Stat 6103, Bayesian Data Analysis

Modern Bayesian methods offer an amazing toolbox for solving science and engineering problems. We will go through the book Bayesian Data Analysis and do applied statistical modeling using Stan, using R (or Python or Julia if you prefer) to preprocess the data and postprocess the analysis. We will also discuss the relevant theory and get to open questions in model building, computing, evaluation, and expansion. The course is intended for students who want to do applied statistics and also those who are interested in working on statistics research problems.

Stat 8307, Statistical Communication and Graphics

Communication is central to your job as a quantitative researcher. Our goal in this course is for you to improve at all aspects of statistical communication, including writing, public speaking, teaching, informal conversation and collaboration, programming, and graphics. With weekly assignments and group projects, this course offers you a chance to get practice and feedback on a range of communication skills. All this is in the context of statistics; in particular we will discuss the challenges of visualizing uncertainty and variation, and the ways in which a deeper integration of these concepts into statistical practice could help resolve the current statistical crisis in science. Statistics research is not separate from communication; the two are intertwined, and this course is about you putting in the work to become a better writer, teacher, speaker, and statistics practitioner.

The communication and graphics course should be no problem; I’ll teach it pretty much how I taught it last year, with 2 meetings a week, diaries, jitts, homeworks, class discussions, projects, etc.

The Bayes class I’ll be doing in a new way. It’ll meet once a week, and my plan is for the first half of each class to be a discussion of material from the book and in the second half for students to work together using Stan, with me and the teaching assistant walking around helping. Also, the homeworks will be more Stan-centered. The idea is for the students to really learn applied Bayesian statistics, as well as to have a chance to grapple with important theoretical concepts and to be introduced to the research frontier. We’ll see how it goes. The key will be coming up with in-class and homework assignments that give students the chance to fit Bayesian models for interesting problems.

The first class sounds great (the second too, but the first one would help me a lot with stuff I’m working in now). Any plans for a related MOOC in the future?

Sorry comment meant to appear here:

Oh to be a student again — both courses look fabulous Andrew.

REQUIRED COURSES FOR ALL HUMANITIES RESEARCHERS?

I am going on 15 years out of graduate school, where I obtained a PhD in the Humanities. My education was very good, but I also wasted time from not following courses such as your. Sure computers weren’t very good in the late 90s, and it’s easy to romanticise what could have been. However, as I get on in my career I wonder if every — or nearly every — empirical research problem worth pursuing isn’t benefited from a knowledge of Bayesian statistics and a capacity to communicate findings clearly? This isn’t to say that the only training Humanities scholars should have is in Bayesian stats & data viz. Historians need archives, Anthropologists need to learn how to interact with people — there’s lots of assimilation of past knowledge, climbing the shoulders of giants … but when a discipline’s research problems are not conceived in manner for which inference is possible, some movement in certainty or uncertainty in response to evidence, it seems to be a sign that people are settling for unimaginative questions. Too harsh on Humanities scholars? Live and let live?

PROBLEM SETS

Maybe let the students try to come up with their problems, perhaps in conversation with peers/researchers who are not in the course, subject to constraints (some controversy, predictions that might improve inference, data to hand, appropriate simplicity… ) … Where their questions fall down could be interesting, and they might feel more engaged working through each other’s problems. Journey-as-the-destination and all that.

Plus who wants to think up problem sets when its autumn in New York…?!

Hooray for more Stan literacy! Can’t wait to sit on a few of the classes.

My suggestion for an interesting problem is to take the notorious polynomial regression air pollution study from China & apply a Bayesian Analysis to that data-set. For that’s one study with seemingly good primary data & measurement just crappy analysis on top.

Rahul:

That example could work but it might not be so easy to get the raw data. You can’t just take the data that they put in that notorious graph, for example, as it has implausibly high life expectancies for some regions.

Andrew:

Agreed. My observation is that examples tend to be either rich & complex but stand alone or comparative yet just toy cases.

What would make a great set of examples is applying Stan or Bayesian methods to data-sets from existing papers that had been analysed using NHST or other methods. It would be interesting to contrast the final conclusions from the two approaches.