Leveraging density ratio models in a binary instrumental variable inference with a binary outcome: A retrospective approach