<div><br class="webkit-block-placeholder"></div>Analysis of EEG time series falls under the heading of stochastic<div>signal processing. There is an understanding that, even if the system</div><div>response properties, i.e., frequency-domain transfer function, are the </div>
<div>same across trials, the exact data recorded will be different across trials.</div><div><br class="webkit-block-placeholder"></div><div>Statistical estimation of the power spectrum always involves some</div><div>form of averaging. See Numerical Recipes in C, or Spectral Analysis</div>
<div>for Physical Applications by Percival and Walden. When we are fortunate</div><div>enough to have multiple data samples, e.g., trials in cognitive tasks, </div><div>then we average across trials. This averaging can take place in each</div>
<div>short moving window, as is done in time-frequency analysis.</div><div><br class="webkit-block-placeholder"></div><div>So to look for differences in spectral response properties across trials,</div><div>the proper procedure is to estimate the power spectrum in each condition </div>
<div>(and each moving window) by averaging across trials, then subtract the </div><div>two conditions in each window, as Stan and Julie said. That is what </div><div>newtimef does, although EEGLAB also handles the baseline interval, </div>
<div>which I have not addressed here. </div><div><br></div><div>I send this comment mainly to point out that simulations with deterministic</div><div>sine waves are not representative of the situation with stochastic signals </div>
<div>like EEG data.</div><div><br></div><div><div class="gmail_quote">-- <br></div>Thomas Ferree, PhD<br>Department of Radiology<br>UT Southwestern Medical Center<br><br></div>