Climate change is one of the most important global environmental problems of the 21st century. Assessments of climate change impacts generally adopt a single or multiple climate scenario paradigm where the scenario(s) is(are) chosen to represent either a single 'best guess' or a rather arbitrarily determined range. This was the approach taken by the Intergovernmental Panel on Climate Change (IPCC) in its Second and Third Assessment Report (TAR). The TAR concluded that many impact studies fail to consider adequately the uncertainties embedded in the scenarios they adopt, partly due to such uncertainties being poorly quantified in the climate scenarios themselves.

Developments in global climate modelling, in our understanding of the probability density function of the climate sensitivity (a key parameter in climate scenario specification), combined with recent advances in computing power, now make it feasible to define quasi-objectively determined distributions of future climate states. Climate change impact and adaptation studies will then be better placed to move towards a risk assessment paradigm. Such an objective is desirable since it explicitly recognises the uncertainty inherent in climate and, by association, impact prediction. Such a paradigm also fits better with the identification of appropriate management responses regarding adaptation to climate change, now increasingly recognised as a priority response to our changing climate.

We are developing new methods for representing uncertainties in climate change scenarios for a range of spatial scales (from global to local) and for quantifying the frequency distributions of future global-to-local climates. Some of these distributions (for example natural climate variability) can be objectively determined, while others (for example future greenhouse gas emissions) can only be determined subjectively using scenario analysis.

We try to bridge the gap between scenario and uncertainty analysis using a combination of Bayesian and Monte Carlo methods. Such an approach would regard the model output as an unknown function about which we are trying to make inferences. Prior information on the inputs is combined via Bayes Theorem with model output. An emulator (or metamodel) is created which is a simple statistical approximation to the output of the model. Using such an emulator, the probability density functions of the outputs can be sampled and the uncertainty estimated numerically. This approach is being tested using the versatile simple climate model used in former IPCC assessments, MAGICC and in impacts models, for example, a climate-mortality empirical-statistical model developed for the city of Lisbon, Portugal.

This research will lead to a better quantification of uncertainty in descriptions of future climate and hence a better representation of risk in climate impacts assessments. Such methodological advances are a necessary pre-requisite for the determination of optimal climate change mitigation and adaptation policy, both nationally and globally. If the risk of climate change being ‘dangerous’ is truly to be assessed, then impacts and adaptation studies need scenarios that span a very substantial part of the possible range of future climates.