma | 36 Hz (32.5–50) | +5.77 | Occipital | 0.06 (0.030–0.111) | 86–315 |
| High gamma | 60 Hz (51.5–74.5) | –5.37 | Frontal | 0.052 (0.024–0.098) | 97–396 |
| Quadratic effects | |||||
| Left beta | 17 (16.5–17) | –4.62 | LH occipito-temporal | 0.033 (0.036–0.11) | 82–264 |
| Right beta | 18 (16 - 20) | –6.26 | RH occipito-temporal | 0.071 (0.044–0.13) | 76–222 |
| Central high beta | 27 (24–30.5) | –6.47 | Anterior central | 0.076 (0.044–0.12) | 80–218 |
Table 2
Summary of significant spatio-spectral regions containing an age effect in Cam-CAN.
The estimated sample sizes are the range of sample sizes needed to have an 80% chance of detecting the effect in a future sample, assuming that both the effects here are true and that the effect sizes are well estimated. Sample size ranges are computed from bootstrapped 95% confidence intervals for effect size estimate. The sample size forecasts are not informative about the present results and serve only as a guide for future study planning.
| Analysis and descriptive label | Age effect size | Age effect size with GGMV | Change in age effect size from including GGMV |
|---|---|---|---|
| Effect size estimates | |||
| Low freq (5 Hz) | 0.26 (0.19–0.38) | 0.11 (0.065–0.18) | –0.15 (–57%) |
| Low alpha (8 Hz) | 0.075 (0.043–0.13) | 0.013 (0.0016–0.039) | –0.048 (–83%) |
| High alpha (10.5 Hz) | 0.28 (0.20–0.40) | 0.12 (0.080–0.20) | –0.16 (–57%) |
| Beta (15 Hz) | 0.23 (0.16–0.32) | 0.078 (0.045–0.13) | –0.152 (–66%) |
| Low gamma (36 Hz) | 0.06 (0.031–0.11) | 0.071 (0.034–0.14) | +0.011 (+15%) |
| High gamma (60 Hz) | 0.052 (0.024–0.096) | 0.044 (0.015–0.10) | –0.008 (–15%) |
| Sample with 80% power for age | Sample with 80% power for age including GGMV | Change in sample for including GGMV | |
| Sample size planning | |||
| Low freq (5 Hz) | 25–52 | 54–148 | +30 to +98 |
| Low alpha (8 Hz) | 76–225 | n.s. | n.s. |
| High alpha (10.5 Hz) | 24–47 | 49–120 | +25 to +73 |
| Beta (15 Hz) | 30–60 | 73–216 | +43 to +156 |
| Low gamma (36 Hz) | 85–304 | 67–285 | –18 to –19 |
| High gamma (60 Hz) | 100–401 | n.s. | n.s. |
Appendix 1—table 1
A selection of publications from the last 20 years reporting age-dependent differences in alpha power.
A variety of methodological approaches are used by the papers. Note that this is not intended to be a comprehensive list of all publications in this time, but is a representative sample of the variability in methods and results found in the literature.
| Publication | Method/alpha frequency/location | Change with age | Figure in source publication |
|---|---|---|---|
| Babiloni et al., 2006 | Method: MEG source space Alpha: 8–10.5 Hz; 10.5–13 Hz Peak: parietal, occipital, temporal, limbic areas | Decrease | Figures 1, 2 & 3 |
| Lodder and van Putten, 2011 | Method: EEG sensor space Alpha: individual peak (3–18 Hz) Peak: individual peak | Decrease | Table 2 |
| Vysata et al., 2012 | Method: EEG sensor space Alpha: 7–13 Hz Peak: whole head | Decrease | Figure 2 |
| Gómez et al., 2013 | Method: MEG sensor space Alpha: 8–13 Hz Peak: parieto-temporal areas | Quadratic | Figure 3 |
| Scally et al., 2018 | Method: EEG sensor space Alpha: individual peak (6–14 Hz) and upper alpha (10–12 Hz) Peak: all channels averaged | No change (individual) Decrease (upper alpha) | Figure 5 |
| Hoshi and Shigihara, 2020 | Method: MEG source space Alpha: 8–12 Hz Peak: temporal pole; visual cortex V1–V4 | Increase (TP) Decrease (visual) | Table 3 |
| Donoghue et al., 2020 | Method: EEG sensor space Alpha: individual peak (7–14 Hz) Peak: sensor Oz | Decrease | Figure 5 |
| Thuwal et al., 2021 | Method: MEG sensor space Alpha: individual peak (8–12 Hz) Peak: occipito-parietal sensors | Decrease | Figure 6 |
| Rempe et al., 2023 | Method: MEG source space Alpha: 8–12 Hz Peak: bilateral inferior temporal gyrus | Increase | Figures 2 & 3 |
| Stier et al., 2023 | Method: MEG source space Alpha: 8–12 Hz Peak: temporal regions | Increase | Figure 2 |
| Merkin et al., 2023 | Method: EEG sensor space Alpha: individual peak (after 1/f correction) Peak: all channels | No change | Figure 3 |
| Tröndle et al., 2023 | Method: EEG sensor space Alpha:individual peak ± 2 Hz (8–13 Hz) Peak: parietal and occipital | Decrease | Figures 5 & 7 |
| Park et al., 2024 | Method: EEG source space Alpha: individual peak (6–14 Hz; after 1/f correction) Peak: sensorimotor and parietal regions | Decrease | Figures 2 & 3 |
| Ustinin et al., 2025 | Method: MEG sensor space Alpha: 8–13 Hz Peak: whole head | Decrease | Figures 5 & 6 |
| Medrano et al., 2025 | Method: EEG sensor space Alpha: individual peak (8–12 Hz; after 1/f correction) Peak: channel Oz | Decrease | Figure 8 |
Appendix 7—table 1
Summary of regression analyses performed in this article organised by figure.
| Figure | Field | Specification |
|---|---|---|
| 1, 2, 4, 9, 18, 19 | Analysis | Linear effect of age (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 3 | Analysis | Quadratic effect of age (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 5 | Analysis | Linear effect of age (Cam-CAN and MEG-UK) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 6, 17 | Analysis | Linear effect of age (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 7, 12, 13, 14 | Analysis | Linear effect of covariates (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 7, 12, 13, 14 | Analysis | Linear effect of age with covariate included (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 8, 9 | Analysis | Linear effect of GGMV with age (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 | |
| 10 | Analysis | Linear effect of age (Cam-CAN) |
| Dimensions | Source parcels× frequency | |
| Model | ||
| Permutations | 2500 | |
| 11 | Analysis | Linear effect of age without ICA (Cam-CAN) |
| Dimensions | Sensors× frequency | |
| Model | ||
| Permutations | 2500 |
Additional files
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Robust and replicable effects of ageing on resting-state brain electrophysiology measured with MEG
eLife 14:RP108829.
https://doi.org/10.7554/eLife.108829.3
