Rough sets are widely used in feature evaluation and attribute reduction and a number of rough set based evaluation functions and search algorithms were reported. However, little attention has been paid to compute and compare stability of feature evaluation functions. In this work, we introduce three coefficients to calculate the stabilities of feature significance via perturbing samples. Experimental results show that entropy and fuzzy entropy based evaluation functions are more stable than the others and fuzzy rough set based functions are stable compared with the crisp functions. These results give a guideline to select feature evaluation for different applications.