A comparison of statistical methods for estimation of less than detectable ionising radiation exposures.

Methods were developed to estimate the ionising radiation dose below the detection level (DL) of personal monitoring devices for a case-control study of protracted radiation exposure and lung cancer. Exposure data were grouped by dosemeter type and monitoring period. Each group contained dosimetry data that were interval-censored from limitations in measurement precision and included left-censoring of observations below detection. The grouped data were fit to a three parameter hybrid-lognormal distribution by maximum likelihood estimation. Using the fitted distribution, bootstrap samples were either simulated by Monte Carlo or constructed by sampling with replacement. The resulting bootstrap sample distributions were then used to predict the missing dose values and the associated uncertainty in the estimate. Among study subjects, 1357 workers were monitored with film dosimetry. Among the 39,263 dose observations 20,416 were recorded as zero dose, indicating 52% left-censoring. The statistical methods estimated 0.31 person-Sv below the DL or approximately 1% of the total collective dose for this study population.