Missing Doses in the Life Span Study of Japanese Atomic Bomb Survivors

Abstract

The Life Span Study of atomic bomb survivors is an important source of risk estimates used to inform radiation protection and compensation. Interviews with survivors in the 1950s and 1960s provided information needed to estimate radiation doses for survivors proximal to ground zero. Because of a lack of interview or the complexity of shielding, doses are missing for 7,058 of the 68,119 proximal survivors. Recent analyses excluded people with missing doses, and despite the protracted collection of interview information necessary to estimate some survivors' doses, defined start of follow-up as October 1, 1950, for everyone. We describe the prevalence of missing doses and its association with mortality, distance from hypocenter, city, age, and sex. Missing doses were more common among Nagasaki residents than among Hiroshima residents (prevalence ratio = 2.05; 95% confidence interval: 1.96, 2.14), among people who were closer to ground zero than among those who were far from it, among people who were younger at enrollment than among those who were older, and among males than among females (prevalence ratio = 1.22; 95% confidence interval: 1.17, 1.28). Missing dose was associated with all-cancer and leukemia mortality, particularly during the first years of follow-up (all-cancer rate ratio = 2.16, 95% confidence interval: 1.51, 3.08; and leukemia rate ratio = 4.28, 95% confidence interval: 1.72, 10.67). Accounting for missing dose and late entry should reduce bias in estimated dose-mortality associations.

Editor's note: An invited commentary on this article appears on page 569, and the authors' response appears on page 574.

Atomic bombs were detonated over Hiroshima and Nagasaki, Japan, on August 6 and August 9, 1945, respectively. A cohort of survivors was enumerated on October 1, 1950, and has been followed since to ascertain vital status and cause-of-death information. This study, referred to as the Life Span Study of Japanese atomic bomb survivors, has served as the primary quantitative basis for estimates of cancer risk after external ionizing radiation exposure in occupational, environmental, and medical settings. Risk estimates derived from the study can be found in reports by organizations such as the US National Academy of Sciences, the International Commission on Radiological Protection, and the United Nations Scientific Committee on the Effects of Atomic Radiation (1–3).

A large effort was undertaken to derive dose values for these Japanese atomic bomb survivors. This effort included interviews in the 1950s and 1960s with survivors who were within approximately 2 km of the hypocenters in Hiroshima or Nagasaki to ascertain detailed information on location and shielding conditions needed to assign a dose to proximal survivors. However, for reasons that included a lack of interview and the complexity of shielding, we are missing the dose for 7,058 of 68,119 proximal survivors. Recent analyses of radiation dose-mortality associations excluded people with missing dose information and considered the start of follow-up to be October 1, 1950, even for cohort members who could not be assigned a dose until their interviews were conducted years later (1–4). In the present study, we describe the prevalence of missing doses and its association with mortality, distance from hypocenter, city of residence, age, and sex and discuss its potential impact on estimates of dose-mortality associations.

MATERIALS AND METHODS

The Life Span Study encompasses follow-up of 120,200 survivors of the August 1945 atomic bombings of Hiroshima and Nagasaki, Japan. The 68,119 of those survivors who were within 3 km of the hypocenters are called proximal survivors. The 25,550 who were between 3 km and 10 km from the hypocenters are called distal survivors. Another 26,531 people who were not present in Hiroshima or Nagasaki at time of bombings were included in the cohort, but these people were not considered in the current analysis or in most other recent major reports on mortality rates in this cohort (1–4). Note that these people not only survived the bombings but were also alive on October 1, 1950, when the first post-war national census of Japan was conducted; the October 1950 census served as the primary basis for enumeration of the people considered in the current analysis. Information on location at the time of bombing, which was used to demarcate proximal from distal survivors, was also necessary to select the Life Span Study cohort. This information was ascertained by the Atomic Bomb Casualty Commission; the most notable source of such information was the 1949 Atomic Bomb Casualty Commission census of bomb survivors, which provided information on location for approximately 180,000 survivors.

Follow-up of the cohort to ascertain vital status and cause of death was initially done retrospectively by field investigators. Since the 1960s, follow-up has been conducted through triennial searches of death records. We analyzed deaths from all causes, all cancers, and leukemia from October 1, 1950, through December 31, 2000.

