Studies of radiation health effects inform EPA actions

Abstract

In 1970, the US Environmental Protection Agency (EPA) was given the responsibility to provide guidance to other federal agencies in the formulation of radiation protection standards. To carry out its federal guidance responsibilities and protect human health, the EPA must estimate risk at low doses to limit the risk of radiogenic cancers from environmental exposures. These risk estimates are based on models which conform to the linear no threshold (LNT) hypothesis. A cancer risk model conforms to the LNT hypothesis if the excess risk of cancer at low doses increases approximately proportional to dose, with no threshold. Risk models with a linear-quadratic dose response can satisfy the LNT hypothesis. Based on careful review of evidence from epidemiological and radiobiological studies, authoritative scientific bodies have repeatedly endorsed the use of LNT models for estimating and regulating risk and concluded that despite uncertainties at low dose and dose rates, the LNT model remains the most practical and implementable model for radiation protection. This article describes the rationale underlying the use of LNT models for calculating risk for low dose and dose rate exposures and discusses some of the epidemiological evidence which inform on its continued use.

1. Introduction

In 1970, the US Environmental Protection Agency (EPA) was given the responsibility to provide guidance to other federal agencies in the formulation of radiation protection standards (Reorganization Plan No. 3 of 1970 2012). The Atomic Energy Act of 1954 (2012), a US law, instructs the EPA to consult with qualified experts in biology, medicine, and health physics and authoritative bodies including the National Academy of Sciences (NAS) and the National Council of Radiation Protection and Measurements (NCRP). In its mission to protect human health and to carry out its federal guidance responsibilities, the EPA must estimate radiogenic cancer risk at doses that are below levels where epidemiology provides direct statistical evidence of risk (e.g. below about 100 mSv). To address this problem, EPA’s risk projections for environmental exposures are based on linear no threshold (LNT) models which conform to the LNT hypothesis (LNTH). This approach has been endorsed by scientific bodies such as NAS, the International Commission on Radiological Protection (ICRP), and NCRP. They conclude that, despite uncertainties at low doses, the LNT model remains the most practical and implementable model for radiation protection. For almost all cancer sites, EPA’s methodology for calculating radiogenic risks is based on LNT models recommended by the NAS Biological Effects of Ionizing Radiation (BEIR) VII committee (NAS-NRC 2006).

2. Linear no threshold models for estimating cancer risk at low doses and dose rates

2.1. LNT models

A common misconception about the LNTH is that the dose-response must be linear over a wide range of doses. As it is used in radiation protection, a cancer risk model is said to conform to the LNTH if the excess risk of cancer at low doses or low dose rates (LD/LDR) increases approximately in proportion to dose with no threshold (ICRP International Commission on Radiological Protection 2007). Herein, risk models that satisfy the LNTH will be referred to as LNT models. Of the four hypothetical dose response functions depicted in figure 1, the first two are examples of LNT models. The second model is a linear quadratic (LQ) model, for which the dashed line represents the linear component; the slope of the linear component approximates the risk per unit dose for low dose exposures. As will be discussed in this section, LQ models have repeatedly been fit to Japanese atomic bomb survivor data (e.g. data on leukemia incidence or mortality) for the purposes of estimating radiogenic cancer risk. The third and fourth models depicted in figure 1 are examples of threshold and hormetic models, respectively, which do not satisfy the LNTH.

Figure 1.

Four dose-response functions relating excess relative risk (ERR) to dose. The top two satisfy the LNTH. The second is an example of a linear quadratic dose response (linear component shown by the dashed line). These are followed by the dose response for threshold (third) and hormetic (last) models.

The rationale for assuming a risk model which satisfies the LNTH is outlined in section 2.4. That section also refers to the controversy regarding application of LNT models. As discussed in section 3, a substantial number of epidemiological studies (e.g. on effects from medical and environmental exposures) provide reinforcing evidence that the dose response satisfies the LNTH.

2.2. Application of LNT models for estimating radiogenic cancer risk

For most types of cancer, EPA’s estimates of excess risk for LD/LDR low-linear energy transfer (LET) exposures are based on extrapolations or interpolations using LNT risk models derived from the Lifespan Study (LSS) of the Japanese atomic bomb survivors (EPA Environmental Protection Agency 2011, Grant et al 2017). For most of these cancers, the risk models are the same or similar to the ones recommended in the NAS BEIR VII report and ICRP Publication 103 (NAS-NRC 2006, ICRP International Commission on Radiological Protection 2007). The general approach of applying risk models derived from the LSS to estimate risks for non-Japanese populations is well established and has been used for decades in authoritative scientific reports (NAS-NRC National Academy of Sciences—National Research Council 1980, UNSCEAR United Nations Scientific Committee on the Effects of Atomic Radiation 2008).

