Consider this clinical dilemma: A 68-year-old male diabetic has suffered for several years from progressive chronic peripheral vascular disease. After an injury to his foot, a severe infection has developed, with possible gangrene. Two treatments are available: (a) a trial of antibiotic therapy, and (b) immediate below-the-knee amputation. With the first option, the infection might heal and surgery be avoided altogether. But, if the antibiotic therapy fails, the infection will probably spread, and an above-the-knee amputation will be necessary under less favorable circumstances. The surgical mortality from above-the-knee amputation is higher than from below-knee amputation; the quality of life for the survivor, somewhat lower, since learning to walk with an above-the-knee prosthesis is more difficult. While the second option, immediate below-the-knee amputation, would leave the patient with an amputated limb, his chances of surviving this surgery now are better than if the surgery were delayed and more extensive surgery became necessary. On the other hand, surgery now forecloses the possibility that antibiotic therapy might save the whole leg. What should be done? How should we decide?

This clinical vignette (adapted from Weinstein et al., Clinical Decision Analysis, 1980) omits many details but highlights the character of many clinical dilemmas. The patient's diagnosis is known, but the choice of treatment is still a problem because there is no completely safe and effective therapy. Each of the available therapeutic approaches (in this case, surgery or a trial of antibiotic therapy) has potential risks and benefits. Such uncertainty pervades many medical decisions, particularly those where no dominating therapy has emerged. For approximately 15 years, some clinicians and researchers have been studying the application of decision analysis and the theory of decision making which it implements, expected utility (EU) theory, to similar dilemmas. For simple clinical problems, the analyst must identify the relevant risks and benefits, quantify them as probabilities and utilities, and select the strategy that maximizes expected utility. A substantial literature has developed applying the necessary mathematical techniques to a wide variety of clinical situations and health policy problems. (See, for example, Kassirer, J. P et al., Annals of Internal Medicine 106 (1987): 275-291.)

EU theory sidesteps the question of what we ought to value, what is good or right, and instead concentrates on the problem of how we ought to choose among competing risks and benefits so as to maximize benefit while minimizing risks. It is a theory of trade-offs amidst uncertainty. Its roots lie in probability theory and economics. One way of looking at expected utility theory is as an account of the idealized functioning of a perfectly rational actor. The central problems of EU theory are the assessment of uncertainty (probabilities) and value (utilities).

Behavioral decision research describes how people actually behave. Psychological research on decision making describes and analyzes the cognitive processes and principles employed in decision making under uncertainty. It is concerned with what people actually do, not with what they should do; the normative theory (EU) serves as the standard of comparison. Current psychological research in clinical judgment and decision making has been influenced significantly by expected utility theory. In the early days of this research, the expected utility hypothesis was taken as a fairly good approximation of what people actually do or at least of what they would want to do. More recently, psychologists have increasingly criticized the validity of this hypothesis as a first approximation of actual decision making, and there is some concern about whether people want to follow EU principles. Alternative theories have been advanced to account more adequately for human decision processes.

Departures from the normative theory can be demonstrated in each phase of the decision analytic process: problem structuring, probability estimation, probability revision, utility assessment, and synthesizing these estimates to reach a decision. The most troubling psychological problems have arisen in situations where decision makers do not obey expected utility rules and, even after their inconsistencies are pointed out, insist that they do not want to change their decisions. What does this imply? Several interpretations have been advanced: (1) that presumably intelligent people do not always think straight (imperfect or limited rationality); (2) that defining rationality as EU maximization is an error; (3) that people want to obey EU axioms, but their understanding of the tasks is different from what the experimenters intended.

The second and third interpretations of these results explain them away by arguing that they are not really errors. The second interpretation basically claims that people are right in what they choose and the theory is descriptively and normatively incorrect. The third interpretation argues that observed choice behavior is not mistaken but is rational with respect to a different problem than the experimenter had in mind.

Only the first interpretation finds a serious problem with human decision making. It argues that the discrepancies between normative and descriptive theories are caused by imperfect rationality Because we are imperfectly rational, the ways we make decisions often do not conform to EU. EU is a normative theory-an ideal-not an account of how people actually make decisions.

Several important questions arise: Why do people violate the normative model? Is the normative model defective? If so, how should it be changed? Assuming that the normative model is essentially correct, how can we get people to improve their decision making? Based on the psychological research conducted so far, some answers are beginning to emerge:

1) Often people do not try to maximize expected utility because it too hard. The computational burden often far exceeds the capabilities of unaided human thinking. The principle of limited information processing capability means that most people will prefer satisfactory short-cut solutions to more difficult strategies that optimize or maximize expected utility. Sometimes it is difficult to show that the amount of extra benefit gained by an optimizing strategy is worth the extra effort. Still, even when microcomputers make it possible to carry out the computations required for a decision analysis quickly and inexpensively, many decision makers are uninterested.

2) The expected-utility model is often difficult to explain. It is complex and very explicit about uncertainties and risks; many patients and physicians would rather not confront these uncertainties directly. For example, a surgeon once told me that unless he believed he was better at a particular procedure than anyone else in the world, he could not bring himself to operate. To maintain such an attitude, one must generally avoid comparative inquiry into one's own surgical results.

3) A decision-analytic model generally evaluates each outcome independently of the route by which it was reached. For example, it counts each death regardless of whether the patient died from surgical intervention or "natural causes." This approach generally ignores the regret we experience over outcomes for which we feel causal responsibility. Some theorists have endeavored to work regret into the utility function. While this is desirable from a psychological standpoint, it has had the undesirable side-effect of making a complex theory even more complex.

Thus, a decision-analytic approach to clinical dilemmas is both difficult to implement and at times contrary to some deep-seated psychological intuitions. Despite these shortcomings, I believe expected utility analyses of complex clinical issues are more correct than our intuitive analyses of them. My own program for teaching EU strategies to health professionals (physicians, nurses, and so on) begins with trying to get them to be more skeptical about their own intuitions and then moves on to showing how quantification and formal problem structuring can help straighten out their thinking.