table of contents
LIPSIA     Bayesian second-level analysis
vbayes, vbayesgroup

The program 'vbayes' conducts a second-level analysis of groups of subjects based on Bayesian statistics. The program takes as input parameter estimates for single subjects for a contrast of interest and its variance (output of 'vcolorglm'/'vgetcontrast' or 'vwhiteglm') and provides as output a probability map showing the probability for the contrast to be larger than zero.

The method is described in detail in Neumann and Lohmann (2003) (see below) and should be viewed as a possible alternative to 'vonesample_ttest'. However, it provides probabilities of activation for the contrast of interest rather than z- or t-values.

The Bayesian analysis is very robust against outliers, which in the traditional fixed- and random-effects models (!) have a large influence on the statistical significance of the results (see Figure 1). It is therefore particularly effective in the detection of cortical activation caused by small experimental contrasts.




Figure 1: The influence of a single outlier in a group of subjects on both traditional z-values and Bayesian probabilities of activation.


The program 'vbayesgroup' conducts a Bayesian comparison of two groups of subjects. It takes as input the results of 'vbayes' for the two groups. Output is a probability map showing the probability for the contrast to be larger in group2 than in group1:

P(contrast group1 < contrast group2).

Caution: When using 'vbayesgroup', the program 'vbayes' must be called beforehand with the parameter '-level true'.

Practical use

Inputs to 'vbayes' are contrast images of all individual subjects, i.e. outputs of 'vgetcontrast' or 'vwhiteglm'). The programs can be called as follows:

vbayes -in conimg_subj1.v conimg_subj2.v conimg_subj3.v -out group_bayes.v

or

vbayes -in conimg_subj1.v conimg_subj2.v conimg_subj3.v -out group_bayes.v -level true

The first call produces a Bayesian probability map, the second call with '-level true' produces a map which serves as input to 'vbayesgroup'. 'vbayesgroup' is called as follows:

vbayesgroup -group1 group_bayes1.v -group2 group_bayes2.v -out group_bayes.v

Note: The output of 'vbayes' and 'vbayesgroup' looks like a zmap, but it contains probabilities instead of z-values. Within the Bayesian framework there exist no thresholds indicating the significance of a result. Typically, very large values of over 99% indicate a high probability of activation (see Figure 2). However, this is only a rule of thumb!

Note further that, like for zmaps, negative (blue) values indicate the probability for the reversed contrast to be larger than zero or, in case of 'vbayesgroup', that the contrast is larger in group1 than in group2.




Figure 2: A comparison of the outputs from 'vonesample_ttest' (left) and 'vbayes' (right).

Parameters of 'vbayes':
-help
Prints usage information.
-in
Input files (Contrast images, e.g. output of vgetcontrast or vwhiteglm).
-out
Output file. Default: (none)
-level [ true | false ]
Whether to produce output to be used for 3rd level analysis. Default: false
Parameters of 'vbayesgroup':
-help
Prints usage information.
-group1
Input files from first group (Output of vbayes).
-group2
Input files from second group (Output of vbayes).
-out
Output file. Default: (none)
Literature:

J. Neumann and G. Lohmann. "Bayesian second-level analysis of functional magnetic resonance images." NeuroImage, 20(2), 1346-1355, 2003.



Max Planck Institute for Human Cognitive and Brain Sciences. Further Information: lipsia@cbs.mpg.de
Copyright © 2007 Max Planck Institute for Human Cognitive and Brain Sciences. All rights reserved.