Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.
OCR for page 62
7 Epidemiologic Considerations EPIDEMIOLOGIC STUDIES OF populations potentially exposed to ionizing radiation as a result of the release of radionuclides into the environment can be of two general forms, monitoring or formal, and serve several possible purposes. Monitoring studies inform members of potentially affected population groups of the nature and magnitude of the risks that might have been imposed on them. Monitoring studies also can guide people who are responsible for the facilities to identify measures that must be taken to minimize any future risk to surrounding populations. Formal epidemiologic studies can increase scientific knowledge about the quantitative risk that attends exposure. Although, in principle, much could be learned from experience, for example, with the accident at Chernobyl in the Ukraine, it must be emphasized that deriving new scientific knowledge from studies of exposed populations around nuclear facilities is difficult. Radiation doses received by the latter individuals probably cannot be quantified precisely and the number of radiation-induced illnesses generally is smaller than the number one would expect to find from all causes in that population. Carefully done epidemiologic studies can inform those who might have been affected about the approximate magnitude of the risk, which diseases (chiefly, specific forms of cancer) they and their physicians should anticipate, and what amount of health monitoring is appropriate. The design of a study and the data needed will depend on the study's aims. Dosimetric data are essential for any epidemiologic study, but the detail and accuracy needed depend on the purposes to be served. If the
OCR for page 63
need is for a monitoring or scoping study, then general information about doses will suffice; a study that is expected to contribute to scientific information about quantitative radiation risk requires careful individual dose estimates. Just as it would be inappropriate to begin an epidemiologic study without knowledge of the radiation doses involved, so it is ineffective to begin a detailed study of the exposures and doses before knowing what kind of epidemiologic study will be undertaken. Each kind of study will affect the other. Comprehensive dosimetric studies should not be undertaken before the full epidemiologic study is designed unless, of course, it is found in a preliminary dose assessment that exposures are too low or the potentially affected population is too small to make an epidemiologic study statistically worthwhile. As a rule, dosimetric and epidemiologic scoping (screening) studies should be undertaken in parallel, so that, at the conclusion of the dosimetric scoping study, an informed estimate of the expected magnitude of risk and the statistical power of a potential epidemiologic study can be derived. This chapter considers various aspects of epidemiologic studies in the context of dose reconstruction, including strengths and limitations, interactions between epidemiology and dose reconstruction, and study design and methods. QUANTITATIVE RISK ASSESSMENT: STRENGTHS AND LIMITATIONS OF EPIDEMIOLOGIC STUDIES Knowledge of the biomedical effects of exposure to ionizing radiation has expanded enormously since the end of World War II. The research has ranged from the exploration of basic cellular mechanisms to extensive epidemiologic studies of exposed populations, including those exposed occupationally (Kendall and others 1992, Gilbert and others 1993), those exposed to therapeutic and diagnostic medical sources (Boice and Land 1982, NRC 1990), and those exposed to nuclear weapons. Prominent among the last group are the studies of the survivors of the atomic bombing of Japan (Shimizu and others 1990), including those summarized in a special supplement in Radiation Research (Mabuchi and others 1994, Thompson and others 1994, Preston and others 1994, Ron and others 1994). As a consequence of the availability of this large body of knowledge and the need to set radiation standards for public and occupational exposures, radiation protection groups have continuously reviewed the literature to estimate the risk that results from radiation exposure. The greatest interest has been in the risk associated with exposure at low doses and dose rates. The most recent efforts to provide risk estimates are described
OCR for page 64
in reports of the International Commission on Radiological Protection (ICRP, 1991), the National Research Council's Committee on the Biologic Effects of Ionizing Radiation (NRC 1990), and the United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR 1988, 1993, 1994). Because estimates of risk that are based on direct study of persons exposed at low doses have often been too imprecise to be useful, risk estimates have been obtained primarily through extrapolation of data from studies of persons exposed to moderate or high doses and at high dose rates. Most scientists are reasonably confident—because of experimental evidence—that linear extrapolation is not likely to underestimate risk but there is appreciable uncertainty about the risk that results from exposure at low doses and dose rates (NCRP 1980). Because the Japanese atomic bomb survivors have been so important in radiation risk assessment, uncertainty about extrapolation from the Japanese group to another population also must be considered. Despite their limitations, numerical expressions of risk are useful in dose reconstruction and epidemiologic studies because they provide a means to estimate the expected number of health outcomes of a particular kind, such as cancer, and hence to evaluate the probable meaningfulness of an epidemiologic study. Several agencies, including the U.S. Nuclear Regulatory Commission (Gilbert 1991b), ICRP (ICRP 1991, Land and Sinclair 1991), and the U.S. Environmental Protection Agency (US EPA, 1994), have provided estimates of the risk of radiation-related fatal cancers that can be used in scoping or subsequent studies. These estimates are summarized in Table 7–1. The risks are expressed as excess fatal cancers per 104 person-gray for various tissue sites. The differences in risk depend on the different projection models used and differences in the interpretation and adjustment of the basic data derived from studies of the Japanese atomic bomb survivors and other highly exposed populations. It should be noted that the largest differences among risk estimates involve stomach, colon, lung, and breast cancers, and the most discrepant values are usually those of ICRP. These differences are not likely to be important in scoping studies, however, in which the primary objective is to provide a preliminary estimate for use in deciding whether further detailed studies are warranted. As shown in Table 7–1, various interpretations can be made with the basic data on cancer site-specific mortality and the reporting of the estimates with the decimal place given implies a false precision. Since most of the data on the occurrence of excess cancers (with the exception of breast cancer) are derived from high-dose, high-dose-rate studies, the National Research Council and others recommend application of a dose, dose-rate effectiveness factor (DDREF) between 2 and 10. For low-dose,
OCR for page 65
TABLE 7–1 Estimates of the Expected Number of Excess Fatal Cancers in a Lifetime at Specific Sites Among 10,000 Individuals Exposed to 1 Gy or 100 Rad (Males and Females Combined, DDREF = 1) Cancer Site ICRPa NIHb EPAc NRCd Esophagus 16.2 21.3 18.1 14.9 Stomach 29.3 27.4 88.7 74.3 Colon 381.0 109.0 196.4 149.0 Liver 30.0 30.0 30.0 29.7 Lung 265.0 78.7 143.2 149.0 Bone 9.3 9.3 1.9 8.1 Skin 2.0 2.0 2.0 1.8 Breast 116.0 32.7 46.2 46.2 Ovary 47.5 25.0 33.2 32.2 Bladder 64.0 38.9 49.7 49.6 Kidney ]—[e ]—[ 10.9 ]—[ Thyroid 7.5 7.5 6.4 6.4 Leukemia 110.0 97.9 99.1 89.9 Residual 325.0 227.0 246.2 193.0 Total 1403 953 973 844 a International Commission on Radiological Protection (ICRP 1991). b These calculations were made by Land and Sinclair in 1991, but are commonly referred to as the NIH model. c U.S. Environmental Protection Agency (US EPA 1994). d Nuclear Regulatory Commission (Gilbert 1991b) e ]—[ = no estimate. low-dose-rate radiation, if a DDREF of 2 (the most conservative factor) is applied to the data in Table 7–1 for all tissue sites except the breast, the total lifetime risk of a fatal cancer after whole-body irradiation would be 5.1 x 10-2/Gy (5.1 x 10-4/rad). For total body effects based on the atomic bomb survivors, the BEIR V committee determined risk factors of 7.5 x 10-2/Gy (7.5 x 10-4/rad) for females and 7.9 x 10-2/Gy (7.9 x10-4/rad) for males without application of a DDREF (NRC 1990). A range of risk factors can therefore be chosen for computing excess cancer risks from radiation exposures; however, most dose estimates encountered in dose reconstruction studies are usually for specific tissues (thyroid or bone, for example) and at low dose rates for cumulative total doses that would generally be considered as low. The lifetime radiation risk factors for dose reconstruction studies would generally range from 5 to 8 x 10-2/Gy (5 to 8 x 10-4/rad). In evaluating whether a proposed epidemiologic study is likely to increase scientific knowledge of radiation effects, it is important to con-
OCR for page 66
sider the extensive body of scientific data on radiation risks and the models for evaluating risks that have been developed and endorsed by the groups listed in the table. The use of models, together with the use of disease rates obtained from cancer registries and vital statistics, can provide a means of estimating the number of radiation-induced cancers that will occur in a particular population exposed to a particular dose. The age and sex distribution must be specified, and given this information, one can estimate both the number of cancers that could occur as a result of the exposure of interest and the number that would occur in the absence of the exposure. Once these values have been determined for various categories defined by dose and other factors, it is possible to calculate the precision with which a proposed study can estimate risk and the probability that the study can statistically detect an excess of a given magnitude. Usually, if the number of radiation-induced cancers is small both in absolute value and as a proportion of the non-radiation-induced cancers, it will not be possible to determine with confidence whether or not an excess number of cancers has occurred. In addition to assessing the potential of a study to evaluate risks based on the assumption that current models are correct, it also could be of interest to evaluate whether risks that are several times larger than those based on current models could be excluded. A simple example is helpful: If 10,000 persons with an age and sex distribution typical of the United States as a whole are each exposed to 10 mGy (1 rad) and followed over their full lifetimes, use of the BEIR V-recommended risk model (NRC 1990) would lead to an expectation of about 7.5 radiation-induced fatal cancers in the group. One would expect to see about 1,800 fatal cancers in a group that has not been exposed to the additional radiation. The problem for the statistician is to determine whether, given the variability in cancer occurrence, an expected excess of 7 or 8 cancer deaths could be detected with a particular level of confidence. Would the study have the statistical power to detect such an excess? If the answer is yes, then the study might reasonably proceed, although other issues such as the feasibility of identifying the population and of ascertaining the causes of the deaths also would need to be considered. If the answer is no, then the merit of proceeding with a formal study is doubtful. The principal strengths of the epidemiologic studies associated with dose reconstructions come from the fact that humans are the populations of interest. No extrapolations are required from animal to human, from high-dose studies, or from populations with other ethnic, lifestyle, or sociodemographic characteristics. Such studies avoid many of the uncertainties involved in extrapolations and maximize the potential to address public concerns. They also can provide direct information about sensitive
