Data Collection

Data for this study were collected from multiple sources over a period of time. This section describes the data acquired from various data sources.

Cohort Population

The study population consisted of 440,892 workers at eight (six public and two private) shipyards conducting overhauls of nuclear-powered vessels: Charleston Naval Shipyard, Charleston, South Carolina; General Dynamics Corporation, Electric Boat Division, Groton Connecticut; Mare Island Naval Shipyard, Vallejo, California; Newport News Shipbuilding and Drydock Company, Newport News, Virginia; Norfolk Naval Shipyard, Norfolk, Virginia; Pearl Harbor Naval Shipyard, Pearl Harbor, Hawaii; Portsmouth Naval Shipyard, Portsmouth, New Hampshire; and Puget Sound Naval Shipyard, Bremerton, Washington. At the public shipyards, workers were identified either from personnel records, Form W2 lists, and/or from Office of Personnel Management records for the period 1957 through 2004. For the two private shipyards, the workers were identified from hard-copy or computerized personnel records for the same period. Overall, 2,955 workers (0.7%) were excluded because they were lacking critical information such as gender or social security number. The final number of workers included in the study cohort was 437,937.

Although the Navy's occupational monitoring program began in the 1940s, the cohort follow-up period began in 1957, as this was the beginning of nuclear-powered ship overhauls and significant radiation exposure in shipyard radiation workers. For all radiation workers included in the follow-up, who had exposure prior to 1957, their prior dose was provided for inclusion in the study.

All 437,937 workers, including workers initially employed from 1957 to 1981 in the original study (4), workers employed from 1957 to 1981 but not in the original study sample, and all subsequent employees newly-hired from 1981 through 2004, were followed up for deaths from January 1, 1957 to December 31, 2011. Because non-radiation workers were required to have had at least one year of employment to be eligible for inclusion in the study, their follow-up start dates were one year following their hire date or January 1, 1957, for those employed more than one year prior to 1957, whichever was later. There was no one-year-employment requirement for radiation workers since all had previous non-radiation employment; thus, their follow-up start date was the first date of radiation work or January 1, 1957, whichever was later. For each worker, the follow-up ended on the date of death or December 31, 2011, whichever was earlier. In this case, their follow-up started when they became a radiation worker; the employment duration prior to their radiation job was counted in their total employment duration, but not included in their follow-up period, the period during which the deaths were collected. The radiation dose-response analysis was done on radiation workers only, using those workers who have 0 mSv time dependent occupational radiation dose as the reference group, in addition to radiation vs. non-radiation workers comparison.

Collection of Death Data

In the original cohort enrolled through 1981, all sampled workers were followed for vital status (alive, dead, or unknown) through records of the U.S. Social Security Administration, the U.S. Civil Service, the U.S. Health Care Financing Administration, and the Veteran's Benefits Administration, as well as through mortality data from some states and the eight shipyards (4). Causes of death were determined from reviewing and coding of death certificates by the study nosologist. For the expanded study, additional death information through 2011 for all workers in the original cohort who were still living in 1981, all workers employed from 1957 to 1981 but not in the sample of the original study, and all subsequent employees newly hired from 1981 through 2004 was obtained from four sources: the National Death Index, Social Security Administration mortality records, a private organization called Pension Benefits Information, and state mortality databases. Causes of death were coded according to the International Classification of Diseases (ICD) 9th or 10th Revision depending on the year of death. The primary causes of deaths were used in this study.

LHC Grouping

LHC and its types were classified based on ICD coding of deaths using modified grouping criteria from the National Cancer Institute's Surveillance, Epidemiology, and End Results Program categories (33) as shown in Table 1. LHC and its types including lymphoid leukemia (acute, chronic, and other), myeloid leukemia (acute, chronic, and other), and leukemia other than lymphoid or myeloid type, multiple myeloma, lymphomas, and a type of myeloproliferative neoplasms and myelodysplasias were examined separately. Analyses of LHC types were undertaken to identify the specific LHCs that are associated with low-dose, chronic radiation exposure. The current study has included leukemia less CLL as an LHC type in analyses for comparison to other studies.

TABLE 1

Lymphohematopoietic Cancer (LHC) Grouping

Radiation Worker Selection and Exposure History

For this study, the radiation worker population were personnel engaged in shipyard work that could have resulted in potential exposure to radiation and therefore required special training and dosimetry, in contrast to, the non-radiation worker population. The non-radiation worker population were employed by one of the shipyards but did not perform any radiation specific work. Radiation workers were individually monitored by one of three dosimetry systems, comprised of either film or thermoluminescent dosimeters, since the onset of the Navy's dosimetry program, which began in the 1940's. A typical dosimetry issue period in the shipyards was weekly, but other issue periods were used based on need, dosimetry processing cycle or job. Regardless, of the issue period, all dose was background corrected, using background dosimeters, and used to determine the cumulative annual occupational dose to the worker. It is these annual occupational doses that were supplied for each radiation worker in the population. For those shipyard workers that had dose prior to 1957, all exposure information dating back to the origin of the program in the 1940s was included. Radiation doses used in analyses were time-dependent cumulative doses for each year of follow-up. Thus, all radiation doses prior to 1957 were added to the annual cumulative doses for these workers, but no additional person-times were added because no deaths were collected prior to 1957. In other words, the annual accumulative doses in each year of follow-up included doses prior to 1957 and all cumulative doses from 1957 to the year of the cumulative dose calculation without adding additional person-times prior to 1957. All exposure records were considered complete (i.e., no missing dose) as any missing dose (lost dosimeter, exposed without a dosimeter, bad read, etc.) was investigated at the time of the occurrence, and an estimate of dose determined, if required, and entered into the individual's exposure record.

