Estimating Risk of Circulatory Disease from Exposure to Low-Level Ionizing Radiation

The comprehensive meta-analysis of Little et al. (2012) summarized possible circulatory disease risks related to medium and low doses of whole-body radiation exposure in humans. The authors looked at excess relative risk (ERR) estimates from 10 different epidemiological studies. Using two statistical measures to calculate pooled ERR, they determined aggregate measures of ERR for four detrimental health outcomes and they reported mostly significant values for the ERR per unit dose in their Table 2. 
 
Nine of the 10 studies Little et al. (2012) considered included moderate cumulative doses > 0.4 Sv (see their Table 1), and they observed that risk trends in most cohorts were driven by a relatively small number of highly exposed individuals. The authors then fitted a linear ERR model to the data of the meta-analysis and derived mortality risks at low-level radiation by extrapolation. 
 
Linear extrapolation is used in radiation protection if cohort strata pertaining to low doses and dose rates have low statistical power. There are, however, indications for non-linear protective effects of low doses delivered at low dose rates for end points related to athero-sclerosis in mice (Mitchel et al. 2011). Moreover, the recent review of Rodel et al. (2012) showed that low-dose ionizing radiation modulates inflammatory immune reactions mostly with discontinuous or biphasic dose dependencies. These recent findings suggest that non-linear dose responses might also play a role in the determination of the radiation risk for circulatory diseases. 
 
In this context we note that in the 10 studies analyzed by Little et al. (2012), risk estimates were mainly calculated with linear no-threshold (LNT) models (in fact, 7 of the 10 studies applied only the LNT model). Motivated by recent radio-biological findings, we fitted a large number of dose responses, in addition to the LNT model, to the data of the Life Span Study (LSS) cohort of Japanese atomic-bomb survivors, which is among the cohorts considered by Little et al. (2012). We realized that several models fitted the data about equally well (Schollnberger et al. 2012). Instead of picking a single model of choice for risk assessment (here, the LNT model), we allowed for model uncertainty via multi-model inference. By reducing the bias from model selection, we obtained larger uncertainty intervals for risk estimates. The “model-averaged” dose response predicted markedly lower risks than the LNT model for cerebro-vascular disease (CVD) and for cardio-vascular diseases excluding CVD. For example, for CVD an ERR model with a step at 0.6 Sv strongly influenced the average with a weight of 0.55 compared with the LNT model with a much lower weight of 0.26 (see Table 1 of Schollnberger et al. 2012). We did, however, not find any evidence for a protective effect but only for the contribution of pathways that have a threshold. 
 
Our results might have implications for issues of public health in the assessment of risk–benefit ratios for radio-diagnosis or radio-therapy. Thus, we encourage the use of multi-model inference techniques in the analysis of other cohorts. From our experience with the LSS cohort, we would expect lower risk estimates in the lower dose range with a more comprehensive characterization of uncertainties and improved support of the epidemiological data.

