Confronting uncertainty beyond Bayesianism
is an initiative within the Advanced Study Group “Values, decision making and risk – current perspectives and future research” at the Pufendorf Institute in Lund, spring 2015.
The starting point of the theme is view of knowledge-based probability as the first order measure of uncertainty. The Bayesian approach to interpret and assign probabilities makes it possible to treat knowledge-based uncertainty in both parameters and model structure. The Bayesian approach rests upon coherent principles of logic reasoning and inference which easily transfers into decision analysis. It is put forward as the basic approach to assess and communicate knowledge-based uncertainty in risk. In strict Bayesian decision frameworks all underlying knowledge is taken care of by the values and knowledge-based probabilities.
There are at least two challenges to Bayesianism.
First, the knowledge basis to assign probabilities can be more or less strong while the Bayesian framework says nothing about how firmly the decision maker holds her degree of belief. However, representing all underlying knowledge (or information) by a precise probability (unique measure) is neither intuitive nor logically appropriate. This has led to development of alternative ways to represent uncertainty e.g. by expanding the probability measure into imprecise probabilities. We explore how these measures can be used in replacement of, or in combination, with Bayesian probabilities.
Second, parameters and model structure in a risk analysis is just one part of a knowledge production process. The use of any model has been preceded by problem framing and assessors choice of how to limit the problem and what sources of information to use.
The growing demand for science-informed policies to handle climate change and environmental problems has led to an increased attention to the treatment of knowledge-based uncertainty and the establishment of evidence bases for management. Lack of scientific knowledge is distinguished as important for the science policy interface, but how to manage lack of knowledge is more than just about asking for more knowledge.
Instead, qualitative aspects of knowledge provide an additional dimension to risk. The entire knowledge production process might reveal more types and sources of uncertainty to consider. It is relevant to ask if and how new types of uncertainty may affect the rationale and applicability of the Bayesian approach and it can be extended to confront unreliable knowledge-based probabilities.
Somehow the uncertainty of underlying knowledge should have an impact on how a decision analysis is set up and performed. What is a comprehensive quantitative decision theory that encompasses the affixing of probabilities in accordance with quality dimension of knowledge? What kinds of rules are required to adequately guide the choice of action to perform based on the strength of knowledge and other contextual factors?
We are not interested in a list of factors that induce uncertainty into modeling and decision making. Rather we seek to discuss how these factors can be incorporated in a comprehensive decision structure. In what way should they influence the reliability of the basis for decision-making? Put in another way, upon which characteristics can qualitative aspects of knowledge be evaluated and how can these be taken into account in risk and decision analysis?
Finally, how can an evaluation of the quality of knowledge be incorporated in a Bayesian decision framework? When is the Bayesian approach not sufficient to treat uncertainty? In what ways, if all, must the Bayesian approach be developed when new treatments of uncertainty enter the scene?
These and other questions led us to initiate a discussion on the theme “Confronting uncertainties beyond Bayesianism”. The discussion will be spread out to two occasions.
March 12th we discussed the possibilities and limits of Bayesian and alternative decision theories to deal with severe uncertainty.
Invited guests were Nils-Eric Sahlin, Peter Gärdenfors and Rasmus Bååth
May 11th we discussed how to determine and consider confidence in an assessment finding.
Invited guest was Kristie Ebi from University of Washington.
Link to Ebi’s open lecture
Contact Ullrika Sahlin if you are interested to know more