Events

 
Seminar

Unexploited statistical concepts in climate research: mediation and suppression

Tuesday, 14 March 2023, 15:00-16:00
online via MS Teams

Machine learning is transforming the way we do research, but it is often hard to make sense of the results, especially if one wishes to understand the pathways of physical mechanisms. An alternative to ML, which can gobble large amounts of data but requires that the scientific questions are posed by the researcher and not the machine, is the branch of statistics known as meditational analysis, which is widely used in psychology research.

A classic example of mediation relates to the correlation between a young person’s height and their ability to read. It is absurd to think that tall kids are more literate than shorter ones. Height is in reality just a confounder that obscures the role of age, which really is a useful predictor of reading skills. Alternatively one can say that height is a mediator of the effect of age on reading skills.

I will mention some climate-related examples of mediation from my own work. For instance, Dave MacLeod and I have investigated the well-known lagged effect of summer ENSO on the East African short rains in October–November. We showed that the ENSO effect is fully mediated by the Indian Ocean SST Dipole, which influences the short rains through an east-west Walker Circulation across the Indian Ocean. The causal chain is that ENSO has a lagged effect on the IOD, which again has a concurrent effect on the short rains. The proof of this only requires three simple regression equations (but lots of data).

Suppression appears to be exploited even less than mediation. A suppressor has been defined as ‘a variable which increases the predictive validity of another variable by its inclusion in a regression equation’. An example from my own research is drawn from recent work on the effect of North Atlantic SSTs on the NAO, where it can be shown that the lack of latent heat fluxes over cold water suppresses the correlation between SSTs in early winter and the NAO several months later.

The overall aim of my talk is to present an overview of a simple set of statistical tools which can be used to identify pathways in climate dynamics, sometimes replacing hand-waving.

This event is part of the eventgroup ICAS External Seminar
Speaker
Erik Kolstad

NORCE
Organizer
School of Earth & Environment
University of Leeds
Leeds
Mail: M VanDerGucht does-not-exist.leeds ac uk
Service-Menu