As three distinct dosimetry systems were used over the course of the study period, a comprehensive comparison study was performed as each radiation measurement system was replaced. These studies were conducted to evaluate the consistency of dose determination from one system to the next. These study reports are available, to include the manuals and procedures specific to each dosimetry system. Since all eight shipyards included in this study are under the cognizance of the Naval Nuclear Propulsion Program (NNPP), they historically collected and recorded occupational dose in a similar manner (34–36).

The primary source of occupational radiation dose to shipyard workers is external exposure to gamma radiation emitted by activated corrosion products, primarily cobalt-60. Most of the exposure comes from inspection, maintenance, and repairs inside the reactor compartments of nuclear-powered ships (4, 16). External exposure to neutron radiation was negligible, since overhauls generally occurred after reactors were shut down and reactor shielding designs ensured exposure from neutrons during reactor operation were much less than that from gammas(37). Internal exposure was also negligible as only 0.2% of all shipyard workers have ever had a recorded internal exposure. Finally, for all radiation workers that had prior non-NNPP dose (commercial reactor plant, radiographer, etc.) or transferred to another shipyard, all pre-existing exposure was collected from their prior employer, and all historical doses were transferred to the new shipyard location (38).

Job History

The occupational code and description of the last job held by everyone in the study were collected through their site's personnel records. Of 437,937 workers in the cohort, 96.7% had their last job specified, while only 3.30% (14,443) did not. In this study, the detailed job groups were not used in analysis, since the annual radiation doses based on personal monitoring records were available. However, those who worked as painters, welders, machine operators, woodworkers, or electricians and those who worked in transportation jobs were grouped into a “solvent related job” category as one of the covariates to control for, since solvents might be a potential confounding factor to radiation risk associated with LHC (39). Since there was no solvent measurement or rank available, the definition of “solvent related job” might not be precise enough to examine the risk of mortality due to LHC related to solvent exposure. Rather, it was used as a rough control for one of the covariates that might bias the risk analysis of LHC with radiation exposure if not included in the model.

Other Covariates

Personal demographic characteristics adjusted in analyses included gender, race, time dependent attained age, and age at first hire. Race was categorized as white, black, other known race, or unknown race. Workers with missing race were categorized as a separate group of “unknown race” in the analyses. This race group was composed of workers with mixed races. Because of the significant proportion of workers with unknown race (25.7% for radiation workers and 22.4% for non-radiation workers) (Table 3), race could only serve as a controlling covariate for analyzing the LHC mortality risk with radiation. However, the association of the LHC risk with race itself, if any, observed in this study would not be valid due to the unknown distribution of races. Socioeconomic status was coded based on an individual's pay grade, education, skill set, and responsibilities related to the worker's last job as described in the original study (4). The categories of socioeconomic status included blue collar, professional, technical, and administrative. Those whose socioeconomic status could not be classified into these categories were put into an “other” category. These variables were treated either by stratification or by inclusion in models directly in analyses. For radiation workers, the age at hire was the age at the beginning of first radiation job.

TABLE 3

General Demographics of the Shipyard Cohort by Radiation Group

Data Analysis

Construction of Time-Dependent Radiation Dose Matrix. Shipyard-related radiation dose measurements began in 1945 with the advent of radiography for inspecting castings and welds. The first overhaul of a nuclear-powered vessel occurred in 1957. A cumulative dose matrix calculated from the annual dose records summarized cumulative dose for each year for a radiation worker during his/her radiation work years between 1945 and 2011. Thus, this cumulative dose matrix was time-dependent and only the accumulative dose in the last year of their radiation work would be equal to the lifetime dose. This time-dependent dose matrix enabled time-dependent dose and lagged exposure-based analyses weighted by person-years. In a time-dependent analyses, the cumulative radiation doses of workers with an LHC death at a given time point were compared to the cumulative radiation dose of reference workers at the same time point, rather than comparing to the lifetime dose of reference workers. This approach was designed to prevent survival bias, where healthy survivors tend to have higher lifetime doses because they lived and worked longer (40). The software SAS© 9.4 was used for establishing this dose matrix (41).

