I have used G-Theory to estimate variance components (VC) in the past, and I tend to use p x r x i crossed designs.  I was recently approached by a colleague who has a singe-facet design (p x i). She's calculated VCs for p and i, but she is very confident that there is an interaction.  I initially informed her that she'd be unable to generate a VC calculation for the interaction term because G-theory can't parse the interaction term from error (i.e. pxi, e).  I was surprised when she showed me that by using SPSS (the variance component command in the GLM analysis option), she was able to get separate estimates (using MINQUE) for p, i, pxi, and error.  

Does anyone know if this is valid?  Is this somehow a non-addiditive model?  If you can point me to resources that clarify for me I'd appreciate it.