This paper is the first to structurally model how a test taker answers questions in a multiple choice exam. We allow for the possibility of a penalty for a wrong answer which makes risk averse examinees more likely to skip questions. Despite the lack of item response data, we can estimate the model by using the insight that skipping behavior, together with penalties for wrong answers, makes certain scores much more likely than others. Using data from the Turkish University Entrance Exam, we estimate the model and find that candidates’ attitudes towards risk differ according to their gender and ability with females and those with high ability being significantly more risk-averse. However, the impact of differences in risk aversion on scores is small. As a result, a higher guessing penalty increases the precision of the exam, and does so with a minimal impact on gender bias.