NCRP report no 126 uncertainties in fatal cancer risk estimates used in radiation protection

Uncertainties in fatal cancer risk estimates used in radiation protection : recommendations of the National Council on Radiation Protection and Measurements.. uncertainty, but also to

National Council on Radiation Protection and Measurement

7910 Woodmont Avenue / Bethesda, Maryland 2081 4-3095

Measurements (NCRP) The Council strives to provide accurate, complete and use- ful information in its documents However, neither the NCRP, the members of NCRP, other persons contributing to or assisting in the preparation of this Report, nor any person acting on the behalf of any of these parties: (a) makes any warranty

or representation, express or implied, with respect to the accuracy, completeness or usefulness of the information contained in this Report, or that the use of any infor- mation, method or process disclosed in this Report may not infringe on privately owned rights; or (b) assumes any liability with respect to the use of, or for damages resulting from the use of any information, method or process disclosed in this Report, under the Civil Rights Act of 1964, Section 701 et seq a s amended 42 US C

Section 2000e et seq (Title VZZ) or any other statutory or common law theorygovern- ing liability

Library of Congress Cataloging-in-Publication Data

National Council on Radiation Protection and Measurements

Uncertainties in fatal cancer risk estimates used in radiation

protection : recommendations of the National Council on Radiation

Protection and Measurements

p cm (NCRP report ; no 126)

"Issued October 1997."

Includes bibliographical references and index

ISBN 0-929600-57-6

1 Radiation carcinogenesis 2 Cancer Mortality 3 Cancer Risk factors

4 Health risk assessment 5 Radiation Dosimetry

I Title 11 Series

[DNLM: 1 Neoplasms, Radiation-Induced etiology 2 Neoplasms, Radiation -Induced mortality 3 Radiation Protection 4 Risk Factors 5 Radiation Dosage QZ 200 N2745c 19971

Preface

In recent years, the practice of providing uncertainties when formulating estimates of dose and risk in human and environmen- tal exposure circumstances has become recognized as a n important step in expressing the degree of confidence appropriate to stated values Knowledge of the magnitude of uncertainties in the nomi- nal values of the coefficient for risk of fatal cancer per unit dose can

be very helpful in providing perspective to those involved in radia- tion protection practice In human cancer risk estimation, however, only rather tentative approaches to the evaluation of these uncer- tainties have been made, starting with the report of the NIH Ad Hoc Working Group on Radioepidemiological Tables in 1985 and with relatively brief attempts by the United Nations Scientific Committee on the Effects of Atomic Radiation and the National Academy of Sciences/National Research Council's Committee on the Biological Effects of Ionizing Radiation The data on mortality from the Lifespan Study of the Japanese atomic-bomb survivors up

to 1985 are virtually the sole numerical source used for risk esti- mates for low-LET radiation exposure today (Later evaluations of the LSS data to 1987 and to 1990 are of wide interest in epidemiol- ogy but have not so far modified the risks recommended for use in radiation protection.) Other sources of risk information are used mainly to support and complement the data from the LSS In the NCRP Taylor Lecture in 1993, it was pointed out that the singular- ity of the LSS as a source of low-LET risk information simplifies the assessment of uncertainties in the risk estimates Because the LSS risk estimates depend on five distinct components, uncertainties overall can be evaluated by examining the uncertainties in each of these components In the 1993 Taylor Lecture, the evaluation (and discussion) of the five components was quite limited, although an overall picture was outlined The NCRP decided recently to build

on that Taylor Lecture by looking a t each of the five components in more detail and attempting to be more quantitative about their uncertainties This Report is the result It makes clear that the fun- damental basis on which the evaluation of some of the components rests is, itself, uncertain and difficult to quantify Nevertheless, the Report seeks not only to clarify the foundation of estimates of

uncertainty, but also to make a reasonable overall appraisal of the uncertainties in the average risk estimates presently used in low-LET radiation protection Risk estimates for individual organs

involve greater uncertainties than for total cancer and are not dealt with specifically in this Report

This Report was prepared by Scientific Committee 1-5 on Uncertainty in Risk Estimates Serving on Scientific Committee 1-5 were:

Warren K Sinclair, Chairman National Council on Radiation Protection and Measurements Bethesda, Maryland

Members

And& Bouville Charles E Land

National Cancer Institute National Cancer Institute Bethesda, Maryland Bethesda, Maryland

NCRP Secretariat

William M Beckner, Senior S t a f f Scientist

Cindy L O'Brien, Editorial Assistant

The Council wishes to express its appreciation to the Committee members for the time and effort devoted to the preparation of this Report

Charles B Meinhold

President

Contents

Preface 111

1 Intmdudion 1 1.1 Risk Estimates for Radiation Protection 1

1.2 Past Risk Estimates 2 1.3 Present Risk Estimates 3

1.3.1 Age and Sex Dependence 3

1.3.2 Lifetime Risk 3

1.3.3 Risk Estimates for Low Dose and Dose Rate 4 1.4 Uncertainties in Risk Estimates 5

1.4.1 Past Uncertainty Evaluations 5

1.4.2 NCRP Approach to Uncertainty in Risk Estimates for Radiation Protection 6

1.4.3 City Differences 8

1.5 Dose Response 8

2 Epidemiological Uncertainties 10

2.1 Introduction 10 2.2 Specific Epidemiological Uncertainties 13

2.3 Bias in Risk Estimates Due to Errors of Detection

and Confirmation 16 2.4 Biases Affecting Risk Estimates of Cancer

Morbidity 18

2.5 Unrepresentative Population 20

2.6 Bias Deriving from City Differences 21

2.7 Summary of Epidemiological Uncertainty - 2 2

3 Dosimetrical Uncertainty 23

3.1 Random Errors and Biases 23 3.2 Bias Resulting from Random Errors in Dose 24

3.3 Bias in Gamma-Ray Measurements Versus DS86 30

3.4 Uncertainty Due to Survivor Shielding

Characterization in DS86 32 3.5 Uncertainty Due to Neutron Weight (Relative

Biological Effectiveness) 32 3.6 Bias and Uncertainties Due to the Presence of Thermal Neutrons a t Hiroshima in Excess of Those Predicted by DS86 34

3.7 Combination of Uncertainties and Bias for Dosimetry 37

4 Transfer of Risk Between Populations 40

4.1 General Considerations 40

4.2 Factors Modifying Risk in Relation to Transfer Between Populations 42

4.3 Site-Specific Evidence for Selecting the Transfer Model -47

4.4 Uncertainty Due to Method of Transfer 49

5 Projection to Lifetime Risk 51

5.1 Constant Relative Risk Projection Model 51

5.2 Considerations Regarding the Projection to Lifetime Risks in the Lifespan Study 51

5.3 Attained Age Model 53

5.4 Lifetime Risk of Those Exposed a t Young Ages 55

5.5 Uncertainty in Lifetime Risk 57

6 Extrapolation to Low Dose or Dose Rate 60

6.1 Effect of Dose Rate and Dose in Radiobiology 60

6.1.1 The Effect of Dose Rate 60

6.1.2 The Effect of Dose 60

6.2 Human Data and Dose-Rate Effects 61

6.3 The ICRP Choice of a Dose and Dose-Rate Effectiveness Factor 63

6.4 The NCRP Position on the Application of a Dose and Dose-Rate Effectiveness Factor 64

6.5 UNSCEAR Evaluation of a Dose and Dose-Rate Effectiveness Factor and Recent Studies 64

6.6 Uncertainties in the Application of a Dose and Dose-Rate Effectiveness Fador 65

7 Combination of Uncertainties 67

7.1 Sources of Uncertainty 67

7.2 Method to Propagate Uncertainties 69

7.3 Results 71

7.3.1 Population of A11 Ages 71

7.3.2 Adult Worker Population 73

7.4 Conclusions 74

Glossary 77 References 83

TheNCRP 92

NCRPPublications 100

Index 109

1 Introduction

1.1 Risk Estimates for Radiation Protection

This Report is concerned with the evaluation of uncertainties in the risk estimates of fatal cancer induced by low-LET radiation (see Glossary) as presently used in radiation protection, i.e., in esti- mates of the risk of fatal cancer following exposure of individuals

