ETHEL S. GILBERT

Several epidemiologic studies of workers who have been exposed occupationally to low levels of external radiation have been conducted in the United States, Great Britain, and Canada. These studies provide a direct assessment of risk and thus serve as a check on the validity of risk estimates obtained through extrapolation from the Japanese atomic bomb survivor studies. This paper reviews results of the worker studies, giving particular attention to national and international efforts to reduce uncertainties by combining data from several studies. Estimates based on the international analyses, which provide the most comprehensive evaluation of the worker data, are lower than those obtained from linear extrapolation from the atomic bomb survivors, but confidence limits indicate that risks are compatible both with a reduction of risks at low doses and with risks that are about double the linear atomic bomb survivor estimates. Overall, the worker studies confirm the appropriateness of basing radiation protection standards on the atomic bomb survivor data. In addition, the studies illustrate the difficulties in obtaining direct estimates that are sufficiently precise for risk assessment, and thus affirm the continued importance of the atomic bomb survivor data for this purpose.

Although this chapter does not directly address the A-bomb survivor studies, these studies have served as the basis for establishing radiation protection standards for the workers who are the subject of this paper. Because the A-bomb survivor

studies have clearly established that exposure at relatively high doses and dose rates increases the risk of cancer, occupational exposures have been deliberately limited, with the expectation that risks for most workers would be very small and that observable health effects would be unlikely.

Many nuclear facilities began operations in the 1940s and 1950s, and thus the worker populations have matured to a point where the appropriateness of risk estimates based on extrapolation from A-bomb survivors can be evaluated. Obviously there are uncertainties in risk estimates that are extrapolated not only from high to low doses and dose rates, but also from a Japanese population exposed under special circumstances in 1945 to modern-day populations of other races and nationalities. Some have claimed that this extrapolation process has underestimated risks, perhaps by an order of magnitude or more, while others argue that extrapolation from high-dose data has overestimated risks.

This paper reviews the findings of the worker studies, emphasizing results of the recently completed international combined analyses. Special attention is given to what the worker studies tell us about the usefulness of the A-bomb survivor data for estimating risks of exposures at low doses and dose rates.

Several studies of workers in the United States, the United Kingdom, and Canada have addressed the effects of external exposure to radiation (Beral et al., 1988; Douglas et al., 1994; Fraser et al., 1993; Gilbert et al., 1993b; Gribbin et al., 1993; Wiggs et al., 1991; Wiggs et al., 1994; Wilkinson et al., 1987; Wing et al., 1991). The characteristics of the populations included in these studies are given in Table 9.1. Cause-specific mortality was the health endpoint of primary interest in all the studies, although cancer incidence was also evaluated in two of the United Kingdom studies (Douglas et al., 1994; Fraser et al., 1993). Additional information on these studies can be found in the individual papers, in a review by Gilbert (1995), in Cardis et al. (1995b), and in a recent report of the United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR, 1994).

In addition to the studies listed in Table 9.1, workers at the Savannah River Plant have also been studied (Cragle et al., 1988), but results of analyses making use of radiation exposure data have not yet been reported. Studies of workers at the Oak Ridge Y-12 plant (Checkoway et al., 1988), the Linde facility (Dupree et al., 1987), and Pantex (Hadjimichael et al., 1983) are excluded from consideration in this paper because dose-response analyses regarding external radiation exposure were not conducted, either because this was not a major exposure and/or exposure data were inadequate for such analyses.

A major strength of the studies listed in Table 9.1 is that estimates of the whole-body penetrating dose, obtained from personal dosimeters worn by the workers, were available for each worker for each year of employment at the facility of

interest. The information in Table 9.1 is based on workers included in dose-response analyses; in most cases, these analyses were limited to workers who were monitored for external radiation, and at Hanford, they were also restricted to workers employed at least six months at the site. The United Kingdom studies and the Hanford study included both male and female workers, but most of the exposure in all studies was received by males.

