Experimental Design Contrasts Inference EEG MEG Joseph Brooks

Experimental Design, Contrasts & Inference - EEG & MEG Joseph Brooks (ICN) Maria Joao (FIL) Methods for Dummies 2007 Wellcome Department For Neuroimaging 13/02/2008 1

Topics • • Exp. design and ERPs SPM for EEG-MEG 2 D interpolation 1 st level analysis 2 nd level analysis Time as another dimension Time-frequency analysis Conclusion 2

Popular approaches to M/EEG Data Event-Related Potentials (ERP) & Event-Related Fields (ERF) ERP/F Quantification Approaches Peaks, latency, area-under-curve Spectral Analysis (a. k. a. time-frequency) Connectivity 3

What is the ERP/ERF? -Def: the average (across trials/subjects) potential/field at the scalp relative to some specific event in time Stimulus/Event Onset 4

What is the ERP/ERF? -Def: the average (across trials/subjects) potential at the scalp relative to some specific event in time Averaging 5

What is the ERP/ERF? -Def: the average (across trials/subjects) potential at the scalp relative to some specific event in time Reflects reliable changes in potential that are strongly timelocked to stimulus onset (i. e. are synchronous over trials) Non-time-locked activity is lost to averaging 6

Interpreting ERP/ERF Waveforms ERP/ERF waveforms are often interpreted in terms of their constituent components Component (def) - Scalp-recorded electrical activity that is generated by a given patch of cortex engaged in a specific computational operation sensor + + + - - 7

Latent Components Any given electrode/sensor records a series of temporally overlapping latent components Latent Components Observed Waveform 8

Latent Components A given waveform could have arisen from many combinations of latent components Latent Components Observed Waveform OR OR Many others… 9

Important Observation #1 The morphology of a component is not necessarily obvious from the observed waveform when components overlap Latent Components Observed Waveform 10

Important Observation #2 Peaks ≠ Components Local maxima and minima in a waveform are not necessarily the best indicators of a component Latent Components Observed Waveform 11

Important Observation #3 Amplitude and latency of components are not independent A change of amplitude in one component can change amplitude and timing of many peaks Latent Components Observed Waveform 12

Feeling hopeless? Given these observations how can one make valid inferences about latent components from observed waveforms? Experimental design to the rescue! 13

Design Strategies Focus on one component and design experiment to stop other components from varying, especially temporally overlapping components Focus on easily isolated components that are wellknown Focus on large components. Large components are less sensitive to variations in others Test hypotheses that are component-independent 14

ERP/ERF Quantification To Peak or Not to Peak? Peak amplitude & latency are common measures BUT THEY ARE POOR MEASURES 15

ERP/ERF Quantification Amplitude and Latency are NOT independent Apparent amplitude difference is actually a difference in latency variance 16

ERP/ERF Quantification Solution: Use non-peak measures such as Area. Under-the-Curve Area under curves is same in the two average waveforms 17

SPM Approach to M/EEG Preprocessing Raw M/EEG data Projection SPM 5 -stats 2 D - scalp SPM{t} SPM{F} Control of FWE mass-univariate analysis Single trials Epoching Artefacts Filtering Averaging, etc. 3 D-source space Kiebel, S. 2005 18

Preprocessing Projection SPM 5 -stats The transformation of discreet channels into a continuous 2 D interpolated image of M/EEG signals Sensor Space Scalp Space 19

Preprocessing Projection SPM 5 -stats The transformation of discreet channels into a continuous 2 D interpolated image of M/EEG signals 20

Preprocessing Projection With data in 2 D (+time) map form we can now apply similar statistical procedures as used in FMRI Create SPMS of significant effects SPM 5 -stats SPM{t} SPM{F} Control of FWE mass-univariate analysis Use random field theory to control error Kiebel, S. 2005 21

Experimental Design, Contrasts & Inference - EEG & MEG Joe Brooks (ICN) Maria Joao (FIL) Methods for Dummies 2007 Wellcome Department For Neuroimaging 13/02/2008 22

Topics • • Experimental design and ERPs SPM for EEG-MEG Projection to voxel space 1 st level analysis 2 nd level analysis Space-Time SPMs Time-frequency analysis Conclusion 23

Voxel Space (revisited) 2/3 D images over peri-stimulus time bins 2 D scalp projection 3 D source reconstruction (interpolation in sensor space) (brain space) [Next week!] Data ready to be analysed 24

M/EEG modelling and statistics Epoched time-series data Data is analysed using the General Linear model at each voxel and Random Field Theory to adjust the p-values for multiple comparisons. Model specification Time e m Ti Parameter estimation Hypothesis Statistic Single voxel time series Typically one wants to analyse multiple subjects’ data acquired under multiple conditions Intensity 2 -Level Model SPM 25