or populations in occupational, environmental or domestic circum- stances These cancers are the main component of the health detri- ment following radiation exposure identified by the International Commission on Radiological Protection [ICRP (1991)l and the National Council on Radiation Protection and Measurements [NCRP (1993a)l as pertinent in low-LET radiation protection Recent evaluations of the risk of fatal cancer induced by low-LET radiation are numerically based on the 1950 to 1985 mor- tality experience of the survivors of the atomic bombs dropped in Japan, as ascertained by the Lifespan Study (LSS) (EPA, 1994; ICRP, 1991; NASLNRC, 1990; NCRP, 1993a; NRPB, 1993; UNSCEAR, 1988) Other epidemiological studies, although they can be highly informative with regard to particular cancer sites, have served mainly to support the results from the LSS, and to show that the LSS results are not isolated, but are generally and broadly supported by these other sources of data Later evaluations

of induced fatal cancer risk in the LSS include the mortality and incidence data up to 1987 reviewed by the United Nations Scien- tific Committee on the Effects of Atomic Radiation [UNSCEAR (199411 and the more recent mortality evaluations up to 1990 (Pierce et al., 1996a) These new studies provide additional infor- mation, especially on cancer incidence, but they do not alter sub- stantially the risk estimates derived in the 1988 to 1990 reports and, more especially, the derivation procedure, which is the source

of the uncertainty considerations, remains the same Not all aspects of radiation protection (notably those involving high-LET exposures) use risk estimates based on the LSS of the atomic-bomb survivors For example, any consideration of radon exposure to workers or to the public uses risk estimates based on radon

exposures to miners However, many radiation protection situa- tions in which risk is at issue will use the results of the LSS The results of the.LSS indicate that lifetime risk coefficients for fatal cancer derived from the high-dose rate exposures of the atomic-bomb survivors are about 10 x SV-' for a population of all ages and about 8 x S V - ~ for an adult (worker) population

1.2 Past Risk Estimates

I t is worth noting, [see Table 1.1, taken from ICRP (1991), Table B-101 that lifetime risk estimates for an acute exposure [i.e., no dose and dose-rate effectiveness factor (DDREF) applied1] have ranged over t,he period 1972 to 1990 from about 1 to about

Table 1.1-Excess lifetime mortality from all cancer, attributable to

1 Gy (or 1 Sv) acute uniform whole-body low-LET irradiation of the

general population (ICRP, 1991)

Probability of Death

Source of Estimate Additive Risk

Projection Model NASNRC, 1972

-

2.3 to 5.0 5.2 7.0' to 11.0~ 8.gdP.f 7.1d

a Population of Japan

Estimate based on age-specific coefficients of probability

' Estimate based on constant (age-averaged) coefficient of probability United States population

Modified multiplicative model

"LOW-dose" leukemia component multiplied by two

or low-LET radiation, the ratio between the biological effect of high-dose rate radiation to that of the effect of low-dose rate radiation a t the same dose is known a s the dose and dose-rate effectiveness factor, DDREF

1.3 PRESENT RISK ESTIMATES / 3

11 x SV-l with the type of projection model used being one of the largest contributors to the variation of risk estimates Risk esti- mates have been more consistent over time when the multiplicative risk projection model is used

UNSCEAR, BEIR I11 [Committee on the Biological Effects of Ionizing Radiation of the National Academy of ScienceslNational Research Council (NASINRC)], ICRP and NCRP, in the period 1977

to 1980, were all in substantial agreement about estimates of life- time risk coefficients for fatal cancer that were several times lower than those in use now (about 1 to 2 x SV-l, compared with 4 to

5 x SV-~) Evident1y;risk values agreed upon today must still

be considered subject to future change as different information comes forward on any of those aspects on which the estimates are based

1.3 Present Risk Estimates

The ICRP (1991) and the NCRP (1993b) have further derived the nominal values of risk to be used for (low-dose rate) radiation protection as 5 x SV-l for a population of all ages and

4 x ~ o - ~ s v - ' for adult workers, after dividing the average high-dose rate estimate by a DDREF of two It is uncertainties in these estimates of risk, now widely used in radiation protection, with which this Report is concerned

1.3.1 Age and Sex Dependence

These risk estimates apply to the populations specified If the age and sex of the population group is known in more detail, tables such as Table 1.2 can be used to apply risk estimates more accu- rately More detail on age and sex dependencies is provided in ref- erences such as Land and Sinclair (1991) and Pierce et al (1996a)

1.3.2 Lifetime Risk

The term "lifetime risk estimate" is not a unique description of the risk resulting from exposure to a tumor inducing agent For a detailed discussion of some of the issues relating to lifetime risk see Thomas et al (1992) and UNSCEAR (1994) One point only with reference to lifetime risk estimates will be cited here It concerns the choice of "risk of exposure-induced death* (REID) or the choice

of "excess lifetime riskn (ELR) as a measure of radiation-related

Table 1.2-Fatal cancer risk for different ages and sex after low

dose or low-dose rate exposure ( x SU-l) (Sinclair, 1992).a

Age (Y) Male Female Average

'United States population, average of multiplicative and NIH trans- fer models (Land and Sinclair, 1991)

population detriment REID represents the probability of an

"untimely" death due to exposure ELR is the difference between the probability of a cancer death given a specific exposure history and the probability of a cancer death in the absence of the specific exposure history Consequently, the REID includes the expo- sure-induced earlier deaths of those who would have later died of cancer without the exposure Because about 20 percent of the pop- ulation would be expected to die of cancer in the absence of radia- tion exposure, the REID is about 20 percent higher than the ELR for all cancer sites combined for uniform whole-body exposure For exposure limited to a single organ (e.g., salivary gland, for which lifetime mortality rates are considerably below one percent) the REID and the ELR are more closely comparable

It should also be pointed out that single values (point estimates)

of lifetime risk coefficients do not convey the wealth of information already known about sex and age variations in risk, which should usually be accounted for when dealing with specific practical situ- ations Consequently, uncertainties in past estimates are only a beginning to the consideration of uncertainties in many risk cir- cumstances Furthermore, uncertainties in risk estimates for indi- vidual organ and tissue sites are also of practical importance and probably differ among themselves; but these will not be addressed

in this Report

1.3.3 Risk Estimates for Low Dose and Dose Rate

The lifetime risk coefficients presently recommended by ICRP and NCRP were derived from the LSS data of the atomic-bomb sur- vivors taking into account the following evaluations: ( 1 ) the REID values given by UNSCEAR (1988) for the multiplicative projection model for the period of observation to the end of life and a Japanese

1.4 UNCERTAINTIES IN RISK ESTIMATES / 5

population2 (11 x SV-'1, (2) the ELR values given by BEIR V (NAS/NRC, 1990) for a United States population3 (about

9 x SV-'1, and (3) REID values derived by the ICRP for a n average of five populations, 9.5 x SV-l (ICRP, 1991) These val- ues for high-dose rate exposures were averaged and rounded to

10 x 10" SV-l and divided by a DDREF of two to obtain 5 x

SV-' for a population of all ages and for the low-dose rate conditions

of normal radiation protection (ICRP, 1991; NCRP, 1993b)

The nominal lifetime risk value for workers was derived simi- larly from a high-dose rate value of 8 x SV-' CUNSCEAR, 1988) divided by a DDREF of two for a nominal value of 4 x SV-' for adult workers (ICRP, 1991; NCRP, 1993b) Lifetime risk coeffi- cients for individual organs were also derived by ICRP and NCRP for use in radiation protection (ICRP, 1991, Table 4; NCRP, 1993b, Table 7.2) These were also used to derive tissue weighting factors (rounded fractional health detriments) for estimating effective doses used in determining compliance with radiation protection limits (ICRP, 1991, Table 2; NCRP, 1993b, Table 5.1) The values are called nominal values because they apply to averages for the whole population and a worker population They do not apply to a specific individual unless that individual can be considered to fit the average in all characteristics Adjustments for age and sex have already been recommended, see Table 1.2

