There were four natural background radiation studies (32,36,41,48). Three were national studies of the effects of terrestrial gamma rays that estimated dose using existing radiation survey data. Study investigators assessed dose without direct measurement or interview. In all three studies, exposures inside buildings contributed more dose compared with the outdoors. The Great Britain study (study ID 4) had access to individual measurements of gamma rays in buildings but assessed exposures to study subjects as the mean for the county district of birth (36). The Swiss and Finnish studies (IDs 5 and 7, respectively) aggregated results in the form of average outdoor gamma-ray dose rates in geographic grid squares (41,48). The Finnish study converted these to indoor dose rates using house-type specific shielding factors. The Great Britain study considered some sources of dose uncertainty (37) but concluded the effect from these sources was limited to a loss of statistical power. There was some evidence of a potential downward bias in risk estimates in the Swiss study possibly caused by a lack of dose information due to residential mobility (41). None of the national studies included dose from other radiation sources, such as ingestion of naturally occurring radionuclides and medical exposures; however, sensitivity analyses in the Finnish study found no evidence of a potential bias from unmeasured exposures related to computer tomography (CT) examinations. National study investigators did not evaluate other sources of dose uncertainty.

The Chinese background study (study ID 3) applied numerous measurements in the study and control areas to estimate indoor and outdoor doses based on hamlet-specific averages (32). Intercomparisons with groups of residents by dosimeters and biologic dosimetry systems provided some validation of dose estimates (33,108). One intercomparison study suggested that the average coefficient of variation (CV) of the ratio of measured values to the estimated values was less than 22% (33). The potential for confounding by medical exposures appears small based on previous studies of this cohort (109). Internal doses were not assessed. Dose error was not addressed in dose-response analyses.

Human activity studies examined cancer risks in persons exposed to radioactive contamination in and around their place of residence (28,29,44,49). This group of studies required retrospective assessment of time-varying doses under different exposure scenarios. The levels of exposure were, in general, a function of source, occupancy, release rate, environmental transport, exposure pathway, and biokinetics. Given this complexity, the modeling necessary to estimate individual dose was unique to the scenario at hand, and the methods used were generally more complex than those for background studies. A common weakness among these studies was the lack of accounting for dose estimation errors in dose-response analysis.

Study ID 1 is a population-based case-control study of childhood leukemia in regions of Belarus, Russia, and Ukraine contaminated by the Chernobyl accident (28). Davis et al. (28) calculated absorbed dose to red bone marrow (RBM) using residential histories and field measurements for external doses and radionuclide concentrations and assumptions on individual food and water consumption rates for internal doses. The dosimetry addressed time dependence of these parameters through residential changes of study subjects. The dose reconstruction used information on residence and personal information obtained by questionnaires administered after case ascertainment; however, there was no assessment of the potential for differential recall. Investigators used Monte Carlo simulation techniques and mean values of 1000 realizations from internal and external dose sources in dose-response analyses (personal communication with study authors on September 20, 2018). Excess risk was most evident in Ukraine, diminished in Belarus, and not found in Russia. Although investigators mentioned dosimetry errors as a possible cause, they deemed them an unlikely explanation for this heterogeneity because of the common dosimetry methods across countries. Instead, there was evidence of a bias in control selection such that Ukraine controls tended to be selected from less contaminated areas than cases given differences in selection procedures (110).

Study ID 2 examined cancer incidence from 1982 to 1995 among Caucasian adults (10 446 men and 11 048 women) who resided within 5 miles of the Three Mile Island (TMI) nuclear power station during the 1979 partial reactor core meltdown (29). Dose estimation combined self-reported information on occupancy within the 5-mile zone for 10 days following the accident with calculated time-dependent gamma dose-rate distributions (30). The approach intentionally overestimated doses (approximately 40%). Estimate precision was poor, with uncertainty in dose ranging from two- to sixfold. Dose-response analyses did not account for dose uncertainty. Because location data were collected about 2 months after exposure and well before follow-up, the potential for differential recall was small.

