I take a critical look at Bayesian parameter estimation in situations where only weak data is available, focussing on climate sensitivity, and discuss and compare subjective and objective Bayesian approaches. Most probabilistic climate sensitivity estimates have used subjective Bayesian methods, which require a prior probability distribution representing existing knowledge about the values of the parameters being estimated. I explain why this approach is unsuitable for analysing data from scientific experiments, and highlight the serious biases in climate sensitivity estimation that it often results in. I explain the very different nature of 'noninformative' priors used for objective Bayesian inference, which can provide satisfactory results even with weak data. I go on to show that the standard Bayesian method for combining independent evidence, Bayesian updating, is inappropriate. I set out a recently developed objective Bayesian approach to combining evidence, well suited for estimating climate sensitivity and similar variables, and demonstrate its application to combining instrumental and palaeoclimate sensitivity estimates. The new approach produces almost identical results to a frequentist statistical method that can produce confidence intervals reflecting combined evidence.
08.08.2018
13:30 h