I n this Report, low doses will refer to absorbed doses in the range 0 to 0.2 Gy and to equivalent doses of 0 to 0.2 Sv Low-dose rates are those below 0.1 Gy d-' for all radiations

1.4 Uncertainties in Risk Estimates

1.4.1 Past Uncertainty Evaluations

I t is important to address uncertainties in risk estimates for radiation induced fatal cancer in a realistic manner The first seri- ous attempt to do so occurred in relation to the evaluation of prob- abilities of causation in given exposure circumstances, i.e., the production of the "NIH (National Institutes of Health) tables"

(NIH, 1985) A very useful initial appraisal of uncertainties in the

relative probability that a specific cancer was due to a given

2~apanese national mortality patterns, 1980 (see UNSCEAR, 1988,

Appendix F, Table 64)

3 ~ i t a l statistics of the United States, 1980 (PHs, 1984)

radiation exposure resulted Our currently used estimates of the risk of cancer in radiation protection start with the evaluations by UNSCEAR in 1988 The UNSCEAR (1988) treatment of uncertain- ties dealt with such issues a s confounding (by smoking for exam- ple), the "healthy worker effect," different and changing baseline cancer rates in different countries and the dose response pattern, but in a general rather than a specifically quantitative way In the BEIR V Committee report of 1990 (NASNRC, 1990) the uncertain- ties were addressed in a more quantitative manner following broadly the approach of the NIH Committee in 1985 (NIH, 1985) and concentrating on individual tumor sites rather than total can- cer risk The features considered included not only random error associated with sampling variation in the fitted coefficients of the models used but uncertainties in dose estimates, certification of cause of death, population effects, the choice of risk versus time model, sex and age differences, and the shape of the dose response curve Some results were provided in the form of geometrical stan- dard deviations (GSD), a number greater than one (see Glossary) The range of uncertainty, expressed as a confidence interval is com- puted by dividing and multiplying the point estimate by a specified power of the GSD For example, a 90 percent confidence interval for estimate E with GSD G has lower limit E / G ~ ~ ~ ~ and upper limit

E ~1.645 While some GSD's provided by the BEIR V Committee for individual tumors, age groups and time after exposure were estimated to be a s low as 1.24, others ranged up to more than three Until now, neither ICRP nor NCRP has specifically addressed the issue of uncertainties in risk estimates recommended for use in radiation protection

1.4.2 NCRP Approach to Uncertainty in Risk Estimates for

Radiation Protection

The question now arises, W h a t degree of confidence (or uncer- tainty) can be attached to the current nominal values of lifetime risk coefficients for all cancer used for radiation protection a t low doses and dose rates a s recommended by ICRP and NCRP?" This Report examines individual uncertainties in the five modular com- ponents on which the lifetime risk coefficients are based (Table 1.3)

The evaluation of overall uncertainty has been accomplished in the following way For each of the five individual modular compo- nents, a probability distribution, a likeliest value (usually the 50th percentile), and a 90 percent confidence interval (5th to 95th

1.4 UNCERTAINTIES IN RISK ESTIMATES / 7

Table 1.34omponents of risk coeficient derivation from the LSS

of the atomic-bomb survivors

percentile) have been subjectively selected Other choices of confi- dence interval, e.g., 95 percent, could have been made but 90 per- cent is commonly used and is quite appropriate for our purposes There is usually little or no information on the shapes of the prob- ability distributions and therefore the choice has been largely sub- jective A triangular distribution (such as shown later in Figure 3.2 for example) has been chosen when a degree of subjective confi- dence can only be attached to the likeliest value and to the possible range of values Normal or lognormal distributions have been pre- ferred for smooth, symmetric or right-skewed distributions These two distributions are appropriate theoretically when there is no reason to believe that random error represents the sum or product

of independent incremental components However, the overall results for the combined uncertainties are not sensitive to the par- ticular shapes of the probability distributions selected for each component provided the likeliest values and the 90 percent confi- dence intervals remain the same (IAEA, 1989; NCRP, 1996) Finally, the overall uncertainty in the risk estimate and its central value have been estimated using Monte Carlo methods which take into account all the uncertainty estimates for the individual modu- lar components (Section 7)

The text of this Report will rely on customary methods of uncer- tainty analysis as applied to uncertainty in environmental data and other circumstances A useful source of information which includes many of the references to relevant principles and methods

is NCRP Commentary No 14 (NCRP, 1996) The uncertainty eval- uations addressed in this Report relate to the methods of derivation

of risk estimates and concern average or nominal values They do not consider variability in the characteristics of individuals in the population which influences their risk and contributes to individ- ual uncertainty

1.4.3 City Differences

In some ways, it might have been useful to examine the uncer- tainties in the risk of cancer derived from the exposures a t Nagasaki separately from those derived from the exposures a t Hiroshima One reason for this is that the estimates of the dose according to the present DS86 system, may be relatively sound for Nagasaki (although this is not certain and there are some unique dosimetry problems with some specific groups from Nagasaki included in the analysis also), whereas more questions have contin- ued to arise about Hiroshima, especially about the magnitude of the neutron component This and other factors concerning city dif- ferences are discussed later (see Section 2.6) The sample of attrib- utable cancers is small, 339 solid cancers and leukemia, in 5,936

cancer deaths altogether in the LSS up to 1985; consequently, sub- dividing the sample into Hiroshima (about two thirds of the sam- ple) and Nagasaki (one third of the sample) is not considered desirable at this time In this Report, uncertainties are considered collectively in the entire LSS sample

1.5 Dose Response

The evaluation of each of the five components (and in some cases subcomponents as well) of the uncertainty in risk estimates for radiation protection is described in Sections 2 through 6 Section 6

is of special impo~~tance because it discusses and accounts for the effects of dose and dose rate by using the DDREF Inevitably, the choice of DDREF involves a choice in the shape of the dose response curve starting with a simple linear response (DDREF = 1) and pro- ceeding to linear quadratic responses with initial linear portions of lower and lower slope as the DDREF increases Values of DDREF from one up to five are considered in the distribution for the DDREF Only if the DDREF went to infinity would the response be initially independent of dose, i.e., a threshold Those responsible for the analysis of risks in the LSS state firmly " the data for solid cancer, including tumor registry incidence data as well as cancer mortality data, are inconsistent with the notion of a threshold for

1.5 DOSERESPONSE / 9

radiation effects" (Pierce and Preston, 1996) Consequently, for the purposes of this Report, viz the evaluation of uncertainties in the risk coefficients derived from the LSS, the choices of DDREF will include all the reasonable linear and sublinear dose response mod- els for the atomic-bomb survivor data For the more general issue

of linearity versus threshold for radiation effects, the NCRP has a committee addressing this question It is noted that values of DDREF less than one, i.e., a supralinear response, have not been considered here either If they had been, only a small value for the frequency could be assigned to a DDREF of say 0.5 or 0.3, and this would have only a very minor impact on the overall uncertainty

Uncertainties

2.1 Introduction

"Epidemiological uncertainties" is a very genera1 term, used in this Report to refer to random error in observations, and also to sys- tematic errors including the possibility that a model used to esti- mate risk may deviate from the actual (and unknown) pattern of excess risk in some important way Confidence limits, standard errors, and p values for hypothesis tests all reflect random error in the context of a statistical model that is assumed to be true as specified

As a starting point, Table 2.1 gives the number of persons in the

LSS sample as of the analysis of 1985 (Shimizu et al., 1988; 1990)

Table 2.1-LSS, atomic-bomb survivors

(adapted from Shimizu et al., 1988; 1990)

