When you're close to self resonance, that means that parasitic capacitance, either substrate or winding-to-winding, is causing a resonance. If the Q is high, then resonance can be very narrow band! You therefore can't tune both matching and single-ended to differential convertion over a wide-band.
Lets consider the sources of capacitance. Substrate and winding capacitance is distributed over the entire length of the inductor. The closer you are to resonance, the more important it is that you model the distributed nature of that capacitance.
For substrate capacitance, you need to model the distributed substrate below the inductor accuratlly in order to get a correct response. Your primary source of variance will be the inductor to substrate height variation, or the oxide variation. Simulate your inductor for the extremes of the process, and extract the S-parameter repsonse using tools suchs as momentum.
Next is winding-to-winding capacitance. This capacitance generally doesn't vary as much because the spacing rules are generally better controlled then say the oxide thickness (as a percent of tottal). Use momentum again and consider the wrost-case width's of the metal (dl variations).
Lastly is your load impedance variation. The closer to resonance you are, the more sensitive the matching and single-to-differential convertion is.
If you take all these into account and simulate it and it still meets your requirements, then you are o.k. Typically though, the oxide variation is approx +/-10%, the substrate modelling is too difficult to guarantee and accurate result and the load variation is too high to operate close to resonance.
Greg