The LSS cohort includes about 120 000 persons of whom about 40% are male and about 94 000 were in either Hiroshima or Nagasaki at the time of the bombings (Grant et al 2017). The 94 000 survivors received acute doses of radiation, mostly in the form of gamma rays, with a small admixture of neutrons. Of these, about 87 000 have dose estimates; weighted absorbed colon dose estimates range from about 0 Gy to more than 3 Gy. Follow-up began in 1950 for cancer mortality and in 1958 for cancer incidence with the establishment of tumor registries. The most recent comprehensive analysis of cancer incidence included data from about 105 000 LSS members with follow-up to 2009 (Grant et al 2017). The analysis excluded three groups of people: (1) those who died before 1958, (2) those who had been diagnosed with cancer before 1958, and (3) those for whom doses could not be estimated. The LSS has important strengths, including: a nearly instantaneous exposure, which can be pin-pointed in time; a large exposed population encompassing both genders and all ages; a wide range of radiation doses to all organs of the body; and detailed follow-up for about 50 years.

For decades, the same general approach has been used in analyses of LSS data for deriving models for the relationship between cancer risk and radiation dose. More specifically, excess relative risk (ERR) and excess absolute risk (EAR) are assumed to be functions of tissue dose (Gy) and other factors (e.g. the specific tissue, sex, age-at-exposure, and attained age). ERR represents the ratio of the age-specific increase in cancer (incidence or mortality) rate attributable to a radiation dose divided by the baseline rate (i.e. the cancer rate associated with the background radiation level); EAR is simply the difference in cancer rates that can be attributed to radiation. In most recent analyses, the functions that describe ERR or EAR are assumed to be the product of: (1) a function that defines the dose response or the ‘basic’ relationship between ERR or EAR and dose, and (2) a function which describes modification of the dose-response by other factors. Among the dose-response models that are often considered, including models with thresholds, linear and LQ dose response models provide good fits to LSS cancer incidence and mortality data. The resulting models for ERR (and similarly for EAR) are of the form:

$$ERR\left(d,e,s,a\right)=\left(\alpha d+\beta d^2\right)g\left(s,e,a,\dots \right)$$ (1)

where d denotes dose, (αd + βd²) represents the dose-response function (e.g. for females with exposure at age 30 and attained age 70), and g(s, e, a, …) defines how the dose response is modified by factors such as sex (s), age-at-exposure (e), attained age (a), and possibly other factors (e.g. categories based on smoking history) as designated by the ellipsis. The model represented in equation (1) satisfies LNTH and, at very low doses, the ERR would be:

$$\alpha dg\left(s,e,a,\dots \right).$$ (2)

The model could be used to estimate excess cancer rate (i.e. the EAR) by multiplying an estimate of the expression given in equation (2) by the appropriate baseline cancer rate. For example, according to the model, the excess cancer rate at age 60 attributable to a tissue-specific (acutely-delivered) dose of 0.001 Gy to a population of US males at age 40 would be:

$$h_{us\left(\mathit{\text{male}}\right)}\left(60\right)\alpha \left(0.001\right)g\left(\mathit{\text{male}},40,60\right)$$ (3)

where h_{us (male)} (60) denotes the baseline cancer rate for the irradiated tissue for US males at age 60. ‘True’ values for α and other parameters which define the function g are unknown, and estimates for these parameters must be inserted into equation (3). Note that, due to the very large size of the LSS cohort, parameter estimates for some categories of cancer (e.g. all solid cancers) may be reasonably precise. LNT models fit to the Japanese atomic bomb survivor data are functions from which excess risk can be estimated at any dose of interest for exposures to non-Japanese populations.

Examples of LNT models fitted to the LSS data used for the BEIR VII report are depicted in figures 2 and 3. In figure 2, the points represent dose-category-specific estimates of ERR for solid cancer incidence for doses from exposures at age 30 and attained age of 70. These estimates depend on numbers of solid cancers diagnosed during pre-specified time intervals within the follow-up period. A regression procedure (Poisson regression) is used to determine what terms should be included in the model (e.g. whether the dose response is linear or LQ) and to estimate parameter values. A linear dose response provides a reasonable fit to the data (solid straight line in figure 2), and is heavily dependent on data from the approximately 11 000 survivors with colon doses between 0.2 Sv and 2.0 Sv. The fitted dose response depicted by the solid straight line depends to a lesser extent on data from the highest dose categories (i.e. >2.0 Sv) because there are relatively few survivors who received doses >2.0 Sv.

Figure 2.

