It estimates the variance by systematically leaving out one observation at a time.

It requires the data to be normally distributed.

It is a parametric method for calculating bias.

It is computationally more intensive than the bootstrap for large $n$.