Re: package model estimation
Yeah I see the problem.
So you'll have very different S21 if you have used the wrong package model!
You've got the solution: you should fit the measurements with simulations with downbonding, input/output bonding and so on.
In my experience, if you have leadless package (such as QFN) and you are matched to 50 Ohm, up to 3 GHz you can neglet the package parasitics on input/output, but it can also be very different... depends on many parameters.
The usual way to perform the package modeling is to use a 3D electromagnetic simulator, such as Ansoft HFSS or Q3D extractor (it depends on frequency vs. package dimentions), in which you have to draw the package in 3D and simulate it with input/output ports (could be very hard to compute it).
Another way is to develop some special test ICs (for example with short, open, dummy loads and so on), measure them and deembed the package model, but this is less accurate (due to the inaccuracy of your IC loads).
If you don't have the 3D simulator the old "cut and try" way is the best one: if you have accurate measurements on S-parameters, noise, linearity, in 2 or 3 wafer runs you should be able to optimize your design with package.
Good luck
Mazz