Solid cancer incidence dose–response function averaged over sex for attained age 70 after exposure at age 30. The solid straight line is the linear slope estimate, the points are dose category-specific ERR estimates, the solid curve is a smoothed estimate derived from the points. The dotted curves indicate upper and lower one-standard-error bounds on the smoothed estimate. Reproduced with permission from Preston et al 2003. © 2019 Radiation Research Society.

Figure 3.

Primary descriptions of the excess risks of solid cancer incidence. The left panel presents fitted sex-averaged ERR estimates using both attained-age-declining (dark solid line) and attained-age-constant (dashed lines) forms, for age-at-exposure groups 0–9, 10–19, 20–39 and 40. ERR estimates for women are about 25% greater and ERR estimates for men are 25% lower than the values shown. The right panel presents fitted EAR estimates for the same dose groups. There is no evidence of significant sex differences in the fitted EAR. Reproduced with permission from Preston et al 2003. © 2019 Radiation Research Society.

Figure 3 illustrates how the estimated dose response is modified by age-related factors (i.e. how the function g(s, e, a) in equation (2) depends on age-at-exposure and attained age). Estimates of ERR decrease with increasing age-at-exposure and attained age; estimates of EAR decrease with increasing age-at-exposure but increase with attained age. The apparent discrepancy on how ERR and EAR depend on attained age is a consequence of the steep increase in baseline cancer rates with age; estimates of EAR increase with attained age, but not in proportion to baseline rates.

2.3. Rationale for the ‘LNT’ approach

For radiation protection of the public, a key concern is managing risk from environmental exposures at low doses or dose rates. However, current estimates of risk depend primarily on data from atomic bomb survivors with acute exposures involving moderate to high doses of radiation. The most oft-cited rationale for this approach was given almost 40 years ago in an article published in the journal Science (Land 1980). Charles Land showed—through a series of relatively simple statistical power calculations—that ‘precise direct estimation of small risks [associated with radiation exposures of interest] require impracticably large samples.’ His calculations indicated that:

• To detect the excess risk from mammography with at least a 70% probability, a cohort study would likely have required a sample of at least 100 million women with 20 years of follow-up. He assumed half of the women would receive a single mammography at age 35 y with an average dose of about 10 mGy to both breasts, and a resulting excess risk of about 6 breast cancers per 10 000 000 person-years. Remarkably, this estimated excess rate is almost identical to the corresponding BEIR VII model-based estimate (for a dose of 10 mGy to both breasts).

• A case-control study would require about 600 000 cases (assuming a ratio of 4:1 controls to cases) to detect the risk of breast cancer from a 10 mGy dose to both breasts.

• Sample sizes needed to detect an excess male leukemia risk from a bone marrow dose of 10 mGy at age 35 y are considerably smaller than for breast cancer. Land’s calculations indicated a cohort study would require about 16 million men (among whom half are unexposed). Alternatively, a case-control study would require about 1300 cases (matched to 5200 controls) to detect the excess leukemia risk. In general, required samples sizes are heavily dependent on relative risk (i.e. the ratio of the excess rate from radiation divided by the baseline rate), and the relative risk for leukemia tends to be very large compared to the relative risk for many other cancers. Nevertheless, samples sizes would need to be extremely large even for leukemia (e.g. in 1980, 1300 cases ‘[was] about twice the annual number of cases among men in the age range 35–49’).

Dr Land also explained that for an LNT dose response, the required sample size is inversely proportional to the dose squared. At 100 mGy, minimal sample sizes would be about 100 times smaller than at 10 mGy (e.g. the corresponding minimal sample sizes to detect an excess breast cancer risk from radiation would be about 1 million for a cohort study or 6000 cases for a case-control study).

Thus, studies with the statistical power to directly detect and precisely estimate risks for doses much less than 100 mGy could not be implemented in 1980. LNT models offered a promising alternative. Even at low doses, risk estimates of reasonable precision can be derived using data from studies such as the LSS and by assuming a relatively simple structure for the dose response over a relatively wide range of doses (e.g. a linear-quadratic dose response for doses <1 Gy). As stated previously, the exact form of these models and parameters are determined using regression methods.