OCR for page 67
subgroups of the population, and they can offer the chance to incorporate any biologic marker studies that seem warranted. The value of a low-dose study, such as one done near a nuclear site, is increased if it is understood that the study results do not stand in isolation and that the study is performed in the context of adding to the body of information from other radiation studies. A systematic meta-analysis is often useful as a framework for integrating and comparing the results of several such studies (Greenland 1987), although joint analysis of the raw data from several studies is preferable when the data are available and comparable (Gilbert and others 1993). Epidemiologic studies done in conjunction with dose reconstructions near nuclear sites or in the aftermath of accidents are subject to several limitations. The exposed population could be poorly defined, inhomogeneous, and transient. It could be relatively small, so that only large effects could be detected reliably. This problem is exacerbated when the diseases of interest are relatively rare and the risk per unit dose is expected to be small. The exposures could be difficult to evaluate because of uncertainties about exposure amount, duration, and timing. The diseases or medical conditions of interest could have long latencies and might not have been recorded reliably. For some outcomes, public concern could have created dose-dependent variations in disease surveillance, which could bias the study results. There also could be effects of chemical contaminants from the site and from other industrial sites nearby because of the lack of detailed exposure information. Epidemiologic studies also can be confounded by dietary or lifestyle factors (of which smoking habits are particularly important), especially when they deal with low-dose data. EPIDEMIOLOGY AND DOSE RECONSTRUCTION For a dose reconstruction to be useful for epidemiologic purposes, it must be designed to allow for the calculation of annual organ doses. Furthermore, an epidemiologic investigation often involves tradeoffs relating to bias and precision in the estimates of exposure and other information that can be crucial not only to the dose reconstruction study but to the epidemiologic one as well. It follows, then, that the involvement of an epidemiologist in the early phases of the dose reconstruction will ensure attention to the potential for bias in the reconstruction of individual doses. Epidemiologists need to understand the details of dose reconstruction both to collect the appropriate information on dietary and lifestyle factors that could significantly affect dose calculations and to evaluate in the epidemiologic analysis various uncertainties inherent in dose estimation. Epidemiologic evaluation also can contribute to dose reconstruction. In particular, epidemiologic data could be used to validate models and
OCR for page 68
assumptions used in dose reconstruction and to assess the usefulness of different dose estimates, starting from the simplest, such as distance from the site, to the most complex, such as those based on models of environmental pathways and on the use of biologic markers. While in theory epidemiologic observations can contribute to validation of dose estimates, as a practical matter they are not likely to do so except for large populations exposed to high doses. As noted in the Introduction, if a preliminary dose assessment results in estimates of exposures that are potential concern, delaying the epidemiologic investigation until a detailed dose reconstruction is complete could not only diminish the usefulness of the reconstruction but could jeopardize a warranted epidemiologic study. For example, the more time has elapsed from the period of interest the more difficult it usually becomes to assemble the complete population needed and to reconstruct the information required for exposure assessment. Moreover, as previously stated, the likelihood of introducing a variety of biases increases as the publicity surrounding dose reconstruction efforts broadens. Finally, negative public perception about delays in directly addressing the issues of concern could create difficulties and require additional efforts that cannot be justified on scientific grounds. STUDY DESIGN In the introduction to this chapter, two types of epidemiologic studies were distinguished: monitoring studies and formal epidemiologic studies. The decision about whether a monitoring study should be done will depend on the amount of information available about the approximate levels of exposure and the amount of concern about possible effects on health. To determine whether a formal epidemiologic study should be undertaken, it is necessary to evaluate the power of the proposed study (which will depend on the population size and the magnitude of the doses) and to evaluate the feasibility of obtaining the necessary data on health effects, exposure, and potential confounders. Making this determination requires consideration of several design issues. Further information on many of these issues can be found in epidemiology textbooks, such as those by Kleinbaum and co-workers (1982) and Hennekens and Buring (1987). Study Types The primary study designs that are likely to be valuable for addressing the effects of environmental exposures in connection with a dose reconstruction are retrospective cohort studies and case-control studies.
