11
Findings and Discussion

Organization

The core question of this study concerns whether exposure to ionizing radiation at Operation CROSSROADS increased mortality. For reasons described earlier (and discussed again later) in this report, we approached this question from three directions. First, hypothesis A compares all participants to all nonparticipant controls to answer whether being a participant at CROSSROADS is associated with mortality. To hone in on whether ionizing radiation might be a causal factor in the relative mortality experiences of participants and controls, we then divide participants into boarding and nonboarding groups to test the hypothesized ionizing radiation exposure gradient as hypothesis B. Finally, based on the Defense Nuclear Agency (DNA) assertion that Navy personnel assigned to Engineering & Hull occupational specialties may have been exposed to radiation via pipes carrying contaminated water (Appendix B), we also examine, as hypothesis C, the association of Engineering & Hull status (as a radiation exposure measurement surrogate) with mortality.

We present the findings in this section from these three directions of inquiry:



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 62
--> 11 Findings and Discussion Organization The core question of this study concerns whether exposure to ionizing radiation at Operation CROSSROADS increased mortality. For reasons described earlier (and discussed again later) in this report, we approached this question from three directions. First, hypothesis A compares all participants to all nonparticipant controls to answer whether being a participant at CROSSROADS is associated with mortality. To hone in on whether ionizing radiation might be a causal factor in the relative mortality experiences of participants and controls, we then divide participants into boarding and nonboarding groups to test the hypothesized ionizing radiation exposure gradient as hypothesis B. Finally, based on the Defense Nuclear Agency (DNA) assertion that Navy personnel assigned to Engineering & Hull occupational specialties may have been exposed to radiation via pipes carrying contaminated water (Appendix B), we also examine, as hypothesis C, the association of Engineering & Hull status (as a radiation exposure measurement surrogate) with mortality. We present the findings in this section from these three directions of inquiry:

OCR for page 62
--> • Model A. BASIC To test effect of participant status on mortality. Independent variables are: participant status (1 = yes; 0 = no) age on 1 July 1946 (continuous) paygrade categories: junior enlisted (baseline) mid-level enlisted senior enlisted officer • Model B. BOARDER GRADIENT To test effect of boarding status (see Chapter 10) as a stronger radiation exposure surrogate than participant status alone, the model substitutes for the participant status variable two indicator variables creating a hypothesized gradient of radiation exposure: nonparticipant control (baseline) nonboarding participant (level 1 exposure) boarding participant (level 2 exposure) • Model C. ENGINEERING & HULL To test effect of the Engineering & Hull occupational category as a radiation exposure surrogate. Includes parameters of the basic model (A), less participant status—participants and controls are tested separately. A variable for Engineering & Hull status is added: Engineering & Hull status (1 = yes, 0 = no) Major mortality endpoints are: (1) all-cause, (2) all-malignancies, and (3) leukemias and aleukemias, excluding chronic lymphoid leukemia (CLL). Other descriptive mortality rates are presented by level of aggregation within cause of death (e.g., selected major categories, selected causes within selected major categories, etc.). In developing the models displayed, we considered alternative modeling of variables, the role of other available data elements, and potential interactions among exposures and personnel characteristics. The tests that revealed no information are not displayed in these tables but are described later in this section. In Tables 11-1 through 11-8 we display rate ratios calculated from estimated proportional hazards parameters for the first two models (A, participation effects, and B, boarding effects), using survival time as the response variable, censored as necessary at the end of the study follow-up period (31 December 1992). Results of Model C (Engineering & Hull effects) are presented separately in Tables 11-9. Because occupational specialty, of which the Engineering & Hull designation is one, is available only for enlisted personnel, we exclude officers from the analysis in Model C.

OCR for page 62
--> Findings: Displayed in Tables Tables 11-1 through 11–8 display the rate ratios computed from the Cox proportional hazards model comparing participants with nonparticipants. As noted above, Model A considers all participants as an exposed group. Model B attempts to measure a hypothesized exposure gradient, with boarding participants as the most likely to be exposed to radiation hazard, nonboarding participants as less likely to be so exposed, and nonparticipants (controls) as unexposed. The display includes all tested causes of death. Model C generated rate ratios in Table 11-9 associated with Engineering & Hull status for participants and controls separately, when included in a model along with age and paygrade.