Number of

People Dose groups

2.1 INTRODUCTION / 11

Then, Table 2.2 summarizes the relationship between radiation- dose and cancer mortality observed in the LSS sample over the years 1950 to 1985 The tabulated estimates resulted from linear regression of mortality rates on radiation dose and the associated confidence intervals reflect statistical uncertainties Baseline (i.e., zero-dose) risk was allowed to depend upon city (Hiroshima or Nagasaki), age at exposure (i.e., in August 1945), attained age (i.e., age at diagnosis), and calendar year, but the slope of the line expressing excess risk as a function of radiation dose (in this case, shielded kerma? rather than organ or tissue dose), was assumed to

be independent of city, sex, age and year In fact, the estimates given in Table 2.2 are only average values and for many of these sites, excess risk is known to depend on sex, age at exposure, attained age andlor time following exposure However, the average estimates in Table 2.2 are appropriate for our specific purpose Excess risk was expressed in two ways: first, in relative terms in which the relative risk (RR) coefficient is given as a multiplier of baseline risk, i.e., a ratio without units and second, as an average number of deaths per lo4 person-year (PY), i.e., in the same units

as baseline risk over the period of observation [excess absolute risk (EAR)] For all cancers a RR of 1.39 times (average) baseline, was found while the EAR was estimated as 10.0 deaths per lo4 PY Excess relative risk (ERR) is the RR - 1, or 0.39 in this example Uncertainty about these estimates was expressed by confidence limits Briefly, a pair of 90 percent confidence limits (i.e., a 90 per- cent confidence interval) for an unknown parameter a includes all numerical values % for which the null hypothesis, a = a would not

be rejected at significance level p = 0.10 in favor of the two-sided alternative hypothesis, a ;t ao In most applications, it is also the set of values ole for which the null hypothesis would not be rejected

a t significance level p = 0.05 in favor of either of the two one-sided alternative hypotheses, a < a or a > aO Thus, if the EAR a t 1 Gy (EARlGy) for all cancers is estimated to be 10.0 per lo4 PY Gy with

90 percent confidence limits (8.36,11.8), that means that all values between 8.36 and 11.8 per lo4 PY Gy are consistent with the data,

a t the 90 percent confidence level; because the lower confidence limit is greater than zero, it also implies that the null hypothesis of

no radiation effect (EARlGV = 0) is rejected at significance level

4~hielded kerma is the kinetic energy released per unit mass after the incident radiation has passed through intervening shielding material, but before entering the body

Table 2.2-Summary measures of radiation dose response for cancer mortality by site:a Both cities, both sexes (unless otherwise stated),

all ages ATB~, 1950 to 1985 (shielded kerma)

Number Site of Cancer of Estimated Relative Excess Absolute Risk

Deaths Risk a t 1 Gy per lo4 PYe Gy All malignant

aAdopted from Table 2a of Shimizuet al (1988) Additional detail on indi-

vidual organs is given in Table 2b of Shimizu et al (1988)

b~~~ = at the time of the bomb

'PY = person years

d( ) Numbers i n parentheses indicate 90 percent confidence interval Blanks in the Table indicate no lower confidence limit was provided eRisk estimation for these sites is based on either males or females only

2.2 SPECIFIC EPIDEMIOLOGICAL UNCERTAINTIES / 13

p = 0.05 in favor of the one-sided alternative of a positive effect (EAR~G!, > 0)

In science, generally a bias or systematic error is something that may invalidate the results of a study, but more often can be accounted for by a modification of the results, i.e., a correction If, for example, smoking were more prevalent among high dose than among low-dose subjects, an analysis of radiation-induced lung cancer risks that did not adjust for smoking or a n adequate surro- gate, would be biased A recognized bias can be corrected by modi- fying the statistical algorithm for estimation or, if that is not possible, by introducing a rationale for subjective adjustment with uncertainty factors contributing to the overall random error Uncorrected biases should be included among random errors Statistically, a biased estimate is one whose expected value is not equal to the value of the parameter being estimated Thus, whether the estimate is biased or not may depend upon the use to which it is put In the example of breast cancer mortality (Table 2.2), 1.02 excess deaths per lo4 PY a t 1 Gy, and a RR of 2.00 (ERRlGy = 1.00) are unbiased estimates of EAR and ERR, respec- tively, a t 1 Gy as a weighted average over all ages a t exposure for the period 1950 to 1985 But they are biased estimates of the EAR and ERR following exposure at age 10, because other analyses in the same study show that both absolute and RR vary by exposure aze

An example of bias resulting from statistical random error occurs with respect to individual dose estimates, and is discussed later (see Section 3.2)

2.2 Specific Epidemiological Uncertainties

The statistical uncertainty in the risks derived from the 1950 to

1985 LSS mortality data (assuming the doses are known correctly)

is represented by the confidence intervals in the EAR coefficients,

e.g., see Table 2.2 adapted from Table 2a of Shimizu et al (1988) For all cancer deaths, the EAR coefficient is 10.0 per lo4 PY Gy at

1 Gy with 90 percent confidence interval, 8.36 to 11.8, i.e., the 90 percent confidence limits are within about 220 percent of the nom- inal value For leukemia, it is 2.29 (1.89 to 2.73) per lo4 PY Gy (also about 220 percent), and for all solid tumors, 7.41 (5.83 to 9.08) per lo4 PY Gy or about z25 percent Further data are available for some individual tumor sites such as stomach, colon, lung, breast, etc with somewhat larger confidence intervals, often of the order of

*50 percent (see Table 2.2)

In this Report, the nominal value of the lifetime risk coefficient RHN (Rm = Hiroshima and Nagasaki) for all cancers for high-dose and dose rate, is taken to be 10 x SV-' (see Section 7) for a pop- ulation of all ages For the purposes of the uncertainty analysis,

Rm will be assumed to have the same relative statistical uncer- tainty due to sampling as for the solid tumors over the period of observation, viz 225 percent Consequently, a factor, F(RHN), which takes into account the statistical uncertainties associated with RHN, will be assumed to be normally distributed (a reasonable assumption based on a linear response model), with an average of one and a 90 percent confidence interval from 0.75 to 1.25, corre- sponding to a standard deviation of 0.15 The probability distribu- tion of F(RHN) is shown in Figure 2.1

The risk coefficients in Table 2.2, which summarizes the atomic-bomb survivor experience from 1950 to 1985, were obtained with a linear model, parameterized as follows:

(AR) and RR models cannot predict the same risk for all combina-

tions of these factors, i.e., p and ay cannot always be equal A more usual practice, not followed in the calculations leading to the results in Table 2.2, is to model RR and then convert to age-specific

AR by multiplying the estimated ERR by the age-specific baseline risk

The usual practice in estimating the risk coefficient (once the choice between relative and AR has been made) is to use the sim- plest dose-response model consistent with the data Linear esti- mates are given in Table 2.2, a t least partly because, for all solid cancers combined and most single organ sites, no statistically sig- nificant improvement in fit is obtained by adding dose-squared or higher power terms to the linear dose-response model Sometimes, this occurs because linearity fits the data very well, and the esti- mated dose-squared coefficient, E, in a quadratic model, e.g., in

Risk = a (1 + yD + dl2), (2.3)

2.2 SPECIFIC EPIDEMIOLOGICAL UNCERTAINTIES / 15

Parameter Magnitude

Fig 2.1 Distribution of statistical uncertainties in estimates of risk for solid tumors [F(RHN)] (The 90 percent confidence interval is shown by the arrows.)

is close to zero with fairly tight confidence intervals, while the value for the linear coefficient y is little different from its fitted value in the linear model When that happens, there is usually lit- tle disagreement that a linear model estimate is appropriate Where a linear-quadratic dose-response model is used for esti- mation, as in the case of leukemia (see, e.g., NAS/NRC, 1990), the relative contribution of the square of dose in Equation 2.3 relative

to the linear term decreases with decreasing dose The same is true

a t low-dose rate, since risk can be modeled as the sum of risks asso- ciated with low doses received during discrete time intervals The DDREF is used when the available data do not support a multi-parameter model like that given by Equation 2.3 and a sim- pler, linear model is used by default, but for theoretical or other reasons it is believed that the true model is actually linear- quadratic with a non-negligible coefficient for dose-squared For low-dose exposures, or cumulative exposures obtained at a low-dose rate, the linear-model estimated risk is divided by the DDREF This does not consider the problem of saturation, perhaps mainly from cell killing which is, however, reconcilable under cer- tain conditions (Sinclair, 1993) In the case of leukemia, a linear quadratic model is often used to determine the initial slope of the response (e.g., N A S M C , 1990) In this case, no DDREF needs to

be applied

A simple linear response divided by a DDREF is the approach used in this Report The DDREF will have a distribution of values determined by uncertainty considerations (Section 6) implying dif- ferent dose response models