OCR for page 69
A cohort study is one in which an exposed population is identified and then followed to determine whether specified health effects develop. In a retrospective cohort study, the study is not begun until some period after the initial exposure, but health effects are still identified over the follow-up period of interest, and the data are analyzed in a manner similar to that used in a prospective study. The observed health effects are then compared with health effects expected based on an appropriate control population or related to variations in estimated doses. In a case-control study, subjects with the disease of interest are identified, and controls are selected, usually matched to cases on the basis of age, gender, and other factors. Estimates of past exposures for cases and controls are then compared. Further discussion of these study designs and their relative advantages and disadvantages is found in Hennekens and Buring (1987) and other epidemiology textbooks. The choice of study design will depend on several factors, including the health end point of interest, resources for identifying both the exposed population and the cases, whether the exposed population constitutes a large proportion of the population in its geographic location, and the mobility of the population. For exposures that occurred many years ago in mobile populations, it is expected that the retrospective cohort study will usually be preferred over the case-control study. This is because many exposed subjects might have left the area where the exposure occurred, and it is thus likely to be difficult to find a means of identifying cases among these subjects without first identifying the population at risk. Registries, hospitals, and other potential case sources in the area of interest will fail to identify cases who have left the area, and a large proportion of the reported cases will probably be persons who were not in the area at the time of the exposure. However, conducting a case-control study within a retrospective cohort study often is useful for obtaining more detailed information on nonradiation exposures, lifestyle factors, and the like. Another type of study that might be considered under some circumstances is the correlation or "ecologic" study. In such studies, disease rates are compared for groups of subjects for whom exposures are judged to differ, and the groups are usually defined according to locations for which mortality or morbidity rates are available. Because these studies make use of available statistics, they often can be conducted quickly and inexpensively. Ecologic studies are limited to the use of groups for which disease statistics are available, and this often does not provide the best delineation of exposure, particularly as current geographic location might not reflect past exposure. The use of grouped data can introduce biases especially if the regions chosen reflect social and economic differences; it is often impossible to control for potential confounders in an ecologic
OCR for page 70
study because data are not available or because of limitations in the models that can be used to adjust for grouped data. With data on geographic groups there usually is no assurance that the diseased persons were even exposed. For these reasons, ecologic studies are usually regarded as hypothesis generating, at best, and their results must be regarded as questionable until confirmed with cohort or case-control studies. Problems with the ecologic study design are discussed in more depth by Piantadosi and co-workers (1988), Greenland and Morgenstern (1990), and Greenland (1992). Although it is not a formal epidemiologic study, interest in possible health effects resulting from a particular exposure source is sometimes generated through identification of a cluster of cases of a specific disease in a particular location, time period, or both. Most clusters are chance occurrences, but it can be difficult to evaluate whether a particular cluster can reasonably be attributed to chance and even more difficult to communicate to the public the role of chance or the magnitude of the purported risks (Slovic 1987). The evaluation of whether a cluster represents an excess involves comparison with a control; often, this comparison is made inappropriately if it is made at all. Also, clusters are sometimes reported without adequate verification of disease status. Nevertheless, clusters require investigation, and if the cluster cannot be readily explained as resulting from chance or from a problem in methodology, a cohort or case-control study might be needed (Kheifets 1993). Problems with cluster studies are discussed in more detail by Rothman (1990) and Neutra (1990). Statistical Power The potential informativeness of a study often is measured in terms of statistical power, which can roughly be defined as the probability of rejecting the hypothesis of no effect (null hypothesis) when in fact it is false and the alternative hypothesis is correct. An example would be that one would conclude that there is no increased rate of disease in the exposed group when, in fact, the radiation exposure does have an effect of the expected magnitude. For the studies of interest here, there are two aims that are of primary interest. The first pertains to the probability of detecting a dose-related effect if an effect is present, given the expected size of that effect as derived from the risk estimates that form the basis of radiation protection standards (UNSCEAR 1988, 1993, 1994; NRC 1990; ICRP 1991). To gauge the power of a study for increasing knowledge about radiation effects, it is often meaningful to express the expected detectable effect as a multiple of current radiation risk estimates. For addressing the general concerns of the public, it also is useful to state the effect as a
OCR for page 71
