  TASK CORRELATION

  Run the Task Correlation ( the sm_taskcorr programs which are described below).
  Check correlations between motion and task stimuli. If no task correlation
  and there is over 2 mm of motion, it's good to add motion covariate explanatory
  models to the design matrix to reduce false activations.
  If motion is highly correlated with the task, don't add the motion covariates
  because they will degrade the accuracy of the results.
  You can run the task correlation program either before or after the motion
  parameters are included in your model as regressors. In the latter case you will
  be asked to choose the SPM.mat file and the correlation matrix for the entire
  model (including the motion regressors) will be displayed. Look for strong
  correlations between task and motion regressors as these may degrade the accuracy
  of the results.
  When using this feature before the motion parameters are included in the model,
  you will also need to specify not only the SPM.mat file but also the text files
  containing the motion data for each session. For event related designs the program
  displays the correlation matrix between seven motion parameters (x,y,z,pitch,roll
  yaw and the total norm) and the task conditions. Again, you should look for strong
  correlations (or anti correlations) in the bottom rows (i.e., correlations between
  task and motion parameters).
  For block designs, in addition to the same correlation matrix, two other graphs
  are displayed. The top plot shows the results for fitting a blockwise linear model
  to the motion parameters. The assumption is that within  each block the motion is
  roughly a linear drift with significant changes of position and direction of the
  drift (slope of lines) between blocks. The motion parameters x,y,z are plotted in blue,
  green and red, along with the black linear fit. If the fit seems to capture the
  behaviour of the motion data (black segments fit the motion parameters, changes in
  slope occur on the boundaries between blocks) then the task and motion parameters are
  correlated. By comparing the ratio of residuals of the fit for the actual data with
  those of random data (simulated by permuting the difference series of the motion
  parameters) we can quantitatively estimate the goodness of the fit.
  This ratio is printed for each motion parameter on your Matlab screen. Values below 0.75
  suggest a good fit and hence a strong correlation.

  Another method to assess whether there is any overall correlation between the motion
  parameters and the reference waveform (i.e., the task condition) is by performing an
  F-test. This essentially tests whether *any* linear combination of the motion
  parameters correlates significantly with the reference waveform. Even though the
  individual motion parameters may not correlate significantly, together they may.
  The F-scoers for the actual motion data as well as for randomly generated normally
  distributed noise are also printed on your main Matlab screen.
  The bottom plot shows the residuals for both types of data. This is the box car
  orthogonalised with respect to the motion parameters, and is then effectively what
  you'll be testing for when including the motion parameters in the statistical model
  with the box-car. Note that the residuals of the random data (in light green) do
  not change the overall shape of the box car. If the actual motion data is correlated
  with the test condition, the residuals will for the real motion data (in blue) will
  deviate from the overall shape of the box car.