Time-dependent Data Preparation with Lagged Exposure

A lagged radiation exposure period of 2 years is presented in the article. Lagging 2 years of radiation exposure was undertaken to remove the influence of exposure in a recent 2-year period prior to an event of interest, e.g., mortality due to LHC. This assumed that the most recent 2 years of radiation exposure, if any, were not relevant to the event due to a latency effect. Since the radiation doses in this paper were time-dependent, so also should the exposure lagging. Time-dependent exposure lagging was performed at time of comparison during follow-up using everyone's time-dependent cumulative dose matrix, most often before lifetime dose was reached. Thus, exposure lagging was not necessary to remove recent exposure from lifetime dose; a dose, accumulated to a certain year during the follow-up, was available depending on the event year. This complicated construction of a lagged radiation dose matrix was accomplished using the DATAB module of the EPICURE© software (42). This DATAB module created a structure of cross tables by categories of with lagged cumulative dose of 2 years. Variables used for DATAB matrix included sex (male and female), race (white, black, other known race, or unknown race), solvent related job (yes vs. no), birth year (5-year intervals), attained age (5-year intervals), and calendar year (5 year intervals). The birth year (5-year intervals), attained age (5-year intervals), and calendar year (5-year intervals) were time scales used in the tabulations. The dose categories were created in a time-dependent manner, by 25 mSv intervals up to 300 mSv. Doses at 300 mSv and above were categorized in one group due to the very small number of workers under study exposed to a lifetime dose higher than 300 mSv in U.S. shipyards. Workers (2.9% of total) with missing birth dates were automatically excluded from analyses because attained ages could not be calculated for them.

Comparison of Radiation and Non-Radiation Workers

Radiation workers were compared to non-radiation workers using the mortality risks of LHC and its types, controlled for sex, race, calendar time (as a time scale variable in time-dependent analysis), and time-dependent age (an age changing with calendar time). These analyses did not involve radiation dose but were a simple comparison of the disease risks of workers who ever had a radiation job vs. workers who never had a radiation job. Since there were some workers who had radiation jobs but had never had measurable radiation exposure (zero dose), radiation workers, excluding those with zero radiation dose, were also compared with non-radiation workers separately. AMFIT module of the EPICURE© software (42) was used for this analysis.

Dose-response Analysis Based on Time-dependent Categorical Doses and Linear Trend Test

The analyses were performed using the population of 153,930 radiation workers. Due to the concern of comparability of radiation and non-radiation workers, the analysis using radiation workers only was considered more reliable when examining a dose-response relationship. Radiation workers who had 0 mSv time-dependent radiation dose were used as the reference population. A 0 mSv time-dependent radiation dose means that either the worker was hired for a radiation job but had never been exposed to occupational radiation, or at the time of time-dependent analysis, the worker had not been exposed to any radiation yet, although he or she may have a radiation dose later. The final time-dependent radiation dose categories used in Poisson regression models were 0, >0–<25, 25–<50, 50–<100, 100–<200, and > = 200 mSv with 0 mSv as reference. Workers with lifetime radiation dose > = 200 mSv had a median lifetime dose at 249 mSv, which means half of the workers had lifetime radiation dose between 200 and 249 and the other half >249 mSv. The 2 year lagged RRs by dose categories were presented for LHC and its types except for other lymphoid leukemia (7 deaths) and other myeloid leukemia (13 deaths). The number of deaths for these two types were too small to generate stable results in the dose response analysis.

All analyses in this section were stratified by sex, calendar time (5-year interval), and time-dependent attained age group (5-year interval) and controlled for race group, socioeconomic status, solvent related jobs, and age at first hire (for a radiation job) in the model directly. DATAB module of the EPICURE© software (42) was used to create the analytical person-year tables stratified by these variables using the categories described above. AMFIT module of the EPICURE© software (42) was used for categorical dose analyses with the analytical person-year tables created. This categorical risk analysis shows the actual shape of dose-response curves in figures using the midpoint of each radiation dose category at 0, 12.5, 37.5, 75 and 150 mSv, respectively, except for the highest dose category (> = 200 mSv) where the median dose 249 mSv was used. The sample size weighted linear trend tests for RRs over the midpoint/median doses for each dose category were performed using the Chi-squared method with Epi-Info 7.2.4.0 Software (43). The P values of the linear trend tests were presented with the categorical dose analysis results.

Dose-response Analysis based on Time-dependent Continuous Doses

The analyses in this section were performed only for radiation workers for the same reason described above. Instead of using categorical doses, individual time-dependent radiation doses were used as a continuous variable. The dose-response analysis based on continuous doses enable a researcher to assume a desired underlying shape of response curve. Several studies have used Poisson linear ERR models for radiation and cancer research and the underlying dose-response curves for LHC, especially leukemia, were assumed to be linear (1, 18, 22–24). However, some other studies suggest the dose-response curves for leukemia are linear quadratic (13, 24). We performed a series of tests on choosing appropriate dose-response models: linear, quadratic, or linear quadratic, using leukemia less CLL. Based on the goodness of fit indicator, deviance, and comparison with the categorical dose analysis, the testing result did not support a quadratic or linear quadratic dose-response pattern in the present data. The linear model was best among the three. The results generated were presented in ERR per Sv (ERR/Sv) for comparison purposes, as well as in ERR and RR at 100 mSv to avoid misleading information, since 96.4% of person-years followed among radiation workers were below 100 mSv. All analyses were stratified by sex, calendar time (5-year interval), and time-dependent attained age group (5-year interval) and controlled for race group, socioeconomic status, solvent related jobs, and age at first hire for a radiation job in the model directly. AMFIT module of the EPICURE^© software (42) was used for this analysis.