The most recent quantitative estimates of radiation dose for cohort members are referred to as “DS02 doses” (henceforth, doses) (5). These dose values are intended to represent the prompt external gamma and neutron doses from the atomic bombings; individual radiation doses from neutron activation and fallout have not been quantified for cohort members. The primary focus of the current analysis is whether the dose was missing.

Although distance from hypocenter is known for all cohort members, detailed information on location and shielding conditions was needed to assign a dose to proximal survivors. This information was ascertained through interviews with survivors. Death or migration before collection of the necessary information or sufficient complexity of shielding conditions reported in the interview led to missing doses for some proximal survivors. Doses for distal survivors were based on estimates of dose-in-air and did not require detailed location and shielding information from such interviews; distal survivors contributed person-time and events solely to the lowest dose category.

The current analysis uses a publicly available tabulation of persons, person-time, and deaths due to all causes, all cancers, and leukemia, with cross-classification by city (Hiroshima or Nagasaki), sex, attained age (in 5-year intervals), age at exposure (in 5-year intervals), calendar time (1950–1952, 1953–1955, and then in 5-year intervals, the final categories being 1991–1995 and 1996–2000), location at the time of the bombing (within 3 km or 3–10 km from the hypocenter), and dose (missing, <0.005 Gy, 0.005 to <0.02 Gy, 0.02 to <0.04 Gy, 0.04 to <0.06 Gy, 0.06 to <0.08 Gy, 0.08 to <0.1 Gy, 0.1 to <0.125 Gy, 0.125 to <0.150 Gy, 0.150 to <0.175 Gy, 0.175 to <0.2 Gy, 0.2 to <0.25 Gy, 0.25 to <0.3 Gy, 0.3 to <0.5 Gy, 0.5 to <0.75 Gy, 0.75 to <1 Gy, 1 to <1.25 Gy, 1.25 to <1.5 Gy, 1.5 to <1.75 Gy, 1.75 to <2 Gy, 2 to <2.5 Gy, 2.5 to <3 Gy, and ≥3 Gy). The number of persons at risk is reported for the first calendar time period only. For each cell of the cross-classification, the person-year–weighted average values for distance from hypocenter, attained age, and age at exposure are reported for each cell of the table.

Statistical methods

Using the information on the number of persons at risk at the start of follow-up (assumed to be October 1, 1950, even for proximal survivors who did not yet have an interview), a file was created with one record for each person in the study cohort. Each person was classified with respect to city, sex, age at time of the bombings, and whether the dose was missing. Prevalence ratios and Wald-type 95% confidence intervals were calculated. Each person was assigned a distance from the hypocenter based upon the person-year–weighted average distance from hypocenter for the first cell in the data tabulation to which the person contributed follow-up time. Histograms were used to illustrate distributions of distance from the hypocenter for survivors with missing and nonmissing doses, and summary statistics were derived for each distribution.

The full tabulations of person-time and events were used to fit log-linear Poisson regression models to quantify associations between missing dose and mortality due to all causes, all cancers, and leukemia. Estimates of the associations were adjusted for potential confounding by sex, city, attained age, and age at exposure by background stratification (6). All statistical analyses were conducted using the SAS statistical package (7).

RESULTS

City, sex, and age at the time of the bombing were associated with having a missing dose. The prevalence ratio for a missing dose was 2.05 for person from Nagasaki relative to those from Hiroshima and 1.22 for males relative to females (Table 1). Missing doses were most common for survivors who were 10–19 and 20–29 years of age at time of the bombing (Table 1). Missing doses were more common among Nagasaki survivors than among Hiroshima survivors within each stratum of age (Appendix Table 1).

Table 1.

