In a recently published overview of the baseline Cohort (Panel 1) , proportional differences in demographic and military service characteristics between the Cohort and the probability-based original invited sample (n = 256,400) were considered small; therefore, Cohort study findings were expected to be generalizable to the military population as a whole. In this mortality study, the crude death rate for Panel 1 Cohort members was 80.7 per 100,000 person-years (95% CI: 72.2, 89.3). To put this into context, age-specific death rates for the US civilian population over the same time period (2001-2006) were all above 90 per 100,000 for age 19 years and above . The lower crude death rate for our study population is consistent with a healthy worker effect since individuals entering military service must undergo extensive health screening, and accession standards ensure that only fit and healthy individuals from the general population are selected. Furthermore, periodic health examinations and fitness standards that must be met for continued active service plus universal access to medical care at military treatment facilities worldwide result in a healthy active-duty work force. When Kang and Bullman previously investigated mortality in veterans of the 1991 Gulf War era, they found that cause-specific standardized mortality ratios for both deployed and non-deployed veterans were significantly lower compared with the general US population .
The crude death rate for Panel 1 Cohort members is also in line with the crude death rate of 72.9 per 100,000 person-years reported for the active-duty component of the military between 1990 and 2008 by the Armed Forces Health Surveillance Center . The higher crude death rate for the Millennium Cohort likely reflects that our study population includes individuals separated from military service, some potentially due to serious illnesses.
We initially compared the Panel 1 Cohort's crude death rate with that of the invited sample and found a statistically significant difference (80.7 per 100,000 person-years, 95% CI: 72.2, 89.3, vs. 101.2 per 100,000 person-years; 95% CI: 95.4, 107.0). To more fully investigate differences between the two independent groups comprising the invited sample, we compared crude death rates as well as age- and sex-adjusted overall death rates for study participants and non-participants. We also examined the distribution of mortality by demographic and military service characteristics, and finally, compared age-adjusted, category-specific death rates for these two population subgroups.
Despite higher proportions of survey respondents in older birth cohorts, the lower crude death rate for participants compared with non-participants was consistent with healthier segments of the military population having the opportunity and consenting to participate in a survey. Even more apparent was the difference between age- and sex-adjusted overall death rate for Panel 1 participants and non-participants--78.0 deaths per 100,000 person-years (95% CI: 68.3, 87.7) vs. 121.4 deaths per 100,000 person-years (95% CI: 112.9, 129.8), respectively.
In addition to healthy individuals having the opportunity to participate, more health-conscious individuals (e.g., higher educational level, married) might have greater interest in and, thus, motivation to join a health-related study. To further assess the representativeness of Panel 1 Cohort members, we compared mortality distribution by demographic and military service characteristics in Panel 1 participants and non-participants. The distribution of mortality by measured demographic factors was generally similar in the two groups. As expected, older persons and men were proportionally higher among decedents. Caucasians were overrepresented and those "not married" somewhat underrepresented among decedents in the Cohort. Among non-participants, this distribution was less pronounced but trending in the same direction. Educational level of "high school or less" was overrepresented among decedents in the non-participant group and consistent with the association between lower educational achievement and higher mortality described in the published literature , but this was less evident among the participants.
In addition to demographics, we examined military service characteristics to compare their distribution among Panel 1 participants and non-participants. Disproportionately higher numbers of deaths occurred in the following categories: Reserve/National Guard, Army, combat specialists, pre-2001 or no deployment experience, and not separated from active service. Army personnel and combat specialists were overrepresented among decedents, consistent with current combat operations in Iraq and Afghanistan, and may reflect the population at highest risk. Disproportionately higher numbers of deaths in Reserve/National Guard members and those with no deployment history may represent older age or certain medical conditions rendering one unfit for deployment and at the same time at higher risk of death. A similar effect, although less pronounced, was seen in the subgroup with deployment experience prior to 2001. This group included 1991 Gulf War veterans as well as veterans of conflicts in Bosnia, Kosovo, and Southwest Asia. Again, patterns of mortality by military service characteristics were generally similar when comparing Panel 1 participants and non-participants. Pay grade (officer vs. enlisted) was not an influential variable, but mortality was differentially distributed by continuing service or separation from military service. There were disproportionately fewer decedents among those separated from military service among both the participants and non-participants, which may be explained by factors such as age, service branch, and occupation. For example, a substantial proportion of young service members separate from military service after one tour of duty, and these individuals are less likely to die from disease (as opposed to external causes). Army and combat specialists were overrepresented among decedents, and since death was excluded as a reason for separation, this could contribute to disproportionately lower mortality in the separated group.
