There is a document cited somewhere above that has a nice graph of measured results vs. the model. The measured results have quite a large amount of spread relative to the model. I would suggest allowing for a good amount of margin, as much as you can tolerate.