2.3 Bias in Risk Estimates Due t o Errors of

Detection and Confirmation

Cancer statistics s d e r from failures to detect cancer cases (detection error), and from erroneous classification of noncancer cases as cancer (confirmation error) Both errors lead to misclassi- fication This is clear for mortality data derived from the results of studies comparing autopsy findings with independently obtained death certificate diagnoses I t appears that, among LSS sample members, the frequencies of the two kinds of errors are indepen- dent of radiation dose, but depend heavily on cancer site and age a t death Also, some organs, such as the liver, are often targets cf metastases from cancers of other organs, and these metastatic can- cers may be erroneously reported as primary cancers of the receiv- ing organ; this kind of error does not involve confusion between benign and malignant disease, however Errors of cancer detection affect absolute measures of risk but not usually RR Consider that the cancer rates a t 0 and 1 Sv are both too low by 10 percent of their respective true values, their difference (an estimate of AR at 1 Sv)

is also too low by 10 percent, but their ratio (an estimate of RR at

1 Sv) is unaffected The BEIR I11 Committee (NASMRC, 1980) used a multiplying factor of 1.23 to correct its estimated AR coeffi-

cients, for all solid cancers, for errors of detection Errors of confir- mation, on the other hand, affect relative measures of risk: if the estimated rates at 0 and 1 Sv are both inflated by the addition of, e.g., 12 cases per 100,000 PY due to misclassified benign disease unrelated to radiation dose, the difference between rates is unaf-

fected but their ratio will be too low, provided that the actual RR for cancer is greater than unity Sposto et al (1992) in considering all these factors, concluded that estimates of ERR for total cancer mor- tality in the LSS sample should be multiplied by a factor of 1.13 to adjust for reporting errors It should be remembered that this correction factor is itself based on data analysis, and is uncertain The level of uncertainty is difficult to assess from the published paper because, in their sensitivity analysis, Sposto et al (1992)

concentrated on the problem of misclassification of cancer deaths

as noncancer deaths, estimated from autopsy data to be 22 percent for all ages combined, rather than on the opposite form of

2.3 BL4S IN RISK ESTIMATES / 17

rnisclassification, estimated to be 3.5 percent However, parameter estimates, with standard errors, were given for the probability of misclassification as a function of age a t death Based on the point estimates, the average noncancer to cancer misclassification prob- ability of 3.5 percent corresponds to death a t age 72.2 If the esti- mates are assumed to correspond to normal random variables, approximate 90 percent confidence limits for the misclassification probability a t age 72.2 are 2.6 and 4.8 percent, which correspond to correction factors 1.09 and 1.18, respectively5 Finally, if the correc- tion factor of 13 percent corresponding to a rnisclassification prob- ability of 3.5 percent is interpolated linearly, 90 percent confidence limits for the correction factor of 13 percent are roughly 9.5 and 15.6 percent

For the purposes of the uncertainty analysis, the conversion fac- tor corresponding to misclassification [F(R)], is taken to have a

most probable value of 1.1 and a 90 percent subjective confidence interval from 1.02 to 1.18 The probability distribution of F(R), shown in Rgure 2.2, is assumed to be normal; a standard deviation

of 0.05 corresponds to the 90 percent subjective confidence interval that has been selected

9 h e s e limits were obtained by fitting corrected nonleukemia cancer mortality data from a modified 1950 to 1985 LSS distribution disk obtained from the RERF, stratified by city, sex, age ATB and attained age,

to a dose response model linear in weighted intestinal dose [neutron weight = 10 (as used here, and as used in the publications of the RERF,

"weightn is synonymous with the relative biological effectiveness or RBE)] Let n,and n,, be the observed numbers in a given cell for nonleu- kemia cancer deaths and noncancer deaths, respectively, and let N, and

N,,, denote the corresponding true values, after correction for misclassifi- cation With misclassification probabilities of 22 percent for cancer to non- cancer and 3.5 percent for noncancer to cancer, N, and N,, can be

calculated by solving the following linear equations,

for No and N,, The solution for N , is proportional to (1 - 0.035) n, - 0.035

n,,, which means that the estimated value of ERR,, is independent of the postulated rate of cancer to noncancer misclassification Substituting the value of N, for that of n, in each cell of the data set yielded a regression coefficient of 0.442 compared to 0.392 based on n,, a 13 percent increase The same procedure, using 0.026 and 0.048, respectively, instead of 0.035, gave regression coefficients 0.428 and 0.463

Parameter Magnitude Fig 2.2 Distribution of uncertainties due to misclassification of cancer deaths [F(R)I (The 90 percent confidence interval is shown by the arrows.)

2.4 Biases Affecting Risk Estimates of

Cancer Morbidity

Estimates of radiation induced cancer mortality can also be derived from incidence estimates adjusted by suitable estimates of lethality fractions There are some advantages to this approach, which, of course, requires the availability of incidence data Epide- miological studies at the level of cancer incidence, which are per- formed through the RERF Tumor Registry, include ascertainment

of nonfatal as well as fatal cases Examination of clinical and pathology records, and pathology review of borrowed tissue sam- ples and slides are used to refine case ascertainment, including that of cases identified from death certificates To the extent that such supporting materials are available, the frequency of false pos- itives can be substantially reduced It is also possible, through aggressive pursuit of cases coded to conditions that are often con- fused with the diagnosis of interest (e.g., cancer of the minor sali- vary glands is often miscoded as cancer of the oral cavity), to reduce detection errors by identifying cases mistakenly diagnosed Thus, studies a t the level of incidence are inherently more accurate than mortality-based studies using death certificate diagnoses, in terms

of classification of those cases that come to the attention of the investigators In particular, bias due to false positives is a minor

2.4 BIASES AFFECTING RISK ESTIMATES OF CANCER MORBIDITY / 19

problem in incidence studies compared to estimates based on death certificate diagnoses

There is, however, no national system by which nonfatal cases

of possible cancer can be identified by RERF among members of the LSS sample, many of whom migrated to other parts of Japan at sometime after the sample definition date of October 1,1950 Noti- fication of deaths among members of the LSS sample, and access to their death certificates, is virtually complete because of special arrangements between RERF and the responsible Japanese gov- ernment agencies The RERF Tumor Registry, however, is based upon local tumor and tissue registries run by the medical associa- tions of Hiroshima City, Hiroshima Prefecture, Nagasaki City and Nagasaki Prefecture, which cover current residents of their respec- tive areas A nonfatal cancer case diagnosed elsewhere may be reported to the RERF registry years later, if the patient returns to Hiroshima or Nagasaki and enters (or re-enters) the local medical care system, or the cancer may be mentioned on the patient's death certificate if he or she dies elsewhere in Japan

The problem with using such varied sources of case-finding information is in determining appropriate person or PY denornina- tors Because of migration of LSS sample members throughout Japan and, to a much smaller extent, nonparticipation of some smaller hospitals and clinics in the local registries, there clearly is under-ascertainment of cancer morbidity Migration is known to have been greater among younger survivors compared to older ones, among Nagasaki as compared to Hiroshima residents, and among LSS sample members who were not in the two cities at the time of the bombings, as compared to those who were present (i.e., the "exposed") It does not, however, appear to depend upon radia- tion dose among those who were exposed Thus, estimates of dose-related RR are largely unaffected by migration, whereas esti- mates of AR must be adjusted for underascertainment The solu-

tion favored by RERF is to limit calculations of dose-related AR to

cases diagnosed locally, using PY denominators adjusted to reflect the numbers of sample members actually resident in or near Hiroshima and Nagasaki at the times of diagnosis, and to consider only exposed cases when computing both absolute and RR esti- mates No attempt is made here to evaluate uncertainties in the case where mortality data are derived from incidence data since this is not the basis of the current risk estimates used in radiation protection Incidence data in the LSS (Mabuchi et al., 1994; Preston

et al., 1994; Ron et al., 1994; Thompson et al., 1994) will eventually play a much larger role in radiation protection

be suffering from various privations, including undernutrition and the possible stress (or benefit) of cigarette rationing both before and after the bombings Some were also affected by blast and burn injuries as well as radiation The population also was depleted of healthy men of military age, but this influence may be small, because in the known dependence of risk on age at exposure, young males are about "average" (Table 1.2), consequently, subtracting or adding them to a whole population has only a small effect In fact, the population is unusually well represented by all age groups