http: //dx.doi.org/10.1289//dx.doi.org/10. /ehp.1206033 Bellinger (2012 recently estimated the loss of cognitive function in terms of Full-Scale intelligence quotient (IQ) in children exposed to certain environmental chemicals. To ascertain pre natal exposures of methyl mercury (MeHg) in children, he used exposure data on mercury (Hg) concentrations in hair of U.S. women of childbearing age (16-49 years) from NHANES (National Health and Nutrition Examination Survey) 1999-2000(McDowell et al. 2004). Bellinger applied a regression coefficient of -0.18 IQ points per microgram per gram increase in maternal hair as calculated by Axelrad et al. (2007). However, the results of Axelrad et al. (2007) relied on incomplete data from a prospective study in the Faroe Islands and on nonadjusted results from the Seychelles study, later found to be confounded by nutrients from seafood (Strain et al. 2008). Bellinger (2012) then applied the regression coefficient to hair Hg levels > 1.11 µg/g (90th percentile), because this level corresponds to the reference dose of MeHg established many years ago. Assuming a concentration of 1.73 µg/g (95th percentile) as the midpoint (rather than the average, which is higher) for the hair Hg levels of the 10% of U.S. women with a level > 1.11 µg/g, he estimated a total IQ loss of 284,580 points. We believe that Bellinger's general approach is sound but that the doseresponse information is outdated, a caveat that Bellinger noted, although it was not reflected in the summary table. We therefore wish to complement these calculations using updated dose-response data.
Prospective data justify a lower threshold Hg level of 0.58 µg/g hair corresponding to 50% of the reference dose (Grandjean and Budtz-Jørgensen 2007). In addition, a 1-µg/g increase in hair Hg concentration is more likely associated with an average adverse impact of 0.465 IQ points, as discussed by Pichery et al. (2012). Assuming a lognormal exposure distribution, a 75th percentile hair Hg concentration of 0.42 µg/g, and a 90th percentile of 1.11 µg/g as reported by McDowell et al. (2004), we estimate that 18.5% of women exceed a threshold of 0.58 µg/g hair Hg and that the average concentration for 0.58-1.11 µg/g is approximately 0.8 µg/g. For the sake of comparing these values with Bellinger's calculations (Bellinger 2012), we used a median concentration of 1.73 µg/g as the average hair Hg level of the 10% of U.S. women with a level > 1.11 µg/g. On the basis of these assumptions, we calculated a total IQ loss for the U.S. population of children 0-5 years of age (n = 25.5 million) to be 1,590,000 IQ points, or 264,000 IQ points per year.
We recently used similar calculations to estimate the annual costs of Hg pollution in France (Pichery et al. 2012), a country onefifth the size of the United States. At slightly higher exposure levels, the annual loss in IQ points was estimated to be 157,000. Greater losses were obtained using a log-scale effect (Pichery et al. 2012). With an estimated value of each IQ point of $18,000 in terms of lifetime earnings, the current loss of IQ points associated with MeHg exposure represents a very substantial value to society. In my paper (Bellinger 2012), I noted among the limitations that the calculations are only as valid as the data on which they are based. My hope was that those with a special interest in a particular risk factor would be stimulated to provide stronger data on either the exposure distribution or the dose-response relationship so that the calculations could be refined. I am therefore grateful to Grandjean et al. for providing an updated estimate of the dose-response relationship for pre natal methyl mercury, the use of which suggests that the total Full-Scale IQ loss among U.S. children is considerably larger than my initial estimate. All of the estimates listed in Table 2 of my paper (Bellinger 2012) should be considered provisional and should be updated when more precise data become available.  Table 1), and they observed that risk trends in most cohorts were driven by a relatively small number of The correspondence section is a public forum and, as such, is not peer-reviewed. EHP is not responsible for the accuracy, currency, or reliability of personal opinion expressed herein; it is the sole responsibility of the authors. EHP neither endorses nor disputes their published commentary.

David C. Bellinger
highly exposed individuals. The authors then fitted a linear ERR model to the data of the meta-analysis and derived mortality risks at low-level radiation by extrapolation.
Linear extrapolation is used in radiation protection if cohort strata pertaining to low doses and dose rates have low statistical power. There are, however, indications for non linear protective effects of low doses delivered at low dose rates for end points related to athero sclerosis in mice (Mitchel et al. 2011). Moreover, the recent review of Rödel et al. (2012) showed that low-dose ionizing radiation modulates inflammatory immune reactions mostly with discontinuous or biphasic dose dependencies. These recent findings suggest that non linear dose responses might also play a role in the determination of the radiation risk for circulatory diseases.
In this context we note that in the 10 studies analyzed by Little et al. (2012), risk estimates were mainly calculated with linear no-threshold (LNT) models (in fact, 7 of the 10 studies applied only the LNT model). Motivated by recent radio biological findings, we fitted a large number of dose responses, in addition to the LNT model, to the data of the Life Span Study (LSS) cohort of Japanese atomic-bomb survivors, which is among the cohorts considered by Little et al. (2012). We realized that several models fitted the data about equally well (Schöllnberger et al. 2012). Instead of picking a single model of choice for risk assessment (here, the LNT model), we allowed for model uncertainty via multi model inference. By reducing the bias from model selection, we obtained larger uncertainty intervals for risk estimates. The "model-averaged" dose response predicted markedly lower risks than the LNT model for cerebro vascular disease (CVD) and for cardio vascular diseases excluding CVD. For example, for CVD an ERR model with a step at 0.6 Sv strongly influenced the average with a weight of 0.55 compared with the LNT model with a much lower weight of 0.26 (see Table 1 of Schöllnberger et al. 2012). We did, however, not find any evidence for a protective effect but only for the contribution of pathways that have a threshold.
Our results might have implications for issues of public health in the assessment of risk-benefit ratios for radio diagnosis or radio therapy. Thus, we encourage the use of multi model inference techniques in the analysis of other cohorts. From our experience with the LSS cohort, we would expect lower risk estimates in the lower dose range with a more comprehensive characterization of uncertainties and improved support of the epidemiological data.