I consider myself as pretty good at sample size determination, but there is not enough information here to do much. To successfully calculate this, we would need:
Some idea of what the particular marginal structural model looks like (number of factors, relationship between factors)
A statement of what size difference you wish to detect in the various factors
A preliminary estimate of the variability (this can be addressed using multiple runs at various levels)
A preliminary estimate of the correlations between factors (for instance, the off-diagonal entries in a matrix)
The acceptable risk for both Type I and Type II errors
From this, a simulation approach could be used, as I doubt very much that any of the closed-form methods available in PROC POWER are good candidates for structural models. I suspect that you are going to have to go through the literature at a fairly detailed level to set up a reasonable null model and alternative model.
SteveDenham