Association of City of Residence at Time of Bombing, Sex, and Age at Time of Bombing With Missing Dose^a, Life Span Study, Japan, 1950–2000

	No. of Survivors With a Missing Dose	No. of Survivors With a Nonmissing Dose	PRb	95% CI
City
Hiroshima	3,442	58,494	1.00	Referent
Nagasaki	3,616	28,117	2.05	1.96, 2.14
Sex
Female	3,771	50,924	1.00	Referent
Male	3,287	35,687	1.22	1.17, 1.28
Age at time of bombing, years
0–9	862	17,833	1.00	Referent
10–19	2,442	17,563	2.65	2.46, 2.85
20–29	1,144	10,891	2.06	1.89, 2.25
30–39	887	12,270	1.46	1.33, 1.60
40–49	908	13,503	1.37	1.25, 1.50
50–59	538	8,960	1.23	1.11, 1.36
60–69	218	4,422	1.02	0.88, 1.18
≥70	59	1,169	1.04	0.81, 1.35

Abbreviations: CI, confidence interval; PR, prevalence ratio.

^a There were 93,669 people who were present in Hiroshima or Nagasaki at time of bombings and were included in mortality follow-up.

^b Estimates were obtained from a binomial regression model that included a binary indicator variable for city, a binary indicator variable for sex, or 7 indicator variables for the 8 categories of age at exposure shown.

Over the study period (October 1, 1950–December 31, 2000), people with missing doses had higher adjusted cancer mortality rates (rate ratio (RR) = 1.08, 95% confidence interval (CI): 1.01, 1.15) and leukemia mortality rates (RR = 1.20, 95% CI: 0.82, 1.76) than did people with nonmissing doses (Table 2). The magnitudes of the mortality rate ratios comparing those with missing doses relative to those with nonmissing doses was highest among people 60–69 years of age at the time of bombing; in that age-at-exposure group, persons with missing doses had higher adjusted rates of mortality due to all causes (RR = 1.21, 95% CI: 1.05, 1.38), all cancer (RR = 1.49, 95% CI: 1.01, 2.20), and leukemia (RR = 4.95, 95% CI: 0.58, 42.61) than did people with nonmissing doses, though the estimate for leukemia was highly imprecise (Appendix Table 2). When considering variation in the relative rate by calendar period, the excess mortality among those with missing doses relative to those with nonmissing doses was greatest during the first calendar periods of follow-up (October 1, 1950–December 31, 1952 and January 1, 1953–December 31, 1955); during those periods, people with missing doses had higher adjusted rates of mortality due to all causes, all cancer, and leukemia than did people with nonmissing doses (Table 2). Of the 543 proximal survivors who died from cancer from 1950 to 1955, 13% (71 deaths) were missing dose estimates. Of the 39 leukemia deaths observed in the period 1950–1955 among proximal survivors, 26% of the decedents (10 deaths) were missing a dose.

Table 2.

Estimates of Adjusted Mortality Rate Ratios by Calendar Period Comparing Survivors With and Without a Missing Dose^a, Life Span Study, Japan, 1950–2000

Calendar Period	Cause of Death
	All Causes		All Cancers		Leukemia
	RRb	95% CI	RR	95% CI	RR	95% CI
1950–1952	1.50	1.29, 1.74	2.16	1.51, 3.08	4.28	1.72, 10.67
1953–1955	1.26	1.09, 1.46	1.51	1.07, 2.15	1.95	0.56, 6.81
1956–1960	0.95	0.84, 1.08	1.07	0.81, 1.43	0.63	0.15, 2.65
1961–1965	0.94	0.84, 1.06	1.00	0.77, 1.30	0.49	0.07, 3.66
1966–1970	1.00	0.89, 1.12	1.15	0.92, 1.44	1.73	0.49, 6.14
1971–2000	0.99	0.94, 1.03	1.03	0.95, 1.12	0.95	0.55, 1.66
1950–2000	1.01	0.98, 1.05	1.08	1.01, 1.15	1.20	0.82, 1.76

Abbreviations: CI, confidence interval; RR, rate ratio.

^a Mortality due to all causes, cancer, and leukemia among 93,669 people who were present in Hiroshima or Nagasaki at time of bombings.

^b Adjusted for attained age, city, sex, and age at exposure.