OCR for page 62
--> TABLE 11-1. Rate Ratios of Major Mortality Endpoints for Exposed to Not-Exposed Personnel     A: Basica B: Boarder Gradienta   Case Definitionb No. of Deaths All Participantsc Nonboarding Participantsd Boarding Participantse All causesf,g 22,847h 1.046 (1.020–1.074)i 1.043 (1.015–1.073) 1.057 (1.014–1.102) All malignanciesf,g 5,647 1.014 (0.962–1.068) 1.010 (0.955–1.068) 1.026 (0.943–1.116) Leukemia +aleukemiaf 163 1.020 (0.750–1.387) 1.024 (0.737–1.422) 1.007 (0.610–1.663) Notes for Tables 11-1 through 11-8: a Analyses based on 73,704 Navy participants and controls. b Table 10-9 in Chapter 10 provides complete category names and ICD9 codes. c All participants relative to all controls, adjusting for age and paygrade. d Non-boarding participants relative to all controls; adjusting for age and paygrade. e Boarding participants relative to all controls, adjusting for age and paygrade. f An a priori hypothesis based on ionizing radiation literature. g Subsets in later table. h Records of 12 deaths for which a necessary variable was missing or out of range were dropped from this analysis. i Rate ratios from SAS PHREG procedure for proportional hazard ratios with 95 percent confidence intervals; boldface if confidence interval does not contain 1.00 (statistically significant at .05 level)

OCR for page 62
--> TABLE 11-2. Mortality Rate Ratios from Six Selected Major Cause-of-Death Categories (Subset of Table 11-1)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Benign neoplasm 63 0.969 (0.591–1.590) 1.065 (0.635–1.785) 0.654 (0.254–1.683) Circulatory disease 8,447 1.011 (0.968–1.055) 1.012 (0.967–1.059) 1.008 (0.940–1.080) Digestive disease 1,005 1.089 (0.962–1.234) 1.088 (0.953–1.242) 1.094 (0.898–1.332) External cause 2,519 1.025 (0.948–1.109) 1.009 (0.927–1.097) 1.079 (0.955–1.219) Malignant neoplasmf,g 5,647 1.014 (0.962–1.068) 1.010 (0.955–1.068) 1.026 (0.943–1.116) Respiratory disease 1,174 1.070 (0.954–1.201) 1.053 (0.931–1.191) 1.129 (0.943–1.352)

OCR for page 62
--> TABLE 11-3. Mortality Rate Ratios from Remaining Major Cause-of-Death Categories (Subset of Table 11-1)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Bood-forming 37 1.221 (0.637–2.340) 1.353 (0.690–2.653) 0.769 (0.224–2.642) Congenital 24 0.910 (0.408–2.028) 0.778 (0.317–1.905) 1.377 (0.444–4.276) Endocrine 317 1.155 (0.925–1.441) 1.248 (0.990–1.574) 0.848 (0.572–1.258) Genitourinary 181 0.876 (0.654–1.172) 0.899 (0.659–1.228) 0.795 (0.480–1.317) Ill-defined 431 0.769 (0.636–0.930) 0.769 (0.626–0.944) 0.770 (0.559–1.059) Infectious 200 1.045 (0.792–1.380) 0.984 (0.728–1.330) 1.254 (0.827–1.903) Mental 152 0.963 (0.701–1.324) 0.927 (0.657–1.309) 1.084 (0.661–1.775) Musculoskeletal 0 — — — Nervous 236 1.107 (0.856–1.430) 1.105 (0.841—1.452) 1.112 (0.738—1.675) Skin 16 1.209 (0.450–3.247) 1.223 (0.429–3.488) 1.162 (0.241–5.597)

OCR for page 62
--> TABLE 11-4. Mortality Rate Ratios from Selected Cancer Sites within All Malignancies (Table 11-2)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Bladder 90 0.964 (0.638–1.458) 0.868 (0.550–1.369) 1.292 (0.708–2.358) Bone 18 0.912 (0.362–2.299) 0.922 (0.343–2.479) 0.876 (0.189–4.060) Brain/CNS 164 0.827 (0.609–1.125) 0.771 (0.550–1.081) 1.013 (0.634–1.618) Buccal 159 0.983 (0.720–1.342) 1.040 (0.749–1.443) 0.792 (0.455–1.377) Digestiveg 1,247 1.021 (0.913–1.141) 1.039 (0.923–1.170) 0.959 (0.798–1.153) Eye 137 1.100 (0.786–1.540) 1.096 (0.766–1.569) 1.112 (0.651–1.901) Kidney 0 — — — Lymphopoieticg 480 0.975 (0.815–1.166) 0.909 (0.747–1.105) 1.198 (0.915–1.567) Prostate 292 0.767 (0.609–0.966) 0.796 (0.623–1.018) 0.668 (0.441–1.011) Respiratoryg 2,354 1.036 (0.956–1.124) 1.039 (0.953–1.133) 1.028 (0.902–1.171) Skin 128 0.817 (0.577–1.157) 0.870 (0.602–1.257) 0.643 (0.340–1.217) Testicular 24 0.763 (0.341–1.705) 0.814 (0.347–1.909) 0.593 (0.134–2.631) Thyroid 9 3.479 (0.720–16.81) 3.248 (0.628–16.80) 4.235 (0.593–30.24)