simple excess risk, either relative to baseline risk or as an absolute increase in lifetime risk. Second, even when a study does not detect an effect, it can yield valuable information if it establishes an upper bound on risks that will increase confidence in the radiation-standard risk estimates or if it can be used to assure the public that the risks have not been large. Thus, a second null hypothesis of interest is that a health effect is not any greater than a specified size (such as the size expected by extrapolation from high-dose data on which the radiation protection standards are based). In practice, this approach usually yields results similar to those obtained by applying the approach described in the previous paragraph, and the two approaches together can be regarded as distinguishing no effect from some specified large effect. If a study has inadequate power to distinguish no effect from effects that would result if risks were several times those based on high-dose extrapolation, the study usually will be judged unlikely to provide a valid and defensible risk estimate. Greenland (1988) provides a discussion of power calculations for distinguishing among various hypothesis, including their relationship to confidence intervals. To evaluate a study's power, it is necessary to have information on the magnitude of the dose, the size of the population, and the number of cases expected if there were no exposure. The last category requires information on age, sex, length of follow-up, and possibly other factors, and it is determined using available disease rates. At the time the power calculations are made, there is likely to be uncertainty both about the doses and about the size and characteristics of the exposed population. For this reason it is desirable to carry out power calculations based on several alternative assumptions. The power of a study can depend strongly on the quality of the dose estimates. The lowest power will result if only an average dose can be assigned and if analysis consists of a simple comparison of the exposed group with an appropriate control group judged to be unexposed. Because uncertainties in estimated doses are likely to be large, it is probably desirable to include this scenario as one of those evaluated. However, power often will be much greater if the information gained through the dose reconstruction is incorporated by including several exposure categories and testing for an increase in risk with an increase in dose (Shore and others 1992). Thus, in most cases, power calculations based on the use of the available quantitative information on dose also should be made. Ideally, these calculations would account for errors in the dose estimates (and thus misclassifications of dose); if this is not done, random errors in the measurement of doses will reduce power, can lead to an underestimation of any effect, and can introduce a spurious curvilinearity in the apparent dose-response relationship. Other study limitations, such as the
OCR for page 72
failure to secure participation from all potential subjects and the failure to ascertain all cases, must be considered in developing power calculations. Howe and Chiarelli (1988) describe an approach to power calculations that uses information on dose and allows for the possibility of dose mis-classification. The sample sizes required to achieve adequate statistical power increase greatly as the magnitude of the dose decreases. To illustrate the relationship among dose, sample size, and statistical power, the committee calculated the sample sizes needed to detect an excess in total cancer mortality for various possible doses. The same assumptions were used as in the example earlier in this chapter. It is assumed that each member of a population with an age and sex distribution typical of the United States as a whole was exposed to a specific radiation dose and followed over a full lifetime, and the risk model recommended by BEIR V (NRC 1990) was used to project lifetime risk for radiation-induced fatal cancers (about 7.5 per 10 millisieverts (mSv) per 10,000 persons or 7.5 per rem per 10,000 persons) as compared with about 1,800 spontaneous cancer deaths per 10,000 persons. Table 7–2 shows the size of the exposed population needed at various doses to have an 80% chance of seeing an excess in a comparison against general population rates. It is notable that at doses of 20 mSv (2 rem) or less the required sample sizes are prohibitively large, ranging from 500 thousand to 32 million persons. The sample sizes come into the realm of possibility only when the mean dose is above 50 mSv (5 rem). Total cancer is not necessarily the most sensitive health end point or the one of most interest for detecting a radiation-induced excess. Table 7–3 shows the required numbers of exposed persons for two cancer types that could be of concern in dose reconstruction studies: leukemia and respiratory cancer mortality. The assumptions are the same as those used above. The BEIR V (NRC 1990) risk models for respiratory cancer and leukemia were used to generate the expected excess malignancies. The results in Table 7–3 show that the required sample sizes are extremely large if the doses are low. Several caveats should be made regarding these sample size calculations. Normally, one would have a range of doses in the population rathern than a uniform dose. Performing a dose-response analysis would create some gain in statistical power (and a corresponding reduction in the required sample size) relative to the simple comparison of the total exposed group to the general population used above (Shore and others 1992). On the other hand, three factors would tend to diminish statistical power. First, the uncertainty in estimating individual doses tends to diminish statistical power and increase the required sample size (Walker and Blettner 1985, De Klerk and others 1989, Armstrong 1990). Second, the calculations in Tables 7–2 and 7–3 assume a full lifetime follow-up.
OCR for page 73
TABLE 7–2 Size of an Exposed Group Required to Detect an Increase in Total Cancer Mortality with a Lifetime Follow-up, According to Dose Mean Whole-Body Dose (mSv) Excess Cancers per 10,000 Sample Size 2.5 1.9 32,000,000 5.0 3.8 7,900,000 10.0 7.5 2,000,000 20.0 15.0 500,000 30.0 22.5 220,000 40.0 30.0 130,000 50.0 37.5 80,000 60.0 45.0 56,000 70.0 52.5 41,000 80.0 60.0 31,000 90.0 67.5 25,000 100.0 75.0 20,000 120.0 90.0 14,000 150.0 113.0 9,100 200.0 150.0 5,200 NOTE: The calculations are based on the BEIR V (NRC 1990, p. 172) estimate of 7.5 excess cancer deaths per 10 mSv per 10,000 persons (averaged across sex and age). For comparison, the number of ''spontaneous" background cancer deaths that would be expected in the study population per 10,000 persons is about 1,800. This means, for instance, that at 10 and 100 mSv the radiogenic risks are only about 0.4% and 4%, respectively, as great as the spontaneous cancer mortality. BEIR V used Vital Statistics of the United States 1980 for the source of baseline data on cancer mortality. Sample sizes were rounded to two significant digits in view of the approximations involved in their calculation. The table assumes the age and sex-structure and the background cancer rates of the exposed group are comparable to the U.S. general population at the time of irradiation; members of the exposed group are followed up for their remaining lifetimes; the excess cancer risk corresponds to the estimates given by the BEIR V report. The calculations are predicated on achieving 80% statistical power with a 5% alpha level and a one-sided statistical test. Studies will typically have a much shorter average follow-up than this, which will diminish the statistical power considerably (especially because a large fraction of the cancers would occur at older ages). Third, the publishers of radiation risk assessment studies (NCRP 1980, NRC 1990, ICRP 1991) generally agree that gamma or beta irradiation delivered at low doses and low dose rates probably causes from 2 to 10 times fewer cancer cases per millisievert than do higher, acute doses. Hence, for the non-leukemic cancers included in Tables 7–2 and 7–3, the effects could be overestimated, because they are extrapolated from high-dose studies. For leukemia, however, this low-dose effect is already taken into account by
