In May 2018, I was lucky enough to travel to the USA with generous support from the Statistical Society of Australia’s Golden Jubilee Travel Grant. I attended three excellent statistical meetings, ate far too much fried chicken, and was greeted with generous hospitality from both organisers and delegates (for which I am very thankful).
I started my trip by attending the inaugural Conference on Predictive Inference and its Applications at Iowa State University. There was a strong show of novel predictive inference applications and theory. I was very impressed by the overall program put together by Dan Nettleton and colleagues. I was also impressed with the strength of Prof Nettleton’s table tennis game. Edward George (U. Pennsylvania) gave the opening lecture on “High Dimensional Predictive Inference”. It was rich with statistical theory from Stein’s shrinkage estimators to current Bayesian methods, and introduced me to several analogous theories in George’s own work on shrinkage for prediction. The Snedecor Memorial Lecture was presented by David Dunson (Duke) and was given on “Predictive Inference from Replicated Network Data”. A very interesting talk, utilising Bayesian nonparametrics with applications ranging from soccer passing networks all the way to brain data. I was glad to receive feedback from many delegates on my poster presentation “Bayesian regularisation from stochastic constraints” which was recognised as one of the best posters (see the attached prize photo which also captures my fight with jet lag).
I’d like to list a few of the other presentations I enjoyed: Martin Wainwright (Berkley) – “Computational Challenges in High-dimensional Prediction”, Jing Lei (Carnegie Mellon) with “Accounting for Uncertainties in Predictive Inference”, and Stephanie Kovalchik’s (Victoria U. & Tennis Australia) “Predicting the Emotions of Tennis Players from Single-Camera Video”. All slides are available on the conference website (https://predictiveinference.github.io) and I highly recommend looking through them.
The 1st Midwest Colloquium on Statistical Machine Learning was held at Iowa State University directly after the Predictive Inference conference. As the name suggests, it was more focused on computer science and machine learning methods and applications, but with a noticeable statistical flavour. Bertrand Clarke (U. Nebraska-Lincoln) presented on “Interpreting the Uninterpretable: Kernel Methods, Shtarkov Solutions, and Random Forests” demonstrating how to interpret machine learning methods, often criticised as uninterpretable, by translating results to an interpretable asymptotic model and measuring loss in predictive performance. This is an issue that I have heard raised by statisticians regarding machine learning methods, so I was glad to discover principled work on this problem. Sanmi Koyejo (U. lllinois at Urbana-Champaign) presented “Probabilistic Models for Brain Data Analysis” with results from time-varying Gaussian graphical models for brain network data that used Gaussian processes as time-varying weights for subnetworks of the precision matrices. These talks (and more) can be found here: https://register.extension.iastate.edu/msmlc/keynotes.
In early June, I attended the Total Survey Error Workshop in Durham, NC, which was jointly hosted by the Duke Initiative on Survey Methodology and the Odum Institute at UNC Chapel Hill. A big theme emerging from presentations and among discussions with delegates were non-response and non-probability survey designs. In particular, Burton Levine (RTI International) gave a talk entitled “Evaluating Bias in Redirected Inbound Call Sampling Surveys” which used data collected from inbound calls to inactive toll free numbers. A very convenient sample, but not without its challenges as a non-probability survey. In another talk, dynamic Bayesian hierarchical models were used by Christopher Claassen to analyse support for democratic ideals using surveys collected from over 120 countries over a period of 24 years. This work finds evidence for the hypothesis that democracy requires public support to survive, for which previous empirical support has been weak. As noted by the audience, this presentation served as a timely reminder of threats democracy is facing in the US. In my talk, I discussed “Undecided voters in US Presidential elections, 2004 – 2016”, analysing the effect of undecided voters (as measured by pre-election polling) on polling error. It was received well, and I needn’t have worried about being an Australian discussing US politics using Bayesian methods at a survey workshop.
I would like to thank all the academics and students I met, who made my trip so worthwhile. It’s amazing and humbling to see a glimpse of how much research is being carried out all over the world. Thanks again to the SSA for financial support to attend these conferences. I highly recommend other young researchers to apply for this grant next year and keep an eye out for future conferences like these. A great resource for finding such meetings all over the world is https://www.imstat.org/meetings-calendar/.