The dose distributions in these studies were generally highly skewed to the right, with the majority of the population receiving relatively little exposure. For example, in the recent international study, about 60 percent of all workers monitored had cumulative doses (added up over all years of employment) less than 10 mSv; only 20 percent of monitored workers had accumulated doses of 50 mSv or more, and less than 2 percent (1,752 workers) had accumulated doses of 500 mSv or more (Cardis et al., 1995b). By contrast, risk estimates based on the Life Span Study of A-bomb survivors are strongly dependent on survivors with doses exceeding 1 Sv (1000 mSv).

A first step in analyzing data from the worker studies has been to compare death rates of the study population to those of the general population using national vital statistics. For all the studies listed in Table 9.1 death rates were lower than those of the general population, in most cases substantially lower. With the exception of Sellafield, the all-cause standardized mortality ratios (SMRs) ranged from 0.62 to 0.82, and the all-cancer SMRs ranged from 0.71 to 0.86; in all cases (except Sellafield), these SMRs differed significantly from unity. For Sellafield, both the all-cause and all-cancer SMRs were 0.96, and did not differ significantly from unity.

The low SMRs exhibited in most of the nuclear worker studies are typical of groups employed in industries that are free from serious hazard and reflect the "healthy worker effect" (McMichael, 1976), a phenomenon that is not fully understood but probably results at least in part because those too ill to be employed have been selected out of the population. It is not clear why Sellafield differs from the other studies in not exhibiting such an effect.

Factors related to the healthy worker effect have made SMRs difficult to interpret. Because of these difficulties, and also because overall SMRs that do not specifically examine more highly exposed workers may be greatly diluted, this approach is not ideal for examining the effects of occupational radiation exposure. For this latter purpose, dose-response analyses, which make use of available data on radiation

dose, provide a less biased and more sensitive method of analysis. With this approach, death rates are compared by level of radiation exposure, and, in a sense, the very large number of workers with very little exposure serve as the control group.

The statistical methods used for dose-response analyses in the worker studies are similar to those that have been used to analyze data from the A-bomb survivor studies. In both the worker and A-bomb survivor studies, observed and internally based expected deaths by dose category have been calculated, trend tests that are sensitive to an increase in death rates with increasing radiation dose have been conducted for many different disease categories, and estimates that express the risk per unit of exposure have been obtained. A difference in the worker data and the A-bomb survivor data is that dose is protracted over time in the worker studies. This complicates analyses in that it is necessary to account for changes in cumulative dose as workers are followed over time, and makes it especially difficult to examine time-related factors such as the modifying effects of time since exposure, age at risk, and age at exposure.

Because an objective of the worker studies is to compare risk estimates with those that have been obtained through extrapolation from the A-bomb survivor data, it is important to use comparable models for estimating risks. Recent analyses of the A-bomb survivor data, especially analyses used to derive the risk models used by BEIR V (NRC, 1990) and UNSCEAR (UNSCEAR, 1988), have estimated the excess relative risk (ERR) based on a model in which the relative risk is given by 1 + βz, where z is the dose and β is the ERR.

In analyses of the A-bomb survivor data, modification of the ERR by sex, time since exposure, age at exposure, and age at risk has been evaluated. Although some attempts have been made to evaluate these factors with the worker data, power was usually insufficient to distinguish among models. For this reason, analyses of worker data have drawn on findings from the A-bomb survivor studies for treatment of these factors. For example, doses in most analyses of worker data have been lagged for 2 years for leukemia and 10 years in analyses of other types of cancer because these were the choices used in recent risk assessments based on the A-bomb survivors. Also, most worker analyses have been based on a model in which the ERR remains constant as workers age, again an assumption that is reasonably consistent with A-bomb survivor data.

Dose-response analyses have been conducted for all studies listed in Table 9.1, and have generally included analyses of all cancers (or all cancers excluding leukemia), leukemia (usually excluding chronic lymphocytic leukemia), and several other specific types of cancer. Most studies found no evidence of a statistically significant positive association of all-cancer mortality and radiation dose, but such associations were found for both Oak Ridge National Laboratory (ORNL) workers and for

United Kingdom Atomic Weapons Establishment (UKAWE) workers. Leukemia exhibited a statistically significant positive correlation in Mound Laboratory workers, Sellafield workers, and Atomic Energy of Canada Limited (AECL) workers.

