"Root Sum Squared" analysis or sometimes called "the law of propagation of uncertainty" is a method of combining uncertainty components estimated as standard deviations.
If i'm correct then its a method of combining components in a mathematical model, but i dont think tolerances can be simply combined using Root Sum Squared. The maths assume normal distribution and this might be un-conservative as it excludes the possibilites of items being at the max/min end of their tolerance boundaries. It is better suited to a large amount of items where statistically they will follow the standard deviation trend as a whole. However never having carried the process out i'm probably not the best suited to give you full guidance.
Practically speaking though, i would think that the allowable stack-up of a particular group of items would be dependant upon the relative geometric properties of each. I was always told that if it looks good then it probably is, if it looks cr*p then...