General Results of the Cohort Follow-up

Table 2 provides a general description of the study cohort by radiation and non-radiation work status. There are a total of 437,937 participants, comprised of 284,007 (64.9%) non-radiation workers and 153,930 (35.1%) radiation workers. There are 138,912 recorded deaths over the 13,393,399 person-years of follow-up. The average follow-up periods are 29.9 and 31.8 years for non-radiation and radiation workers, respectively. The average employment durations for non-radiation and radiation workers are 10.8 years and 19.4 years, respectively. As described previously, the follow-up for radiation workers began on the date of first radiation work or January 1, 1957, whichever was later. The employment periods prior to radiation jobs contributed to the longer duration of radiation worker employment as seen in Table 2. The median, 75th percentile, and 95th percentile lifetime radiation doses are 0.82 mSv (0.082 rem), 10.71 mSv (1.071 rem), and 83.63 mSv (8.363 rem), respectively. Although the average lifetime occupational radiation dose was 15.36 mSv, most radiation workers have extremely low doses with a small proportion (∼25%) of radiation workers with total lifetime doses greater than 10 mSv (1 rem).

TABLE 2

General Information of the Shipyard Cohort Follow Up by Radiation Group

Table 3 provides selected demographics of the study cohort, including gender, race, and age at death or at the end of follow-up period if still alive (i.e., 2011), birth cohort, socioeconomic status, and solvent-related jobs. Males constituted 97.1% of radiation workers and 80.0% of non-radiation workers. Both radiation and non-radiation workers are predominantly white. The distribution of age at death or at the end of follow-up (if alive) shows that 45.7% of non-radiation workers and 32.8% of radiation workers are under age 60. Also, 46.4% of non-radiation workers were born in or after 1950, while only 31.3% of radiation workers were born in or after 1950. More than half of workers are classified as blue collar; the proportion of blue-collar workers for radiation workers is slightly lower than non-radiation workers. Proportions of solvent-related jobs are similar between the two groups. All variables presented in Table 3 are considered potential confounders and are adjusted in the multivariate analyses.

Table 4 shows the breakdown of cause-specific LHC deaths relative to time-dependent person-years of follow-up and time-dependent dose groups lagged for 2 years. There is a total of 3,699 LHC deaths (2,225 from non-radiation and 1,474 from radiation workers) recorded during the 54 years of follow-up (1957-2011), which represents 0.84% of the total cohort population. 96.4% of radiation workers have a time-dependent radiation dose, lagged 2 years, below 100 mSv. Among the LHC types, 39.2% (1,450) of the LHC deaths in the total cohort are due to Hodgkin's or non-Hodgkin's lymphoma, 37.4% (1,382) to leukemia, 17.3% (641) to multiple myeloma, 5.9% (219) to myeloproliferative and myelodysplastic disease, and 0.2% (7, not shown in the table) could not be classified into types. When evaluating only radiation workers, 38.3% (564) of the LHC deaths are due to Hodgkin's or non-Hodgkin's lymphoma, 37.0% (545) to leukemia, 17.8% (263) to multiple myeloma, and 6.7% (99) to myeloproliferative and myelodysplastic disease. Table 4 also shows that with increasing time-dependent radiation doses, deaths for LHC and its types and time-dependent person-years decrease. Therefore, in order to ensure that there would be enough study power to obtain more precise estimates of the categorical RRs, the time-dependent radiation doses are collapsed into the following five dose categories (>0–<25, ≥25–<50, ≥50–<100, ≥100–<200, and ≥200 mSv) for subsequent categorical dose-response analysis compared to 0 mSv category.

TABLE 4

Causes of Death by Time-dependent Dose Category Lagged 2 Years, 1957–2011

TABLE 4

Extended.

