Providing relevant uncertainty information to decision makers: Subjective post-processing of rigorous Bayesian uncertainty assessment of model projections

Abstract:

Conveying information on the uncertainty of a prediction to decision makers is essential. However, predictions of dynamic systems (e.g., about an excess of a threshold) include uncertainty about model choice, parameter values, and temporal/spatial aggregation. When it comes to forecasts into the future, further sources of uncertainty accumulate. Even trying to capture all existing uncertainties often seems to obscure the information in the results and might hinder interpretability by the decision maker. Subjective filtering of results is a practical way out, but without consultation of decision-makers may lead to biases in reported uncertainties.

To promote and improve the communication of uncertainties, we propose a two-step evaluation scheme: in step (1), our framework rigorously assesses uncertainties in data and model simulations; and in step (2), it provides a means to extract the essence of the estimated uncertainty based on the decision maker’s preferences and uncertainty attitude. The approach is founded on a Bayesian framework, which quantitatively traces data and model uncertainty through the forecasting model chains. This full set of projections is then re-evaluated according to the decision maker’s information needs. Hence, a more uncertainty-averse person will receive a differently filtered outcome than a risk-seeking one. Integration over all possible risk attitudes would finally reproduce the whole ensemble of outcomes obtained from step (1). Hence, the framework yields objective and reproducible results, but is subjective where needed – because decision-making is subjective and needs adequately-tailored support.

A main contribution of our work is to make subjective filtering rigorous and transparent. We thereby hope to spark discussions about different possible interpretations of the same data and how to best cope with existing uncertainties in different decision-making contexts.

We present an application of our proposed framework to regional forecasts of climate change impacts on drainage and crop production in a coastal landscape. One example is the drainage capacity in a region, which is referenced to an exceedance of a given uncertain threshold (Figure 1).