Characterizing Aleatory and Epistemic Uncertainty

As noted above and in Chapter 3, uncertainties are an integral part of the modern risk-based analysis. To evaluate and model aleatory and epistemic uncertainties, a characterization can be made in terms of their effect on models and estimates of model parameters (Table I-1). Modeling epistemic uncertainty represents differences between a physical process (hurricane surge, embankment failure) and prediction models. Modeling epistemic uncertainty can be estimated by comparing model predictions to observed events/performance. Parameter uncertainty is the epistemic uncertainty associated with the estimates of model parameters as derived from available data. Parametric uncertainty is quantified by observing the variation in parameters inferred (either in a direct or an indirect manner).

The distinction between aleatory and epistemic uncertainty is not intuitive and can be difficult to ascertain. Furthermore, the assessment of these uncertainties is model dependent. For example, a simple engineering model of an event (levee performance during a flood) may have higher model aleatory variability than a more complex model that addresses more details of the physical process of a levee dealing with the loads it is exposed to during the flood (i.e., increased hydrostatic loading, wave action, seepage forces, etc.). At the same time, the more complex model may have larger parametric epistemic uncertainty; there are more parameters to estimate and there may be limited data to estimate them. Thus, the characterization of uncertainties is model dependent, making the distinction between different types of uncertainty difficult. Nonetheless, making a distinction between the sources of uncertainty in a logical manner helps ensure that all uncertainties are identified and quantified. In principle, epistemic uncertainties are reducible with the collection of additional data or the use/development of improved models.

A Conceptual Framework for Flood Risk Analysis

As described in the main report, a starting point for the development of an NFIP flood risk analysis is a conceptual framework for flood risk analysis that can be defined as:


where risk (R) is a function of flood hazard (H), vulnerability (V), and consequence (C). In this context, risk is a function of the flood hazard a community is exposed to, the vulnerability of flood protection systems and the potential that their failure will contribute to flooding, and the consequences associated with system failures and the damage to a community exposed to flooding, including economic impact and life safety.

As noted in Chapter 3, this framework can be extended to include the concepts of uncertainty. This extension

TABLE I-1 Characterization of Aleatory and Epistemic Uncertainty on Models and Model Parameters

Element Type of Uncertainty
Epistemic Aleatory
Modeling Uncertainty about a model and the degree to which it can predict events or outcomes (e.g., levee performance), that is, to what extent a model has a tendency to over- or underpredict observations Aleatory modeling uncertainty is the variability that is not explained by a model. For instance, this variability is attributed to elements of the physical process that are not modeled and therefore represents a variability (random differences) between model predictions and observations.
Parametric Uncertainty associated with the estimates of model parameters, given available data, indirect measurements, etc. This uncertainty is similar to aleatory modeling uncertainty. This is a variability that may be due to systematic, but random, variations associated with parameters of a model.

SOURCES: Abrahamson et al. (1990), USR/JBA (2008), IPET (2009).

The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement