TAR file including the code and the data.
Notes:
It is a self-containing MATLAB function that has been tested on MATLAB 2015a.
The Model parameter (line 13) specifies which scenario to run. The models differ in terms of whether the data gets z-scored in the process and whether the voxels have "positional sensitivity". The positional sensitivity is actually always explicitly modelled, however, for the cases which it's not needed, the variance of the positional likelihood function is set an absurdly great number (e.g. 10^10) so that in practice the voxels respond uniformly to all positions.
For LDA it currently uses the MATLAB's own fitdiscr function which is absent from MATLAB 2013 and below. This also means that the LDA function doesn't run on Octave. You can change the LDA function with any of the freely available ones on the web.
Similarly, z-scoring is done with MATLAB's zscore function from the Statistics Toolbox, which might not be available for all MATLAB versions.
For the sake of speed the script uses parfor instead of for -- change that if necessary.
Please email me for any questions.