2024

Kozlova Mariia; Piano Samuele Lo; Yeomans Scott Julian

For a computational model to be useful for decision-making purposes, its foundations need to be “shaken” sufficiently so that one can ascertain what might happen if its underlying assumptions ever changed. This type of robustness assessment is performed via the techniques of sensitivity analysis and uncertainty analysis. These techniques supply many alternative pathways under which such evaluations can be conducted. In this chapter, we consider a full spectrum of sensitivity analysis tools – from setting up computer experiments, to analyzing their results, to discussing the variety of methods at each phase, up to investigating how the overall process might actually be deployed effectively in practice. Our conclusions on the current “state of the practice” are worrying. The most commonly used one-at-a-time sensitivity analysis methods can severely mislead decision-makers. Further to that, simple Monte Carlo simulations and the portraying of their output distributions are not sufficient to understand the actual drivers behind the model behaviour. Advanced global sensitivity analyses are often so focused on quantification that they miss any insights into the nature of the effects. Conversely, the novel Simulation Decomposition (SimDec) method incorporates the best components from both sensitivity analysis and uncertainty analysis – which enables the production of holistic insights in a very straightforward fashion – thereby reinforcing its considerable potential for widespread adoption anywhere a global sensitivity analysis is performed.