Dear Ramesh,

I would recommend adjustment for age. If the distribution of ages differs between your groups, then you probably want to try to estimate what the group difference would be for hypothetical groups of matched age, which is what adjustment is attempting. Even if age is relatively well balanced across your groups, then adjusting for it should be helpful, since it should explain some otherwise unexplained variability, which should increase statistical power. You might also like to consider adjusting for gender and total intracranial volume, see e.g.
  http://dx.doi.org/10.1016/j.neuroimage.2...

To use cluster-level p-values for VBM, you need to correct for non-stationary smoothness. It sounds like you are using Christian Gaser's VBM8 toolbox for SPM8, which I believe does this for you. In that case, either is fine. Cluster-level inference should be more sensitive to spatially large contiguous effects, while peak-level should be more sensitive to very focal effects. Cluster-level only lets you conclude that a particular cluster as a whole is significant (which can be awkward if the cluster is very large) but not to interpret individual voxels within it as significant. Peak-level lets you do the latter.

Regarding the overlays, I think you will find it helpful to look at chapter 28 of the SPM8 (not VBM8) manual,
  http://www.fil.ion.ucl.ac.uk/spm/doc/spm...
even though it is for fMRI rather than VBM, the results machinery including overlays etc. is the same. In particular, see pp.226-228.

Best wishes,
Ged


P.S. You might like to direct further SPM related questions to the SPM mailing list:
  http://www.fil.ion.ucl.ac.uk/spm/support...
or if you take up Cartik's suggestion of using FSL, then the equivalent mailing list is:
  https://www.jiscmail.ac.uk/cgi-bin/webad...