I t has been argued also that the survivors were healthier indi- viduals (Stewart and Kneale, 1990) Little and Charles (1990) found some evidence of a selection effect but only in the first follow-

up period of the survivors They also analyzed the possible effects

of the Stewart and Kneale hypothesis assuming it to be true and concluded the effect would underestimate the risk but at most by a small amount in the range 5 to 35 percent The influences of these factors on radiation-related cancer risk years later are unknown, although they seem likely to be small

Thus, the population has certain unrepresentative features but overall these may be less important than those that exist in many other exposed populations (such as those in medical treatment groups) also used for risk estimation

Care must be taken, however, in applying LSS-based risk esti- mates to populations distributed differently with respect to age and sex Risk coefficients often vary significantly by sex and by age a t exposure, and summary coefficients for the LSS population reflect its particular age and sex distribution It is therefore important to apply risk coefficients on an age and sex-specific basis to other pop- ulations when an overall estimate is desired (Land and Sinclair, 1991) and see Table 1.2 given earlier No allowance has otherwise been made here for possible nonrepresentative features of the population

2.6 BIAS D E R M N G FROM CITY DIFFERENCES / 21

2.6 Bias Deriving from City Differences

Another possible source of bias is the fact that the LSS sample consists of survivors from two different types of atomic bombs dropped on two cities, Hiroshima and Nagasaki The survivor pop- ulations in the two cities have different age distributions, Nagasaki survivors were, on the average, 5 y younger than Hiroshima survi- vors They have different baseline cancer rates (depending on site) and are genetically somewhat different [for example, an HLA hap- lotype predisposing carriers to adult T-cell leukemia~lymphoma is markedly more frequent in Nagasaki than in Hiroshima (Preston

et al., 1994)l Also the two cities have had very different histories of postwar industrial and economic development These city differ- ences would not be a serious problem for risk estimation, requiring only that each city be modeled in addition to age, sex, dose, etc., in order to produce separate risk estimates for each city However, the Hiroshima and Nagasaki bombs were of different types (uranium versus plutonium) and construction, produced different radiation spectra, and were exploded over vastly different terrains Any attempt to use the LSS data to estimate the weight of neutrons rel- ative to gamma rays, given that neutrons made up more of the total dose in Hiroshima than in Nagasaki, is necessarily confounded with other city differences, making it difficult to tell to what extent possible neutron effects actually reflect differences in the epidemi- ological data between the cities With the current DS86 dosimetry, neutron doses in both cities are generally so low that estimates of neutron weight (of acceptable uncertainty levels) cannot be derived from the epidemiological data, even ignoring the other city differ- ences (see also Section 3.5) Another problem (Straume et al., 1992)

is that the DS86 probably underestimates the neutron component

at Hiroshima (see Section 3), but by an uncertain amount It seems less likely, but not certain, that the same problem exists a t Nagasaki I t is difficult to use city differences in risk to contribute

to the solution of this dosimetry problem because of the many con- founding factors noted above and because there are other dosime- try problems a t Nagasaki involving the assignment of doses to certain individuals and groups If eventual revisions in the dosim- etry should increase the estimated neutron dose in Hiroshima to the extent that the LSS data could become informative about neu- tron weight, the problem of confounding due to other city differ- ences will need to be addressed

2.7 Summary of Epidemiological Uncertainty

Given accurate radiation doses, complete and accurate ascer- tainment of cases, and a risk model that corresponds to the true dose-response relationship and to the use to which the estimates are to be put, epidemiological uncertainty is adequately repre- sented by confidence limits for sex- and age-specific risk coeffi- cients of interest Bayesian methods (Lindley, 1972) can be used to incorporate externally-derived information about details of the dose response that cannot be estimated adequately from the dose-response data at hand An upward correction factor of 13 per- cent seems appropriate for dose-related ERR for all-site cancer mortality, to adjust for case ascertainment errors, mainly false pos- itives More information is needed about the effect of this adjust- ment upon uncertainty For site-specific cancer mortality, different adjustment factors, which are yet to be determined, will be appro- priate For cancer incidence based on tumor registry data or site-specific incidence studies, any adjustment required should be considerably less than for mortality but such adjustments have not been evaluated here

Thus, for all cancer, the lifetime risk coefficient RHN for acute

exposure, i.e., without DDREF, has a most probable value of

10 x SV-l Statistical uncertainties are expressed by the factor [F(RHN)] which is normally distributed with the 90 percent confi- dence interval from 7.5 to 12.5 x S V - ~ corresponding to a stan- dard deviation of 0.15, as shown earlier in Figure 2.1 There is also

a bias due to ascertainment errors or misclassification which is taken also to have a normal distribution F(R), a most probable value of 1.1 and 90 percent confidence intervals from 1.02 to 1.18

as shown earlier in Figure 2.2 No account will be taken of possible confounding factors including those involved in city differences or the possible unrepresentative nature of the population

Dosimetrical

Uncertainty

3.1 Random Errors and Biases

The random and systematic errors in the presently used dosim- etry system iDS86) (Roesch, 1987) have been estimated to be rep- resented by a coefficient of variation of the order of 25 to 40 percent

in both the DS86 final report itself (Woolson et al., 1987) and also

in later re-evaluations of the uncertainties in DS86 (Kaul, 1989; Kaul and Egbert, 1990~) In the report by Kaul and Egbert (1990)~ the uncertainty model is designed around the calculation of kerma (kinetic energy released per unit mass) in a specific organ of a sin- gle survivor Under equilibrium conditions, kerma and absorbed dose are equal The kerma in question is the sum of kermas from eight radiation components, as follows:

prompt neutron kerma (Drip)

fission product (delayed) neutron kerma illnd)

early (prompt and airlground secondary) gamma-ray kerma (Dm)

fission product (delayed) gamma-ray kerma (D*)

prompt neutron-house secondary gamma-ray kerma (Dhp) delayed neutron-house secondary gamma-ray kerma (Dhd) prompt neutron-body secondary gamma-ray kerma (Dbp)

delayed neutron-body secondary gamma-ray kerma (Dbd)

There is little difference (if any) in the values of kerma and absorbed dose in most organs of the survivors For many purposes

it is convenient to express field quantities such as free in air kerma

or shielded kerma in air, and t o express localized energy deposition

in terms of organ or tissue absorbed dose In the remainder of the

'KAUL, D.C and EGBERT, S.D (1990) "DS86 uncertainty and bias analyses," unpublished (see reference list)

Report, the term "dose" will be used for absorbed dose and ICRP and NCRP define a mean absorbed dose in an organ and base the quantity equivalent dose (ICRP, 1991; NCRP, 1993b) on this mean absorbed dose times the value of the radiation weighting factor, w ~ , for the radiation in question At RERF the term weighted dose is used to describe the gamma-ray absorbed dose plus 10 times the neutron absorbed dose When the neutron component is small this

is approximately equal to an equivalent dose

In the DS86 final report (Roesch, 1987), the most important con- tributors to the total dose are the prompt and delayed gamma-ray doses, Dyp and Dd In recent years, however, measurements have bolstered the case for a significant prompt neutron kerma (Drip) at

Hiroshima, which remains to be confirmed (Straurne et al., 1992;

1994) (see Section 3.6)

The calculation of dose for each organ in DS86 involves the source, a free field, and house-shielding and body-shielding trans- fer functions Uncertainty and possible bias in the shielding, yield, air transport and phantom models receive special attention because of their dominance in the overall uncertainty Table 3.1 presents estimates of the main sources of uncertainty in the doses for Hiroshima and Nagasaki (Kaul and Egbert, 1990)