To illustrate why this approach results in more precise estimates of risk, consider the recent LSS cancer incidence data given in table 1. Crude cancer rates (unadjusted for age) and underlying data are given for female atomic bomb survivors: 59.9 and 89.3 cases per 10⁴ person-years for doses <0.005 Gy and 0.5–1.0 Gy, respectively (Grant et al 2017). A simple estimate of ERR for the ‘high’ dose group is 0.50 = (89.3/59.9)−1. Since the average dose for that category is 0.703 Gy, and assuming that the dose response is linear up to 1.0 Gy, a reasonable ‘back-of-the-envelope’ estimate of the slope would be (0.50/0.703) Gy⁻¹ = 0.7 Gy⁻¹. At 0.01 Gy, the estimated ERR would be 0.007 and a corresponding estimate of the excess cancer rate at age 60 would be 0.007h_(female) (60). The relative precision of these estimates depends primarily on the precision of the estimated cancer rates for the high dose category and the difference in cancer rates for the two dose groups. The 95% confidence interval of (82, 97) cases per 10⁴ person-years for the high dose group suggests the estimated cancer rate (89.3 cases per 10⁴ person-years) would likely be within about 10% of the true value. An estimated one-third of the rate is attributable to the exposure to radiation, and it follows that the corresponding estimate of excess risk would be within about 30% (3 times 10%) of its true value. As stated in section 2.2, models are fit to almost the entire LSS data (e.g. for survivors with dose estimates <4 Gy). The models account for potential modifying factors such as sex, age-at-exposure, and attained age, and the data from intermediate dose groups allows for tests relating to the shape of the dose response (e.g. whether an LQ model would result in a significantly better fit to the data). Reasonably precise estimates of excess risk for low dose exposures are made possible through model assumptions on how excess risk at moderate-to-high doses (for which estimates tend to be more precise) are related to excess risk at low doses.

Table 1.

Selected statistics for LSS female solid cancer incidence data from 1958–2009^a.

Weighted colon dose (Gy)	No. of Females	Person-years	Cases	Cancer rate (per 104)	ERR (dose group)	ERR per Gy
<0.005	21 404	653 836	3918	59.9	0	0
0.5–1.0	1854	56 790	507	89.3	0.49	0.70 Gy−1
Mean = 0.703				CI: (82, 97)	CI: (0.36, 0.63)	CI: (0.51, 0.90)

^aBased on results given in table 3 of Grant et al (2017); the CI denote 95% confidence intervals.

2.4. Potential for bias from incorrect model assumptions

LNT-based estimates of excess risk from LD/LDR exposures may be sufficiently precise for many types of environmental exposures, but that does not mean they are accurate. The analyses of moderate or high dose data, such as from the LSS, might be based on incorrect model assumptions, and this could lead to bias in estimates of risk associated with LD/LDR exposures. The two most obvious potential sources of such bias are: (1) deviations (if they exist) from the LNTH (i.e. excess risk at doses <100 mGy might not be proportional to dose), and (2) incorrect assumptions on how risk depends on dose rate. More subtle biases can occur for a variety of reasons such as: (1) the functional form for the basic dose-response relationship for doses >100 mGy differs from what is assumed, (2) incorrect functional forms are used to model modification (e.g. by smoking) of the dose response, (3) the risk model does not include all modifiers of the dose response, and (4) baseline cancer rates are not adequately modeled.

The rationale for assuming a LQ form for the basic dose-response function for the radiation exposures for the atomic bomb survivors is summarised in EPA Radiogenic Cancer Risk Models and Projections for the US Population (2011):

…radiobiological data suggest that the probability of a mutational damage in a cell’s DNA from an acute exposure to low-LET radiation can be expressed as a linear-quadratic (LQ) function of dose… The linear term is assumed to reflect the effect of single tracks, the quadratic term the added effect of two tracks traversing the cell close together in space and time, or perhaps the saturation of repair mechanisms at higher dose… It is presumed that the probability of carcinogenesis induced in an organism from an exposure to radiation is proportional to the number of induced mutations remaining after repair is complete. This has led scientists to model the excess risk as a LQ function of dose for relatively high acute dose…

For very small doses, the linear term will dominate (i.e. the quadratic term reflecting the effect from two closely spaced tracks can be ignored), and the excess probability of cancer will be approximately proportional to dose. For doses delivered in small fractions, ‘each dose fraction produces an incremental effect, … the quadratic terms associated [with each dose fraction] will be negligible, and the overall summed effect will be linear with dose’ (EPA Environmental Protection Agency 2011). The dose response for chronic exposures is also expected to be linear since ‘a chronic exposure can be thought of as a sequence of very small fractionated exposures’ (EPA Environmental Protection Agency 2011). According to this argument, the slope of the dose response for chronic exposures would be the same as slope for very low dose acute exposures.

However, results from relatively recent research in radiobiology have fueled controversy about the validity of LNTH. The existence of phenomena such as the adaptive response, genomic instability, and bystander effects indicate that ‘radiation may induce DNA damage indirectly, by affecting non-targeted cells and that the processing of DNA damage by cells may be strongly dependent on dose, even at low doses’ (EPA Environmental Protection Agency 2011). If such phenomena result in a markedly different slope of the dose response at very low doses than the estimated slope based on LNT (e.g. the linear component of the estimated LQ dose response function), low dose radiation risk estimates based on LNT would be invalid. Furthermore, power calculations, such as those by Dr Land, suggest data from the LSS cannot be used to test for meaningful deviations from the LNTH. For that, different sources of radiobiological or epidemiological data—as discussed in the next section—are needed.