A few other diseases showed statistically significant associations in single studies, but these can reasonably be explained as chance findings given the number of statistical tests conducted. Diseases showing statistically significant associations in more than one study include Hodgkin's disease in Hanford and Los Alamos National Laboratory (LANL) workers, lung cancer in ORNL and UKAWE workers, and cancer of the esophagus in ORNL (Gilbert et al., 1993a) and LANL workers. Earlier analyses of both Hanford and Sellafield workers had indicated a statistically significant association with multiple myeloma, but the association was no longer significant in the most recent analyses of data from either study.

Table 9.2 shows estimates of the ERR for all cancers excluding leukemia and for leukemia excluding chronic lymphocytic leukemia (CLL) for those studies where such estimates were calculated. These estimates were those presented in the papers on the individual studies listed in Table 9.1. For comparison, it is noted that linear estimates for male A-bomb survivors exposed between the ages of 20 and 60 were 0.18 per Sv for all cancers excluding leukemia and 3.7 per Sv for leukemia excluding CLL; these estimates were calculated at IARC as described in Cardis et al. (1995b). For leukemia excluding CLL, a linear estimate from a linear-quadratic dose-response function of 1.4 per Sv was also calculated.

For all cancers, the confidence limits from most studies included the possibility of zero or no risk, and also included estimates several times the linear estimates from male A-bomb survivors exposed as adults. For two of the smaller studies, ORNL and UKAWE, statistically significant correlations were identified, and the confidence limits excluded the A-bomb survivor-based estimates. The UKAWE correlation resulted primarily from a correlation for lung cancer, and nonmalignant respiratory disease showed a similar correlation; thus, smoking-related bias may explain this finding.

The ORNL estimates shown in Table 9.1 are those calculated by Wing et al. (1991). Gilbert et al. (1993a) also analyzed the ORNL data and obtained a smaller all-cancer ERR with confidence limits that included zero. The difference in results can be attributed to several differences in the statistical methods employed, particularly the choice of dose categories used in the analyses; a detailed discussion of these differences is found in Gilbert et al. (1993a). Also, the ORNL correlation was strongly influenced by deaths from cancers of the esophagus and larynx; both diseases have been linked with smoking (Doll and Peto, 1981).

As noted above, leukemia showed a significant positive correlation with radiation dose in Sellafield, Mound, and AECL workers. It can be seen that the confidence limits for all studies in Table 9.2, including both Sellafield and AECL, included both the linear and linear-quadratic extrapolations from the A-bomb survivor estimates. Although no risk estimates were presented for Mound workers, the correlation for

this study resulted from two deaths with doses exceeding 50 mSv, one of which was CLL.

A broad assessment based on the totality of evidence from all studies is obviously needed, and, for this reason, analyses based on combined data from several studies have been carried out at both national and international levels. A major objective of these combined analyses is to obtain more precise estimates of risk than those obtained from individual studies, which have generally exhibited very wide confidence limits (Table 9.2).

In the United States, analyses of data on workers at Hanford, ORNL, and Rocky Flats were published in 1989 (Gilbert et al., 1989) and later updated (Gilbert et al., 1993b). Data on Mound and LANL workers were not available at the time these combined analyses were conducted. In the United Kingdom, results of combined analyses of the three facilities listed in Table 9.1 were published in 1994 (Carpenter et al., 1994), and these analyses are subsequently referred to as "combined UK analyses."

In addition, a study of National Registry of Radiation Workers (NRRW) was published in 1992 (Kendall et al., 1992). This is in a sense a combined analyses in that it includes workers in the facilities listed in Table 9.1, as well as workers in several smaller facilities. The total number of workers in the NRRW study was more than double the number of workers in the combined UK study. However, because many of the additional workers in the NRRW study were recent entrants to the industry, and because of the longer period of follow-up for some workers in the combined UK study, the number of deaths in the two studies was very similar. Results from the two studies were also very similar, and subsequently only the combined UK analyses as reported by Carpenter et al. (1994) are considered in this paper.