Random errors are very similar at Hiroshima and Nagasaki, and vary very little as a function of distance from the hypocenter They also vary little if expressed on a relative scale Table 3.2 pre- sents recent estimates of overall uncertainties due to random errors in the assessment of exposures to survivors from both cities (Kaul and Egbert, 1990); for indoor locations, the fracticnal stan- dard deviation (FSD) or coefficients of variation are about 35 per- cent of the means, while the 90 percent confidence intervals indoors, assuming normal distributions, are estimated to be about

60 percent of the means These random errors give rise to bias in the doses causing a need for an upward correction to the risk with increasing dose which is discussed in Section 3.2

3.2 Bias Resulting from Random Errors in

Dose

Most risk estimates are not adjusted for random errors in dose estimates The principal effect of such errors is a downward bias

(Gilbert, 1982; Jablon, 1971; Pierce et al., 1990) in the slope of a fit-

ted linear dose response, i.e., a lowering of the estimated risk coef-

ficient If the random error increases with increasing dose, as in a lognormal error model, a related effect is a bias toward less

Table 3.2-Uncertainties due to random errors in the assessment of exposure for survivors a t Hiroshima

and Nagasaki a t various distances from the hypocenter

(adapted from Kaul and Egbert, 1990)

a Averages of results obtained using several models

Assuming normal distributions

positive, or more negative curvature of the fitted curve (To see this, imagine that the individual doses are measured with less and less accuracy as the dose increases In the limit, measured dose would have nothing to do with the response, the risk of excess cancers, i-e., the response to the measured dose is flat However zero dose would still be zero, resulting in a curve rising steeply from zero excess a t zero dose and quickly flattening out.) This type of bias can in prin- ciple be characterized, given an assumed distribution of the actual dose within the population and, for each possible value of the actual dose a statistical error model for the measured dose Given these assumptions, and a measured dose value, a conditional probability distribution can be constructed for the unknown actual dose under- lying that particular measurement; the correction for a linear response model consists of substituting the mean of that distribu- tion for the measured dose (Pierce et al., 1990) These effects are illustrated by a numerical exercise calculated for this purpose and summarized in Table 3.3 in which RERF cancer mortality data for

1950 to 1985 (RERF, 1990; Shimizu et al., 1988; 1990) were regressed on unadjusted and adjusted dose The left-hand side of the Table pertains to leukemia and bone marrow dose, while the right side pertains to solid tumors and intestinal dose For each, models linear in weighted dose, Dw = D, + WD,, where W in this

example is 1 or 10 or linear in Dw and D: (i-e., quadratic in gamma dose), are summarized as fitted to DS86 or DS86 adjusted dose (i.e.,

adjusted for dose error) over the dose range 0 to 4 Gy (Dy = gamma dose; D, = neutron dose; W = neutron weight, i.e., neutron relative biological effectiveness (RBE); D, = total weighted dose) Thus, the tabulated values also illustrate the effects of dose adjustment on fit and parameter estimates for different dose-response models and neutron weights

This exercise is intended only to show effects of dose adjustment

in some average sense: radiation related tumor risk also depends

on sex, exposure age, attained age and time following exposure The dose adjustment is a matter of scaling, which preserves the ordering of doses within cities and changes i t very little for the com- bined cities For any given model and W value, therefore, there is little difference between the degree of fit obtained using adjusted and unadjusted dose values, as measured by tabulated values for deviance [the lower the deviance, the better the fit (see Glossary)l, although changes in the parameter estimates are evident The dif- ferences in fit between the linear and quadratic models were sub- stantially higher for leukemia, but not for solid cancer, when the analysis was based on adjusted rather than unadjusted dose esti- mates For example, the deviance difference for leukemia (W = 10) between the fitted linear and quadratic models was 849.09 - 840.33

= 8.76 based on unadjusted dose, but 851.29 - 839.99 = 11.30 based

on adjusted dose Moreover, using adjusted dose, the point estimate 5.15 for the parameter P for leukemia, and its lower confidence limit of 0.81, suggest a DDREF higher than the ICRP value of two which would correspond to a P value of about 0.67 if DDREF were determined only by the degree of curvature of the dose-response function for gamma rays For solid cancers as a group, the corre- sponding analysis showed essentially no difference in fit between linear and quadratic models for either unadjusted or adjusted dose for either value of W However, the 90 percent confidence limits for

p using unadjusted doses, (-0.33, 0.72) when W = 1 and (-0.27, 0.69) when W = 10, while consistent with the value P = 0.67 corre- sponding to the ICRP DDREF value of two (P = 0.67, etc.), suggest

a smaller value while the corresponding confidence limits obtained using adjusted dose, (-0.21, 1.59) for W = 1 and (-0.25, 1.55) for

W = 10, are easily consistent with P = 0.67 and with values consid- erably higher

The dose adjustment recommended by Pierce et al (1990)

requires no change in the model for the variance of the estimated coefficient, that is, the adjustment itself is assumed to be without

3.2 BIAS RESULTING FROM RANDOM ERRORS IN DOSE 1 29

error But different parametric assumptions about a lognormal model for dose measurement error would yield different values for adjusted dose, with consequent effects on calculated risk coeffi- cients It would appear that further work is needed to quantify this uncertainty In the meantime, the proposed adjustment substan- tially corrects a known bias, and represents an important improve- ment in our ability to quantify radiation-related risk; the remaining, unresolved uncertainty, while not unimportant, seems relatively small in comparison

The estimated bias errors in the risk estimate to which random errors in the dosimetry give rise, are as follows: for the dose range

0 to 4 Gy, the bias errors cause an overestimate of the dose and an underestimate of the risk by 7 to 11 percent for solid tumors and 4

to 7 percent for leukemia (Pierce et al., 1990) The overall under- estimate of the risk coefficient due to random errors in the dosime- try, F(RE) is taken to be 10 percent on average (3 to 15 percent with

a subjective distribution assumed to be normal, as shown in Figure 3.1, with 90 percent confidence interval of 0 to 20 percent) These uncertainties do not include errors due to the magnitude of the neutron component and its weight, which are considered in Sec- tions 3.5 to 3.7

In addition, other biases associated with the dosimetry system have been identified The primary sources of other bias are the

Parameter Magnitude

Fig 3.1 Distribution of uncertainties in risk from random errors in dosimetry [F(RE)I (The 90 percent confidence interval is shown by the arrows.)

errors in both the gamma ray and neutron fields and the survivor shielding models

3.3 Bias in Gamma-Ray Measurements

Versus DS86

Measurements by thermoluminescence dosimetry of ceramic bricks that were present and exposed at the time of the bombings allow a comparison to be made with estimates of gamma-ray free fields obtained using transport models (Maruyama et al., 1987) There is no substantive evidence that the Nagasaki gamma-ray free field, as estimated by DS86, is subject to any bias On the other hand, it is very likely that the gamma-ray free field at Hiroshima,

as currently described using DS86, systematically underestimates the true value, and that bias increases with distance from the hypo- center (Maruyama et al., 1987) It is probable that revised fission product gamma radiation and air/ground secondary gamma radia- tion calculations may result in very little change in the total gamma-ray free-field kerma at Hiroshima near the hypocenter but

to increase by approximately 20 percent at 1,400 m In this Report, the bias in the gamma-ray free field for the two cities F(Dy) is taken

to range from 1 to 1.4 with a most probable value of 1.1

This value and the distribution of values is triangular as shown

in Figure 3.2 The 90 percent subjective confidence interval ranges from 1.04 to 1.32

Among the criticisms of DS86 is another paper which finds that the gamma-ray kerma (based on biological dosimetry) is overesti- mated at distances shorter than 0.8 k m (Scott, 1994) However, this

is of little relevance for low-dose risk estimation since it refers to doses of 4 Gy and above

It may also be noted that the risk estimates derived from expo- sures at Hiroshima and Nagasaki are for relatively hard gamma rays, in the range 2 to 5 MeV In radiation protection, more biolog- ically effective lower energy gamma, x or beta rays may be encoun- tered No distinction or correction is usually made in radiation protection for differences in effect resulting from differences in the energy of the photons or beta particles However, these differences

in effects in some biological systems can be appreciable (e.g., Bond

et al., 1978; Straume, 1995) although changes with energy just over

1 MeV do not seem to be rapid in the plot given by ICRU Report 40,

3.3 BIAS IN GAMMA-RAY MEASUREMENTS VERSUS DSS6 / 31

Parameter Magnitude

Fig 3.2 Distribution of uncertainties due to bias in gamma-ray dose estimates [F(Dy)] (The 90 percent confidence interval is shown by the arrows.)

