Reply-To: | | [log in to unmask][log in to unmask] > 909.607.0914 > > and > > Visiting Associate Scientist > Psychology & Neuroscience > Division of Humanities & Social Sciences > Caltech > Pasadena, CA 91125 > [log in to unmask] >>>> Vladimir Litvak 02/02/11 3:19 AM >>> > Dear Michael, > > On Wed, Feb 2, 2011 at 7:02 AM, Michael Spezio > wrote: >> Hi Vladimir, >> >> I saw in your post 041502, dated 2010-05-07, that for some reason bandpass >> filtering with wide bands (e.g., 0.1 to 30 Hz) elicits bad performance by >> the Matlab SP Toolbox and causes displays of the subsequent data to fail. >> >> Can you specify why this is occurring? It doesn't seem to make sense given >> that wide bandpass filters are used in signal processing routinely. >> > > It's not a display issue but actual data corruption issue. There are > numerical stability problems that I have since also observed in > high-pass filters for some combinations of filter settings and > sampling rate. These problems are common for wide bandpass filters, > especially for ones with low cut-off close to DC (e.g. 0.1-40 Hz). > Therefore, I recommended that people use separate high-pass and > low-pass in these cases. In the next SPM8 update there is a change in > the code that detects automatically when a filter is unstable and > gives an error so people will not have to figure it out later when > their display crashes. > >> Also, for the Multimodal Face tutorial, why isn't filtering done prior to >> downsampling? Downsampling, especially to 200 Hz or lower, should only be >> done after filtering out higher frequency contaminating signals, measured >> at >> a high sampling rate (500 Hz or above), to avoid aliasing from muscle >> signals at frequencies of 90 Hz to 200 Hz. Can you help me understand what >> is going on with the bandpass processing and why filtering is left out of >> the tutorial? >> > > Aliasing can occur when downsampling is done by simple decimation in > the time domain (e.g. taking every other sample). SP toolbox > downsampling routine is smarter than that and it pre-filters data to > avoid aliasing so it is not necessary to do it explicitly. In the > absence of SP toolbox we use or own routine that downsamples in the > frequency domain by truncating the DFT coefficients. This way there is > also no aliasing. The only problem that can occur is that if there are > large DC offsets in the data, low-pass filter will cause ringing at > the edges. Therefore I'd suggest to apply high-pass filtering before > downsampling especially for epoched data. In the case of the faces > tutorial this is not a problem because downsampling is done on > continuous data (for EEG) or on baseline-corrected data (for MEG). > > Best, > > Vladimir > >> Thanks so much. >> >> Best, >> >> Michael >> >> Michael L. Spezio, Ph.D. >> Assistant Professor of Psychology >> Department of Psychology >> Scripps College >> Claremont, CA >> [log in to unmask] >> 909.607.0914 >> >> and >> >> Visiting Associate Scientist >> Psychology & Neuroscience >> DiviĀD/ |