University of Sydney statistics professor, Gillian Heller, delivers the New South Wales branch 2020 annual lecture on her career-long journey though regression modelling and computing, followed by a vibrant end-of-year dinner at a pub around the corner.
Professor Gillian Heller spent 28 years in the statistics group at University of Macquarie before recently moving to the University of Sydney Clinical Trials Centre. Prior to that she was a graduate student at the University of Cape Town in her home country of South Africa. During this time flexible regression modelling and computing has been a theme that permeates her work – most recently driven by clinical trials studies and their data. Gillian described her career-long journey on this topic in a talk titled "The New Normal: Flexible Regression Models". It was the 2020 Annual Lecture of the New South Wales branch - and the first one to have a tele-audience because of pandemic restrictions.
An early figure plotted log(computing efficiency) against time for the period 1946-2010. The signal was linear and consistent with exponential increase since the ENIAC computer of the late 1940s. Gillian told us about being exposed to the package GLIM as a graduate student and made its use
sound quite painful from a 2020s standpoint. Later in the talk R packages such as gamlss, VGAM and mgcv were mentioned - which allow very flexible regression analyses to be carried out on laptops for large data sets.
A key paper that was highlighted early on is the 1972 one by Nelder and Wedderburn titled "Generalized Linear Models", that extended Gaussian response models to those with general response and has had a profound effect on many of us, including Gillian. A quote from esteemed Australian statistician, Murray Aitken: "Theoretical and applied statistics were both convulsed by the publication of the GLM paper by Nelder & Wedderburn (1972)" was shared with the audience.
The theme of parameter orthogonality ran through most of the talk. It is well-known that location and scale parameters are orthogonal in Gaussian response models. This matter is less clear in generalized response models and speaker Heller showed the importance of this aspect via analysis of some clinical trials data with the Poisson Inverse Gaussian family and a particular reparameterization.
Gillian also told us about her work with Mikis Stasinopoulos and Bob Rigby on the gamlss package. This has involved yet another extension: nonparametric regression approaches for flexible functional location, scale and shape effects. The talk finished with a statistical version of the song "It ain't necessarily so" such by Gurdeep Stephens and fellow South African statistician Michael Greenacre from a conference held in Spain in the mid-200s.
The speaker and almost the entire live audience then headed down to the Duck Inn pub in Chippendale and had an appetizing and lively three-course meal together. It was a great way for the branch to say good-bye to 2020.
University of Technology Sydney