International combined analyses have also been carried out, which included the studies listed in Table 9.1 with the exception of Mound and LANL (data from Mound and LANL were not available at the time the international analyses were initiated). This effort was coordinated by the International Agency for Research on Cancer (IARC), but all investigators for the individual studies participated in the planning and interpretation of the data. Comparability of dosimetry for different studies and in different time periods received special attention from a dosimetry subcommittee comprised of dosimetry experts from the three countries. The main findings from the international analyses were published in 1994 (IARC, 1994), and a more detailed description of these results was published recently (Cardis et al., 1995b).

Table 9.3 shows characteristics of the study populations included in the national and international combined analyses; the single Canadian study is also shown for completeness. The numbers from the three countries do not add up exactly because criteria for inclusion of subjects were not identical. For example, international analyses included only subjects that had been employed at least six months at the facility, while the UK and Canadian studies did not have such a restriction.

Table 9.4 shows estimates and confidence limits for the ERR based on the national and international analyses. For comparison, estimates for male A-bomb survivors exposed between the ages of 20 and 60 are also shown; as noted above, these estimates were calculated at IARC as described in Cardis et al. (1995b).

For both the national and international combined analyses, the estimates for all cancers excluding leukemia were very close to zero. However, the upper confidence limit based on the international analyses was lower than those for any of the individual studies. That is, in this case, the objective of obtaining more precise estimates and tighter confidence limits was achieved. These confidence limits included both negative values and values larger than the linear A-bomb survivor-based estimate.

For leukemia excluding CLL, the United States estimate was negative, but both the United Kingdom and Canadian studies showed significant positive correlations. Given the wide confidence limits, the findings were not necessarily incompatible. A test for homogeneity of the leukemia ERR estimates among the individual facilities included in the international analyses yielded a p-value of 0.08; such tests, however, have limited statistical power.

The estimate for leukemia excluding CLL, based on the international data, was intermediate between the combined US and combined UK estimates, with confidence limits extending from very close to zero to about twice the linear A-bomb survivor-based estimates. The leukemia ERR estimate differed significantly from zero, with a one-tailed p-value of almost exactly 0.05. The result was strongly driven by six leukemia deaths with cumulative doses exceeding 400 mSv, compared with 2.3 deaths expected in this category; four of these deaths occurred in Sellafield workers. Statistical significance was not achieved when analyses were restricted to doses below 400 mSv.

The lengths of the confidence limits from the A-bomb survivors may be contrasted with those from the international worker study. Risks have been much more

clearly demonstrated in the A-bomb survivor study, and the A-bomb survivor estimates are more precise than those from the workers. Also, A-bomb survivor estimates based on all the data including females and those exposed in childhood are of course more precise that those shown in Table 9.4. Nevertheless, the precision of the worker-based estimates is beginning to approach that for the A-bomb survivors. Of course, uncertainties resulting from dosimetry and confounding are not included in these confidence limits, and the latter are likely to be more severe for low-dose data than for high-dose data.

Both national and international analyses included statistical tests for the association between radiation exposure and development of cancer. In the US combined analyses, these tests yielded statistically significant results (one-tailed p<0.05) for cancers of the esophagus, cancer of the larynx, and Hodgkin's disease. In the

UK combined analyses, melanoma and other skin cancers, and ill-defined and secondary cancers showed statistically significant associations with radiation dose. For none of these disease categories is there any a priori reason based on results from high-dose studies to expect an especially strong correlation with radiation dose.

In the international analyses, 30 specific cancer types (subcategories of the all-cancer-excluding-leukemia category) were tested for an association with radiation dose. Of these, multiple myeloma was the only type of cancer to exhibit a statistically significant result, with a p-value of 0.037. With the standard ten-year lag period, this disease had not exhibited a significant correlation with radiation dose in any of the national analyses, but had shown a significant correlation in earlier combined US analyses (Gilbert et al., 1989) and in the combined UK study with a 20-year lag. Given the number of tests conducted, obtaining such a correlation for a single type of cancer is not inconsistent with chance. Recent analyses of cancer incidence data on A-bomb survivors have not shown a statistically significant correlation of multiple myeloma and radiation dose. Thus, there is no a prior reason to expect a correlation for multiple myeloma that is stronger than that expected for other types of cancer. Further discussion of the multiple myeloma correlation is found in Cardis et al. (1995b) and in Gilbert et al. (1993b).