Although it is not possible to directly investigate whether the proportion of missing doses is differentially distributed with respect to the dose value that would have been assigned (if the doses were nonmissing), it is possible to indirectly evaluate such an association because distance from hypocenter is known for all cohort members and is correlated with dose. Figures 1 and 2 illustrate the distribution of survivors with respect to distance from hypocenter, stratified by city (Figure 1 shows data for Hiroshima and Figure 2 shows data for Nagasaki) and by whether the dose was missing. Those proximal survivors with a missing dose tended to be closer to the hypocenter than did proximal survivors with a nonmissing dose. Among Hiroshima survivors, none of the 16,113 distal survivors had missing doses, whereas doses were missing for 3,442 of 45,823 proximal survivors. The mean distance from the hypocenter among those with a missing dose was 1,523 m (standard deviation = 86 m); the minimum and maximum distances for survivors with a missing dose were 1,019 m and 2,564 m, respectively. Among Nagasaki survivors, dose was missing for 3,616 of 22,296 proximal survivors. The mean distance from the hypocenter among those with a missing dose was 1,663 m (standard deviation = 132 m); 1 survivor with a missing dose had a distance less than 1,000 m, and the minimum and maximum distances for survivors with missing doses were 953 m and 2,477 m, respectively.

Figure 1.

Distribution of Hiroshima survivors with respect to distance from hypocenter, Life Span Study, 1950–2000. There were 42,381 proximal Hiroshima survivors with a nonmissing dose (A), 16,113 distal Hiroshima survivors with a nonmissing dose (B), and 3,442 proximal Hiroshima survivors with a missing dose (C).

Figure 2.

Distribution of Nagasaki survivors with respect to distance from hypocenter, Life Span Study, 1950–2000. There were 18,680 proximal Nagasaki survivors with a nonmissing dose (A), 9,437 distal Nagasaki survivors with a nonmissing dose (B), and 3,616 proximal Nagasaki survivors with a missing dose (C).

DISCUSSION

Although the Life Span Study of atomic bomb survivors is the major source of quantitative estimates of cancer risk from ionizing radiation, little attention has been given to the potential impact of missing doses. Recent major Life Span Study analyses have handled missing doses by exclusion (1–4). Work on causal diagrams has contributed to the understanding of the conditions under which the effect of interest can be estimated without bias by analyzing only the people with complete data, sometimes called a complete-case analysis (8, 9). For example, if exposure is the only cause of a missing dose, then a complete-case analysis will yield an unbiased estimate of the association of interest because the analysis excluding people with a missing dose is equivalent to a random sample within strata of exposure. However, if there are unmeasured common causes of missing doses and death (Figure 3A), excluding people with a missing dose will induce a bias path between exposure and mortality. In the Life Span Study, mortality differences between participants with and without missing doses suggest that such factors might include unmeasured causes of morbidity or mortality before shielding interview (because a missing dose may occur if a person fell ill or died before their shielding interview) that are associated with later mortality risk.

Figure 3.

Directed acyclic graphs illustrating some scenarios in which a complete-case analysis leads to biased estimate of an exposure-disease association. E denotes exposure, D denotes death, R denotes a binary indicator of missing dose (R = 0 indicates missing dose, otherwise R = 1), U denotes unmeasured causes of R and D, and V denotes unmeasured causes of R and E. A complete-case analysis is one that excludes people with missing doses; the restriction to R = 1 is indicated by a box. A) Restriction to R = 1 conditions on a collider; this induces an association between E and U. Complete-case analysis will be biased. B) Restriction to R = 1 induces an association between V and U. Complete-case analysis will be biased.

Figure 3B illustrates a scenario in which both exposure and death are associated with having a missing dose. Having a missing dose is associated with exposure due to a common cause of exposure and missing dose. One reason for such an association in the Life Span Study is that doses for distal survivors were assigned without detailed interview-based shielding information, whereas such information was required for assignment of doses for proximal survivors (10). A common cause of missing dose and the exposure of interest, therefore, is proximity to ground zero. For some proximally exposed individuals, the dose is missing simply because the assignment of dose has been regarded as too complicated or too uncertain, and hence the dose was treated as missing; for example, 205 Nagasaki factory workers are missing doses because of the complexity of shielding (9). Dose is not missing for distal survivors. In Figure 3B, as in Figure 3A, death is associated with missing dose through a shared common cause. Under the scenario illustrated in Figure 3B, conditioning on missing dose may induce bias in estimates of the association between the exposure of interest and mortality.

