Exploiting the exciting interface between mechanistic and statistical models

Speaker: Dr Matthew Adams
Affiliation: Queensland University of Technology

Abstract

When modelling real-world phenomena, there is always uncertainty. But where does this uncertainty come from? The uncertainty could be irreducible, and thus our best bet is to use the available data to devise a statistical model. We may, given time, be able to obtain process-based understanding of the system, and then our best bet might be to devise a (usually complicated!) mechanistic model.

This talk explores research work which bridges the divide between mechanistic and statistical models. Such work typically requiring cross-training at the postgraduate or postdoctoral level and is a potential route to take best advantage of the strengths of both modelling strategies. Analogous to the broader tension felt between fundamental and applied research, hybrid modelling strategies may also have the potential to balance the two scientific progress goals of mechanistic insights and practical predictions.

I will present some of the lessons my colleagues and I have learned about navigating this mechanistic/statistical interface with examples of surprising and unexpected outcomes we have seen in the fields of biology, ecology, agriculture and elsewhere. These examples include applications where: (1) statistical techniques can inform strategic simplification of mechanistic models, (2) mechanistic and statistical models have equal predictive power, but each has a preferred usage depending on the model’s goal, (3) emulation of a mechanistic model will be utterly essential to practically inform decision-making, and (4) statistical assumptions vastly alter conclusions about parameter identifiability in both model types.