The Risk Communicator: February 2010

by Stephen Hora

As mathematical models of social and human behavior become more realistic and complex, data requirements mushroom. Yet the most sophisticated models require data that may not exist in the traditional sense of statistical information, or the data may be so sparse that making statistical inferences is impossible. It is therefore not surprising that experts are often called upon to interpret available information and quantify models.

Expert judgments have been successfully used in a wide range of activity, from risk models of nuclear power generation to diagnostic programs for medical applications. But there are also limits. If reliable data is available in sufficient quantity, then the data should provide the quantification -- not the experts. At the other extreme, if there is no basis from which to make judgments, then the conclusions will be no more than guesses. It is the intermediate range -- between nearly perfect knowledge and near ignorance -- that is the ideal realm for expert judgment processes.

One of the earliest formal applications of expert judgment to risk modeling was undertaken as part of the Reactor Safety Study, produced in 1975 for the Nuclear Regulatory Commission. The NRC's NUREG-1150 study of 1991 updated these methods and provided more structure to the process of acquiring judgments. These methods were fashioned for physical models of operating systems that, while complicated, are more straightforward than models of terrorist behaviors. Moreover, the subject matter experts were invariably engineers or physical scientists who felt relatively comfortable expressing their knowledge through probabilities and probability distributions.

The situation facing risk modelers of terrorist threats, however, is different and may require some rethinking. As opposed to the reactor safety world, which emphasizes exact expression, experts from the intelligence community have worked in a culture where verbal expressions of uncertainty are the rule. It is worthwhile, therefore, to step back and examine the important issues that arise when developing a successful expert judgment process.

The first of these issues is the definition of the quantities or events that are the subject of the expert judgment. It is deceptively difficult to develop statements that are crystal clear and mutually well understood. The problem may be that the risk modelers make implicit assumptions that are different from those made by the subject matter experts. Or it may be that the subject matter experts don't agree with the model that the risk analysts are trying to quantify. Testing of questions may help sort out these differences. The testing should involve the use of stand-in experts who are not part of the staff conducting the risk study. A dry run of this kind can often lead to needed changes in the questions that would otherwise seriously interfere with obtaining the best expert analysis.

Along with defining the issues, the selection of experts is critical. Preparing some criteria for choosing them is a good first step. After all, experts need to have special knowledge or access to such knowledge. But they also must be free of prejudices that might lead them to answer in a less-than-straightforward manner. Even the appearance of bias can lead to an analysis being discredited.

It is also possible that the selection process itself is subject to motivational distortions. An individual choosing subject matter experts can inadvertently suppress diversity of opinion by selecting experts who share the chooser's own methods or opinions. Alternatively, differences among judgments given by multiple experts will often convey more information about the real uncertainty than do the individual judgments. Diversity of opinion is helpful and should not be repressed either through the selection process or through some well-intentioned method of aggregating judgments.

After an appropriate slate of experts has been selected, the forum for collecting their judgments can take many forms. The experts may work in isolation from one another, or they may come together physically or electronically to share information. The choice made here can implicate the dual issues of independence and shared knowledge. When experts meet together to share information, they become less independent but more knowledgeable, and vice versa. Decisions need to be made about the degree of interaction allowed among the experts.

Another challenge is that subject matter experts may not be skilled in expressing their judgments through probabilities. This is a cognitively difficult task that improves with practice. Thus, some training in probability judgments that involves practice questions and feedback on the quality of the judgments should be mandatory. Training is not only known to improve the quality of the judgments, but it also instills confidence in the experts concerning the appropriateness of the expert judgment process.

Finally, some random thoughts. Documentation is critical. Not only the judgments but also the rationales and reasoning processes by which the experts reach their judgments should be documented. Those who will be using the judgments as input for models need to understand how the judgments were developed. The experts' names should also be attached to their conclusions. Doing so helps ensure that the experts will be diligent.

Often times judgments will be aggregated to the level of a single probability distribution that will be used to quantify a risk model. After all, one distribution is much is easier to deal with than many distributions, and using a single distribution avoids the possibility of contradictory results caused by disagreement among the experts. Whenever aggregation is employed, care should be taken not to suppress diversity through mathematical methods that shrink uncertainty or through behavioral consensus methods that force agreement when in reality there are unresolved differences of opinion.

The science, and art, of obtaining judgments as probabilities is still in development. There is no single right way to do things but there are some good rules of thumb: two (or more) heads are better than one; develop unambiguous questions; provide training; and do not suppress diversity of opinion. All are principles that should aid the development of an effective expert judgment process.

Stephen Hora is Director of the USC Homeland Security Center for Risk and Economic Analysis of Terrorism Events (CREATE).