The literature on expert clinical judgment and decision making illuminates some of the strengths and problems of naturalistic decision making (NDM). It also has striking similarities to ®ndings with other groups of experts. Therefore, this comment ®rst reviews some of the major ®ndings of research on medical reasoning and decision within an NDM framework. Then I shall argue that although descriptive accounts of experts' behavior are needed and helpful, they do not obliterate the prescriptive±descriptive problem, any more than descriptive accounts of moral behavior eliminate the need for normative ethical theories.
NATURALISTIC DECISION MAKING AND MEDICAL REASONING
According to Lipshitz, Klein, Orasanu, and Salas (this issue) the distinctive features of NDM research are: (a) The subjects are pro®cient decision makers, experienced in the decision domain. (b) The research objective is a set of decision rules that match actions to situations, not an abstract formalism like `Maximize expected utility'. (c) The decision model is informal, not quantitative, and context-speci®c, not abstract. NDM models depict the information decision makers actually attend to and the arguments they actually use. (d) The contexts of interest are ill-structured problems, characterized by high levels of uncertainty in dynamic environments. The goals are shifting, ill-de®ned or competing; time constraints exist, and the stakes are high.
These characteristics are all present in clinical medicine, where highly trained and experienced physicians work in an environment characterized by ill-structured problems, high levels of uncertainty, and unclear or competing goals. Sometimes there are ample opportunities to collect information prior to action and to sift through the data to determine meaning. On other occasions, the clinician is expected to recognize emergencies when deliberation is inappropriate and to act urgently, matching an action (treatment) to the situation. Further, few physicians use formalisms like decision analysis to select diagnostic tests or to choose a treatment. Instead, clinical reasoning is usually pattern matching to categorize situations and rule-governed choice of action. There are many similarities to the military situations that have been the context of much of the work undertaken by NDM researchers, and human factors researchers before them. And it is surprising that in taking stock of NDM, relevant studies of clinical medicine were overlooked or excluded.
Researchers from psychology who became interested in medical judgment, problem solving and decision making were motivated initially by educational considerations. Medical practice is a highly skilled set of cognitive and motor acts that are learned by a combination of didactic teaching and repeated practice with feedback. This training is long, arduous, and expensive, and it was natural to cast about for ways to help students acquire expertise more ef®ciently. These studies frequently examined expert±novice differences, usually by stratifying physicians by level of training, with a view to determining what novices needed to do or know to be more like experts. In a very in¯uential early study, Elstein, Shulman, and Sprafka (1978) studied experienced clinicians using simulated patients to standardize, at least partly, the clinical stimulus situations. They reported that diagnostic problems were solved by generating a set of hypotheses to structure the ill-structured problem, and then testing the hypotheses by selective data collection. This formulation became known as the hypothetico-deductive method and became the object of much subsequent research and debate.
Early studies lacked statistical power because of within-group variability and the small sample sizes used in protocol analysis. Hence, it was dif®cult to identify important and consistent expert±novice differences. Subsequent research has introduced methodological variations to make larger samples practical. Most researchers have switched to controlling the data base, using case presentations that are less elaborate than simulated patients, and have abandoned using verbal