OCR for page 62
--> TABLE 11-5. Mortality Rate Ratios for Selected Cancer Sites within Digestive, Respiratory, and Lymphopoietic Cancers (Subset of Table 11-4)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Esophagus 177 1.165 (0.866–1.568) 1.227 (0.898–1.677) 0.962 (0.582–1.588) Hodgkin's Disease 45 0.748 (0.415–1.347) 0.536 (0.264–1.090) 1.444 (0.673–3.097) Large intestine 422 0.931 (0.769–1.127) 0.946 (0.771–1.160) 0.883 (0.642–1.215) Leukia+aleuk.f 163 1.020 (0.750–1.387) 1.024 (0.737–1.422) 1.007 (0.610–1.663) Liver 55 1.492 (0.866–2.572) 1.533 (0.866–2.712) 1.355 (0.576–3.188) Lung 2,252 1.048 (0.965–1.139) 1.056 (0.966–1.153) 1.024 (0.895–1.170) Lymphos.+Reticul. 71 0.998 (0.627–1.591) 0.767 (0.449–1.313) 1.787 (0.973–3.283) Other lymphaticg 169 1.021 (0.755–1.381) 0.998 (0.721–1.380) 1.100 (0.686–1.763) Pancreas 240 1.114 (0.864–1.436) 1.122 (0.856–1.470) 1.085 (0.720–1.635) Rectum 88 0.844 (0.556–1.283) 0.781 (0.493–1.237) 1.060 (0.561–2.002) Stomach 194 1.038 (0.783–1.376) 1.139 (0.849–1.529) 0.703 (0.413–1.196) TABLE 11-6. Mortality Rate Ratio for Selected Subset of Other Lymphatic Tissue (Subset of Table 11-5)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Multiple myeloma 65 0.893 (0.549–1.453) 0.978 (0.588–1.628) 0.608 (0.237–1.559)

OCR for page 62
--> TABLE 11-7. Mortality Rate Ratios for Subset of External Causes (Subset of Table 11-2)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants All accidentsg 1,600 0.975 (0.884–1.076) (0.865–1.069) 1.019 (0.873–1.190) All suicides 525 1.030 (0.867–1.223) 0.958 (0.795–1.156) 1.262 (0.980–1.624)

OCR for page 62
--> TABLE 11-8. Mortality Rate Ratio for Subset of All Accidents (Subset of Table 11-7)     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Motor vehicle accidents 778 0.929 (0.807–1.070) 0.922 (0.792–1.073) 0.953 (0.762–1.192) TABLE 11-9. Rate Ratios for All-Cause Mortality of Engineering & Hull to Non-Engineering & Hull Groups by Status as Participants or Controls Case Definitionb No. of Deaths Engineering & Hull PARTICIPANTSa,j No. of Deaths Engineering & Hull CONTROLSa,j All causesf 10,811 0.995 (0.953–1.038)h 9,562 0.979 (0.936–1.024) All malignanciesf 2,620 1.051 (0.965–1.144) 2,403 1.032 (0.945–1.127) Leukemia+aleukemiaf 74 1.515 (0.940–2.442) 69 1.292 (0.786–2.125) Notes for Table 11-9: a Analyses based on 66,872 Navy enlisted personnel, participants and controls. b Table 10-9, Chapter 10 provides complete category names and ICD9 codes. j Enlisted personnel in Engineering & Hull specialties relative to enlisted personnel in other occupational specialties, adjusted for age and paygrade only. f An a priori hypothesis based on ionizing radiation literature.

