Real data. I always use real data, and I never use data sets that are in the book, or that I've used in class as examples, or that I've ever used on an exam previously. I want every data set to be both real and new to the student. Imposing these constraints on myself means that I spend quite a bit of time, during the weeks leading up to the exam, looking for suitable data sets.

Point values. Questions are grouped by data set. A typical exam might involve 3 or 4 different data sets. Some data sets may be rich enough to support several questions. Others may lend themselves to just one. To help students plan their allocation of effort, I try to create a set of main questions that count equally. (The number varies from exam to exam, but is usually between 6 and 12 main questions.) Some questions may have several parts, but here, also, I try to create divisions into parts that count equally.

Conceptual vs. ritual. For the most part, I rely on HW and weekly quizzes to check for computational ritual. I want the mid-term exam to deal with conceptual understanding. Here I have in mind two main kinds of understanding: interpreting results (going from the abstract numbers or plots to their concrete meaning in relation to the applied context), and recognizing abstract structures (going from the concrete applied scenario to the relevant abstract models or methods). Although that's my main goal, I try to start the exam with a first problem that I hope all prepared students will find straightforward. This one tends to be somewhat more mechanical. Also, unavoidably, most problems have a somewhat more mechanical element to them, and this strikes me as perfectly OK.

See all answers to this question

See all of George's answers