[en] The shape of the cancer mortality dose-response in the atomic bomb survivor data is analyzed in the context of linear-quadratic (LQ) models. Results are given for all cancers except leukemia as a group, for leukemia, and for combined inferences assuming common curvature. Since there is substantial information aside from these data suggesting a dose-response concave from above, the emphasis here is not on estimating the best-fitting dose-response curve, but rather on assessing the maximal extent of curvature under LQ models which is consistent with the data. Such inferences are substantially affected by imprecision in the dose estimates, and methods are applied which make explicit allowances for biases due to this. The primary means used here to express the extent of curvature is the factor by which linear risk estimates should be divided to arrive at appropriate low-dose risk estimates. In the past, influential committees have recommended ranges of 2-10 and of 1.5-3 for such a factor. Results here suggest that values greater than about 2 are at least moderately inconsistent with these data, within the context of LQ models. It is emphasized, however, that there is little direct information in these data regarding low-dose risks; the inferences here depend strongly on the link between low-dose and high-dose risks provided by the assumption of an LQ model. (author)