Keynote speakers

Francois-Xavier Briol, University College London

Bayesian global sensitivity analysis almost for free

How much does your prior actually matter? This is one of the most natural questions a Bayesian analyst can ask, yet it remains surprisingly difficult to answer in practice. Bayesian sensitivity analysis has grappled with this problem since the eighties, but despite a rich theoretical literature, it is rarely used in modern practice, primarily because existing methods are far too expensive to scale to realistic models. In this talk, I will present a new approach that is both flexible and computationally cheap. The key insight is that measuring sensitivity via the Fisher divergence allows us to quantify how much posterior conclusions could change over a rich class of priors, at essentially no extra cost beyond running a standard sampler. I will illustrate this across several modern inference problems, including Bayesian neural network regression, simulation-based inference, and Bayesian time series modelling.

Alain Vandal, University of Auckland

Notes on a problem gambling trial

Maria E. Bellringer, Alain C. Vandal, David C. Hodgins, Kirsten van Kessel, Ben Riley, David Smith, Stéphane Janicot, Takayoshi Ikeda, Emma Wang, Bea Cooke, Jason Landon, and Malcolm Battersby
A two-arm parallel randomised controlled trial was commissioned by the New Zealand Ministry of Health and conducted by AUT’s Gambling and Addictions Research Centre between 2015 and 2019, comparing two active counselling interventions (Motivational Interview with workbook vs Cognitive Behaviour Therapy, the latter deemed superior by the primary alternative hypothesis). At least three aspects of the trial posed challenges for analysis and interpretation: 1) the two primary outcomes (self-reported monthly average money spent and days gambled at 12 months) were heavily zero-inflated; 2) there was no control arm; and 3) adherence to the intervention, in the form of number of attended sessions, was highly variable. We will review the originally planned intention-to-treat analyses, based on trichotomised outcomes. We will then consider, going forward, alternative models with an atom at zero for the original outcomes as well as ways to obtain unbiased estimates of the average difference in intervention effect, assuming equal levels of adherence. Amongst the specific topics considered will be zero-inflated and Tweedie models, instrumental variable approaches and ways to block unobserved confounding.

Rachel Owens, Air New Zealand

I, for one, welcome our new statistical overlords.

Several years ago, I helped deliver a COVID-19 hospitalisation risk model intended to help health workers prioritise callbacks. We spent a lot of time demonstrating how it compared with other ways of ranking calls and preparing people to use it. Once deployed, however, the score acquired meaning we had never intended. It was a lesson in what happens when statistical models leave the hands of their developers. Since then, my work has taken me through healthcare, AI governance and Privacy. I have encountered a forecasting model where the demand we want to predict was only partially observed, an optimisation tool where getting the model into the operational workflow has been much harder than building it, and AI initiatives where enthusiasm for the technology has run well ahead of our ability to establish its value. Now generative AI is what people in the business think of when they hear “AI”, and agents are beginning to automate parts of the analytical process itself. This raises some interesting questions about the future of statistical work, the expertise we need to preserve, and how we evaluate and govern increasingly autonomous systems. Drawing on my experiences, I will look at what it takes to make AI useful and dependable in organisations, and why statistical thinking has an important role to play in shaping what comes next.

Chanatda Somchit, New Zealand Institute for Bioeconomy Science

From Forestry, Farming, and Sensors to Scientific Discovery: How Statistics Contributes to Modern Forestry and Agricultural Science

Modern forestry and agricultural research generate increasingly diverse and complex datasets ranging from studies of forest health and biosecurity to livestock pastoral farming systems, animal welfare, and high-volume sensor data. Transforming data into reliable scientific knowledge with which we inform our farmers, foresters and horticulturalists requires more than analytical tools alone. It requires human statistical thinking to ensure that conclusions are robust, defensible, reproducible, and relevant to real-world decision-making. In this keynote, I will present examples from my work as a statistician collaborating in forestry, agricultural, and environmental research. These examples cover a wide range of disciplines including forest entomology, invasive tree control, animal behaviour and welfare, digital agriculture, and the application of emerging technologies such as accelerometer sensors, wearable devices, and drone-based aerial spraying systems. Together, these examples demonstrate how statistics enable researchers to discern meaningful biological signals from random variation, supporting both scientific discovery and evidence-based decision-making. Through these case studies, I will illustrate how statistics contributes at every stage of the scientific process, from experimental design and data collection to modelling, interpretation, and the translation of research findings into practical solutions for real-world challenges facing New Zealand’s farmers and growers. Whether optimising aerial herbicide treatments for the control of dense wilding conifer infestations, improving biosecurity responses to forest pests, enhancing animal welfare through behavioural monitoring, or evaluating new technologies for precision agriculture and aerial insecticide application, statistics provides the foundation for turning data into actionable knowledge. Perhaps the most rewarding aspect of this work is seeing research outcomes translated into practical solutions that support New Zealand’s primary industries and environmental management. Equally rewarding is collaborating with scientists from diverse disciplines. This presentation will highlight how statistics serves as the bridge from data to discovery, helping modern forestry and agricultural science transform observations into knowledge, and knowledge into real-world impact.

Workshops

Both workshops are subject to minimum enrollment thresholds to go ahead. Make sure to indicate your intention to attend on your registration form!

Issues in Survival Analysis, presented by Thomas Lumley

Now that we understand it better, survival analysis is largely just the same as any other sort of regression modelling. There are still some important conceptual issues. This workshop will primarily cover the Cox regression model, but will include discussion of left truncation, interval censoring, competing risks, time-varying effects, time-varying predictors, and recurrent/multivariate events. A reasonable knowledge of logistic regression will be assumed. There will be lab sessions for you to do stuff: bring a computer. Most of the examples will be in R, but Stata translations may be available.

Recent Advances in the VGAM package, presented by Thomas Yee

Will cover variance-factorized loglinear models; generally altered, inflated, truncated, and deflated regression and parameter space inference.

Women in Data Science Event

Women in Data Science New Zealand will run a free event 12:30-3:30, in parallel with the workshops. Please register at WiDSNZ2026AKL. All genders are invited to attend all WiDS Worldwide events.