OCR for page 62
--> Narrative Summary of Findings Participant Status All-Cause Mortality (Table 11-1) The participants in CROSSROADS had a 4.6 percent higher mortality in comparison to the nonparticipants (p = .0006), relative risk (RR) = 1.046, 95% confidence interval, 1.020–1.074. Age at CROSSROADS and paygrade confirmed known mortality risks: increasing age and decreasing paygrade are associated with increasing mortality. Mortality from All Malignancies (Table 11-2) Participants experienced a slightly higher (1.4 percent) mortality from malignant neoplasms, RR = 1.014 (0.962–1.068), but this could be due to chance (p = 0.2579). Mortality from Leukemia (including leukemias and aleukemias, excluding CLL) (Table 11-5) Participant mortality was 2 percent higher than for the comparable controls RR = 1.020 (0.750–1.387). However, these results could again have been due to chance (p = 0.8992). Mortality from Other Selected Causes (Tables 11-2 through 11–8) The excess in all causes of mortality did not appear to be concentrated in any one of the subcategories of mortality studied, and none of the subcategory differences was statistically significant. We note that both nonmalignant respiratory disease and respiratory cancer showed elevated but not significant relative risks. In addition, ill-defined causes and prostate cancer showed significantly decreased mortality risks of 0.769 (0.636–0.930), p = 0.0067, and 0.767 (0.609–0.966), p = 0.0243, respectively, but this could be a result of chance, given the large numbers of comparisons made.

OCR for page 62
--> TABLE 11-10. Mortality Rate Ratios of Participants Relative to Controls (Marines, n = 1,137) Case Definition No. of Deaths All Participants All causes 346 1.059 (0.857–1.308) All malignancies 81 1.711 (1.092–2.680) Leukemias 4 0.358 (0.037–3.489) The increased risk of death from malignant neoplasms should be explored in a larger cohort to confirm this apparent increase. The standardized mortality ratios for all-malignancies were 0.5858 (0.3796–0.7921) for Marine controls and 0.9524 (0.6801–1.225) for the participants. For the complete Navy cohort they were: for controls, 0.8156 (0.7846–0.8465), and for participants, 0.8196 (0.7895–0.8497). The Marine controls appeared significantly healthier than either the Navy participants or the Navy controls. The apparent increase in allmalignancy mortality may be a deficit in the control population; the small size of the Marine cohort may exaggerate instabilities in the data. We cannot dismiss the possibility that this is a chance finding among the many comparisons made in this analysis. More definitive information will be available at the conclusion of the ongoing Five Series Study, which has a Marine cohort 10 times the size of the one in CROSSROADS. TABLE 11-11. Mortality Rate Ratios of Participants Relative to Controls (Army, including Army Air Corps, n = 6,482) Case Definition No. of Deaths All Participants All causes 2,454 0.754 (0.694-0.818) All malignancies 542 0.777 (0.652-0.926) Leukemias 15 0.775 (0.270–2.219) Deficits in both all-cause mortality and all-malignancy mortality are statistically significant. Preliminary explorations of Army data suggest problems with the control cohorts assembled. Although we based control selection on the rank distribution of the participant cohort, we did not have the education information available that would permit a finer selection. Because Army officers were selected for Operation CROSSROADS to perform high-level technical and scientific tasks, they were probably more highly educated than a random set of Army officers—a characteristic associated in other studies with better health outcomes. Thus, there could well be a mismatch in education level between Army participants and controls. A comparison of the two groups might yield an apparent and erroneous impression that CROSSROADS participation was associated with better health, a ''protective'' effect. While this may be true of Navy officers also, the effect would be larger in the Army, where 28.3 percent of the cohort were officers, compared to the 8.7 percent in the Navy.

OCR for page 62
--> In addition to these problems, the Army at the time of CROSSROADS included the Army Air Corps. There is some ambiguity in the rank structure between the mid-level and senior-level Army enlisted personnel. As a result there is an imbalance in these ranks between the control and participant cohorts for a relatively small number of men. Finally, we are still missing dates of birth for approximately 21 percent of the Army participants. Because we have imputed missing dates from randomly selected cohort members with the same rank and paygrade, misclassifications in rank assignments could influence the age variable. For all these reasons, we are not well assured that the findings for the Army subcohort of this study are valid. We present them here for completeness. Discussion Boarders Navy personnel who, according to records, boarded target ships during the CROSSROADS period following either or both detonations are the most clearly identifiable group of participants exposed to ionizing radiation on contaminated target ships. Hence, as we discussed in Chapter 8, those 8,996 participants represent a more highly exposed surrogate group. If radiation following the atmospheric nuclear tests at CROSSROADS influenced the subsequent mortality of participants, we would expect its influence to be most concentrated among the boarders. A dose-response relationship could be hypothesized with controls at no dose, nonboarding participants at some dose, and boarding participants at a higher dose. (It also is possible that "boarding" could have entailed other, unidentified risks that affected later mortality. We are not aware of such factors, nor have others proposed examples.) Such a dose-response relationship is not observed in these data. A clear indication of a radiation effect would be a gradient with the somedose group (participant nonboarders) having significantly higher risk than the no-dose group (controls) and the higher-dose group (participant boarders) having significantly higher risk than the some-dose group. This lack of association could indicate: (1) radiation from CROSSROADS detonations did not affect mortality, (2) the precision in assigning boarder status was inadequate to ensure accurate exposure classification, making a finding of no association uninformative, or (3) risks were present but were too small to be detected in a cohort of this size. Boarders did, however, show increased mortality risk estimates relative to the control cohort—though not attaining statistical significance—in the ICD9 categories for Lymphosarcoma and Reticulosarcoma (RR = 1.787 [0.973–3.283]) and Hodgkin's Disease (RR = 1.444 [0.673–3.097]), both within the