A somewhat more tractable problem is that deviations from model assumptions about the shape of the dose response at higher doses (e.g. 0.2–2.0 Gy) can also result in bias in estimates of low dose cancer effects. For example, if analysts decide data is insufficient to include a term to account for cell killing effects at higher doses, upward curvature in the dose response might not be detected, a purely linear dose response model might ultimately be fit, and an overestimate of the low dose risks could result. On the other hand, neglecting effects from cell killing could, instead, result in underestimation of risk (e.g. if the ‘true’ dose response up to about 1 Gy is approximately linear and cell killing effectively flattens out the fitted dose response). An evidently more serious problem is that the shape of the dose response might depend on cancer type (e.g. for a wide range of doses, the dose response might be essentially LQ for some cancers but linear for other cancers). In contrast to potential deviations from the LNTH at very low doses, LSS data provides information for evaluating dose response assumptions at higher doses. However, even in large studies such as the LSS, statistical power is often not sufficient to detect small to moderate deviations from dose response assumptions for specific cancers. Inevitably, there will also be model misspecification of effect modification associated with gender, temporal, and lifestyle factors, which further complicates matters.

Results from the recent analysis of LSS solid cancer incidence data illustrate the potential for bias associated with the shape of the dose response (Grant et al 2017). Grant et al found that the shape of the estimated dose response was significantly different among males and females. For females, the results suggest an essentially linear dose response over the full dose range, whereas, for males there was significant upward curvature over the same range. At low doses the estimated ERR for females at 0.1 Gy was 0.064 with 95% CI: (0.052–0.077), but for males, the estimated ERR of just 0.01 (based on the fitted LQ dose response) was not significantly different from 0. Estimates of EAR at low doses were also substantially larger for females. These results were based on the pooling of data over a variety of different cancer types. Questions remain as to the extent to which excess risks truly are greater for females than for males. The extent to which the results reflect the dependence of the dose response shape on cancer type and differing distributions of cancers for males and females is yet undetermined, and the extent to which model misspecification may be a contributing factor is unknown. Analyses investigating these issues are ongoing.

3. Low dose and dose rate epidemiological studies

In part because of technological advances such as the implementation of computerised medical record keeping, it has become possible to conduct larger studies than considered feasible by Dr Land in 1980. Relatively recently, very large studies have resulted in reasonably precise estimates of leukemia risk for doses well below 0.1 Gy (e.g. the studies of cancer effects from childhood computerised tomography (CT) scans (Pearce et al 2012, de Basea et al 2015) and from childhood exposure to natural background radiation (Demoury et al 2016, Kendall et al 2013)). There have also been studies which include moderate-to-large cumulative doses, and have the statistical power to examine risks of solid cancer and leukemia from low dose rate exposures. These include studies of health effects from long-term occupational and environmental exposures associated with activities of the Mayak Production Association (MPA), a nuclear weapons production complex established in 1948 in the former Soviet Union. One study of occupational exposures is based on the follow-up of Mayak workers for whom doses from both external gamma radiation and internal alpha radiation were often substantial (Sokolnikov et al 2015). Additionally, a cohort was established of residents of villages next to the Techa River, which was heavily contaminated through a series of accidents and the dumping of radioactive waste from the MPA into the river from 1949–1956 (Kossenko et al 2005). Information also can be combined from several studies to obtain enough statistical power to detect risks at low doses. For example, the ‘INWORKS’ study pooled data from studies of nuclear workers in the US, UK, and France, and found an approximate linear positive dose response for cancer mortality from relatively small, chronic occupational exposures (Leuraud et al 2015, Richardson et al 2015).

These types of studies often provide direct evidence of risk for LD/LDR exposures and estimates of risk that are consistent with estimates derived from the LSS. For example:

• The estimated ERR Gy⁻¹ for leukemia excluding chronic lymphocytic leukemia (CLL) for occupational exposures derived from INWORKS (2.96, 90% CI: 1.17–5.21) was similar to a linear dose-response estimate of ERR at 1 Sv from LSS for males exposed between ages 20 and 60 y (2.63, 90% CI 1.5–4.2) (Leuraud et al 2015). For INWORKS, the mean dose cumulative red marrow dose was 16 mGy and mean annual doses were 1.1 mGy.

