To increase the computational speed of the algorithm.

To handle problems associated with separation (where MLE does not converge or is biased).

To allow for non-linear relationships between predictors and the logit.

To automatically perform variable selection.