[Homer-users] Nonparametric analysis

[Huppert, Ted]

The challenge with non-parametric testing like those in that paper for EEG and MEG is that the neural response is much shorter then the hemodynamic response (e.g. NIRS).  They effectually a blocked design task or slow event-related where the stimuli are far enough apart to prevent the responses from overlapping.  They also have a lot of trials (it looked like 600).     In their parametric test, you take the 600 trials and take random subsets of this data and compute the histogram of test statistics.    The more trials per condition you have, the more stable this test will be.  To get the same number of trials for NIRS (where the hemodynamic response is 8-12s long), that’s a two hour data file having to wait 12s between trials. I don’t know how well it would work with the sorts of sample numbers we typically have in NIRS (which might be something like 50 trials at best).

I only quickly read the paper just now, based on their section 2.3, the pseudo-code for NIRS might look something like this:

Given a study with two experimental conditions (say number of trials is M and N for these two conditions)
1) Collect the trials from both conditions (now, you have a pool of N+M total trials)
2) Select a random partition of  M and N trials of the data (randomly shuffle the labels so the the conditions are now all mixed up- that’s the point)
3) Preform the GLM, regression, block average (etc) to get the separate response for the randomly selected TWO partitions
4) Calculate the test statistic for the random partitions (e.g. The difference of partition A vs B over some time window or time-to-peak, presence of an initial dip, etc etc etc).
5)  Repeat steps 2-4 to make a histogram of the stats value of interest
6)  Test the real data (e.g. The actual two conditions) against the histogram in step 5.

In practice using HOMER, you could create a bunch of copies of the *.nirs files where you have changed the stimulus labels around randomly.  Then run the analysis in HOMER and make a histogram of the results from all the random partitions.  Of course, there are a lot faster and more efficient ways to do this if you are able to just write your own matlab code.  Currently, HOMER would not natively support this sort of analysis.


----------------
Theodore Huppert, PhD
Associate Professor
Departments of Radiology and Bioengineering
Center for Neural Basis of Cognition
University of Pittsburgh
Email: huppertt at upmc.edu
Phone: (412) 647-8459



"Insanity: doing the same thing over and over and expecting different results"-  Einstein

From: Cécile Issard <cecile.issard at etu.parisdescartes.fr<mailto:cecile.issard at etu.parisdescartes.fr>>
Reply-To: "homer-users at nmr.mgh.harvard.edu<mailto:homer-users at nmr.mgh.harvard.edu>" <homer-users at nmr.mgh.harvard.edu<mailto:homer-users at nmr.mgh.harvard.edu>>
Date: Friday, March 6, 2015 at 9:13 AM
To: "Homer-users at nmr.mgh.harvard.edu<mailto:Homer-users at nmr.mgh.harvard.edu>" <Homer-users at nmr.mgh.harvard.edu<mailto:Homer-users at nmr.mgh.harvard.edu>>
Subject: [Homer-users] Nonparametric analysis

Hello Homer users,

Is there a way to conduct non-parametrical analysis in homer, similar to those proposed by Maris & Oostenveld (2007) for EEG data ?

Best regards,

Cécile Issard
Doctorante
01.42.86.43.20
Laboratoire Psychologie de la Perception - UMR8242
45 rue des Sts Pères
75270 Paris cedex 06
http://lpp.psycho.univ-paris5.fr/index.php
Labo bébé : http://recherche.parisdescartes.fr/LBB