Introducing Bayes and Computation in Eight Short Weeks

I'll be teaching Applied Bayesian and Computational Methods to the Masters-level students here at CMU in the spring, though as the course is scheduled for a "mini-term", I'll only have a short time to present many of these ideas to the group. My plan is to teach from Gelman and Hill ("ARM", or as AG wishes he'd called it, "Regression") since it's filled with useful tips and it's inexpensive.

At this point, my plan is to do what most practitioners do when teaching: show the students how to avoid the mistakes you've made. At a minimum, this involves

• Think about the model before you code it.
• MCMC may be flashy, and powerful, and the root of a lot of my work, but it's also really, really easy to screw up.
• Posing a stochastic model is the beginning of scientific wisdom, not the end of it. Therefore, think about what other approaches may be equally valid.

One of my least favourite classroom experiences was from a lecturer whose approach to teaching problem solving was to say "well, you could try this", as if teaching was a laundry list. So making sure to put these methods in perspective is at the top of my agenda.

For coding, my plan is to go with R and WinBUGS using the R2WinBUGS package; as much as I'd like to go pure R, I'm sure I'd spend too much time worrying about the little details and missing the point of the modelling, which in an eight-week course I can't afford.

For those students who come into the course with really good skills in R already, I'd consider introducing C or C++ coding of routines if it were a longer course or one purely focused on the details of stochastic model programming, but right now I hesitate to bring it in.