Excerpt One -
Chapter One - "Background Information and
Scientific Principles," pp. 44-51.

EPIDEMIOLOGICAL STUDIES: SPECIAL CONSIDERATIONS

Epidemiologic studies are a critical tool in assessing radiation
risks, since they alone provide data directly applicable to
humans. However, epidemiologic studies of individuals exposed to
radiation have methodologic limitations which should be kept in
mind when assessing the results of such studies. This section
briefly summarizes these concerns. Further discussion of these
issues can be found in standard textbooks on epidemiologic
methods (Ma70, Ro86). Most epidemiologic studies of low-LET
radiation have focused on cancer as the outcome. This discussion
of epidemiologic methods and their limitations also focuses on
cancer, although most of the considerations also apply to studies
of other outcomes.

High-Dose Studies

The use of high-dose studies to quantify risk estimates involves
a two-stage process. First, risk parameters that apply to the
particular high-dose group under observation must be estimated
from the empirical data. Second, mathematical models must then be
used to extrapolate from the experience of the specific high-dose
population to that of the low-dose population of interest, taking
into account differences both in exposure factors such as dose
and dose rate and host factors such as age, sex and race. Both
steps are, of course, subject to error, and the assumptions and
limitations involved in the second step will be discussed in
detail later in this chapter. The problems and limitations
involved in the first step are discussed here.

Studies reported to date have essentially been of the
retrospective cohort type. Populations receiving high doses of
low-LET radiation are rare, and exposure to such doses is
unlikely to occur in the future, apart from the therapeutic
irradiation of patients. Such studies are subject to both
sampling variability and bias. Sampling variability should
generally be adequately expressed by the confidence intervals
around the parameters estimated by the particular mathematical
model, but bias represents a greater problem. Biases in
epidemiology are generally classified as resulting from
selection, information, or confounding.

Selection bias can be defined as arising from any design problem
that tends to make the study subjects unrepresentative of their
source population. Such a bias can prevent generalization of the
results. For example, if the survivors of the atomic bombings at
Hiroshima and Nagasaki were healthier than the general
population, their susceptibilities to radiation carcinogenesis
could be different from those of the general population. In
addition, selection may lead to internally biased results when
the follow-up is selective. This occurs when those individuals
selected for follow-up are different for differing categories of
exposure and when that difference is associated with a differing
underlying cancer risk. For example, if only 50% of the
atomic-bomb survivors had been followed, and there were more
smokers in the high-dose group that were followed than in the
low-dose group, there would be an excess of lung cancer in the
high-dose group that was not caused by radiation. Such a
selection bias is likely to occur only when there is substantial
loss to follow-up. It is unlikely that this plays a role in the
major high-dose epidemiologic studies on which risk estimates are
currently based, since follow-up has been essentially complete
for these studies.

Information bias, which refers to any process which distorts the
true information on either exposure or disease status, is likely
to be of more importance than selection bias. Misclassification
of exposure is likely to be a major potential source of error in
making risk estimates. Nondifferential misclassification with
respect to exposure level leads to an underestimation of risk and
tends to reduce any upward curvature in the dose-response
relationship. This occurs, for example, when the distribution of
errors in dose estimates is the same in the diseased and the
nondiseased, as will generally be the case for most cohort
studies. Other biases may be more subtle. Misclassification of
disease status is particularly important when such status is
determined from death certificates which are often unreliable for
a number of cancer types. These errors are more likely to be
differential, i.e., dependent upon a subject's exposure status,
and could bias a dose-response curve away from the null.

Finally, confounding--i.e., distortion of risk estimates due to
the association of both exposure and disease with some other
covariate, such as smoking--is unlikely to be of substantial
importance in affecting risk estimation based on comparison of
groups of individuals with varying degrees of exposure, but it
could be of importance when an unexposed control group is also
used in the estimation procedure. For example, the
characteristics of the "not in the city" group in the Japanese
atomic-bomb survivor study may be somewhat different from the
group exposed to the radiation, and if these characteristics are
associated with differing cancer risks, such confounding would
have an effect on the risk estimates. This may be a particular
problem with studies of patients irradiated for medical
conditions if risk estimation is carried out with an unexposed
comparison group, such as the general population: the condition
for which irradiation is used could well be associated with an
altered cancer risk.

The three types of bias discussed above could all play roles in
affecting the internal validity of risk estimates (i.e., the
validity of the results for the particular population being
studied). However, even in the absence of such biases, there
remains a fundamental problem in extrapolating the risks from one
population to another, for example, from the Japanese to North
Americans. The method of such extrapolation depends on the
mathematical model chosen; and, although empirical evidence may
be available from studies carried out in both countries, there
often is considerable uncertainty about the validity of the
procedure that is used.

The quantitative risk estimates developed in Chapter 4 of this
report are based primarily on extrapolation from studies of
populations exposed to high doses of radiation over relatively
short periods of time. The rationale for this approach is that
only these studies provide sufficiently precise estimates of risk
at any dose. Risk estimates for low doses and protracted exposure
could therefore be in error because of (1) an inappropriate
mathematical model, or (2) biases in the high-dose epidemiologic
studies used to estimate the parameters of the chosen model, as
discussed above.