The confidence limits presented above include only uncertainty from sampling variation. Additional uncertainty derives because of potential bias resulting from confounding and from the fact that dose estimates are subject to errors.

The national and international combined analyses, and analyses of data from individual studies, were adjusted for age, calendar year period, sex, and sometimes additional variables. This was usually accomplished through stratification on these variables.

For the international analyses, a special effort was made to develop a measure of socioeconomic status for each contributing study, although the measure differed from study to study depending on the available data. It was hoped that this socioeconomic measure might serve as a surrogate measure for smoking, diet, and other lifestyle factors that might confound analyses. For the Hanford study, job category data were used to assign socioeconomic status in a manner that was reasonably comparable to that used to assigned social class for the UKAEA and UKAWE studies. For ORNL, socioeconomic status was determined by paycode (monthly, weekly, hourly); for Rocky Flats by education; and for Sellafield by industrial or non-industrial job category. For AECL, no measure of socioeconomic status was

available. In addition, the international analyses explored the role of variables such as length of employment and time since initial employment, although the final analyses did not include adjustment for these variables.

In the international analyses, supplementary analyses were performed with various alternative treatments of confounding factors. The overall conclusion—that results were consistent with a range of estimates from zero to estimates a few times larger than linear A-bomb survivor estimates—was unaltered by these alternative analyses; however, the precise quantification of confidence limits was affected. For example, adding adjustment for duration of employment increased the upper confidence limit on the all-cancer-excluding-leukemia estimate for 0.3 to 0.5 per Sv and led to a negative lower confidence limit for leukemia excluding CLL.

Unfortunately, adequate data on every factor that might influence cancer mortality analyses were not available. Thus, results might be biased, positively or negatively, because of inadequately measured or unidentified confounders. Bias is a concern in all epidemiologic studies, but is less important in studies in which exposures of interest increase the risk of disease by a factor of two or more. In nuclear-worker studies, where for most workers and diseases the increase is much less than a factor of two, biases may be larger than the expected effects, and this must be kept in mind in interpreting results of these studies.

A major strength of the worker studies is the availability of objective quantitative measurements of exposure, and the quality of these measurements greatly exceeds exposure data usually available in occupational studies. Nevertheless, the recorded estimates obtained from personal dosimeters are subject to several sources of potential bias and uncertainty.

For the purpose of comparison of worker-based estimates with those obtained from the A-bomb survivor data, estimates should ideally be expressed in terms of absorbed dose to various organs. The objective of current dosimetry systems, however, is to estimate deep dose (energy absorbed at a depth of 1 cm in tissue), and in earlier years the objective dose may not have been as clearly defined. The bias in recorded dose as an estimate of bone marrow dose or doses to other organs depends on the dosimetry system in use, the energy of the radiation, and the geometry (direction from which the radiation was received). A difficulty in estimating the bias is that the energy and geometry are not known for any individual exposure, and even average values for a facility are difficult to determine.

At Hanford, extensive efforts have been made to document historical dosimetry practices (Wilson et al., 1990) and, recently, to estimate the bias in recorded doses as estimates of bone marrow dose and lung dose (as a typical internal organ for use in evaluating risks from all cancer excluding leukemia), and also the uncertainty associated with this estimated bias (Gilbert et al., 1996). This latter effort estimated that the recorded dose overestimates bone marrow dose by about 50%, and overestimates lung dose by about 10%, but with large uncertainty for both factors.

In addition, a variety of sensitivity analyses have been carried out on the Hanford data, addressing potential dosimetry biases for several different sources (Gilbert and Fix, 1995a). Overall, the results of these sensitivity analyses were reassuring and did not greatly modify overall conclusions of the study. Nevertheless, as was the case with different approaches to adjusting for confounders, the precise quantification of confidence limits was affected.