OCR for page 62
--> broader category of lymphopoietic cancers. When comparing the approximately 9,000 boarders to the controls and the nonboarding participants combined, 26 the rate ratios increase to 2.447 for lymphosarcoma and reticulosarcoma and 2.971 for Hodgkin's Disease, both reaching statistical significance. This difference in risk estimates reflects the inclusion in the comparison group of the much lower risk nonboarding participants. Possible explanations—none of which can be tested within the available data—are: (1) radiation is associated with increased risk of lymphosarcoma and reticulosarcoma and Hodgkin's Disease; (2) some unidentified nonradiation exposure associated with boarder status is also associated with increased mortality risk for those cancers; or (3) the boardercancer association is observed by chance, quite possible in this study with the large numbers of comparisons considered. While earlier medical literature on radiation and cancer posed a possible relationship for these types of cancers (Upton 1982), a recent compilation of radiation effects specifically states that these types of cancer are not likely to be radiogenic (NRC 1990). Pierce (1996) found no significant dose effect for lymphoma in the atomic bomb survivor cohort. However, clinicians and researchers have identified problems with these diagnostic categories and created new disease classification schemes by which to describe types of lymphomas. These newer classifications replacing lymphosarcoma and reticulosarcoma use both histologic and morphologic criteria. We cannot determine what revised diagnoses might be coded from the observed lymphoma deaths in our dataset using the newer, more descriptive terminology. It is possible that these cancers would then fit into categories considered radiogenic; we may have identified a hitherto undescribed association or, as mentioned earlier, this may be a chance finding. Engineering & Hull In its attempt to assign doses to CROSSROADS participants, DNA considered personnel assigned to Engineering & Hull occupations as a potential high exposure group (see Chapter 8 and Appendix D for more detail). We tested whether that dichotomy was associated with mortality among enlisted personnel (these assignment data were not available for officers). Our analyses were designed to test whether Engineering & Hull personnel had higher mortality, after mathematically adjusting for paygrade and age. There were no statistically significant elevations of mortality from all causes, malignancies, or leukemias in the Engineering & Hull group of participants compared to the non-Engineering & Hull group. More important, comparisons 26    In the model comparing the 9,000 boarders to the controls and nonboarding participants combined, lymphosarcoma and reticulosarcoma and Hodgkin's Disease were the only 2 of 44 causes of death tested to reach statistical significance at the 0.05 level.

OCR for page 62
--> within the controls also yielded similar and nonsignificant results. These findings do not support a hypothesis that Engineering & Hull personnel experienced higher mortality due to radiation exposure at CROSSROADS. This lack of association could indicate: (1) radiation from CROSSROADS detonations did not affect mortality, (2) Engineering & Hull status was an inadequate exposure surrogate for radiation dose, making a finding of no association uninformative, or (3) risks were present but were too small to be detected in a cohort of this size. Discussion of Other Findings The tables provided earlier in this section display some risk estimates that, while not achieving statistical significance, may be useful topics for discussion and further exploration. We display in Table 11-12 the risk estimates that indicate at least a 40 percent excess (RR ≥ 1.4) or a 40 percent deficit (RR ≤ 0.6) risk related to CROSSROADS participation or boarder status. Most of these are based on small numbers of deaths and are not statistically significant. It should be noted that when many comparisons are made, a few will show a large rate ratio by chance alone. By mathematical definition, a random 5 percent of numerous comparisons might show statistical significance at the 0.05 level regardless of biological causation, with about half of the point estimates being raised and the other half lowered. Similarly, some numerical comparisons may indeed represent true associations but not reach statistical significance. By the standards of scientific reporting, these associations would not qualify as study findings. However, understanding the limitations of inadequate statistical power and the importance of many different studies observing the same nonstatistically significant patterns, we choose to call attention to some of the observed data. Of the eight rate ratios noted at the 40 percent increased or decreased risk thresholds, seven were elevated. If we eliminate the two "all participants" relative risks, considering them duplicative of the boarder and nonboarder data, we still have five elevations out of the six independent rate ratios considered. One might entertain the possibility that this is a nonrandom pattern. We note with interest the thyroid cancer risk ratios. First, thyroid cancer is known to be radiogenic. Second, it is rarely fatal, so the incidence of thyroid cancers would not be reflected in the mortality data on which this study is based. Third, because there are only nine thyroid cancer deaths identified among CROSSROADS participants and controls, statistical comparisons do not have the power to pick up a statistically significant difference. Yet, the risk estimate for participants relative to controls is the highest found in this study. As discussed in the preceding paragraph, however, chance cannot be eliminated as an explanation.

