This post is from Bob

Mitzi Morris and Bob Carpenter, two of Stan’s developers, will be presenting a tutorial on Stan and Bayesian data analysis aimed at psychometricians this summer.

• Modern Modeling Methods Conference (M3), Fordham University Lincoln Center Campus, Manhattan, June 22–24
• Program here

Abstract

This workshop is a full day, hands-on introduction to Bayesian modeling and statistical inference using the probabilistic programming language Stan.

The course will be organized around the key properties of Bayesian statistical modeling for science, including the nature of uncertainty, modeling a generative process through a data generating distribution, modeling existing knowledge through a prior, and pushing uncertainty through inference. As we do this, we will show how Stan can be used to both code the models and perform statistical inference for quantities of interest, be they retrospective parameter estimates or prospective predictions or forecasts. We will concentrate on full Bayesian posterior inference, including a discussion of calibration, model checking for both prior and posterior inference, and model comparison with cross-validation. We will spend some time showing how some structural equation models (SEM) can be translated directly to Stan and will also introduce psychological models for educational testing, crowdsourcing, rating and ranking, and real-time decision processes.

This class will require a notebook computer with a network connection (Wifi will be available in the classroom). We will use the Stan Playground, which runs Stan in the browser, which we will pre-populate with models of interest. We will probably also break into R or Python at various points to demonstrate methods not yet supported by the Playground, such as the brms regression expression language.

Andrew on psychometrics

Andrew once told me that any model you could come up with was probably invented by a psychometrician 50 years ago (make that 60—he said it at least 10 years ago). I have evidence that he’s right form the project that drew me into Bayesian statistics—crowdsourcing. Andrew and Jennifer Hill helped me formulate a crowdsourcing model where raters give you noisy measurements of underlying categorical variables (e.g., they answer survey questions about whether a word in context is a noun, for example, to use something I was working on at the time). Turns out Phil Dawid and A.P. Skene published the same model in 1979 in one of the earlier applications of the expectation maximization (EM) algorithm and they used natural language data (drawn from medical records).

The rest of the conference

The rest of the program looks really great—it’s just the kind of applied wrestling with real data that I like.

Speaking of Jennifer Hill, she’s one of the keynote speakers at M3. Every talk of Jennifer’s I’ve attended has been great. You may know her as Andrew’s co-author on the regression books, which I cannot recommend highly enough if you’re interested in this kind of applied modeling.

Beyond the conference

You see the same kind of Bayesian modeling focus for real data at venues such as ISEC (international ecology conference to which I went to once just because I like these models and these kinds of conferences), StanCon (see you in Uppsala in August!), and GeoMed (which Mitzi attends). I’m sure there are more in other fields. I’m always disappointed that there’s almost nothing like these kinds of nitty-gritty applied papers at ISBA (Nagoya this summer) or BayesComp (somewhere next year). The conferences about Bayesian statistics or computing that I’ve been to have all been super theoretical.