Comparison of Radiation and Non-Radiation Workers

Table 5 presents RRs of LHC and its types in two comparisons: 1. radiation vs. non-radiation workers, and 2. radiation workers with greater than 0 mSv lifetime dose vs. non-radiation workers, stratified by sex, calendar time (5-year interval), and time-dependent age group (5-year interval) and controlled for race group in the Poisson regression models. The second comparison excludes radiation workers without measured occupational radiation dose, thereby representing a “true” radiation worker group to compare with non-radiation workers. The results in Table 5 show that when the radiation worker group and the non-radiation worker group are compared, LHC (RR = 0.927, P = 0.030), lymphoid leukemia (RR = 0.748, P = 0.011) and lymphomas (R = 0.886, P = 0.031) have RRs that are significantly lower than that of the non-radiation workers' group. Radiation workers show protective RRs' or less chance of mortality due to LHC (6.3% less), lymphoid leukemia (25.2% less), and lymphomas (11.4% less) than non-radiation workers after adjustment for potential confounders. The other types do not present any difference with statistical significance. To evaluate whether the large number of radiation workers with 0 mSv lifetime dose within the radiation worker group might have diluted the mortality risk due to radiation, the RRs between radiation workers with greater than 0 mSv lifetime dose is compared to the non-radiation workers’ group. The radiation workers with greater than 0 mSv lifetime dose also show lower risk for LHC (RR = 0.919, P = 0.028) with statistical significance compared to non-radiation workers. Among types, leukemia (RR = 0.880, P = 0.044), especially lymphoid leukemia (RR = 0.658, P = 0.002) show RRs lower than 1 with significance. For the rest of LHC types, the differences between the two comparison groups are not statistically significant (Table 5).

TABLE 5

Adjusted Relative Risks for Radiation and Non-Radiation Workers, 1957–2011

Dose-response Analysis Based on Time-dependent Categorical Doses and Linear Trend Test

Tables 6a–c display the results for the following two categorical analyses: 1. a one-to-one comparison of each dose category group versus the 0 mSv group for LHC and its types to identify any risk in the subset and 2. the P values for categorical dose linear trend (i.e., RR) test for LHC and each of its types weighted by sample sizes in groups in comparison.

TABLE 6a

Relative Risks (RR) of Categorical Cumulative Time-dependent Radiation Dose, Lagged 2 Years among Radiation Workers – 1/3

TABLE 6b

Relative Risks (RR) of Categorical Cumulative Time-dependent Radiation Dose, Lagged 2 Years among Radiation Workers – 2/3

TABLE 6c

Relative Risks (RR) of Categorical Cumulative Time-dependent Radiation Dose, Lagged 2 Years among Radiation Workers - 3/3

The RRs for the following five LHC dose categories relative to the 0 mSv control group when lagged for 2 years were 0.960, 0.914, 1.104, 1.067 and 1.143, respectively, with a linear trend P value of 0.380 (Table 6). A graphical representation of the LHC categorical results are shown in Fig. 1. The RRs for leukemia less CLL for the same evaluation criteria were 0.930, 1.040, 1.132, 1.208 and 1.356, respectively, without statistically significant linear trend (P = 0.253). A graphical representation of the leukemia less CLL categorical results are shown in Fig. 2. The remainder of LHC types did not present any significant linear trending, either.

FIG. 1

Relative risks (RR) and 95% confidence intervals (CI) of LHC with time-dependent dose lagged 2 years.

FIG. 2

Relative Risks (RR) and 95% confidence intervals (CI) of leukemia less CLL with time-dependent dose lagged 2 years.

None of the RRs for any of the dose categories vs. 0 mSv, for any of the cancer types or diseases, show any statistical significance, except for chronic myeloid leukemia which shows a significant elevation of RR at 50–<100 mSv category (RR: 2.746; 95% CI: 1.002, 7.521; P = 0.049) and Non-Hodgkin's lymphoma which shows RRs with borderline significance in categories of 50–<100 mSv (RR: 1.391; 95% CI: 0.954, 2.027; P = 0.087) and 100–<200 mSv (RR: 1.499, 95% CI: 0.975, 2.305; P = 0.065) vs. 0 mSv, respectively. This has resulted in the RR in 100–<200 mSv category for lymphoma (most of them are Non-Hodgkin's type) to be borderline significant (RR = 1.459; 95% CI: 0.965, 2.205; P = 0.073) and an elevated RR for the category 50–<100 mSv (RR = 1.323; 95% CI: 0.919, 1.905; P = 0.132).

One important finding in this section is that for LHC, leukemia less CLL, acute myeloid leukemia, lymphomas, and non-Hodgkin's lymphoma, there are often RRs lower than 1.000 observed in the categories >0–<25 or/and 25–<50 mSv vs. 0 mSv. Although the difference of RRs in these categories are not statistically significant from the reference group (0 mSv), they are consistently lower than 1.000 across LHC and its types (Table 6).

Dose-response Analysis based on Time-dependent Continuous Doses

The results of the dose-response analysis based on time dependent continuous radiation doses lagged 2 years are presented in Table 7. The results presented in ERR/Sv are for purposes of comparisons with other studies, since ERR/Sv (or ERR/Gy) is the most widely used risk unit in published literature related to radiation and cancer and demonstration of time dependent continuous radiation dose-response relationship with LHC and its types. ERRs/Sv are 0.941 for LHC (95% CI: –0.368, 2.250; P = 0.159) and 1.775 (95% CI: –0.851, 4.401; P = 0.185) for leukemia less CLL, similar to what has been observed in categorical dose-response linear trend tests.

