My editorial accomplished its purpose to generate vigorous debate about screening. The large benefit of treating diseased patients should not be confused with the exquisitely small benefit of screening patients who are very unlikely to have disease. In contrast to neonatal screening programs for phenylketonuria [1, 2], screening for ovarian cancer is not without risk. Follow-up transvaginal sonography may be uncomfortable, and exploratory laparotomy will produce some complications; both will consume scarce health care resources.

As Drs. Feagin and Burris note, clinicians care for individuals, but we often use average data about populations to estimate prognosis. Even acknowledging variation, a well-controlled series is substantially better foundation for clinical care than is personal experience or a case report. Although some physicians are troubled if we use averages in models to project future or expected benefits and risks, the point is that they are more stable and make better projections!

Dr. Burris appears to have chosen statistics selectively and perhaps incorrectly to support what he acknowledges to be an emotional argument. When deciding whether to screen for preclinical disease, one must take the perspective of each patient offered screening, not just selected patients with early disease. If we consider the best case gain, we must also consider the worse case loss—healthy women who have falsely positive test results and die during laparotomy. In Dr. Burris' calculus, the death of a young woman with undetected preclinical cancer is unnecessary and premature; yet we cannot simply discount the death of a healthy woman during laparotomy by referring only to the uneventful and reassuring outcomes. Good clinical care considers the patient's emotions, but emotional arguments that neglect logic and statistics do not help our patients or policymakers make their best decisions.

When few treatable diseases were identifiable in their preclinical phase, logic and emotion pushed us to "go for the gold," even when few affected patients were found. The real debate must now focus on numbers—not only the economics and benefit accrued to each patient screened or affected, but the number of identifiable diseases (and disease propensities). The possibility exists of screening for hundreds if not thousands of preclinical diseases, each with imperfect tests that will have independent false-positive results. Almost two decades ago, Drs. Gambino and Galen pointed out that a panel of 12 tests, each with a 5% false-positive rate, produces one false-positive result for every two patients screened [3]. Extending his argument, we should also anticipate that 600 tests, each with a 0.1% false-positive rate, will also produce one false-positive result for every two patients screened. As Dr. Feagin emphasizes, screening for low-prevalence diseases must rely on informed consent and negotiation between the physician and the patient. That ideal may not be achievable, however, when we screen for many diseases.

Focusing on patients with preclinical ovarian cancer, Dr. Lindow seems to downplay the burden of "unnecessary" testing that will befall 99 986 out of 100 000 women screened. What would happen if the benefit to affected women held steady at 11.6 years but the prevalence of preclinical cancer fell from 14/100 000 to 1/100 000 or 1/1 000 000, as might be the case if young women were screened? At some point, the prevalence of preclinical disease will be too low to justify screening—but how low is too low?

As technologies for identifying preclinical disease proliferate, physicians must engage in explicit, rational discussion so that we and our patients can decide when screening for which diseases is appropriate, and for whom.