OCR for page 62
--> Other Models and Parameters Considered During Analysis As mentioned in the opening section of this chapter, we considered—using all-cause mortality—various interactions and combinations of variables in the building of the statistical models upon which we based this report. Here, we describe attempts at modeling and decisions made regarding paygrade, age, and Engineering & Hull status, and the interactions of participant status with each of them. We use age as an independent variable to adjust the data for variations in age in the study cohorts. We considered a model in which both age-at-shot and age-at-shot-squared were variables. The estimated parameter coefficient for the age-at-shot squared variable was not statistically significant; we did not include the squared term in the final models. We examined all-cause mortality, all-cancer mortality, and leukemia mortality in models testing a range of time-related variables. Neither time-sinceshot nor age-at-shot interacted with exposure variables (participant status, boarder status, Engineering & Hull assignment) to produce statistically significant or otherwise observable relationships. Using the LIFETEST program in SAS to analyze survival times in participants and controls, we looked at data by age-at-shot strata, again finding nothing statistically significant and observing nothing that seemed to indicate time-dependent differences between the participants and the controls. We checked whether the use of imputed dates of birth might have affected the results. Chapter 9 describes the procedure we had followed to impute missing dates of birth, using matched rank and paygrade. The all-cause mortality model run without these imputed dates, thereby excluding those 1,447 records (and 186 deaths) from the analysis, did not reveal different information. See Table 11-13.

OCR for page 62
--> TABLE 11-12. Nonsignificant Risk Estimates from Tables 11-4 and 11-5 that Indicate (boldface*) at Least a 40 percent Excess or Deficit in Mortality     A: Basic B: Boarder Gradient   Case Definition No. of Deaths All Participants Nonboarding Participants Boarding Participants Thyroid cancer 9 3.479 (0.720–16.81) 3.248 (0.628–16.80) 4.235 (0.593-30.24) Hodgkin's Disease 45 0.748 (0.415–1.347) 0.536 (0.264–1.090) 1.444 (0.673–3.097) Liver cancer 55 1.492 (0.866–2.572) 1.533 (0.866–2.712) 1.355 (0.576–3.188) Lymphosarcoma and reticulosarcoma 71 0.998 (0.627–1.591) 0.767 (0.449–1.313) 1.787 (0.973-3.283) * In this table (distinct from earlier tables) the boldfaced values mark rate ratios ≥1.4 or ≤0.6. TABLE 11-13. All-Cause Mortality Rate Ratios, Including and Excluding Records with Imputed Dates of Birth   Model Aa Model Bb   Imputed Dates of Birth All Participants Nonboarding Participants Boarding Participants Including imputed DOBs 1.046 (1.020–1.074)c 1.043 (1.015–1.073) 1.057 (1.014–1.102) Excluding imputed DOBs 1.051 (1.024–1.079) 1.052 (1.023–1.082) 1.049 (1.006–1.094) a Model A compares all participants to all nonparticipants, adjusting for paygrade and age at shot. b Model B uses gradient of potential exposure and compares nonboarding participants and boarding participants to all nonparticipants (controls), adjusting for paygrade and age at shot. c Boldface represents statistical significance at p < 0.05 level.