After adjustment for age, a comparison of category-specific death rates for Panel 1 participants and non-participants showed consistently higher death rates for those not participating in the study, although substantial variability in rate estimates for some categories might be the case due to small numbers. As in other health-related studies, non-participants may have less opportunity to participate due to serious illness or injury, institutionalization, or other life circumstances. Individuals who engage in unhealthy behaviors may also decide not to participate. These age-adjusted rate comparisons highlight some important differences between the two groups that warrant further investigation, notably in the categories of race/ethnicity, marital status, branch of service, and deployment experience. Next steps would include the examination of causes of death to better understand rate differences. Over time, any continued meaningful divergence in the mortality experience between the participants and non-participants should become more apparent as greater numbers accrue. Adjustment for additional covariates (survival analysis) and a focus on cause-specific rates, as well as selective subgroup analyses would be very informative.
Our final study objective was to examine the utility of selected data sources for ongoing mortality ascertainment in the Millennium Cohort. Out of the 202 total deaths identified by at least one of the four data sources, the VA databases identified the largest number (196 or 97.0%), presumably because the VA dataset also includes the SSA-DMF. In a recent study of mortality data in the VA, the BIRLS Death File alone was reported to be accurate but not complete, but with the SSA-DMF, it was viewed as the best choice if just one source had to be selected for mortality ascertainment in veteran populations . The VA data files would seem to be highly suited to our study population, with all potentially eligible for death benefits from the VA and presumably all accounted for in SSA databases. If deaths were related to combat deployment, families of decedents may be more likely to claim VA death benefits, including burial at national cemeteries. Notably, the VA dataset identified two deaths not included in any of the other three data sources.
The number of deaths captured by NDI was 164 of 202 (81.2%). While we expected NDI to capture a larger proportion of the Cohort deaths, we did consider it possible that many of the deaths not captured by NDI may have occurred outside the United States and its territories (e.g., combat deaths in theaters of operations or other overseas assignments). In fact, using information from the DoD-MMR, as well as electronic military deployment data, 31 of the 38 deaths (82%) not identified by NDI were confirmed to have occurred overseas. The other seven (18%) unidentified deaths may be a result of non-matches due to inconsistencies in death certificate information and possibly longer-than-expected lag time in posting deaths. Despite missing a considerable number of Cohort deaths, NDI did identify four deaths not captured by any of the other three data sources.
The SSA-DMF was examined separately from the combined VA dataset because it is currently the main source of updates to mortality information for the Cohort (accessed through DMDC). It is also more generally accessible for mortality ascertainment than VA files. One SSA-DMF death was, in fact, not identified by the VA dataset that includes the SSA-DMF. Because mortality ascertainment is more complete in older age groups, the utility of the SSA-DMF is expected to increase over time. However, inclusion in the SSA-DMF is more likely among eligible beneficiaries either claiming death benefits or reporting a death to discontinue Social Security benefits to the deceased; thus, deaths may be incompletely captured in some population subgroups .
Potential sources of bias in recording or reporting of deaths by data source need to be understood and monitored over time. Additionally, changes in relative importance of currently used data sources are likely to occur as the Cohort ages and a greater proportion of individuals separate from military service, and other sources may gain importance, such as the Centers for Medicare & Medicaid Services. Finally, the success of long-term follow-up, including mortality ascertainment, is often affected by time since last follow-up and the type and amount of information available for use . Thus, ongoing efforts to improve retention in the Millennium Cohort, including regular contact with participants through biennial postcards , provide multiple opportunities to maintain contact and correct any inconsistencies in personal identifying information.
There are several study limitations. First, due to the short period of observation (2001-2006), the total number of deaths in this relatively young, healthy occupational cohort is small (n = 341 through 2006), based on three frequently updated data sources (excluding NDI). When NDI was included, mortality could only be assessed through 2004, thus reducing the number of deaths available for the comparison of all data sources to 202. On the other hand, the large size of this Cohort with oversampling of Reserve/National Guard personnel, as well as the current involvement of US troops in combat theaters of operation, make a strong case for early assessment of mortality as an important health outcome. Over time, the number of deaths will increase, providing statistical power for cause-specific and other analyses. A second limitation is that slightly different methods were used to obtain death data because mortality ascertainment occurs on a regular basis, including prior to each survey administration, to avoid contacting family members of decedents. However, the same criteria were applied to count deaths (exact match on two of three personal identifiers). Finally, because our Cohort is composed of relatively young, healthy individuals subject to unique occupational experiences or exposures, it may be difficult to generalize our findings to other cohort studies. For example, higher likelihood of combat deaths overseas among relatively young individuals affects the degree of capture in various national mortality data sources.
Strengths of our study include using four data sources to conduct a baseline assessment of all-cause mortality in this large, population-based cohort of active and former military service members to be followed through 2022 in order to monitor any trends over time. We were able to link deployment, demographic, and separation data to the mortality data, enabling us to examine category-specific differences in the mortality experience between the participants and non-participants, as well as the four data sources. The Millennium Cohort is of particular interest to policymakers and the public because of ongoing military deployments to Iraq and Afghanistan. Many questions are being raised about potential short-term as well as long-term health effects of military service, particularly deployment. As statistical power increases, any effect of deployment or other military service-related exposures on all-cause and cause-specific mortality can be evaluated with adjustment for important covariates.