Abstract

Statisticians are trained to deal with uncertainties that can be quantified by probability distributions or confidence levels (depending on one’s philosophy). But when modelling, for example, finance, epidemics, or climate change, there may be substantial doubts about how the world works and what might happen. In 1921 Frank Knight distinguished ‘risk’ (quantifiable) from ‘uncertainty’ (unquantifiable)

[http://en.wikipedia.org/wiki/Knightian_uncertainty]

and this distinction continues in critiques of the modelling process, where such deeper uncertainties provide an argument for the ‘precautionary principle’ see, for example, Stirling (2007)

[http://www.pubmedcentral.nih.gov/picrender.fcgi?artid=1852772&blobtype=pdf] .

In climate change, the IPCC attempted to distinguish different ‘types’ of scientific uncertainty in their 4th assessment report

whereas the recent UK Climate Impact Programme assessments are fully probabilistic [see

http://ukclimateprojections.defra.gov.uk/content/view/1394/543/

to see how East Anglia might be in a few years].

I would like to discuss how statisticians can deal with the limits of quantification in uncertain models of complex phenomena. I will tentatively suggest an analytic structure and hope people might suggest improvements. It may help to have a look at the links provided.