Come to Seattle to work with us on Stan!

Our colleague Jon Wakefield in the Department of Biostatistics at the University of Washington is interested in supervising a 2-year postdoc through this training program.

We’re interested in finding someone who would with Jon and another faculty member (who is assigned on the basis of interests) on exciting projects in spatio-temporal modeling and the environmental health sciences; the successful applicant will also be working with us as a core Stan developer. There are different ways this project can go, based on your interests and expertise, but we emphasize the training aspects. Before applying to the program, we recommend you first email me with a CV and brief cover letter. This position is only open to US citizens and green card holders.

P.S. Steven Johnson sent in the above picture of a cat who would like to come inside and start working on Stan full time.

7 thoughts on “Come to Seattle to work with us on Stan!

  1. Jon’s doing great work on the statistical, methodological and computational aspects of spatio-temporal modeling with a fantastic group of grad students and postdocs. I hung out with them for a couple hours when I was in Seattle for ISEC to talk about spatial modeling in Stan.

    In addition to Jon, we’re working on the broader problem with Dan Simpson (of INLA fame, Stan NumFOCUS board member), Aki Vehtari (co-author of BDA and author of GPstuff, Stan ore developer), Rob Trangucci (Stan developer who wrote all of the recent matrix algebra analytic derivative improvements; he also just rewrote the GP chapter of the manual for the next release), and Michael Betancourt (chief of smooth manifolds and MCMC for Stan and a computational GP expert), and Andrew Gelman (who needs no introduction on this blog!).

    This is an amazing opportunity to work with some of the best researchers in the world on spatio-temporal models. As Jonathan Auerbach asked me the other day, “When’s the last time you saw truly i.i.d. data set?”. He and Rob are building spatio-temporal models in Stan for pedestrian traffic deaths, very much like what Mitzi did in her recently published case study. Because so many of us are focused on improving this area, a candidate could look forward to joining the core Stan dev team and working closely with us on modeling and implementation issues.

    From my perspective as a Stan developer, it would be perfect if we could find someone who was excited about applied modeling who would motivate work from the models down through the Stan programs to sparse matrix algorithms and densities in C++. There’s a lot of room for improvements in our current implementation of these models and there’s a huge demand around the world for more efficient and robust versions of spatial models. Jon was the one behind PK-BUGS (like Andrew, he has worked on lots of application areas) and he has a long track record of working in cutting-edge modeling areas that require fundamental MCMC work on the back end.

    • +1 for all the above

      i suspect sparse methods for STAN would blow the lid off of a lot of the space – time ‘big N location’ kind of bottlenecks.

      • We’re certainly hoping we can add a couple of orders of magnitudes to what can be fit with full Bayes. Mitzi fit an intrinsic autoregressive model to a 700-location data set in 10 minutes that BUGS/JAGS couldn’t make any progress on.

        This should also give us a way to evaluate how well the approximations in INLA work and whether BUGS/JAGS are getting the right answers for smaller models they can handle like the classic Scottish lip cancer data set. There are case studies on our web site (under docs; two in the main set of case studies and a GP case study in the StanCon 2017 set).

        • The interesting thing is that it’s not actually very hard to write a very efficient, bespoke MCMC scheme for this class of model (where “not very hard” = there was *a lot* of research on this about 10 years ago that pretty much answered the question). But they turn out to be much harder for non-specialised MCMC algorithms to solve (hence BUGS/JAGS having problems, and Stan not being there yet).

          If you want to evaluate the approximations in INLA, you can probably just use the MCMC sampler that’s hiding in it (which is one of these specialised methods)…

      • Just because I hear a lot of weird things about size in the space-time world, the largest problem that I know of that’s being solved has O(10^11) space-time locations (done by the always impressive Finn Lindgren in Edinburgh). For that you’ve got to do some really really interesting things numerically.

  2. Sounds like a great job. I am interested in using spatio-temporal models in PyStan to look at the spread of fungal disease across the US. Currently we have it implemented in INLA.

Leave a Reply

Your email address will not be published.