Logistical requirements limited shielding interviews to survivors exposed within 2,200 m of the hypocenter in Nagasaki and within 2,000 m in Hiroshima. In the face of a large backlog of interviews for Hiroshima survivors, an effort was made to interview all subjects who were less than 1,200 m from the hypocenter and a 10% random sample of those exposed at or beyond 1,200 m. When interviews with persons who were less than 1,200 m away were finished, a 100% sampling was extended up to those who were 1,300 m away and so on, ending at 1,600 m. The shielding interview program was terminated in 1965 (5). Because missing doses occurred among proximal survivors (particularly in the range of 1,500–1,600 m from hypocenter, for which dose values tended to be moderate), excluding people with a missing dose primarily may impact the magnitude of risk estimates in the moderate-dose range. Although the Life Span Study is large, the exclusion of observations as in a complete case analysis may impact radiation risk estimates of interest, particularly for findings regarding death from leukemia.

The protracted collection of information necessary to assign doses to proximal survivors has implications for considering the date of start of follow-up for cohort members. Because one criterion for inclusion in recent analyses of this cohort is a nonmissing dose and because the necessary location and shielding information were collected in interviews that continued into the 1960s, there is the potential for immortal person-time accrued among proximal survivors. In principle, follow-up for person-time at risk of the study outcome should not be included if that person-time precedes an event required for entry into the study population or satisfaction of an exposure definition. The exposure categories that include immortal person-time will yield downwardly biased mortality rates (11). Distal survivors would not have this problem, because interviews were not required to assign dose to them. Such immortal person-time during the early years of follow-up could influence the shape of the dose-cancer associations; examination of changes in dose-mortality trends with follow-up has been a topic of recent interest in analyses of this cohort (12, 13).

Given the suggestion that early deaths cause missing doses, it is tempting to consider excluding the early period of follow-up from analysis of the Life Span Study. However, dealing with missing dose by excluding all person-time and events for the early years of follow-up restricts what could be learned about radiation-mortality associations from these data (to a period of observation relatively distant in time after exposure to radiation). Moreover, as illustrated in Figure 2A and 2B, when exposure is associated with missing dose, a bias path is induced in an analysis that conditions on missing dose if there is an unmeasured common cause of missing dose and mortality during the period of follow-up under study. Such bias operates even if the analysis excludes the first calendar periods at risk. Although selective survival before enumeration of the Life Span Study cohort in 1950 has received attention as a potential source of bias (14, 15), it is plausible that a similar bias is produced by missing doses.

Missing doses are relatively most common among survivors who were 10–29 years of age at time of bombing in both cities and particularly in Hiroshima (Appendix Table 1). These are the ages at which mortality rates are low and migration related to work and marriage would have been relatively common in the 1950s and early 1960s. Migration before interview could also be a cause of missing dose, which may suggest selection related to low mortality risk if migrants tended to be healthier than nonmigrants (12, 16). Death before interview to collect detailed information on location and shielding for proximal survivors and inability to complete an interview due to illness are reasons for missing dose that could be especially important for older participants, who show the highest rate ratios of mortality from all causes, all cancers, and leukemia (Appendix Table 2). If missing dose is relatively more common among healthy out-migrants and workers at younger ages and relatively more common among unhealthy persons at older ages, the direction of bias in dose response estimates due to missing dose could differ by age at exposure.

Several analytical methods have been proposed for analysis of longitudinal data with missing information, including multiple imputation, maximum likelihood methods, and inverse probability of missing data weighting (17–20). The validity of such methods depends upon the correctness of unverifiable assumptions, and therefore sensitivity analyses would be appropriate to characterize the impact of model assumptions regarding missing data on dose-response estimates. Among the assumptions used in the Poisson regression analyses in the present article (similar to what was seen in other recent analyses of these data) is independence of outcomes as person-time is censored by competing risks. The issue of competing risk of death due to other causes is another topic for further attention in analyses of these data. In the present article, we describe the problem of missing doses but do not report radiation dose-mortality associations obtained using contemporary methods to account for missing data in longitudinal analyses. Available data for this cohort are limited to tabulations of person-time and events; analyses making use of detailed information, such as individual-level covariate and outcome information, date of interview, or an indication of whether a person's dose is missing due to complexity of shielding conditions, mortality, morbidity, or migration before interview, are not possible with these publicly available data.

