Risk Assessment and Decision Making Using Test Results

Decisions are most often based upon the results of experiments coupled with the knowledge of experts. When sampling or experimental results are available, they often constitute the major factualinformation input into the decision-making process. For example, clinicians and physicians use diagnostic tests and clinical findings along with their expert knowledge to diagnose their patients’ problems. When the physician estimates that the “risk” of a disease is high enough (here we define “risk” as both the probability and the severity), expert knowledge is again used to develop appropriate treatment plans. Toxicologists (in industry, government, and academia) use test results on live animals as well as short-term in vitro tests to study the carcinogenic potential of chemicals. If the risk of carcinogenicity for a particular chemical is high, then a pharmaceutical or chemical company may decide to stop or delay the development of the chemical, a regulatory agency may decide to ban the development or restrict the use of such a chemical, or a researcher in academia may decide to study this chemical further to examine its modes of action. In industry, quality control managers interpret results from multiple “inspectors” (humans or machines) to identify defective parts and to decide whether a defective part should be destroyed or sent to a rework station. Water resources and environmental engineers utilize results from well sampling to decide what type of action is warranted on an aquifer found to be contaminated.

In this chapter we will examine four basic methodologies and decision tools that are utilized in the CPBS approach to decision making. The first is Bayesian decision analysis, which forms the heart of the CPBS approach. Tests and measurements that are used to identify or detect a property of interest are generally not perfect. When tests are biased or inaccurate, it is often advantageous to use more than one test. The interpretation of a combination of test results can be problematic because there often exists a variable amount of information overlap (positive dependence) and differences (negative dependence) among the tests. It is a difficult problem to account for both the imperfection of the individual tests as well as their interdependencies in their joint interpretation.

In Chapter 3 we described several analyses that can be used to summarize the performances of the individual tests and the interdependencies among the tests from a data base containing test results on objects with known properties. With the availability of this summary information, whether it is obtained from the preliminary analyses or from other sources, we are now in a position to try to determine which combination of tests would be best to use for a given decision problem.

Consider the situation in which we would like to determine whether some object has a certain property. For example, the object might be a chemical and the unknown property might be the potential carcinogenicity of the chemical. Suppose, further, that we have a set of tests that we can use to help us determine whether the property is present in the object. In Chapter 3 we discussed several analyses for computing the performances of the tests and the interdependencies between the pairs of tests, and in Chapter 4 we discussed how one can select the “best” battery of tests to use for a particular application. In this chapter, it is assumed that we have chosen a battery to use, we have applied this battery on the object, and now we must interpret the results of this battery.

It is a common practice to use multiple tests and sources of information in making risk-based decisions; however, the manner in which the information is integrated and interpreted differs. Well-established statistical methods for interpreting the results of a single test or a single repeated test — tests that are affected by random errors — are available. A large number of methods have been proposed for combining information from various sources, with the field being dominated by statistical decision theory and with a particular emphasis on Bayesian inference. The CPBS approach is grounded on Bayesian theory and provides a systematic mechanism — a synthesis of systems, statistical, and decision analyses — for transferring information from tests and experts into results which can aid decision making.