Post-traumatic stress disorder (PTSD) is a mental health condition which severely affects society on many levels. However, PTSD remains often undiagnosed and untreated. To improve screening access and acceptance, a fully automated virtual human was designed to conduct many PTSD screening interviews (N=198). Here, we investigate which questions elicit the most indicative multimodal behaviors for PTSD. To this effect, we employ an information gain driven ranking procedure to identify the most informative questions. Further, we look into question dependent behaviors. To evaluate the question ranking, we investigate the discriminative faculty of the top questions using support vector machines. Our results reveal that only a subset of posed questions are required to robustly detect symptoms of PTSD in subject-independent machine learning experiments. We observe strong performance and can confirm that many of the identified behaviors correspond to commonly found behavioral indicators related to PTSD and depression.