For in-class exams, I try to use real data from questions that I generally find in statistics texts other than the one I am using in class. I try to have every question reflect real data or at least realistic data. For the take-home exams, the data is definitely real. It comes from my research with scientists in a variety of disciplines. I, however, adapt the questions for the exams to be appropriate for introductory statistics. So the questions might not be the actual questions asked in the research paper, but they are definitely realistic questions. I do not think the distinction between real and realistic data/questions is important. I DO think it is important to use data in the correct context. There are paper now in the literature that use random number generators to generate "real data" so that students have individual data sets, but I do not like that approach because it is devoid of context and context is very important. I generally have questions valued according to how long I think it will take the student to complete the question. Questions taking more time, have higher point values. I like to have each in-class exam total 100 points. I think right now, my exams are about 40% calculation and mechanics and 60% conceptual and interpretation. I think I have been good at asking interpretation questions, but only recently have I tried to construct conceptual questions.