Figure 3 (ICRU, 1986) No specific allowance is made for the effects

of gamma-ray energy here

One additional uncertainty relates to the fact that the intestinal dose is used as a surrogate for the actual organ dose for each organ Since cancers of the gastrointestinal tract constitute a large portion (45 percent) of the total risk (ICRP, 1991) this is not an unreason- able choice However, it does introduce an additional uncertainty The two principal components are the prompt and delayed gamma rays In the case of prompt gamma rays (Hiroshima), the intestinal dose has a transmission factor relative to free-field kerma of about 0.80 (Roesch, 1987) and for other deep organs (e.g., bladder, bone marrow) it varies by 26 percent For delayed gamma rays, this range is larger, of the order of k10 percent (Roesch, 1987) For shal- low organs such as breast, the difference is larger too but the con- tribution of the breast to the overall risk is relatively small When considering the exposure to and risk in the whole body, these differ- ences will tend to balance out, consequently, the small additional uncertainty is not accounted for here A more thorough evaluation would be more important in evaluating the uncertainties in the risk coefficients for individual organs

3.4 Uncertainty Due to Survivor Shielding

Characterization in DS86

Doses could be lower or higher by 20 to 50 percent for survivors

in certain classes of shielding (Roesch, 1987) For example, frontal shielding by a building external to the one occupied by the survivor

is only taken into consideration in DS86 if the survivor and the building are separated by less than twice the building height How- ever, more refined calculations indicate that any building in the line-of-sight from survivor to hypocenter provides some shielding Survivors for whom shielding by an external building was not con- sidered may have received doses up to 30 percent less than esti- mated using DS86 This effect is compounded if there were several buildings in the line-of-sight from survivor to hypocenter Also, it seems that the heights of internal walls were overestimated; this correction would lead to a dose increase of 4 to 10 percent for a sin- gle house However, the DS86 model may have overestimated the dose received by the workers a t the Mitsubishi Heavy Industries Plant a t the Nagasaki Shipyard, located approximately 1,700 m from the hypocenter In addition, changes in the shielding condi- tions may have occurred in the first several seconds following the detonation of the bomb; light Japanese houses may have been blown away, or, conversely, the collapse of solid walls may have pro- vided more or less shielding Since delayed radiation from the fire- ball makes a relatively large contribution to the total dose, the loss

or gain of shielding as a result of the blast effect could be important for individuals who were indoors a t the time of the detonation Time-dose dependencies have not, however, been taken into account in DS86 (UNSCEAR, 1988) All of these sources of uncer- tainties in the shielding conditions affect individuals, or categories

of individuals, and it is very difficult to estimate whether there is a bias and what the average bias would be Thus, although additional uncertainty results, in this Report it has been assumed that there

is no average bias in the shielding factor

3.5 Uncertainty Due to Neutron Weight

(Relative Biological Effectiveness)

Changes in the neutron weight (from 1 to 10) have little effect

on the fit (as reflected in the deviance values) of dose-response models to the LSS leukemia and solid tumor data based on DS86 (Table 3.3) That is, these data provide very little information about neutron weight (W) or RBE, whose value must therefore be based

3.5 UNCERTAINTY DUE TO NEUTRON WEIGHT / 33

on other data It must also be emphasized that, with such a small neutron component in DS86 for either Nagasaki or Hiroshima (nominally one to two percent of absorbed dose at Hiroshima, less

at Nagasaki), even large differences in the values chosen for the W will have a relatively small effect on uncertainties in the total weighted dose

Let us consider that the neutron absorbed dose for given organs

at a given location is as estimated in DS86 and consider the contri- bution of errors due to uncertainty in W Note that values of W (i-e., RBE) and their relationship to values of Q (or wR) for neutrons have been reviewed in such texts as Rossi (1977), Sinclair (19851, ICRU (1986), Straume (1988) and NCRP (1990) Table 3-1-1 of Shimizu et al (1988) provides neutron and gamma doses sepa-

rately (for the combined sample, not Hiroshima and Nagasaki alone) For orientation purposes, consider the organ dose range 1.0

to 1.99 Gy: the average gamma dose to bone marrow is 1.369 Gy and the average neutron dose is 0.019 Gy or 1.4 percent of the total absorbed dose A W of 10, which has been a preferred value in recent publications (Thompson et al., 1994; UNSCEAR, 1994),

raises the equivalent dose due to neutrons to 12 percent of the total (0.19 Sv out of 1.559 Sv) and a W of 20 raises the neutron equiva- lent dose to 22 percent of the total (0.38 Sv out of 1.749 Sv), i.e., whether W is 1,10 or 20 causes an uncertainty of about k10 percent about the median value of the weighted dose for the value of W of

10 As noted earlier the weighted dose approximates the equivalent dose

Note that some estimates of weighted dose have been made by considering a W which is an increasing function of declining neu- tron dose (Shimizu et al., 1988; 1990) This subject has been dis- cussed in some detail with regard to the incidence data by Thompson et al (1994) The changes made by these more precise

approaches to the problem are minor compared with a simple appli- cation of W = 10 Thus, it appears appropriate to apply a neutron W

of 10 throughout the range of interest and an assumed uncertainty

of +I0 percent in the total equivalent dose due to this assignment

whether the slope of the linear portion of the gamma-ray curve is reduced by larger contributions from the neutrons (Straume, 1996) This possibility is dependent on the potential presence of more neu- trons a t Hiroshima, a possibility that we have treated separately in Section 3.6 Until some of these questions are clarified by ongoing new studies and measurements, at which time a greater range of uncertainty in W may need to be considered, it seems appropriate

to proceed as indicated earlier

Thus, finally, in this Report, the uncertainty in the nominal equivalent dose due to choice of W is taken into account by multiplying by a factor F(NR) ranging from 0.9 to 1.1 with a most likely value of one (for W = 10) The distribution of values a t F(NR)

is assumed to be triangular, as shown in Figure 3.3 The 90 percent subjective confidence interval ranges from 0.93 to 1.07

3.6 Bias and Uncertainties Due to the

Presence of Thermal Neutrons at Hiroshima

in Excess of Those Predicted by DS86

I t was known prior to the completion of DS86, that there were some measurements of 6 0 ~ o resulting from neutron activation in steel that did not agree well with the calculations of neutron flu- ence used in DS86 (Chapter 5, Figure 3, Vol 1 of Roesch, 1987) The discrepancies between calculation and measurement were appar- ently present a t both Hiroshima and Nagasaki Subsequent exam- ination of the 6 0 ~ o points a t Nagasaki has resulted in revisions of the uncertain locations of the measurements which seem to cloud whether a discrepancy actually existed for that city More recent measurements with 3 6 ~ 1 (Straume et al., 1994) indicate good agree- ment with calculations for Nagasaki, although the measurements

do not extend beyond 1,261 m slant range The neutron dose a t Nagasaki is about 0.5 percent or less of the gamma dose For Hiroshima, calculation and measurement of neutron activation and therefore presumably of neutron dose are in agreement at

800 m, but differ by up to a factor of 10 at about 1,600 m (Figure 3.4) (Straume et al., 1992) One scenario is that a modified Hiroshima spectrum (actually closer to a true fission spectrum) which will cause some increase in the estimated fast neutron kerma must be invoked in order to account for this Precisely what combination of spectrum change and increase in fast neutron kerma will be required has not yet been determined and the dosim- etry committees of the United States and Japan have yet to make

a recommendation