In summary, inferences from the Life Span Study regarding mortality are complicated by missing data. A tremendous effort has been made to address technical issues regarding the radiation dosimetry systems used to estimate dose values for members of this cohort. However, major recent updates of the dosimetry system have not substantially reduced the number of people with missing doses (4). Life Span Study analyses that exclude survivors with missing doses and start follow-up of all survivors on October 1, 1950, should be expected to yield biased estimates of the radiation-mortality association (particularly for the early years after enumeration of the cohort). These considerations underscore the challenges in estimating the radiation effects on health and mortality rates of the Japanese atomic bomb survivors, particularly in the first decades after the atomic bombings. By design, the Life Span Study does not permit evaluation of radiation risks in the 5 years after the atomic bombings of Hiroshima and Nagasaki, and inferences regarding radiation effects on mortality during the study period may be impacted by selective survival into the cohort (14, 15). In contrast to the challenges of selection bias, however, the problems of late entry and missing data that are the focus of the current paper are tractable using contemporary methods for analysis of longitudinal data, suggesting a clear path forward to address these issues in future analyses of radiation dose-response associations and attributable risk.

APPENDIX

Appendix Table 1.

Association Between Missing Dose and Age at Time of Bombing, Life Span Study, Japan, 1950–2000

Age at Exposure, years	Hiroshima Survivors (n = 61,936)					Nagasaki Survivors (n = 31,733)
	No. of Survivors With Missing Dose	%	No. of Survivors With Nonmissing Dose	PRa	95% CI	No. of Survivors With Missing Dose	%	No. of Survivors With Nonmissing Dose	PR	95% CI
0–9	239	2.1	10,989	1.00	Referent	623	8.3	6,844	1. 00	Referent
10–19	1,172	9.7	10,922	4.55	3.97, 5.22	1,270	16.1	6,641	1.92	1.76, 2.11
20–29	626	7.6	7,574	3.59	3.10, 4.15	518	13.5	3,317	1.62	1.45, 1.81
30–39	491	5.2	8,966	2.44	2.09, 2.84	396	10.7	3,304	1.28	1.14, 1.45
40–49	452	4.4	9,725	2.09	1.79, 2.44	456	10.8	3,778	1.29	1.15, 1.45
50–59	299	4.5	6,344	2.11	1.79, 2.50	239	8.4	2,616	1.00	0.87, 1.16
60–69	131	4.0	3,145	1.88	1.52, 2.32	87	6.4	1,277	0.76	0.62, 0.95
≥70	32	3.7	829	1.75	1.22, 2.51	27	7.4	340	0.88	0.61, 1.28

Abbreviations: CI, confidence interval; PR, prevalence ratio.

^a Model: log(P(dose = missing) = age at time of bombings.

Appendix Table 2.

Estimates of Adjusted Mortality Rate Ratios by Age at Time of Bombing Comparing Survivors With and Without a Missing Dose^a, Life Span Study, Japan, 1950–2000

Age at Time of Bombing, years	Cause of Death
	All Causes		All Cancer		Leukemia
	RRb	95% CI	RR	95% CI	RR	95% CI
0–9	1.06	0.87, 1.30	1.27	0.92, 1.75	1.00	0.24, 4.18
10–19	1.03	0.94, 1.12	1.07	0.93, 1.24	0.67	0.27, 1.67
20–29	0.95	0.86, 1.04	0.98	0.83, 1.15	1.86	0.93, 3.73
30–39	1.01	0.93, 1.09	1.13	0.98, 1.30	1.60	0.75, 3.40
40–49	0.98	0.92, 1.05	1.01	0.87, 1.17	1.10	0.39, 3.06
50–59	1.01	0.93, 1.11	1.15	0.94, 1.42	N/A	N/A
60–69	1.21	1.05, 1.38	1.49	1.01, 2.20	4.95	0.58, 42.61
≥70	1.13	0.87, 1.47	1.09	0.26, 4.48	N/A	N/A

Abbreviations: CI, confidence interval; N/A, not applicable; RR, rate ratio.

^a Mortality due to all causes, cancer, and leukemia among 93,669 people who were present in Hiroshima or Nagasaki at time of bombings.

^b Adjusted for attained age, city, sex, and age at exposure.