1 st Level Analysis Epoched time-series data • Similar to f. MRI analysis. The aim of the 1 st level is to compute contrast images that provide the input to the second level. • Difference: here we are not modelling the data at 1 st level, but simply forming weighted sums of data over time At the 1 st level, we select periods or time points in peri-stimulous time that we would like to analyse. Choice made a priori. Time is treated as an experimental factor and we form weighted-sums over peristimulus time to provide input to the 2 nd level Example: if we were interested in the N 170 component, one could average the data between 150 and 190 milliseconds. 1 0 26

1 st Level Analysis Epoched time-series data Example: EEG data / 8 subjects / 2 conditions For each subject 1. Choose Specify 1 st-level 2. Select 2 D images 3. Specify EEG file SPM output: 2 contrast images average_con_0001. img 4. Specify Time Interval Timing information 5. Click Compute 27

2 nd Level Analysis Epoched time-series data Given the contrast images from the 1 st level (weighted sums), we can now test for differences between conditions or between subjects. 2 nd level contrast 2 nd level model = used in f. MRI -1 1 SPM output: = second level + Voxel map, where each voxel contains one statistical value The associated pvalue is adjusted for multiple comparisons 28

2 nd Level Analysis Epoched time-series data Example: EEG data / 8 subjects / 2 conditions 1. Specify 2 nd-level 2. Specify Design SPM output: Design Matrix 29

2 nd Level Analysis Epoched time-series data Example: EEG data / 8 subjects / 2 conditions 3. Click Estimate Output: Ignore brain outline: 4. Click Results 5. Define Contrasts “Regions” within the 2 D map in which the difference between the two conditions is significant 30

Space-Time SPMs (Sensor Maps over Time) Time as another dimension of a Random Field We can treat time as another dimension and construct 3 D images (2 D space + 1 D peri-stimulus time) We can test for activations in space and time Both approaches available: choice depends on the data Advantages: • If we had no a priori knowledge where and when the difference between two conditions would emerge. Weighted sums of data, over time, not appropriate in this case • Especially useful for time-frequency power analysis Disadvantages: • not possible to make inferences about the temporal extent of evoked responses 31

Space-Time SPMs (Sensor Maps over Time) How this is done in SMP 5 Example: EEG data / 1 subject / 2 conditions (344 trials) 1. Choose 2 D-to-3 D image on the SPM 5 menu and epoched data: e_eeg. mat 3. Statistical Analysis (test across trials) 2. Choose options 32 x 161 images for each trial / condition 4. Estimate + Results 5. Create contrasts 32

Space-Time SPMs (Sensor Maps over Time) How this is done in SMP 5 Example: EEG data / 1 subject / 2 conditions (344 trials) Ignore brain outline!!! Overlay with EEG image: More than 1 subject: • Same procedure with averaged ERP data for each subject • Specify contrasts and take them to the 2 nd level analysis 33

Time-Frequency analysis Transform data into time-frequency domain Useful for evoked responses and induced responses: Not phase-locked to the stimulus onset – not revealed with classical averaging methods SPM uses the Morlet Wavelet Transform Wavelets: mathematical functions that can break a signal into different frequency components. [Tallon-Baudry et. al. 1999] The transform is a convolution The Power and Phase Angle can be computed from the wavelet coefficients: 34

Time-Frequency analysis How this is done in SPM 5: Example: MEG data / 1 subject / 2 conditions (86 trials) 1. Choose time-frequency on the SPM 5 menu and epoched data: e_meg. mat 2. Choose options t 1_e_eeg. mat and t 2_e_eeg. mat power at each frequency, time and channel (t 1*); phase angles (t 2*) 3. Average mt 1_e_eeg. mat and mt 2_e_eeg. mat 4. Display 5. 2 D Time-Frequency SPMs 35

Summary Projection to voxel space (2 D interpolation or 3 D source reconstruction) 1 st Level Analysis (create weighted sums of the data over time) (contrast images = input to the 2 nd level) 2 nd Level Analysis (test for differences between conditions or groups) (similar to f. MRI analysis) Time-Space SPMs (time as a dimension of the measured response variable) Time-Frequency Analysis (induced responses) 36

References • S. J. Kiebel: 10 November 2005. ppt-slides on ERP analysis at http: //www. fil. ion. ucl. ac. uk/spm/course/spm 5_tutorials/SPM 5 Tutorials. htm • S. J. Kiebel and K. J. Friston. Statistical Parametric Mapping for Event. Related Potentials I: Generic Considerations. Neuro. Image, 22(2): 492 -502, 2004. • S. J. Kiebel and K. J. Friston. Statistical Parametric Mapping for Event. Related Potentials II: A Hierarchical Temporal Model. Neuro. Image, 22(2): 503 -520, 2004. • Todd, C. Handy (ed. ). 2005. Event-Related Potentials: A Methods Handbook. MIT • Luck, S. J. (2005). An Introduction to the Event-Related Potential Technique. MIT Press. 37

Thank You! For difficult questions: j. kilner@fil. ion. ucl. ac. uk (James Kilner) 38