Penalized likelihood estimation via data augmentation