The radio occultation experiment on the CHAMP satellite has been collecting observations of the Earth's atmosphere since April 2001. Previous work has shown that diffraction effects not accounted for by geometrical optics processing can be partially corrected by back-propagation and canonical transform methods, such as implemented at the Jet Propulsion Laboratory. In the present paper we evaluate the bias and standard deviation of refractivity differences between Data Assimilation Office global analyses and observations processed by these three methods. In the tropics at 2--5 km altitude, the refractivity biases range between -2.5% and -0.5% for geometrical optics, between -1% and 0.5% for back-propagation, and between -0.5% and 1.5% for canonical transform. We also assess the methods by performing one-dimensional variational temperature retrievals and comparing them with close radiosondes. Our final conclusion is that canonical transform is a better candidate than geometrical optics and back-propagation for generating GPS radio occultation datasets for data assimilation.