• The estimated ERR mGy⁻¹ for leukemia from childhood CT scans in the UK (0.036, 95% CI: 0.005, 0.12) was similar to the estimated ERR mSv⁻¹ from the LSS for ages of exposure 0–15 y (0.045, 95% CI: 0.016, 0.188). Bone marrow doses from CT scans depended on procedure, age, and calendar year; estimated mean doses per scan were <10 mGy (Pearce et al 2012).

• The estimated ERR Gy⁻¹ for solid cancer mortality for occupational exposures derived from INWORKS (0.47, 90% CI: 0.18–0.79) was similar to the estimate of ERR Sv⁻¹ from LSS for males exposed between ages 20 and 60 (0.32, 90% CI: 0.01–0.50) (Richardson et al 2015). The mean estimated colon dose among exposed workers was about 21 mGy, but for some workers colon doses were greater than 500 mGy. However, the estimated ERR Gy⁻¹ for the dose range of 0–100 mGy was similar to that for the entire dose range.

• The estimated ERR Gy⁻¹ for solid cancer incidence for environmental exposures derived from the Techa River cohort (0.077 Gy⁻¹, 90% CI: 0.013–0.15) was similar the LSS estimate of ERR Sv⁻¹ = 0.05 averaged over sex for age 70 and age-at-exposure 30 (Davis et al 2015).

These results raise important questions:

• Do sources of non-sampling error (i.e. sources other than those associated with sampling) in these studies invalidate findings of a significant dose-response for LD/LDR exposures?

• Can information from some of these studies provide reasonable alternatives to the LNT approach for estimating radiation risk?

3.1. Non-sampling errors in epidemiological studies

In addition to the type of model misspecification discussed in the previous section on LNT models, there are other sources of error that can lead to bias in estimates of radiation risk. Two of the most important potential sources of bias are: (1) errors in estimates of organ or tissue doses, and (2) confounding.

3.1.1. Dosimetric errors.

In the LSS, dose estimates for individual survivors are based on a sophisticated dosimetry system with many individual inputs. Among these are determinations of: (1) source terms of the bombs and (2) survivor location and orientation relative to the explosions at the time of the bombings based, in part, on information gathered from survivor interviews. Dose errors associated with these inputs have been classified as predominantly systematic (e.g. those relating to yields, neutron outputs, burst heights) or random (e.g. those related to survivor location). Statisticians have developed techniques to ascertain the effect of dose errors on estimates of excess risk. Some types of random and systematic dosimetric error can lead to bias in estimates of excess risk. Effects of dosimetric error on risk estimates are often complex (e.g. some types of random and systematic dosimetric error can lead to bias in estimates of excess risk, but would have no effect on the validity of a significant finding that there is an excess risk from radiation.)

In the LSS, effects of dose errors on estimates of risk are relatively well characterised compared to many of the LD/LDR studies. For the INWORKS study, although workers wore dosimeters, ‘the dosimetry is complex and presents a technical challenge to researchers because of the long time-period covered, changes in dosimetry technology, evolution of administrative exposure and recording policies, mixed radiation fields including varying gamma/x-ray energies, internal exposure, neutron exposure, and missed dose’ (NCRP 2018). For the UK study of childhood CT scans, dosimetric uncertainties and their effect on risk estimates has been cited as a concern (NCRP 2018), since, for example, individual doses were not available. Doses for individual children had to be constructed based on information on age, sex, type, and year of scan, typical machine settings obtained from surveys, and computations based on human phantoms. Extremely complex dose reconstructions have also been required for the Techa River cohort study, for which doses were received from internal exposures via consumption of contaminated water, milk, and food, and external exposures predominantly from gamma rays originating from the river shoreline and flood plain. For the Techa River cohort, there has been a long and extensive effort to improve dosimetry and to evaluate the effect dosimetric errors may have on estimates of risk.

3.1.2. Confounding.

Confounding occurs when the observed relationship between exposure and disease differs from the true relationship because of the influence of a third variable (confounder). For a variable to be a confounder, it must be: (1) a risk factor for the disease, and (2) associated with the exposure—but not as a result of the exposure. Examples of confounders include smoking in occupational studies of cancer effects for which smoking is correlated to exposure. Confounding can be controlled for through modelling, if the study provides adequate information (e.g. on the confounder and how it relates to disease and the exposure). The INWORKS study collected data on socioeconomic status to partially control for smoking and other potential risk factors that could be related to worker category (i.e. professional and technical workers, administrative staff, skilled workers, unskilled workers, and uncertain). Effects of confounding can also be evaluated through sensitivity analyses. For the INWORKS analysis of cancer mortality, models were refit after excluding lung cancer, and the authors reported little difference in ERR Gy⁻¹ with the corresponding estimate for all solid cancers. The same type of sensitivity analysis—based on the exclusion of pleural cancers—indicated the bias associated with occupational asbestos exposures would likely be relatively small. The impact of confounding in epidemiological studies of radiogenic leukemia is unclear. For example, the overall impact may be small in studies of childhood leukemia because, other than rare congenital conditions and hereditary diseases, there are very few well-established risk factors other than radiation for leukemia (Belson et al 2006). However, it is difficult to measure the effect of potential environmental risk factors (e.g. parental benzine and insecticide exposure), and these effects would likely depend on leukemia subtype (Whitehead et al 2016).