OCR for page 74
TABLE 7–3 The Required Size of an Exposed Group to Detect an Increase in Leukemia or Respiratory Cancer Mortality with a Lifetime Follow-up, According to Dose Levels Mean Target Organ Dose (mSv) Sample Size Leukemia Respiratory Cancer 2.5 74,000,000 >100,000,000 5 19,000,000 44,000,000 10 4,700,000 11,000,000 20 1,200,000 2,700,000 30 520,000 1,200,000 40 300,000 680,000 50 190,000 440,000 60 130,000 310,000 70 99,000 220,000 80 76,000 170,000 90 61,000 140,000 100 49,000 110,000 120 25,000 77,000 150 11,000 50,000 200 3,900 28,000 NOTE: The calculations are based on the BEIR V (NRC 1990, p. 175) estimates of excess cancer deaths per 10 mSv per 10,000 persons (averaged across sexes and age) of 0.95 for leukemia and 1.7 for respiratory cancer. It is assumed that the BEIR V estimate of leukemia risk at 100 mSv extrapolates linearly (i.e., there is no quadratic term) downward to lower doses. This has the effect of increasing the estimated low-dose risk and yielding smaller sample sizes than strict adherence to the model would do. See footnote to Table 7–2 for other assumptions. the linear-quadratic model that was used. Considering all these factors together, the sample sizes in Tables 7–2 and 7–3 probably err on the side of underestimating the required sample sizes. Outcomes The health outcomes to be studied should be chosen in keeping with information available about which health effects are expected derived primarily from other studies of exposed populations. As previously noted there is a considerable body of literature on radiation effects, based on animal and human studies. The radionuclides involved in the exposure and the organs of the body that are likely to be most highly exposed are clearly important determinants of the health outcomes chosen for study. With whole-body exposure, cancers of many types could be possible outcomes; in this case, it is important to select a small number, associated
OCR for page 75
with the most radiosensitive sites, as those of primary interest. Public pressure to study effects that have not been strongly linked with radiation in high-dose studies should be resisted because they are likely to be subject to the same biases as cluster studies (Rothman 1990) and to yield spurious results. Population Identification and Follow-Up In a cohort study, it is necessary to define the population to be studied carefully to develop the exposure data for this population and to determine whether subjects have developed the health effects of interest. If a population was exposed many years ago, and if a substantial proportion of the population has subsequently moved to other locations, identification of the group will be extremely difficult, if not impossible. Methods for ascertaining health effects will depend on the effect being studied. For diseases with high fatality rates, the use of mortality data from the National Death Index, state death records, and Social Security Administration records could be an option. Tumor registries are another potential source, although with mobile populations and no national registry, this is difficult. For some end points, it will be necessary to locate all members of the population to determine whether health effects have occurred; in some cases (for example, in the study of thyroid disease), it could be necessary to conduct physical examinations of the study population to allow reasonably unbiased evaluation of health effects. Before undertaking an epidemiologic study, it is essential to determine the feasibility of identifying the study population and of ascertaining whether the health effects of interest have occurred. These difficulties could preclude a study regardless of the study's potential statistical power. Bias and Confounding Results of an epidemiologic study can be biased for several reasons (bias that results from uncertainties in dose estimates is discussed in Chapter 5 in Uncertainty). For studies conducted in connection with a dose reconstruction, the information needed to estimate dose and the information on confounders often is collected through interviews with subjects or relatives of subjects. The information can be biased both by the magnitude of exposure and by whether the subject has developed the health effect of interest. This could be particularly true if the study has received wide publicity and subjects are aware of the expected health effects and the parameters that affect exposure. It is thus important that questionnaires be designed and administered to minimize bias. Careful attention should be given to the wording of questions and to ensuring, to the extent
OCR for page 76
possible, that interviewers do not know the exposure or disease status of subjects. Bias also can result if case ascertainment differs by magnitude of exposure, and the study must be designed to minimize such bias. This means that case ascertainment efforts and methods cannot depend on exposure, and it generally rules out the use of volunteer subjects, or information volunteered about disease status. Also, it is important to achieve a high rate of participation in the study; incomplete participation increases the likelihood of bias. When the study involves physical examinations, an effort should be made to prevent the examiners from knowing the exposure status of the examinees. Even if results are not biased for the methodological reasons noted above, epidemiologic studies are always subject to the possibility of confounding; that is, bias can result from differences among subjects in risk factors other than the exposure of interest. Data on known risk factors for the disease of interest should be collected if possible and taken into account in the analysis. Even when this is done, the possibility of bias cannot be excluded because all possible biasing factors have not been measured. This is an inherent limitation of observational as opposed to experimental studies. (Randomization in experimental studies ensures that the study groups will be comparable, on average.) In the studies considered here, exposed