As part of the international analyses, dosimetry experts from each of the three countries met several times to evaluate the comparability of recorded doses from different time periods and from different facilities (Cardis and Esteve, 1991). For the most part, dose estimates were judged to be reasonably comparable and reliable. However, it was concluded that doses from low-energy photons and from neutrons were not reliably measured, especially in earlier time periods. Also, external dose estimates do not reflect doses that might have been received from internal depositions. Fortunately, for most workers, most exposure is from high-energy photons, and such exposure was judged to be reliably estimated. Some efforts were made in the international analyses to identify workers with potential dosimetry problems and to conduct supplementary analyses with these workers excluded; these analyses did not greatly modify the estimated ERR, but did increase the upper confidence limits, primarily because of the smaller sample size. The dosimetry experts also estimated that recorded dose would overestimate bone marrow dose by about 20%; however, this evaluation was not as detailed or extensive as that noted above for the Hanford study.

The radiation worker studies are probably the most informative of the various low-dose studies that have been conducted. These studies have a wide range of doses that have been objectively measured through the use of personal dosimeters. The recent international analyses included nearly 100,000 workers and a collective dose of about 3800 Sv. What do these studies tell us about the importance of the A-bomb survivor data for estimating risks of radiation exposures at low doses and dose rates?

First, the worker studies have clearly demonstrated the continued need for the use of high-dose data to obtain estimates that are sufficiently precise for risk assessment purposes. Although combining data from studies of workers in three countries provides more precise risk estimates than individual studies, the confidence intervals for these combined estimates were still large. For cancer excluding leukemia, the lower confidence limits were negative, while for leukemia excluding CLL, the lower bound was very close to zero; for both disease categories, the upper confidence limits were about twice as high as estimates obtained from linear extrapolation from the A-bomb survivor data.

The precision of estimates obtained from worker studies can be expected to improve in the future. Only about 15% of the workers in the international study

were dead by the end of the follow-up period included in this study, and, thus, there is a great deal more that can be learned from these data. Also, a collaborative study of additional nuclear workers in several countries is being conducted, with IARC serving as the coordinating agency (Cardis and Esteve, 1992). To realize the full potential of these new studies, these workers will need to be followed for many years, but they may eventually help us to overcome some of the sample size problems, and thus reduce the statistical uncertainty in risk estimates. Uncertainty resulting from confounding is not so easily overcome and is especially troubling in low-dose studies. Although the worker studies can contribute important supplemental information, it seems likely that the A-bomb survivor studies will continue to be the main source of information on low-level radiation exposure.

A second way that the worker studies have demonstrated the usefulness of the A-bomb survivor data is by confirming that risk estimates obtained through extrapolation are reasonably appropriate. The direct study of workers, especially the combined international analysis, has shown that it is unlikely that linear extrapolation has seriously underestimated risks. This is important because claims have been made by some that risks might be underestimated by an order of magnitude or more (Stewart and Kneale, 1990). The upper confidence limits from the international analyses fairly effectively rule out this possibility.

There is of course interest in whether risks at low doses and dose rates might be less than those predicted by extrapolation from the A-bomb survivor data, and whether or not there might be a threshold below which there is no risk. the statistically significant correlation of leukemia risk and radiation dose observed in the international study might be interpreted as providing direct evidence of risk at low doses and low dose rates. However, the fact that the result is just barely significant (one-tailed p = 0.046) and the possibility of bias resulting from either confounding or dosimetry mean that this conclusion should be drawn somewhat tentatively.

Unfortunately, the worker studies are uninformative on the issue of whether extrapolation might have overestimated risks. Although worker results are compatible with reduced risks at low doses, they are also compatible with risks that are higher than those predicted by high-dose studies. Sampling uncertainty alone makes it extremely difficult to distinguish between no risk and the very small risks that would be predicted in workers if linear extrapolation were correct. Additional uncertainty from confounding and dosimetry add to the difficulty.

To conclude, results of studies of workers thus far demonstrate both the importance and appropriateness of the A-bomb survivor data for the purpose of estimating risks of radiation exposure at low doses and dose rates. Worker studies have already provided useful upper confidence limits on risks. Although there will always be limitations to the worker data, continued follow-up, new studies of nuclear power workers, and combined analyses should lead to more precise estimates of risk based on a direct assessment at low doses and dose rates, and thus provide important information supplemental to that obtained from the A-bomb survivors.