3.2. Other potential sources of bias

It is beyond the scope of this article to provide a comprehensive discussion of all sources of bias that occur in epidemiological studies (for more information, see Rothman et al 2008, UNSCEAR United Nations Scientific Committee on the Effects of Atomic Radiation 2008, and NCRP 2018). However, two additional potential sources of bias of particularly serious concern in CT scan studies are reverse causation and confounding by indication. At issue are whether CT scans were performed because of a suspected cancer (reverse causation), or whether they were ordered to monitor conditions associated with an increased cancer risk (confounding by indication). Confounding by indication would occur if, for example, children with Downs syndrome—a very strong risk factor for leukemia—were more likely to receive CT scans. To address these issues, investigators have made extensive efforts to collect information on the reasons for the CT examinations and how they may be related to cancer risk factors (e.g. Journy et al 2015, de Basea et al 2015).

3.3. Low dose versus high dose epidemiological studies

Given the very small risks expected at low doses, investigators need to be especially careful in low dose studies to control for potential sources of bias. Bias associated with errors in dosimetry, confounding and other sources of non-sampling error could be comparable in magnitude to the excess risk from low doses of radiation, and thus invalidate findings from low dose studies. As Dr Land noted: ‘Increasing the sample size cannot compensate for such bias, and may in fact add to the difficulties of maintaining strict control over the observations. On the other hand, when the excess risk due to radiation is high [as might be expected in high dose studies], such biases often can be safely ignored’ (Land 1980).

3.4. Other ‘direct’ epidemiological information on low dose and low dose rate risks

Numerous other epidemiological studies, beyond those already mentioned, provide relevant information on radiation risks for LD/LDR exposures. Among these are studies of historical cohorts exposed to medical x-rays that feature patients exposed to moderate to large fractionated doses of radiation or patients (e.g, because of their age) who are considered to be particularly sensitive to radiation. For example, results from the last four publications shown in table 2 provide evidence of an excess cancer risk from doses delivered in small fractions to the breast and thyroid, but not lung cancer. This suggests that there is an excess cancer risk from LD/LDR low-LET exposures, but that the excess risk is strongly dependent on cancer site. The Oxford Survey of Childhood Cancer (OSCC) case-control study of prenatal x-rays found an elevated radiation risks for doses of about 10 mGy for leukemia and all cancers combined. However, in all such studies, particularly those involving low doses (see section 3.3), careful consideration regarding the potential for bias is warranted. For example, questions have been raised as to why the observed relative risk in the OSCC case-control study is about the same for solid tumors as leukemia; the increase may be due to some unaccounted for type of confounding (Boice and Miller 1999).

Table 2.

Selected publications on studies of patients exposed to medical x-rays.

Reference	Study population	Radiation exposure	Excess cancer risk estimate
Bithell and Stewart (1975)	Case-control study from the Oxford Survey of Childhood Cancer: about 8500 children who died from cancer with equal number of matched controls.	Examined effect of diagnostic x-rays in utero. Mean dose of exposed: 6–10 mGy	RR (mothers irradiated versus not irradiated): 1.47 (95% CI:1.34, 1.62)
Ron et at (1989)	10 834 Israeli children (age <16 y) matched to 5392 nonexposed siblings	Treated for tinea capitis. Mean total dose to thyroid: 93 mGy Mean dose per (5-day) course of therapy: ~84 mGy.	ERR cGy−1 for thyroid cancer: 0.3 (95%CI: 0.1, 0.8)
Howe (1995)	64 172 Canadian tuberculosis patients (Miller et at 1989)	Multiple fluoroscopies for monitoring lung collapse from pneumothorax treatment. Mean lung dose: 11 mGy (per treatment) 1.0 Gy (total cumulative)	ERR at 1 Gy = 0.00 95% CI: (−0.06, 0.07)
Howe and McLaughlin (1996)	Canadian tuberculosis patients (Miller et at 1989)	Multiple fluoroscopies for monitoring lung collapse. Mean total breast dose depended on province: 2.1 Gy (Nova Scotia), 0.8 Gy (elsewhere)	ERR Gy−1 depended on province: In Nova Scotia: 3.6 95% CI: (1.9, 6.8) Elsewhere 0.4 95% CI: (0.1, 0.8)
Ronckers et at (2010)	5573 US females (age <20 y) evaluated for scoliosis	Multiple diagnostic examinations. Mean cumulative dose: 109 mGy (to breast tissue) Mean number of exams: 22.9	ERR Gy−1 for breast cancer: 3.9 (95%CI: 1.0, 9.3)