Study ID 6 examined cancer incidence in Techa River residents exposed to releases from the Russian Mayak Radiochemical Plant in the Southern Urals (44). Dose accrued externally from fission product contamination in river sediment and surrounding soil and accrued internally from the consumption of contaminated water, milk, and food. Interviews were used to assess group occupancy factors but not for individual dose assessment. Dose was treated as a time-dependent continuous variable calculated as the 5-year–lagged absorbed dose to the stomach. Doses were estimated using the Techa River Dosimetry System (TRDS-2009). As in most models, estimates relied on specific choices for uncertain modeling parameters, which is a source of shared uncertainty. Model validation efforts were noteworthy, including multiple intercomparisons with results from other models, radionuclide assays, and opportunistic dosimetry (45, 111–116). In particular, two studies of stable chromosome aberrations assayed via fluorescence in situ hybridization (FISH) suggest trends with dose a bit lower than (although comparable with) those in the Japanese atomic bomb survivors (115,116); these and various studies of electron paramagnetic resonance in tooth enamel have been summarized in a recent review (113). A recent assessment yielded realization distributions with geometric standard deviation values for internal absorbed dose to the stomach of about 2 to 3, with external dose uncertainty slightly less (117). Other reports have indicated dose uncertainties on the order of four- to fivefold (108). Investigators did not account for dose uncertainty in dose-response analysis.

In the early 1980s, a large quantity of steel reinforcing bars contaminated with cobalt-60 was used in the construction of schools and residential buildings in Taiwan. This contamination was not discovered until 1992, when measurements revealed dose rates of 0.5–270 μGy/h (118). Breast cancer and leukemia were examined in a cohort of 6242 building residents with adequate information for dose assessment. As in previous studies of this cohort, doses were assessed using the Taiwan Cumulative Dose system (50,51). Occupancy factors were assessed by interviews, with some taking place after case ascertainment; therefore, some bias from differential recall appeared possible. Model validation procedures involved comparisons of radiation survey data and personal dosimetry measurements in a sample of residents. Dose-response analyses did not describe or account for dose estimation errors. Analysis of chromosome aberrations in these individuals has shown excess micronuclei (119), as well as dicentric chromosomes (120), in the exposed group, but no dose-response analyses were carried out in either study. Doses are likely too low to yield statistically significant trends, in view of the difficulty in detecting signals much below 100 mGy (121).

Eisenberg et al. (study ID 9) (52) examined cancer risks from radiation exposure in 82 861 adult patients undergoing fluoroscopically guided procedures (with or without contrast media) or nuclear medicine procedures following acute myocardial infarction. The authors calculated the cumulative effective dose for each patient by summing average values per procedure abstracted from the literature (53,54). This was accomplished by linking patient billing codes to procedures of interest, which were myocardial perfusion imaging (15.6 millisievert [mSv]), diagnostic cardiac catheterization (7.0 mSv), percutaneous coronary intervention (15.0 mSv), and cardiac resting ventriculography (7.8 mSv). These estimates did not consider heterogeneity in dose within broad groups of procedures or between centers or individuals. The investigators acknowledged that the variability of the doses between centers and operators was a limitation; however, they did not carry out a formal quantitative evaluation. Another limitation was the use of effective dose, which can greatly differ from the absorbed dose to organs of interest under partial-body irradiation and nuclear medicine procedures.

The French and United Kingdom CT cohorts (study IDs 10 and 11) were launched in the early 2000s and included 67 274 and 180 000 pediatric patients, respectively (55,57). These studies represent a new source; there were no previous CT studies in BEIR VII. In both studies, the authors collected information from the radiology information system of participating radiology departments. The radiology information system is devoted to the administrative recording of the radiology activities but includes only limited information on the type of examination performed (ie, body region scanned). Dose reconstruction therefore involved typical protocols defining image-acquisition parameters rather than individual data. Doses were based on typical values obtained at the national level in study ID 11 (58), whereas an extensive two-step survey in participating hospitals allowed hospital-based protocols to be used for dose reconstruction in study ID 10 (55,56), with imputation of median values from other radiology departments in cases of missing data. In both cases, assigned doses do not reflect interindividual variability. The Picture Archiving Communication System, which provides systematic recording and archiving of all images from the CT machine as well as a summary of the machine settings associated with each image taken, was progressively introduced worldwide after the mid1990s and could be used to derive more individualized organ dose estimates. Dose uncertainties were not quantified and therefore not considered in dose-response analysis.