TABLE 7

Excess Relative Risks (ERR) and Relative Risks (RR) at at 1 Sv or 100 mSv Cumulative Time-dependent Radiation Dose, Lagged 2 Years, Radiation Workers

As mentioned earlier, 96.4% of person-years followed among radiation workers under study were below 100 mSv. Table 7 presents the ERRs and RRs for LHC and its types at 100 mSv. RR for LHC at 100 mSv is 1.094 (95% CI: 0.963, 1.225; P = 0.159). Leukemia less CLL has an RR at 1.178 (95% CI: 0.915, 1.440; P = 0.185) at 100 mSv. Other types with RRs above 1.000 included myeloid leukemia at 1.283 [95% CI: 0.921, 1.644 (acute at 1.308, 95% CI: 0.883, 1.732, and chronic at 1.356, 95% CI: 0.484. 2.227)], multiple myeloma at 1.004 (95% CI: 0.731, 1.277), and lymphomas at 1.233 [95% CI: 0.988, 1.478 (Hodgkin's at 1.317 and non-Hodgkin's at 1.227) at 100 mSv.

Lower LHC Risk in Radiation Workers vs. Non-Radiation Workers

The first key finding is that shipyard radiation workers had a significantly lower risk of LHC mortality relative to non-radiation workers when adjusted for sex, calendar time, time-dependent age, and race (P = 0.030). Among the LHC types, the radiation workers showed an 11.4% less chance of mortality from all lymphomas combined, compared to non-radiation workers (P = 0.031). The analysis also showed a 25.2% less chance of mortality from lymphoid leukemia (P = 0.010).

This finding is consistent with a study that used prior to 1981 data from the same cohort, in which the non-radiation workers had higher risk than radiation workers (4). Several plausible explanations for the deficit in LHC mortality in radiation workers compared to non-radiation workers are 1. it is a real deficit effect; 2. the “healthy worker effect”, or 3. a combination of both. Prior to entry into the radiation worker program, radiation workers must pass an initial medical screening looking for the existence of cancerous conditions. If a cancerous condition is found, the worker is precluded from becoming a radiation worker until the condition is treated and the worker is medically cleared. This medical screening is performed periodically during employment with the frequency of screening increasing with age. This is a form of selection bias, as non-radiation workers do not receive this same service. To rule out the “healthy worker effect” in subsequent radiation dose related analyses, only evaluations from within the radiation worker population were carried out, thus, all results in Tables 6 and 7 were not impacted by the “healthy worker effect”.

LHC Risks in Higher Radiation Dose Subsets of Shipyard Radiation Workers

Although radiation workers had a lower risk of mortality due to LHC than non-radiation workers, further analyses were performed using only the radiation workers to examine if any subset of radiation workers exposed to higher doses have an elevated risk of mortality due to LHC. Radiation workers with a 0 mSv radiation dose were used as a reference group when calculating categorical RRs.

LHC as a group did not show any statistically significant elevated RR in any categorical time-dependent doses lagged 2 years at >0–<25, 25–<50, 50–<100, 100–<200, and > = 200 mSv vs. 0 mSv. Among types, chronic myeloid leukemia showed a significant elevation of RR for the 50-<100 mSv category only (RR: 2.746; 95% CI: 1.002, 7.521; P = 0.049) vs. 0 mSv. It might be possible that the risk observed in a single subset may not be valid if the risk did not increase with dose. Any increased or decreased RR in a single dose category could be a result of uncontrolled covariates in that category rather than radiation exposures. Thus, dose-response analysis would be still needed to further evaluate the LHC risks.

Sublinear Categorical Dose-Response Relationship

As seen in Figs. 1 and 2, a curved dose-relationship was not supported. Given the fact that RRs in mid dose categories did not fall into a perfect straight line and had wide confidence intervals, we define this as a quasi or “sublinear” dose-response relationship, in which the point that the risk could fluctuate around the straight line and their linearity trends can be tested with linear regression models. However, the radiation doses in the present study were limited to low levels (<250 mSv) and the result of the analysis did not preclude that the dose-response relationship may become quadratic, linear quadratic, or other shapes at higher doses beyond the present study. The problem is that the sample sizes (person-years and deaths) were heavily skewed toward low doses. Even if RRs were observed increasing with dose, the one-to-one categorical dose comparisons and the linear trend tests are not statistically significant, because the number of person-years with mid and high doses were limited. The other problem might be that our dose range among radiation workers was not broad enough to reveal the curvature of a dose-response relationship at higher dose range; thus, the linear pattern identified in our present paper, although not significant, would not preclude a curved dose-response relationship in a higher dose range beyond the doses in this study.