The committee has attempted to mitigate the first problem by
using sufficiently general model classes that include most of the
widely accepted alternatives and by providing estimates of the
range of uncertainty in the estimates. In general, the estimates
of risks derived in this way for doses of less than 0.1 Gy are
too small to be detectable by direct observation in epidemiologic
studies. However, it is important to monitor the experience of
populations exposed to such low levels of radiation, in order to
assess whether the present estimates are in error by some
substantial factor.

Low-Dose Studies

A number of low-dose studies have reported risks that are
substantially in excess of those estimated in the present report.
These include risks to populations exposed to high background
levels of radiation, diagnostic x rays, and fallout from nuclear
weapons testing or nuclear accidents, and to individuals with
occupationally derived exposures. Some of these studies are
discussed in more detail subsequently. Although such studies do
not provide sufficient statistical precision to contribute to the
risk estimation procedure per se, they do raise legitimate
questions about the validity of the currently accepted estimates.

The discrepancies between estimates based on high-dose studies
and observations made in some low-dose studies could, as
indicated above, arise from problems of extrapolation. An
alternative explanation could be inappropriate design, analysis,
or interpretation of results of some low-dose studies. This
section discusses the particular methodologic problems which can
arise in such studies, and the section on low-dose studies in
Chapter 7 summarizes a number of these studies and assesses their
results, taking into account the methodological limitations
discussed here.

The problem of random error caused by sampling variability is
relatively more important for low-dose than for high-dose
studies. (Sampling variation means the range of results to be
expected by exact replication of the study, if this were
possible; its major determinant is sample size and its
distribution across exposure and disease categories.) To
understand why this is so, suppose that two studies were
conducted, one in a population exposed to 1 Gy and one in a
population exposed to 0.01 Gy, in which similar sample sizes and
designs were used, and suppose that the resulting standard errors
on the log relative risk were the same. Thus, suppose the
relative risk in the high-dose population was 11 with 95%
confidence intervals of 5.5 and 22 and the relative risk in the
low-dose population was 1.1 with confidence intervals of 0.55 and
2.2. The point estimates on the relative risk coefficient from
the two studies would be identical at 10/Gy, but the confidence
intervals on the high-dose estimate are 4.5 and 21 and on the low
dose estimate are -4.5 and 12.0. This comparison emphasizes the
importance of considering sampling variability in assessing the
results of low-dose studies. In fact, the problem of sampling
variation is even more serious than this simple example would
indicate. The standard error of the relative risk in a simple 2 x
2 table of exposure by disease status is determined primarily by
the size of the smallest cell in the table, which is usually the
number of exposed cases. In most studies of low-dose effects,
this cell may be quite small, so the resulting standard error is
larger than that for high-dose studies, even if the overall
sample sizes were the same.

In general, systematic biases are also relatively more important
for the objectives of low-dose studies than they are for those of
high-dose studies. Because of the existence of more and larger
populations exposed to low doses, low-dose studies are often
ecological (correlational) or case-control studies rather than
cohort studies. The ecological and case-control studies are
particularly prone to bias in their design.

Selection bias is a major potential problem in case-control
studies: the major concern is over the appropriateness of the
control group. This is a particular problem for those studies in
a medical setting.

Information bias leading to misclassification of either exposure
or disease status, if random, leads to underestimated risk, and
several low-dose studies could well involve substantial
systematic misclassification, for example, misclassification
because of recall bias by cases in case-control studies.
Similarly, tumors which can be induced radiogenically could be
overestimated in radiation-exposed individuals.

Confounding may be more important for low-dose than for high-dose
studies. An observed relative risk of 2 is much more likely to be
produced solely by confounding than a relative risk of 10 (Br80).
The possibility of confounding can only be judged on a
study-by-study basis, but some generalizations are possible.
Ecological correlation studies, such as the studies of areas with
high levels of background radiation, are probably the most
susceptible to confounding. Residents of areas with high levels
of background radiation are likely to differ in many ways from
those in areas with low levels of background radiation. This
could affect cancer rates, but data on the relevant
characteristics are unlikely to be available for analysis. As an
example, exposure to radiation from terrestrial sources may vary
with housing structure, which, in turn, may reflect a
socioeconomic status that correlates with such factors as smoking
and alcohol use. This possibility alone generally makes such
studies uninterpretable, and when the ecological fallacy
discussed below is also considered, these two problems alone are
enough to make such studies essentially meaningless. Case-control
studies, on the other hand, generally offer the greatest
opportunity to control for confounding by matching or obtaining
information on definable covariates for use in analysis. However,
the extent to which this has been done varies from study to
study. It is necessary, of course, to collect data on such
confounders, and, if the confounders are not recognized in
advance, the appropriate data may not be available.