The low-dose PIRATES study (study ID 12) examined thyroid cancer risk following exposure to low doses (<200 mGy) of ionizing radiation in childhood by pooling data from nine studies with individual estimates of thyroid dose (59). The Tinea Capitis cohort included 10 834 children treated in the 1950s in Israel, who represent most children involved in the pooled analysis (124). Individual doses used on-phantom measurements simulating x-ray prescriptions implemented in treatment centers (62). In reanalysis of the Tinea Capitis cohort, the authors assessed multiple sources of dose estimation error and developed a predictive model to account for major sources of uncertainty in dosimetry (60,61). The predictive model used information collected from three studies of anthropomorphic phantoms to estimate dose, using age at first irradiation, x-ray filtration, prescribed dosage, and the number of treatments. The model accounted for missing data by averaging the prediction equation over the probability distribution of the required variables, given the available data. The model also accounted for random errors representing intraindividual effects (due to motion during the treatment or peculiarities in positioning of the body), interindividual effects (distribution of physical characteristics of the head), and other sources of random error. Researchers combined errors to compute the expected true dose from the available patient data and the prediction equation using a Monte Carlo approach. Thus, for dose-response modeling, Poisson regression calibration was accomplished by using expected true dose categorization.

Dose-response regression parameter estimates, standard errors, and inferences were essentially unchanged after accounting for measurement error, which study investigators attributed to the linearity of relative risk in dose and the predominance of Berksonian error (60,61). There was also little evidence of influence on the estimated potential effect modifiers attributable to dose uncertainty. This assessment accounted for most major sources of uncertainty; however, it did not account for measurement error associated with the phantom studies.

The NW studies (n=11; study IDs 13–26) comprised the largest group of occupational studies (Table 2). Study populations primarily comprised workers employed in research, weapons and fuel production, commercial power, or military operations. Doses encompassed exposures beginning as early as the mid-1940s and ending in 2005 (Table 2). Annual exposure patterns largely follow Cold War weapons production, with the bulk of the collective dose in studies occurring in the mid-1960s (Appendix B). There was overlap between studies stemming from the United Kingdom, United States, and French NW studies (66,81,95) comprising subcohorts pooled in the International Nuclear Workers Study (INWORKS) (82). Also, some Korean workers in a cancer incidence study (study ID 16) (68) were included in a previous mortality study (study ID 13) (63).

INWORKS (study ID 23) examined mortality patterns among 308 297 nuclear workers employed in France, the United Kingdom, and the United States between the years 1944 and 2005 (82,83,125,126). This study updated the dosimetry system used in the previous collaborative study coordinated by the International Agency for Research on Cancer, which included a comprehensive assessment of dose uncertainties (84). Investigators made considerable efforts to acquire complete dose histories and to derive unbiased estimates of absorbed dose to target tissues, considering known sources of systematic errors in facility dosimetry over time (17,84). In particular, investigators derived facility- and time-specific “bias factors” to estimate absorbed dose from recorded dose. For RBM dose, recorded values were divided by bias factors ranging from 1.4 to 2.2, with corresponding CV values ranging between 0.2 and 0.6. For colon dose, which was used in analyses of all solid cancers combined, factors ranged from 1.2 to 2.1 (CV values 0.3–0.8). Investigators also quantified uncertainty in dose conversion; however, dose-response analyses did not use this information.