Risk Observation at Low Dose

One interesting observation that we could not ignore in this study was that among LHC and some of its types (leukemia less CLL, acute myeloid leukemia, and lymphomas), there were often RRs lower than 1.00 in the categories >0–<25 or/and 25–<50 mSv vs. 0 mSv. Although the difference of RRs in those categories were not statistically significant compared to the reference (0 mSv), they were consistently observed across LHC and its types. For instance, for LHC as a whole group, the RRs in categories of >0–<25 and 25–<50 mSv vs. 0 mSv are 0.960 and 0.914, respectively. RRs of Leukemia less CLL vs. 0 is 0.930 in dose category of >0–<25 mSv. The RR of acute myeloid leukemia is 0.885 in dose category of >0–<25 mSv. The RR of Non-Hodgkin's lymphoma in category of 25–<50 mSv vs. 0 is as low as 0.737, which is 26.3% lower than the reference group. Although, these results hinted a hormesis like effect below 50 mSv, all individual categorical dose results were not significant and therefore not conclusive. The addition of more radiation worker mortality data in future studies will add more statistical power to the results and further clarify this observation.

Risk Analysis Based on Time-dependent Continuous Doses

Further analysis was done using radiation dose as a continuous variable, so that the dose response analysis could be done at an individual level rather than aggregated as in categorical analysis. Table 7 presents the result of this analysis using continuous dose by EPICURE in ERR/Sv, ERR at 100mSv, and RR at 100 mSv. The results for linear trends were not statistically significant for LHC or any of its types, although lymphomas came close (P = 0.062). In the shipyard population, an RR in continuous dose analysis would be heavily weighted towards dose-response patterns at low doses, less than 50 mSv, since 91.7% of radiation workers had a dose below this level (Table 4). The small sample sizes in higher dose groups lead to low statistical power of the analysis might be the main problem of the insignificance of the dose response.

The small sample size in the higher dose groups lead to low statistical power and greater uncertainty. The lack of adequate data points may be a main contributor in the lack of any significance for this analysis. The inclusion of additional mortality data in future analysis may provide further clarification; however, the magnitude of risks in the continuous analysis were still comparable to that in the previous categorical analyses. For example, leukemia less CLL had an RR of 1.178 (95% CI: 0.915, 1.440; P = 0.185) at 100 mSv in the continuous analysis, which was between the 1.132 for 50–<100 mSv and 1.208 for 100–<200 mSv lagged 2 years in the categorical analysis. Therefore, this study will be repeated when addition mortality data becomes available.

Comparison with Other Epidemiologic Studies

Table 8 summarizes ERR/Sv and ERR/Gy for leukemia less CLL, the most common LHC type reported in published studies on radiation risk among workers. The ERRs per Sv range from 1.70 (90% CI –0.22, 4.70) to 2.96 (90% CI: 1.17, 5.21) from several large-scale epidemiological studies around the world (1, 6, 18, 23). In these studies, various lag periods from 0 to 10 years were used. For comparison purposes, the ERR/Sv after lagging 2 years for leukemia less CLL observed in the present study is used in Table 8. The INWORKS (France, UK and U.S.) (23) had the highest ERR/Sv at 2.96 (90% CI: 1.17, 5.21). In 2021, the U.S. pooled INWORKS reported an even higher ERR/Sv at 3.15 (90% CI: 1.15,5.72) for leukemia (including CLL) (8). However, the U.S. pooled INWORKS study (44) showed lower and statistically nonsignificant ERR/Sv at 1.70 (95% CI: –0.22, 4.70) (44). ERR/Sv for leukemia less CLL is 1.775 (95% CI: –0.851, 4.401) in the present study, which is consistent with the other studies in the table and particularly close to 1.712 (95% CI: 0.06, 4.29) in the 3rd UK NRRW study by Muirhead et al. (6) and 1.70 (95% CI: –0.22, 4.70) in the U.S. pooled INWORKS study by Schubauer-Berigan et al. (18). This comparison suggests that the slope of the dose-response relationship in U.S. shipyard workers is similar to that in other radiation workers in those studies.

TABLE 8

Excess Relative Risk of Leukemia less CLL in Comparison with Other Studies

While the present study analyzed LHC types in detail, three other published studies looked at some but not all the same LHC types on low-dose radiation exposures (9, 18, 23). A comparison of the findings of these studies is presented in Table 9. While the INWORKS study (23) showed that the ERR/Sv of chronic myeloid leukemia was much higher than that for acute myeloid leukemia (10.45 versus 1.29), the present study showed comparable results (3.555 vs. 3.077). The ERR/Sv for multiple myeloma was found to be elevated in the U.S. pooled INWORKS study at 3.90 (95% CI: 0.60, 9.60) (18), but reduced to a much lower level when combined with UK and France, 0.84 (90% CI: –0.96, 3.33) (23). The most recent update on NRRW-3 cohort showed that ERR/Gy for multiple myeloma was 1.50 (95% CI: –0.42, 5.60) without statistical significance (9). The INWORKS study observed ERR/Sv of 2.94 for Hodgkin's lymphoma and 0.47 for non-Hodgkin's lymphoma (23), while NRRW-3 cohort showed negative association for Hodgkin's lymphoma but positive association for non-Hodgkin's lymphoma at 1.31 (95% CI: –0.25, 3.77) (9). None of the values in either study achieved statistical significance. In the present study, ERR/Sv was 3.167 (–1.000, 12.110) for Hodgkin's lymphoma (very close to INWORKS 2.94 ERR/Gy) and 2.266 (–0.280, 4.812) for non-Hodgkin's lymphoma (close to NRRW-3 1.31 ERR/Gy), both without significance. Berrington de Gonzalez et al. compared mortality rates of 43,763 radiologists and 64,990 psychiatrists who were followed from 1979 to 2008 (45). This study reported a significantly increased RR for non-Hodgkin's lymphoma of 2.69 (95% CI 1.33–5.45) for radiologists vs. psychiatrists. These findings might indicate some need for further investigation regarding the association of non-Hodgkin's lymphoma with radiation exposure.