Finally, three other potential biases of low-dose studies should
be mentioned (Be88). The first is the ecological fallacy, that
is, that in correlational studies, any excess risk occurring in a
population with increased exposure may be occurring in
individuals other than the individuals who are actually receiving
the excess exposure. Second, is the possibility of selective
reporting. Epidemiologists are more likely to report and journal
editors are more likely to accept positive findings than null
findings. Thus, information in the literature on populations
exposed to low doses of radiation may be slanted in favor of
those studies that show higher risks than the conventional
estimates, since those that show estimates consistent with the
accepted values would not be seen as significant. The magnitude
of this potential effect is unquantifiable, but it almost
certainly exists and plays a role in the plethora of low-dose
studies with a reported positive risk. Third, there is the
problem of multiple comparisons. This arises if a number of tests
of significance are made with respect to elevated risks for a
number of cancer sites. Such a process invalidates the
conventional value quoted for the test of significance and leads
to more significant results than nominally would be expected by
chance. For example, in following a cohort of occupationally
exposed individuals, if comparisons are made for 10 cancer sites
with a p value of 0.05, which nominally would be expected 5% of
the time by chance for a single comparison, significant excesses
would arise 40% of the time by chance for at least one of those
outcomes. Interpretation of such results must be guided by prior
hypotheses, and by consistency of results among studies, a major
criterion for causality.

RISK ASSESSMENT METHODOLOGY

Need for Models in Risk Assessment

One of the major aims of this report, as of previous BEIR
reports, is to provide estimates of the risks of cancer resulting
from various patterns of exposure to ionizing radiation. In
principle, such estimates could be derived by identifying a group
of individuals with similar exposures and similar backgrounds and
following them to compare the proportion of the group who
eventually developed cancer with the proportion who developed
cancer in a comparable unexposed group or in the general
population. For situations in which it is not possible to measure
the risks directly, statistical models must be used to derive
estimates.

Large sample sizes are needed in any such comparisons, to
minimize random variation; the rarer the disease and the smaller
the effect of exposure, the larger the sample needs to be. For
example, the BEIR III report estimated that a single exposure to
0.1 Gy (10 rads) of low-LET radiation might cause, at most, about
6,000 excess cases of cancer (other than leukemia and bone
cancer) per million persons, as opposed to a natural incidence of
about 250,000. To identify this number as a statistically
significant excess, a cohort of about 60,000 people with the same
exposure would have to be followed for a lifetime, or an even
larger number of people would have to be studied if follow-up
were for a shorter period of time. Under ideal conditions, a
case-control study to identify the same excess would have to
consist of at least 120,000 cases and 120,000 controls. It is
unlikely that such large groups with similar exposures could be
identified, let alone feasibly studied. Furthermore, even if the
random variation could be overcome by the large sample sizes
needed, estimates of such small excess risks (2%) could easily be
biased by confounding, misclassification, or selection effects.
Epidemiologists generally agree that excess risks of less than
50% are difficult to interpret causally (Br80). In practice,
therefore, it is necessary to obtain risk estimates by
extrapolation from smaller and less homogeneous groups who have
been exposed to larger doses by using statistical dose-response
models.

The second problem is that there are many other factors that are
known to contribute to cancer risks or to modify the effects of
radiation on cancer risks, and these factors need to be taken
into account. While it is theoretically possible to control for
such factors by cross-classifying the data into subgroups that
are homogeneous with respect to all relevant factors, it is again
unlikely that sufficiently large subgroups will be available to
allow for stable estimates, particularly if the number of factors
is large. For investigating lung cancer, for example, it might be
necessary to control for sex, age, time since exposure, and
smoking habit; if four levels were used for grouping each factor
other than sex, a total of 128 subgroups would be needed, each of
which would need to be the minimum size if risk estimates
specific to each group were to be observed directly. Since this
is not generally feasible, it is necessary to rely on
multivariate statistical models to identify the consistent
patterns across the variables simultaneously and to predict the
risks for subgroups in which the sample sizes are inadequate.

The third problem is that direct estimates of lifetime risk can
only be obtained after an exposed population has been followed
for a lifetime. Few populations have been followed so long, and
even the atomic-bomb survivors, one of the populations followed
for the longest period, has been followed only for just over 40
years. As the risks for many cancers in this population are still
elevated, it is an open question whether the excess risk will
continue for the remainder of the population's life and, if so,
at what rate. It is not appropriate to wait until follow-up is
complete, however, since interim estimates of risk are needed now
for public health purposes. Again, to provide such estimates, one
must fall back on statistical models that adequately describe the
data available so far and the range of uncertainty around them.

Epidemiologic data have increasingly been called on to help
resolve claims for compensation by exposed individuals. Because a
radiation-induced cancer is clinically indistinguishable from
cancers caused by other factors, such claims must be settled on
the "balance of probabilities," in other words, by determining
what was the most likely cause, given the individual's history of
exposure to radiation, and taking into account confounding and
modifying factors. The calculation of these probabilities of
causation depends on the availability of suitable multivariate
exposure-response models. A recent National Institutes of Health
working group (NIH85) has provided tables of such probabilities;
these were based on data that were available at the time.