INWORKS comprised several subpopulations in previous epidemiologic investigations spanning decades. Over the course of these studies, there have been a number of improvements in dose estimates afforded through multiple records reviews. The extended follow-up also enabled dosimetry to incorporate improvements in measurements over time; however, this did not alleviate errors in early dosimetry that were carried forward. By design, the selection of similar study populations reduced heterogeneity and the potential confounding from unmeasured high linear energy transfer radiations and incorporated radionuclides. Nevertheless, data were inadequate to quantify contributions from internal dose and neutron exposures. Instead, dose-response analyses indirectly examined effects from other radiations in alternative models (82,83).

INWORKS and the US atomic veterans study (study ID 24) estimated absorbed dose (82,85), whereas others used unadjusted doses in units of whole-body equivalent dose (66,69,74), personal dose equivalent at a tissue depth of 10mm [H_p(10)] (81,95) or effective dose (63,68,71,80). Most data originated from measurements using film meters in the early years (1940s–1980s) and thermoluminescent dosimeters thereafter. The US atomic veterans study also used available measurement data from film meters; however, relatively few individuals were assigned personal dosimetry (85). Only 25% of participants had film badge records accounting for at least 80% of their dose (86). Therefore, estimates of RBM-absorbed dose stemmed primarily from group radiation measurements or by using time-motion models based on job descriptions and area radiation levels. There was no information on validation methods, although there was reasonable agreement with estimates from a detailed dose reconstruction involving a subset of workers (86). The average CV in the subset analysis was about 0.4–0.5, and values ranged upward of threefold in some exposure scenarios. In this comparison, doses were consistently lower in the detailed dose reconstruction compared with the cohort-assigned values; therefore, a scaling factor of 0.64 was used in the epidemiologic study to correct cohort doses (85).

Some NW populations were susceptible to dose from incorporated radionuclides and neutron exposures (Table 4). Recorded whole-body doses included contributions from neutrons (63,66,68,69,71,80) and 50-year committed effective doses from incorporated radionuclides (63,68,71,80) in several studies. The Rocketdyne NW study (study ID 17) quantified internal dose from 16 different radionuclides; however, quantification was limited to those workers judged to have a 50-year committed effective dose of 10 mSv or greater (69,70). In that study, there were 46 970 cohort members, including 5801 monitored for radiation and 2322 monitored for internal dose. Study investigators estimated annual doses from internally deposited radionuclides for 292 workers. The inclusion of internal dose showed no meaningful effects on the dose-response. The Canadian and US NW studies (IDs 19 and 22, respectively) estimated dose from tritium uptakes (74,81). Dose-response analyses conducted with and without tritium dose or treating tritium as a separate model term did not suggest substantial tritium-related effects. In the French NW study (study ID 26), investigators examined confounding by internal exposures and concluded that neglecting internal dose did not substantially bias risk estimates in this cohort (104). Studies without adequate quantitative data on neutron exposures or incorporated radionuclides examined the effects in various sensitivity analyses using markers of exposure potential (66,82,95). These analyses did not reveal evidence of a strong bias in risk estimates resulting from excluding dose from neutrons and internal emitters.

Doses below detection limits (BDL) were explicitly addressed only in the UK National Registry for Radiation Workers (UKNRRW) study (study ID 15) (66); however, three other studies (study IDs 22, 23, and 26) used dosimetry systems that have addressed detection limits in previous reports involving full and subcohort populations (22,23,25,67,106,127,128). In studies of the UKNRRW study, results with and without BDL dose adjustments revealed no evidence of meaningful bias in risk estimates (67,129). Similarly, there was little evidence of a strong bias from BDL doses in other studies (22,23,127,128). Among these, a “worst-case” example found a 22% drop in the linear excess relative risk per sievert (ERR/Sv) for all cancers in a previous study of Oak Ridge National Laboratory workers, also included in the US NW study and INWORKS, after adjusting for BDL doses between 1943 and 1956, when film badge dosimeters were processed weekly (128). In subsequent examinations of these data, effects on risk estimates were more modest and potentially completely offset by other errors (22,127).

