Some effects of random dose measurement errors on analyses of atomic bomb survivor data.

The effects of random dose measurement errors on analyses of atomic bomb survivor data are described and quantified for several analytical procedures. It is found that the ways in which measurement error is most likely to mislead are through downward bias in the estimated regression coefficients and through distortion of the shape of the dose-response curve. The magnitude of the bias and the power for testing the hypothesis of no effect are evaluated for several dose treatments including the use of grouped and ungrouped data, analyses with and without substituting 600 rad for estimated doses exceeding this value, and analyses which exclude doses exceeding 200 rad. The calculations are based on a model in which the error distributions are assumed to be log normal with standard deviations that are 0, 30, and 50%, respectively, of the true dose values. Results are limited to a dose-response function which is linear on total dose. It is found that the commonly applied practice of substituting 600 rad for doses exceeding this value definitely reduces bias in the presence of error. Restricting analyses to doses less than 200 rad reduces bias even more but at the price of considerable loss of power. Both the bias and the power for analyses based on grouped data are very close to the respective bias and power with ungrouped data.