OCR for page 62
--> We looked at various ways of handling occupation, paygrade, and potential interactions of paygrade and participant status and boarder status and Engineering & Hull. These did not yield informative results; the earlier described models were not changed. Standardized Mortality Ratios Analysis of the data using standardized mortality ratios (SMRs) yields results consonant with the findings we have discussed thus far in this section. Please refer to Chapter 8 and Appendix C for rationale and detailed results. In this study for the Navy, all-cause SMRs were 0.8715 for participants and 0.8322 for controls. For all malignancies and leukemias, they were, respectively, 0.8196 and 1.004 for participants and 0.8156 and 1.079 for controls. Discussion Rrelating these Findings to those of Similar Studies The data and findings we present in this report of CROSSROADS mortality are consistent with findings of earlier studies of nuclear test participants. Findings in Other Studies of Atomic Veterans In 1988, Darby and colleagues reported on a study of 22,347 men involved in the British nuclear test program from 1952 to 1967. The participants for the study were identified from a multitude of records by the British Ministry of Defense (MOD) and are distributed among the services as follows: 29 percent Navy, 27 percent Army, 40 percent Air Force, and another 4 percent civilian. In determining the completeness of the participant list, the investigators identified 2,121 "independent respondents" who were found through solicitation of sources apart from MOD (e.g., veterans' groups) and were verified as participants. Of these, 1,707 were found to be on the main study list, suggesting a completeness of 83 percent. Using only the main study list (MOD-generated), Darby et al. reported mortality of participants relative to controls of 1.01 (90% confidence interval, 0.95–1.07). However, when adjusted for the independent respondents who were not included in the main study, all-cause mortality increased to 1.05 (90% confidence interval, 0.97–1.13). The CROSSROADS study differs from the Darby et al. study in that it covers only one nuclear test series whose nuclear detonations were completed within one month, six years before the beginning of the British testing period. The CROSSROADS cohort on which we base our analysis and conclusions is entirely Navy, but is 73 percent larger than the all-service Darby study cohort.

OCR for page 62
--> We estimate (see Appendix E) that the CROSSROADS participant roster is 93–99 percent complete. The apparent difference to Darby's 83 percent completeness is attributable mostly to differing U.S. and U.K. policies in assembling the lists. The identification methodology used by Nuclear Test Personnel Review (NTPR) includes self-identified participants who subsequently were verified as test participants through official records. The Darby list, however, is purely a MOD records-based compilation; it does not include self-identified participants. Thus, it is likely that our study is more similar to the Darby analysis with adjustments for missing participants than to the main study. Given that, our findings of a 1.046 (95% confidence interval, 1.02–1.074) relative risk for all-cause mortality in the participants over the controls seems to be in consonance with the Darby results. In 1993, Darby et al. completed a second analysis of British atomic veterans that added an additional seven years of follow-up. At that time the relative risk for all-cause mortality was still near unity—0.99 (90% confidence interval, 0.95–1.05). In this study, Darby et al. did not report all-cause mortality for the main study adjusting for men not included in the main study. They did, however, study the "independent respondent" group separately and determined that the self-identified participants who were not found in the main MOD study list did have elevated mortality for both all causes and all cancers. This they attributed to a self-selection effect. Watanabe, Kang, and Dalager (1995) reported finding increased relative risk for all-cause mortality in a study of 8,554 U.S. Navy personnel who participated in the HARDTACK nuclear test in 1958. Their study used the same NTPR database as this one, but dates of birth were not available for controls and many participants and relative risks were adjusted only for rank. Their comparison group of similar Navy nonparticipants numbered 14,625. In participants with an assigned dose of less than 2.5 mSv (0.25 rem), they observed a relative risk for all causes of death of 1.09 (95% CI, 0.98–1.21, n = 3345); for 2.5–10.0 mSv (0.251–1.00 rem), 1.08 (95% CI, 0.98–1.19, n = 4115); and for greater than 10 mSv (1.00 rein), 1.23 (95% CI, 1.04–1.45, n = 1094). Overall, the crude rate ratio for all-cause mortality in the study was 1.10 (95% CI, 1.02–1.19). The all-cause relative risk of 1.047 that we have observed is in consonance with these findings. Selection Bias as an Explanation for Increased Participant All-Cause Mortality Is it possible that a selection effect of having "independent respondents" included in our study by virtue of the NTPR process is alone responsible for our finding of a 5 percent increase in mortality for the participants? It is possible, but unlikely based on the discussion that follows. We have no data with which