Arguably the most compelling evidence in support of EPA’s application of LNT for evaluating effects from chronic radiation exposures relates to the risk of lung cancer from radon and its progeny in homes. EPA’s risk estimates for radon in homes are based on extrapolation using models derived by the NAS BEIR VI committee from data on under-ground miners. Although the miners were exposed to levels of radon many times larger than those typical of residential exposures (NAS-NRC 1999), EPA’s risk estimates have been validated by results from the pooled analyses of residential case-control studies in North America (Krewski et al 2005), Europe (Darby et al 2006), and China (Lubin et al 2004). These studies provide estimates of excess risk per unit radon concentration that are remarkably consistent with the BEIR VI models. These three pooled analyses comprise the main results on radon identified by the United Nations Scientific Committee on Effects of Atomic Radiation (UNSCEAR United Nations Scientific Committee on the Effects of Atomic Radiation 2008).

4. Challenges associated with low dose and dose rate cancer risks

4.1. Addressing uncertainty

The previous sections describe many of the uncertainties inherent in epidemiological studies of low doses and dose rate radiation risks. Our discussion focused on cancer risks, but many of the same issues are pertinent in studies of radiogenic non-cancer risks (e.g. circulatory diseases and cataracts). Difficulties in interpreting study results can result from biases that are either not identified or not fully accounted for in the study design and analysis.

To date, no reliable biomarkers for radiogenic cancer have been identified to distinguish them from other cancers of the same type. For example, if we were able to tell which lung cancers of a given type are caused by smoking versus radiation exposure, then the uncertainty around the estimate of lung cancer risk per unit dose would significantly narrow.

Much exciting work is ongoing to combine epidemiology with biology, and there is reason to hope that our understanding of human health risks following low dose exposure to radiation may improve as a result. For example:

• Kitahara et al (2015) describe how ‘next generation genomic and epigenetic sequencing of germline and somatic tissues could soon propel our understanding further regarding disease risk thresholds, radiosensitivity of population subgroups and individuals, and the mechanisms of radiation carcinogenesis.’

• An NCRP committee is currently evaluating the potential for applying adverse outcome pathways, an approach used in the field of risk assessment for environmental chemicals, to integrate epidemiological results with those from laboratory animal, molecular, and cellular studies (NCRP National Council on Radiation Protection and Measurements 2015, NCRP 2019).

Important results from several epidemiological and radiobiological studies of low-dose radiation effects have been recently published. Several studies, including the NCRP’s Million-Person study (Boice et al 2019) are underway or are being planned. The EPA encourages and has supported several of these studies which seek to clarify the effects of low-dose exposure to radiation. It is anticipated that many of the practical limitations that Dr Land described must continue to be dealt with in the design, analysis, and interpretion of epidemiological studies.

4.2. Implementation of LNT

The EPA acknowledges that there are numerous issues with implementation of LNT-based dose-response models with respect to risk communication and optimisation. However, there is no consensus on a scientifically-sound and feasible alternative to the LNT model for radiation protection purposes. The following are the main conclusions from NCRP’s comprehensive review of recent epidemiological studies and their implications for the LNT model and radiation protection (NCRP 2018):

‘… based on current epidemiologic data, the LNT model (with the steepness of the dose-response slope perhaps reduced by a DDREF [dose and dose-rate effectiveness] factor) should continue to be utilised for radiation protection purposes’.

‘… no alternative dose-response relationship appears more pragmatic or prudent for radiation protection purposes than the LNT model’.

‘The most recent epidemiologic studies show that the assumption of a dose-threshold model is not a prudent pragmatic choice for radiation protection purposes. The consistency of the better-designed and larger studies with dose-response functions that are essentially linear or LQ, argues for some risk at low doses.’

5. Conclusion

The EPA uses LNT models to interpolate between the absence of radiation cancer effects at zero dose and the observed effects of radiation (mostly at high doses) to estimate risks at doses that represent small increases above natural background radiation. Many scientists agree that the available data best support use of LNT for estimating such risks. The EPA has consistently sought advice and opinions from authoritative bodies such as the NAS BEIR committees, ICRP, NCRP, and UNSCEAR. Based on careful review of the evidence, these authoritative scientific bodies have repeatedly endorsed the use of LNT to estimate and regulate risks.