What is the dominant factor of the mixer linearity?

I use inductive degeneration in the RF transconductance stage,but the P1dB is still low(about -25dB).The RF transconductance stage alone has a P1db about -10dB.

It's said in some papers that the overall Gilbert mixer linearity is controlled primarily by the transconductance stage.So why does the total P1dB is so much lower than that of the transconductance stage's?How to improve it?