So you want to run an MVPA experiment…
Lindsay MorganApril 9, 2012
Overview
• Study Design• Preprocessing• Pattern Estimation• Voxel Selection• Classifier
Study DesignBlocked design
• Smaller # of conditions• Better estimate of the
average response pattern
Event Related Design• Larger # of conditions– Similarity analyses
• Better estimate of the response distribution across exemplars
• Psychologically less predictable
• Requires sequence optimization (e.g., OptSeq, de Bruijn)
Study Design Suggestions
• Multiple runs– Independent data sets for training & testing– Many short runs preferable to a few long runs
(Coutanche & Thompson-Schill NeuroImage 2012)• Equal # of exemplars per stimulus class– Or use subsamples of more numerous class
Pre-processing
• Pre-process each run separately• Slice time correction• Motion correction• Smoothing?
To Smooth or Not to Smooth?
Op de Beeck NeuroImage 2010
Pattern Estimation
Raw signal intensity values• Suitable for block or
slow event-related
Betas (parameter estimates) or t values
• Suitable for all designs• Derived from GLM– Accounts for overlap in
HRF– Can remove motion
effects and linear trends
Mur et al., Soc Cog Affective Neurosci, 2009
Data transformation so far…
Kriegeskorte et al., Frontiers Sys Neurosci, 2008
Ungrouped design• 96 images • Each image
presented 1x/run• 3 comparisons• Inanimate vs.
animate• Face vs. body• Natural vs.
artificial
Betas or t values?
Misaki et al., NeuroImage, 2010
Pattern Normalization
Misaki et al., NeuroImage, 2010
Pattern Normalization
Misaki et al., NeuroImage, 2010
Data transformation so far…
Mur et al., Soc Cog Affective Neurosci, 2009
Voxel Selection
• Typically, performance decreases as the # of voxels increases
• Data must be independent of classifier– Anatomically-defined region– Functional localizer– Training set from your experimental data• E.g., ANOVA for all conditions at each voxel select top
N voxels
The Classifier
Misaki et al., NeuroImage, 2010
Which classifier should you use?
Misaki et al., NeuroImage, 2010
Data transformation complete!
Mur et al., Soc Cog Affective Neurosci, 2009
How to implement the classifier
• AFNI 3dsvm• Princeton MVPA toolbox• PyMVPA toolbox• LIBSVM toolbox
General Conclusions
• Design your experiment to yield as many independent patterns as possible
• Estimate your patterns using t values (or z scores)
• Use a linear classifier