Skip to content
Cyril Pernet edited this page Jun 28, 2024 · 1 revision

Should you downsample data?

The answer boils down to how many trials do you have. At 1st glance, LIMO MEEG is just doing a GLM and thus it seems like you always have enough trials since you will (almost) always have more trials than regressors (conditions and covariates). This is true only when using an Ordinary Least Squares (OLS) estimation.

Downsampling and Weighted Least Squares (the default method)

In the paper describing how LIMO MEEG derives a single weight per trials we also looked at what is the most efficient way to decompose trials over time or frequencies for data at 1000Hz, %00Hz and 250Hz (see supplementary table 4 at the very end of the paper) and results showed that:

  • having more trials than time frames always gives better classification results
  • down-sampling gives better results than using ill-conditioned data
  • in very noisy conditions, downsampling is not a good option and it is better to let use method do it's thing
Clone this wiki locally