TABLE 9

Excess Relative Risk (ERR) Comparison among Different Large-Scale Studies on Low-Dose Radiation

It is important to point out that this population has had very low radiation exposure. Although some of the dose categories presented some risks with certain cancers, the actual number of workers exposed in those categories were small. In addition to what has been presented in Table 2, the 95 percentile of lifetime radiation dose was at 83.63 mSv. Figure 3 shows the average annual radiation dose distribution and number of workers exposed by calendar year from 1957 to 2011. This graph shows two periods of peak employment: one large peak from 1964 to 1970 and a smaller peak from 1982 to 1992. The average annual radiation doses were between 0.0025 and 0.0060 Sv before 1980, decreased to below 0.0020 Sv in 1980 and further decreased to below 0.0010 Sv in 1995 and thereafter. Most importantly for this study, this suggests that the radiation exposures among workers hired after 1979 were at relatively lower levels compared to workers hired before 1979.

FIG. 3

Distribution of average annual radiation dose and number of workers exposed by calendar year.

Analysis of Covariates

RRs for each covariate are not presented in this study as some covariates were not sufficiently well-defined to generate reliable RRs. For instance, race was categorized as white, black, other known race, or unknown race. Ideally, only three categories should exist (white, black, other) based on shipyard worker demographics and number of workers in each category. However, since a quarter of the population had unknown race, any analysis results for this covariate would have been unreliable. Although specific covariate results are not presented, it did not prevent their use in the statistical analysis for controlling purposes only.

Background Radiation Exposure

As described in the methods section, the radiation data in this study refers to occupational radiation exposure after background (i.e., environmental) correction. Thus, the occupational radiation dose is not total annual dose (occupational, background, medical, etc.) to the individual. It was assumed that, for other than occupational exposure, that both radiation and non-radiation workers were exposed to similar background exposure scenarios while working in the shipyard (e.g., environmental, medical, industrial). As some of these exposure scenarios can vary based on age or year of employment, for example, advances and increased use of medical diagnostics and nuclear medicine, controlling for time-dependent age and calendar time was imperative to minimizes these effects in the analysis.

Strengths and Limitations

The strengths of the present study include: 1. a total of 437,937 workers enrolled, representing one of the largest cohorts with the longest follow-up time among studies to date for LHC risk among radiation workers. The resultant increased reliability and statistical power are critical for low mortality cancers such as LHC; 2. detailed annual radiation dose measurements were available for each radiation worker during their entire follow-up period, thereby allowing sophisticated exposure analyses; 3. all occupational radiation exposures examined were from the single industry of shipyard workers, thereby avoiding the issue of variation in type of radiation examined, which complicated other past studies; 4. time-dependent radiation doses were used in our analyses, which dramatically increased the validity and accuracy of the associated measurements, since the radiation dose varied by time; 5. the exposure lagging analyses were used to deal with the latent period issue of cancers, which further improved the reliability of our study; and 6. the present study used both categorical and linear dose-response models using continuous individual doses, an approach widely accepted by cancer epidemiologists (1, 6, 18, 22–24, 46) so that the results are comparable to other major studies.

The limitations of the present study include: 1. missing race information on a quarter of the cohort population, which prevented reliable covariate analyses; 2. approximately three percent of radiation workers had to be excluded from the analysis due to missing birth dates reducing the statistical power of the analysis; 3. use of the last job, instead of the longest job or full job history, to assign the solvent-related jobs, which might have led to uncontrolled residual risk; 4. deaths, rather than incident cases, were used, which leads to an underreporting of cases, since treatment for LHCs, has been increasingly successful; 5. cancer causes were based on National Death Index reports/Death Certificates, rather than on pathology reports, which might have limited the ability to examine LHC types. LHC information might be missing due to lacking details; 6. low number of person-years with doses in the upper range of the dose analysis; 7. the potential influence from unknown background radiation; and 8) the completeness of follow-up, since the death search was the only method to follow up workers for LHC cancers.

In summary, this study found that: 1. the risks of LHC and its types were not elevated in radiation workers compared to non-radiation workers as a group in U.S. nuclear shipyards; 2. ERR at 100 mSv for LHC and leukemia less CLL was in line with other low dose radiation studies and 3. although this was a relatively large study, that additional mortality data is needed to draw any further significant conclusions on low dose LHC mortality risks.