The advent of hospice programs and their funding under Medicare has recently made eligibility for substantial insured services turn on whether a patient has less than three or six months to live. The implicit assumption is that physicians can provide this prediction accurately. To test this assumption and to improve predictions, the life spans of 108 consecutive applications for inpatient hospice care were estimated independently by two oncologists, an internist, an oncology nurse, and a hospice social worker, based on data in a ten-page multidisciplinary application packet. The applicants were followed up until death. Actual life span was correlated with predictions. The median (+/- SD) life span was 3.5 +/- 12.4 weeks. The predictions as a group were overly optimistic about survival by an average of 3.4 weeks. The best prognosticator's prediction was only moderately correlated with actual life span, and no two prognosticator's predictions correlated closely with one another. Predicting actual interval until death was more accurate than predicting a 90% confidence interval around the time of death, though the latter procedure was better at avoiding the error of unpredicted long-term survivors. This imprecision in "expert" estimation of life span poses substantial problems for hospice programs and policymakers.