One cannot say whether EKF will perform better than UKF in any case. There is theoretical proof though, that the UKF will perform better than the EKF in estimating the mean of the state estimates by one order in the Taylor Series expansion of the nonlinear transformation. In practice, however, the answer to which estimator will perform better than others will be system dependent, I believe.
If the noise levels in states and measurements are are very large (signal to noise ratio), both EKF and UKF will give same results. Even for systems with near Gaussian pdfs also you can achieve the same kind of performance. EKF can never be better than UKF, because both converges to KF in mild conditions, i.e linear state and measurement models.
If the system is linear or linear nearly, and if the numerical condition is bad, like the state is small floating value: 1e-3, ukf will lose precision in square root computation seriously. In this condition, I think EKF performs better (just based on my experience :) ).