The dose-response relationship for radiation-induced leukemia was examined in a pooled analysis of three exposed populations: Japanese atomic bomb survivors, women treated for cervical cancer, and patients irradiated for ankylosing spondylitis. A total of 383 leukemias were observed among 283,139 study subjects. Considering all leukemias apart from chronic lymphocytic leukemia, the optimal relative risk model had a dose response with a purely quadratic term representing induction and an exponential term consistent with cell sterilization at high doses; the addition of a linear induction term did not improve the fit of the model. The relative risk decreased with increasing time since exposure and increasing attained age, and there were significant (P < 0.00001) differences in the parameters of the model between datasets. These differences were related in part to the significant differences (P = 0.003) between the models fitted to the three main radiogenic leukemia subtypes (acute myeloid leukemia, acute lymphocytic leukemia, chronic myeloid leukemia). When the three datasets were considered together but the analysis was repeated separately for the three leukemia subtypes, for each subtype the optimal model included quadratic and exponential terms in dose. For acute myeloid leukemia and chronic myeloid leukemia, there were reductions of relative risk with increasing time after exposure, whereas for acute lymphocytic leukemia the relative risk decreased with increasing attained age. For each leukemia subtype considered separately, there was no indication of a difference between the studies in the relative risk and its distribution as a function of dose, age and time (P > 0.10 for all three subtypes). The nonsignificant indications of differences between the three datasets when leukemia subtypes were considered separately may be explained by random variation, although a contribution from differences in exposure dose-rate regimens, inhomogeneous dose distribution within the bone marrow, inadequate adjustment forcell sterilization effects, or errors in dosimetry could have played a role.