Description
Average many signal epochs locked to a repeated event so a small consistent response emerges from noise, and estimate how much noise remains so the average can be trusted.
Signal averaging recovers small responses buried in noise, such as event-related potentials in EEG or phasic pupil and skin-conductance responses. It assumes the signal is consistent across trials, its timing is known, and the noise is random, zero-mean and uncorrelated with the signal. Averaging N trials reduces residual noise variance by a factor of N, so noise amplitude falls with the square root of N. Van Drongelen gives three ways to estimate the noise left in an average: the prestimulus baseline, bootstrap averages built from random triggers, and the plus-minus average, which inverts every other trial so the signal cancels and only noise remains. Separate odd and even averages check consistency. Periodic noise such as mains hum is enhanced rather than averaged away when the stimulus rate divides evenly into 50 or 60 Hz, so intervals should be jittered.
Attributes
- Discipline
- Neuro, Biometrics & Affect
- Loss type
- Siloed
- Stage
- Evaluation
- Data
- Quantitative
- Effort
- Weeks
Steps
- Record repeated trials, each cut as an epoch time-locked to a trigger (stimulus onset or user event).
- Randomise the inter-stimulus interval or use a non-integer rate (for example 7.7 Hz rather than 10 Hz) to avoid locking to mains interference.
- Average the epochs sample by sample to form the average waveform per condition.
- Form separate odd-trial and even-trial averages; real components must appear in both.
- Estimate residual noise from the prestimulus period, random-trigger bootstrap averages or the plus-minus average.
- Compute SNR as root-mean-square of the average over that of the residual noise, and plot it against trial count to decide how many trials are enough.
Sources
- Signal Processing for Neuroscientists (van Drongelen, 2018)
- Biopsychology (Pinel & Barnes, 2022)
- How Do Brains Work? Papers of a Comparative Neurophysiologist (Bullock, 1993)
Also known as
- Event-related averaging
- Evoked potential averaging
- ERP averaging
- Plus-minus average
- Odd-even replicate averaging
- Random-trigger bootstrap control average
Categories
- Analytics & Behaviour
- Neuro, Biometrics & Affect