Some scientists have expressed concern that scientific investigation requires a long sequence of observational records, replicable trials, or model runs (e.g., Monte Carlo simulations) so the results can be specified by a formal statistical characterization of the frequency and frequency distribution of outcomes being assessed. In statistical terms, "objective" science means attempting to verify any hypothesis through a series of experiments and recording the frequency with which that particular outcome occurs. The idea of a limitless set of identical and independent trials that is "objectively out there" is a heuristic device that we use to help us rigorously quantify uncertainty by using frequentist statistics. Although there may be a large number of trials in some cases, however, this is not the same as a "limitless" number, and these trials rarely are truly identical or independent.
Most interesting complex systems cannot possibly be put to every conceivable test to find the frequency of occurrence of some socially or environmentally salient event. The popular philosophical view of "objective science" as a series of "falsifications" breaks down when it confronts systems that cannot be fully tested. For example, because climate change forecasts are not empirically determinable (except by "performing the experiment" on the real EarthSchneider, 1997), scientists must rely on "surrogate" experiments, such as computer simulations of the Earth undergoing volcanic eruptions or paleoclimatic changes. As a result of these surrogate experiments and many additional tests of the reliability of subcomponents of such models, scientists attain confidence to varying degrees about the likelihood of various outcomes (e.g., they might assign with high confidence a low probability to the occurrence of extreme climate outcomes such as a "runaway greenhouse effect"). These confidence levels are not frequentist statistics but "subjective probabilities" that represent degrees of belief that are based on a combination of objective and subjective subcomponents of the total system. Because subjective characterization of the likelihood of many potentially important climatic eventsespecially those that might be characterized by some people as "dangerous"is unavoidable, "Bayesian" or "subjective" characterization of probability will be more appropriate.
Bayesian assessments of probability distributions would lead to the following interpretation of probability statements: The probability of an event is the degree of belief that exists among lead authors and reviewers that the event will occur, given the observations, modeling results, and theory currently available, all of which contribute to estimation of a "prior" probability for the occurrence of an outcome. As new data or theories become available, revised estimates of the subjective probability of the occurrence of that eventso-called "posterior probability"can be made, perhaps via the formalism of Bayes theorem (see, e.g., Edwards, 1992, for a philosophical basis for Bayesian methods; for applications of Bayesian methods, see, e.g., Howard et al., 1972; Anderson, 1998; Tol and de Vos, 1998; Malakoff, 1999)
Other reports in this collection