subjects generally reside in specific geographic areas and can differ in various ways (other than exposure) from subjects in other locations. In this regard, comparisons by magnitude of exposure could be less subject to bias than will be comparison with a control population that resides in a different community. However, there might have been socioeconomic gradients by distance from a plant, for example, that could produce bias in the results. A special concern in dose reconstruction studies is exposure to chemicals that correlate with radiation exposures. Smoking is an example of a particularly important confounder for lung cancer and some other cancers. Because even small differences in smoking habits can have a greater influence on lung cancer risks than does the exposure of interest, it is almost never possible to be certain that one has fully adjusted for smoking even when reasonably detailed smoking histories are available. For this reason, if lung cancer is the health effect of primary interest, it might be necessary to apply stricter criteria in determining whether to conduct a study, and such studies should probably not be conducted if smoking data cannot be obtained. On the other hand, if detailed and reliable smoking data are available, then it should be possible to test if the prevalence of upper respiratory diseases in the cohort under study compared to nonsmokers meets statistical significance. Bias that results from confounding is especially troublesome in study-
OCR for page 77
ing small increases in risk, where the magnitude of the bias could exceed the magnitude of the excess risk of interest. Furthermore, bias does not decrease with increased sample size (Boice and Land 1982, Monson 1990). If the expected magnitude of the excess risk resulting from radiation does not exceed the baseline risk by more than a few percent, the results of an epidemiologic study cannot be interpreted unambiguously and could therefore have little value regardless of the size of the population. Statistical Analysis Statistical analyses should be designed to make optimal use of available information on doses. If doses have been estimated for each study subject individually, this will usually mean that the analysis will be based on several dose categories (or possibly on ungrouped doses) using methods that are sensitive to an increase in risk with an increase in dose. However, because of uncertainty in dose estimates, it also might be desirable to make overall comparisons of the exposed population with an unexposed control population. Analyses should be adjusted for age, sex, and other potential risk factors for which data are available (Gilbert 1982). Preliminary analyses could determine those risk factors that need to be controlled for in the subsequent analyses. Further discussion of statistical methods for cohort studies is given by Breslow and Day (1987). Although in the interest of thoroughness it might be desirable to conduct more than one type of analysis, it is important to specify in advance a "primary" approach that is to be emphasized in reporting and interpreting results. Similarly, if more than one health end point is to be studied, it is important to specify the main effects that are to be considered in interpreting study findings; this specification would generally be based on effects seen in other studies of radiation effects. Results should be presented as estimates of risk, and confidence limits should be used to express uncertainty in the estimates. For the purpose of increasing scientific knowledge about the magnitude of radiation effects, risks need to be expressed per unit of exposure in a form that can be readily compared with estimates from other sources, such as those obtained through extrapolation from high-dose studies. It also is useful to provide estimates of excess risk for subjects in various exposure categories, and possibly for the exposed group as a whole. These risk estimates, which may be particularly helpful for those not familiar with the radiation literature, can be expressed relative to the baseline risk or as increases in absolute risk. In presenting risk estimates and confidence intervals, it is important to emphasize that the conventional confidence intervals do not include uncertainty resulting from bias or uncertainties in dose esti-
OCR for page 78
mates, or from unidentified confounders that have not been measured or used in the analyses. These issues are discussed by Gilbert (1989). Uncertainty and Misclassification Statistical analyses should account for uncertainties in dose estimates (Gilbert 1991a, Clayton 1992). In practice, this is extremely difficult, in part because statistical methods for doing so are complex and often require extensive software development. In addition, it is necessary not only to have information about the uncertainty distributions for doses of individual subjects but also to understand the extent to which these uncertainties are correlated across subjects. Because neither uncertainties nor their correlations are known exactly, it is desirable to conduct dose-response analyses based on several assumptions regarding uncertainties. To accomplish this, it is important that statisticians or epidemiologists work closely with dosimetrists to achieve an understanding of the various uncertainty sources and to help to ensure that uncertainties are expressed in the form needed for use in epidemiologic analyses. Even though it might not be possible to develop a perfect understanding of the correlation structure, it often is possible to separate uncertainty sources that are highly correlated across subjects from uncertainties that are independent across subjects and that primarily reflect subject variability. These two kinds of uncertainty must be treated differently in epidemiologic analyses, and analyses that incorporate highly correlated uncertainties (such as might result from uncertainties in the magnitude of the source term) are generally much simpler to implement than are analyses that address random uncertainties for individual doses. Although analyses that consider dose estimation uncertainties are desirable, the primary analysis to be emphasized in presenting results probably should not be adjusted for such uncertainties; the results adjusted for uncertainties would then be presented as an elaboration of the main results. If an uncertainty-adjusted analysis is to be used as the primary analysis, this should be clearly stated in advance, and prior specification of the exact form of the uncertainty adjustment to be used should be included. SUMMARY AND RECOMMENDATIONS Radiation dose reconstruction provides detailed quantitative information about individual exposures for epidemiologic study of populations near nuclear facilities. The quality and quantity of the dose information are central to any good epidemiologic study. Because different levels of epidemiologic investigation require different amounts of detail and