At some facilities, notional doses were assigned to periods of unmonitored exposure (eg, because of a lost badge or doses at a previous facility) as a means of assuring compliance with dose limits. The UKNRRW study accounted for notional doses, which lowered the collective dose from pro-rata assignments from 295 person-Sv to 15 person-Sv but did not meaningfully change risk estimates given a small change in the total collective dose (4260 person-Sv) (25). The lack of substantial bias from notional doses was also evident in a study of shipyard workers included in the US NW study and INWORKS (26).

Most studies expended considerable efforts to gain complete exposure histories; nevertheless, investigators of the Canadian NW study raised concern over missing data (74). The study may have omitted exposures to a group of early workers because of information lost during transfer to the central dose registry. The authors speculated that the missing data might explain the dose-related risk of solid cancer mortality observed among these workers that was absent among other workers; however, there was no attempt to examine the plausibility of the error to fully explain the risk difference. Recently, the missing data have been found and researchers have initiated an update to the study; therefore, lingering questions on bias in risk estimates from the missing data may be resolved soon (personal communication with study authors on October 17, 2018).

Work-related medical x-ray examinations (WRX) are a potential source of unmeasured occupational exposure in some cohorts. The majority of WRX dose stems from fluoroscopic or photofluorographic chest exams in the 1940s and 1950s; therefore, studies including workers employed prior to 1960 appear more susceptible to dose error from this source (130–133). In some US NW studies, WRX data were abstracted from medical records to estimate dose (101–103,134–137). Of these, three examined the potential for bias from unmeasured WRX, with two reporting no effect (134,135) and one showing attenuation of the association between lung cancer mortality and external dose, including WRX (136). In this review, WRX was examined in the French NW cohort (105). Medical records were unavailable; therefore, doses were estimated as the product of assumed yearly exams and dose per procedure. Because it could not be ruled out, fluoroscopy (1.5–3.0 mSv per exam) was assumed prior to 1955 and radiography (0.1–0.3 mSv per exam) thereon. Risk estimates without WRX were imprecise in this cohort; adding WRX doses led to modest attenuation (7–47%) and further reduction in precision, yet positive but nonsignificant dose-response associations persisted.

The NW studies did not account for random measurement error in dose-response analyses; however, doses that are the sum of many measurements (ie, as in cumulative dose) likely have relatively small random error. Previous examinations have provided little evidence of a substantive bias from random error. Xue et al. (127) examined the simultaneous effects of BDL doses and random error on findings in a study of Oak Ridge National Laboratory workers included in the US NW study and INWORKS. They concluded that random errors in measurements were unlikely to substantially bias risk estimates. Similarly, a detailed examination of the relative error in cumulative doses from film badges worn by Hanford workers (also included in study IDs 22 and 23) suggested that the increase in the total variance of measured cumulative doses is unlikely to be more than 1% of the total variance in true cumulative doses, leading to negligible bias in the risk estimates (138). A similar conclusion was reached regarding the Canadian NW study (75).

Three publications on cancer in the USRT (study ID 25) were eligible for consideration (89–91), with each using identical dosimetry but differing target organs of interest (92). For brevity, the narrative is limited to a study of breast cancer incidence and mortality patterns in females certified by the American Registry of Radiologic Technologists for at least 2 years from 1926 through 1982 (89). Exposures were reconstructed for the period 1916–1997 for technologists conducting medical diagnostic and therapeutic procedures using ionizing radiation. The mean cumulative breast dose was 37 mGy, ranging from 0.058 to 2500 mGy. (92).

