To quantify the risk of radiation-induced leukemia and provide further information on the nature of the relationship between dose and response, a case-control study was undertaken in a cohort of over 150,000 women with invasive cancer of the uterine cervix. The cases either were reported to one of 17 population-based cancer registries or were treated in any of 16 oncologic clinics in Canada, Europe, and the United States. Four controls were individually matched to each of 195 cases of leukemia on the basis of age and calendar year when diagnosed with cervical cancer and survival time. Leukemia diagnoses were verified by one hematologist. Radiation dose to active bone marrow was estimated by medical physicists on the basis of the original radiotherapy records of study subjects. The risk of chronic lymphocytic leukemia, one of the few malignancies without evidence for an association with ionizing radiation, was not increased [relative risk (RR) = 1.03; n = 52]. However, for all other forms of leukemia taken together (n = 143), a twofold risk was evident (RR = 2.0; 90% confidence interval = 1.0-4.2). Risk increased with increasing radiation dose until average doses of about 400 rad (4 Gy) were reached and then decreased at higher doses. This pattern is consistent with experimental data for which the down-turn in risk at high doses has been interpreted as due to killing of potentially leukemic cells. The dose-response information was modeled with various RR functions, accounting for the nonhomogeneous distribution of radiation dose during radiotherapy. The local radiation doses to each of 14 bone marrow compartments for each patient were incorporated in the models, and the corresponding risks were summed. A good fit to the observed data was obtained with a linear-exponential function, which included a positive linear induction term and a negative exponential term. The estimate of the excess RR per rad was 0.9%, and the estimated RR at 100 rad (1 Gy) was 1.7. The model proposed in this study of risk proportional to mass exposed and of risk to an individual given by the sum of incremental risks to anatomic sites appears to be applicable to a wide range of dose distributions. Furthermore, the pattern of leukemia incidence associated with different levels of radiation dose is consistent with a model postulating increasing risk with increasing exposure, modified at high doses by increased frequency of cell death, which reduces risk.