Event Details

Making sense of multiple models

Making sense of multiple models

Much of modern statistics, and science more generally, is underpinned by the use of models that attempt to represent the key features of the systems being studied. Some, such as statistical regression models as they are often used, are primarily descriptive in the sense that they aim to find a way of representing observed variation – albeit often in an interpretable way. Others, such as dynamical models of physical systems, are mechanistic in nature: they attempt to represent explicitly the mechanisms / processes operating in the system. One feature that all models share, however, is that no model perfectly replicates the system that it aims to represent.

A consequence of this imperfection is that multiple models are often available for well-studied systems. For example, much of the science in the most recent report of the Intergovernmental Panel on Climate Change was informed by the outputs of around 50 global climate models; while management of the Covid-19 pandemic in the UK was informed by the results from several different modelling groups. The absence of a unique model presents challenges for those wanting to use modelling to inform their own research or decision-making.

A user might be tempted to ask whether one or more of the available models is obviously “better” than the others; or whether, instead of working with the output of a single model, it may be preferable to average the projections, perhaps after weighting the models according to some measure of their performance. A “statistically advanced” user might seek a more principled approach, treating the model outputs as sources of imperfect information and using these to make defensible statements about the phenomena of interest accompanied by rigorous assessments of uncertainty.

This masterclass will explore these issues in more depth. We will discuss the purpose of models; why they differ from each other; and the extent of data availability and its implications for the interpretation of multiple model outputs. We will review some standard techniques that are often used when working with multiple models, including those from other application areas, and discuss their limitations. Finally, we will discuss some of the more principled techniques that arise from careful consideration of the statistical characteristics of the problem. There will be opportunities for practical work, probably focused mainly on climate projections

---

Schedule

Tuesday 6 May

Masterclass session 1

14:30-16:30 Charles Carter A18

Wednesday 7 May

Masterclass session 2

14:30-16:30 Charles Carter A18

Thursday 8 May

Masterclass session 3

10:00-12:00 Fylde C48