OCR for page 79
precision in dose estimates and because both the epidemiologic study and the dose reconstruction will require refinements, dosimetry and epidemiologic screening efforts will be most informative if they are done interactively and in parallel. For example, during the scoping phase of the dose reconstruction, epidemiologists could collect information on the population with potential exposures and provide information about the importance of various sources that contributed to the dose. Acceptance of an epidemiologic study by the public and the scientific community will hinge on the quality of the science and on the perceived thoroughness of the study. These ends will be achieved only if the following general conditions are met. First, an epidemiologic study must be justified scientifically, and the decision to undertake it should be based on a careful, preliminary scoping study of the site. The scoping study can define, in outline at least, the design, scope, and methods of the epidemiologic study if one is begun. Second, if a study is justified, it must begin in a timely manner—delaying the epidemiologic investigation could jeopardize its completeness and hence its credibility. If a study cannot be justified, alternative health surveillance strategies that address public concern should be considered. This will usually be the case if a study does not have reasonable statistical power, either to detect effects that are approximately as large as those predicted from high-dose extrapolation or to set useful upper confidence limits if, in fact, there are no statistically demonstrable effects. For instance, developing a cancer registry would permit continued health monitoring of the long-time residents at the site, so as to detect any unusual cancer incidence. Or a screening program could be a useful surveillance mechanism to provide assurance to the target population that its health concerns are being considered. Third, if the study is to be credible, there must be close interaction between dosimetrists involved in the dose reconstruction and the epidemiologists who will conduct the study. The epidemiologists must understand the details of the dose reconstruction sufficiently well to collect the appropriate information on dietary and lifestyle factors that could contribute to health effects. It is also necessary for statisticians and epidemiologists to work closely with dosimetrists to define the various sources of uncertainty and to help ensure that uncertainties are expressed in a form that will be useful in epidemiologic analyses. Fourth, the particular health end points to be targeted in an epidemiologic study will be defined primarily by the organ doses that members of the public received (which may vary appreciably when there are radionuclide exposures) and by the radiosensitivity of various organs and tissues. The primary end points will normally be cancer, although in
OCR for page 80
some circumstances other health end points might be considered (such as adverse reproductive outcomes when pregnant women are exposed). Fifth, the design of a study and the data needed will depend on the purposes to be served. If an epidemiologic study is conducted, the retrospective cohort or case-control designs are usually the methods of choice. Ecologic (correlational) or cluster studies have large potential for bias and can, therefore, be misleading (Stidley and Samet, 1994). They should be avoided wherever possible. Bias resulting from confounding in epidemiologic studies is especially troublesome in studying small increases in risk, because the magnitude of the bias can exceed the magnitude of the excess risk of interest. Careful attention should be given to ways to control potential biases. Sixth, the statistical analyses should make optimal use of the available dose information, including uncertainties in the dose estimates. A primary approach to analyzing, reporting, and interpreting results should be specified in advance, to avoid the pitfalls of a posteriori analyses. Similarly, the main health end points to be considered in interpreting the study findings should be specified in advance, to avoid a posteriori "chance" findings; this specification would generally be based on the effects that have been seen in other radiation epidemiology studies. Finally, as previously noted, studies of exposed populations demand that those populations be involved in their design and implementation. Citizens should be involved in every phase of the scoping and decision-making process and in any studies that are performed. Every effort should be made to communicate scientific issues in a fashion that is understandable to lay audiences. An advisory committee, consisting of citizens and impartial scientists, should be established at the outset and invested with oversight powers. This committee should be maintained throughout the study or surveillance program. These requirements have led the committee to make the following five recommendations: Dosimetric and epidemiologic scoping studies around the sites of nuclear facilities or accidents should be considered, although the extent of such studies might vary from one site to another based on preliminary evidence about exposures, population sizes, and public concern. These studies should be performed interactively and in parallel, because both are needed to inform a decision about further study of the site or for establishing priorities among sites. Epidemiologic and dosimetric assessments should be closely coordinated. It is important to have epidemiologists involved from the outset of any dose reconstruction activity to ensure that the dosimetric
OCR for page 81
information developed is appropriate for epidemiologic decisions and planning. A full-fledged dose reconstruction and epidemiologic study should be proposed only if the scoping studies show that adverse health effects are likely to be statistically detectable, given the probable dose distribution and size of the exposed population. Studies of health end points for which high-dose studies give no clear evidence of an excess should be avoided, because observing true excesses of these end points is biologically implausible. Such studies tend to waste resources and they are uninformative at best and misleading at worst. A statistical power assessment based on a realistic set of assumptions about the dose distribution, population size, and radiation risk coefficients should be part of the scoping phase.
Representative terms from entire chapter: