Online Course on Model Averaging

There is now an online course to go with my book on model averaging. This course provides an up-to date overview of the topic, from both the Bayesian and frequentist perspective, including methods that are popular in machine learning, such as bagging and stacking. Recent developments, published since the book came out in 2018, are also covered, including confidence distributions, the frequentist analogue of posterior distributions. There is emphasis on understanding the theory underlying the methods and on experience using the methods in RStudio.

Cost: US$200 (+ applicable taxes) for 30 days access. You can find more information and registration details here

Faster bootstrap intervals for indices of relative abundance

I recently provided advice on the use of “single-fit” bootstrapping to obtain confidence intervals for indices of relative abundance, when fitting a delta-lognormal model to fisheries data. The key idea is that of resampling the model parameters from a multivariate normal distribution. This is computationally nice as it allows a parametric bootstrap confidence interval to be calculated without refitting the model. You can find more details about “single-fit” bootstrapping in these seminar slides.

Estimating percentiles in air dispersion modelling

I have recently been providing advice to STIMBR (Stakeholders in Methyl Bromide Reduction) on the estimation of percentiles when modelling the dispersion of chemicals in the atmosphere. It was interesting to see just how unreliably the highest upper percentiles of a skewed distribution are estimated, even from a very large sample. This has implications for the use of such percentiles in setting environmental health regulations.

Testing lack-of-fit in phylogenetics

I am currently working with David Bryant at Otago University on testing lack-of-fit of a model in phylogenetics.

It’s great to be able to use my research experience in estimating overdispersion on an important scientific problem. One of the key aspects of the problem is the sparseness in the data, which are multinomial with a very large number of categories.

Joint work with Farzana Afroz and Matt Parry will be particularly useful in this setting, as we were able to derive an estimate of overdispersion that works really well for sparse multinomial data: the paper we published can be found here.