Climate change is an issue of risk management. The most important causes for concern are not the median projections of future climate change, but the low-probability, high-consequence impacts. Because the policy question is one of sequential decision making under uncertainty, we need not decide today what to do in the future. We need only to decide what to do today, and future decisions can be revised as we learn more. In this study, we use a stochastic version of the DICE-99 model (Nordhaus WD, Boyer J (2000) Warming the world: economic models of global warming. MIT Press, Cambridge, MA, USA) to explore the effect of different rates of learning on the appropriate level of near-term policy. We show that the effect of learning depends strongly on whether one chooses efficiency (balancing costs and benefits) or cost-effectiveness (stabilizing at a given temperature change target) as the criterion for policy design. Then, we model endogenous learning by calculating posterior distributions of climate sensitivity from Bayesian updating, based on temperature changes that would be observed for a given true climate sensitivity and assumptions about errors, prior distributions, and the presence of additional uncertainties. We show that reducing uncertainty in climate uncertainty takes longer when there is also uncertainty in the rate of heat uptake by the ocean, unless additional observations are used, such as sea level rise.

© 2008 Springer