Investigators used available personal dosimetry measurements and information on work procedures, protection practices (eg, apron use), x-ray imaging technology, and other factors to estimate whole-body dose equivalent and subsequently absorbed dose to breast tissue. Parameters used in estimation procedures were treated as expected values with an associated probability distribution. Monte Carlo methods were used to generate multiple dose realizations for the full cohort. These methods accounted for sources of shared and unshared errors, including treatment of BDL doses and errors associated with changing dosimetry and work practices over time. Regression calibration was used to account for random error. The CV in cumulative breast dose from 1000 dose realizations was about 2.4. The dosimetry system was partly validated by comparison with badge reading at five major US hospitals and biodosimetry methods linking this study to observations in LSS participants (92). Validation was also conducted by a study of stable chromosome aberrations assayed in 238 of the USRT participants using FISH suggesting that the trends of stable chromosome aberrations with dose are comparable with those in the Japanese atomic bomb survivors and in various other groups (139).

Dosimetry strengths were the collection and integration of individual film badge dose data available between the years 1960 and 1997, extensive efforts to account for sources of dose estimation error, and methods used to validate dose estimates. However, relying on indirect estimation because of the unavailability of measurement data prior to 1960 was an important limitation. Film badge measurements were available for 39% of the years worked, with about 25% of the collective dose derived from film badge data. The proportion of film badge–based dose estimates varied by the year of first exposure, ranging from essentially none for the earliest workers to 60% for those who began working after 1980. Moreover, birth cohort and total cumulative dose were associated, and the dose-related excess risk of breast cancer incidence was strongest in women born before 1930 (ERR/Gy = 1.6, 95% confidence interval [CI] = 0.3 to 3.9). In fact, the experience of early workers was the primary determinant of cancer incidence and mortality risk. Because of the large uncertainty in early doses and the strong birth cohort effect, the authors interpreted results cautiously.

Two case-control studies were reviewed. The first (study ID 14) was nested within cohorts of mostly male (>95%) liquidators from Belarus, Russia, and Baltic countries who took part in recovery activities between the accident date (April 26, 1986) and December 31, 1987 (64). The second (study ID 20) was nested within a cohort of 110 645 male Ukrainian workers who were 20 to 60years of age during cleanup activities between the years 1986 and 1990 (78). The exposure to liquidators was predominantly penetrating whole-body gamma radiation emitted from radionuclides (primarily ¹³⁷Cs) on contaminated surfaces. Dose from ingesting contaminated food and drinking water was plausible, especially in those living in Belarus; however, investigators posited that the dose contribution from incorporated radionuclides was at least an order of magnitude lower than the external contribution (140).

Studies of Chernobyl liquidators reported in BEIR VII relied primarily on dose data in the Russian National Medical and Dosimetric Registry, which was known to have several gaps and problems (79,140). To account for the Russian National Medical and Dosimetric Registry shortcomings, researchers combined information from available radiation measurements with self-reported data to develop the Radiation Dose Reconstruction with Uncertainty Estimates (RADRUE) software package used to estimate absorbed dose to RBM for workers in both studies. In general, RADRUE dose estimates are the product of exposure rate and irradiation time, given a number of exposure scenario parameters. Input data included work histories, exposure rates, adjustment factors for protective equipment used, and dose conversion coefficients. The output included point estimates of dose and associated uncertainties, the latter reported as geometric standard deviation values across all liquidator categories between 1.7 and 3.4 (65). The methods used to estimate uncertainty enabled examination of the effects of random errors on dose-response analyses, which suggested negligible effects on risk estimates, although confidence intervals were slightly wider (64). Validation of RADRUE was accomplished by intercomparisons with data from personal dosimeters and biological dosimetry, including both ESR and FISH (64,65,79,141). RADRUE does not provide information on internal exposure; however, separate estimation of doses and attendant uncertainties from consumption of contaminated food were calculated for study participants who resided in Belarus.

The dosimetry systems relied on self-reported information from a selected set of cases and controls, using in-person and proxy interviews. As a result, differential recall may have introduced a bias in dose estimates. To reduce potential biases, the interviewers were blinded to disease status. Similarly, dosimetrists estimated doses without knowledge of disease status. Some efforts to assess recall accuracy in these workers have suggested that a large bias in dose estimates was unlikely; however, an examination sufficient to exclude a bias in liquidator doses is difficult, if not impractical (79). Thus, the potential for spurious dose-response results from differential recall cannot be ruled out.