OCR for page 62
--> to determine exactly what fraction of the NTPR data set is self-identified. However, we estimate that we are missing between 1 and 7 percent of the participants. That allows us to estimate the magnitude of the self-identification problem. Let us assume that all of the deaths among the "missing participants" would have appeared in the NTPR data set as a result of self-reporting after the onset of disease or from reporting by a member of the family during a posthumous filing of a claim for compensation. In other words, we assume our present number of deaths includes those deaths that would otherwise have gone uncaptured in a completely unbiased participant data set. Thus, our count of deaths would accurately reflect all participant deaths, but our count of participants would be low by 1–7 percent (approximately 400–2,700 people). If all these assumptions are true, our risk estimate for participants has used a correct numerator with a too-small denominator and is therefore biased upward. Were we to correct for that—by adding 400–2,700 people to the participant denominator and leaving the numerator as it is—we would get a 1–7 percent lower risk estimate. This is again consistent with Darby's findings. The number of CROSSROADS personnel missed by the NTPR is thus critical to establishing the magnitude of the self-selection bias. In Appendix E we reported several estimates for missing participants. The highest of these was based on the match of our participant list to the National Association of Atomic Veterans (NAAV) list of CROSSROADS participants. All inexact matches were declared as missing participants, leading to a missing rate of 7 percent. If we adjust the crude mortality rate based on our assumptions above (i.e., that the deaths for the 7 percent missing participants are actually in our mortality count), the crude mortality in the participants drops by 7 percent, effectively wiping out the increased rate we have reported in the participants. The data on missing participants from the write-in study (see Appendix E) suggested a much lower missing rate of 3.7 percent; which would also reduce the crude participant death rate by 3.7 percent. After completing more detailed research on each of the suspected missing participants, we determined that from the NAAV list we would estimate only 1.5 percent participants to be missing and 1.1 percent to be missing if we use the write-in data. Given those estimates, our crude rates would drop only 1–2 percent in the participants if the self-selected deaths, who would otherwise have been missing, were removed. The latter two estimates are probably more accurate than the former two and suggest that, at most, half of the excess relative risk found in the participants for all-cause mortality would be attributable to self-selection in the NTPR process. Furthermore, one would expect at least as large a self-selection effect on cancer mortality as on all-cause mortality. After all, preparation of claims for cancers thought to be a consequence of test participation would be one of the primary reasons a veteran would have for contacting the NTPR and becoming a member of the atomic veteran cohort. In this study, the all-malignancy relative

OCR for page 62
--> risk, 1.014 (95% CI 0.962–1.068), was not statistically different from the all-cause relative risk, 1.046 (95% CI 1.020–1.074), and indeed the point estimate was smaller. Thus, the cancer mortality data do support the argument that the all-cause relative risk was elevated in the participants largely due to self-selection bias. In summary, selection bias may have contributed to some degree to our finding of increased relative risk for all-cause mortality among CROSSROADS participants. Given the completeness of the participant roster and the lack of a concomitant increase in all-malignancy risk, however, it is unlikely that this bias accounts for all of the increase observed. Avenues of Further Exploration Because the value of the work invested in this study of CROSSROADS may not be fully apparent, we close this report with our views of what has been accomplished and what can still be accomplished using this extraordinary database. We now have a dataset with the records of 80,000 individuals who served in the U.S. Navy in 1946. Most typographic and administrative recording errors have been corrected; missing data have been, when appropriate, estimated, with documentation to provide justification; the participant cohort is 93–99 percent complete; vital status follow-up across almost 50 years is around 90 percent—this is approaching, in fact, a total follow-up rather than a study sample. Many conceptual and practical limitations recede with complete mortality ascertainment in a cohort. At some point, regrettably, all of the participants at CROSSROADS will have died. Even if that is another 50 years from now, we should maintain the data resource and update vital status follow-up. A study of the complete mortality experience of the CROSSROADS participant and comparison group cohorts would carry minimal marginal cost relative to what has been invested so far. We hope that resources could be provided to maintain, permanently document, and update this database. Questions that might suitably be asked of these data—with supplemental information collected as necessary—are: What more can we learn about the associations military paygrade and rank have with mortality? Does short-term versus career military participation provide clues? To what extent might career-mortality associations be the result of selection decisions about military service, career success and promotion, or occupational exposures? What more can we learn about the mortality experience of the Army and Marine personnel who participated in Operation CROSSROADS? Using a larger sample, an IOM committee and staff are now building a comparison group to study the military participants in five other U.S. atmospheric test series.

OCR for page 62
--> It would be informative to consider test-specific duties as well as rank/occupation of Army and Marine personnel in building an appropriate comparison cohort, as the CROSSROADS study was able to do in significant numbers for Navy personnel. What can we learn by analyzing, in addition to the underlying causes used in this report, the associated causes of death listed on death certificates for the CROSSROADS participants and controls? Might there be clues to reporting bias? Might there be information that gives us a better understanding of the morbidity experience of these individuals? Finally, can the contribution of a self-selection bias in the participant cohort be quantified? And, how best can we control for it in ongoing and future studies of atomic veterans?