Comparing strategies for United States veterans' mortality ascertainment

Clinicians, healthcare administrators, researchers, regulators and policymakers are concerned with optimizing mortality ascertainment using administrative data. In addition to its clinical importance, mortality informs program planning, quality assessment and improvement, and public reporting [1‐8]. Veterans are an important, vulnerable population in which mortality has been examined as a function of race / ethnicity, service characteristics, access, and quality of care. Valid, complete reporting is critical to the success of such endeavors, and limitations in using death certificates have been acknowledged [9, 10], although VA mortality data is generally regarded as accurate [11‐15]. To understand the limitations of single source ascertainment, we described decedents who would not have been identified by a strategy using either VA decedent files alone or death certificates alone. We compared cases that would have been missed using either single source with the rest of the cohort based on their demographic and clinical attributes and the settings in which they received care.

In order to evaluate the implications for improving veterans' end-of-life care, we constructed a population-based decedent cohort [16]. For such purposes, it is particularly important to understand whether death was recorded elsewhere for veterans who were under VA care since the VA system may be responsible for much of their end-of-life care even if they do not die while receiving health care in a VA facility.

From 345,380 deaths during the period 30 September 1999 to 1 October 2000 identified in BIRLS, we distinguished 6,071 decedents who were users of VA inpatient, outpatient, or long term care services in California. California death certificates included 227,308 deaths during the same period, including 3,580 additional users of VA inpatient, outpatient, or long term care services in California. Using SSN and other identifiers to match decedent cases to VA utilization data, we excluded non-veterans (n = 365), users of only non-clinical care such as laboratory tests (n = 251), those possibly alive based on subsequent VA encounter data (n = 229), and 3 cases for other reasons. Of the final cohort of 8,813 veteran decedents, 5,698 (65%) cases were identified in both source files, while 689 (8%) were only identified in VA decedent files, and 2,426 (28%) additional cases were only identified through death certificates (Figure 1).

We examined potential biases associated with veteran decedents missed by either single source of mortality ascertainment (e.g., VA BIRLS or California death certificates). Ninety-nine percent of decedents missed by using VA data alone were California residents (vs. 92% of the remainder cohort, p < 0.001); whereas, 62% of those missed by using death certificates alone were out-of-state residents (vs. 1% of the remainder cohort, p < 0.001). Relatively fewer veterans of white or black ethnicity and relatively more veterans of missing ethnicity were represented among decedents missed by either single source strategy. The proportion of married or previously married veterans was higher and single or missing marital status lower among those decedents missed using only BIRLS, and relative proportions were reversed for a strategy using only death certificates.

Decedents missed by either single source approach were less likely to have been diagnosed with an advanced chronic illness than the identified cohort. Veteran decedents missed by using only BIRLS were less likely to be diagnosed with any condition except HIV and dementia, and those missed by using death certificates alone were less likely to be diagnosed with any condition except HIV. With a BIRLS only approach, 37% of missing cases vs. 35% of the remainder cohort (p < 0.001) had no diagnosed chronic illness (death certificate only approach; 69% vs. 32%, p < 0.001). Veteran decedents missed by either single source approach were equally or more likely to have been users of the outpatient setting, but missed cases were less likely to have been users of inpatient healthcare settings (Table 1).

Veterans' mortality ascertainment was significantly improved by using both VA and death certificates as source files. Our findings indicate that either single source approach for mortality ascertainment may misrepresent veteran mortality based on comparisons of race / ethnicity, marital status, severity of illness, and settings of care. Diagnoses associated with serious medical co-morbidity and the likelihood of receiving any inpatient services (e.g. hospital or long term care) were both significantly lower among veterans missed by either single source approach.

Our findings are consistent with Washington State where the deaths of 25% of 533 veterans who only used outpatient services were only identified with death certificates, and 5% were only identified in BIRLS. [9] Using BIRLS only for mortality determination, it is unclear why generally healthier, primarily outpatient users are less likely to be noted. Death notification is typically triggered by benefit claims (e.g., burial assistance, pension and related benefits). Affluent veterans whose families might be less likely to file benefit claims were drawn to the VA recently [27]. However, poverty or low social support might also make it harder to file claims. On the other hand, a death certificate only approach to ascertainment misses relatively fewer non-resident veterans. Such veterans may be homeless or mobile, retired veterans, and they may seek care transiently in California, or their deaths may be recorded elsewhere.

One limitation of our study is that we did not identify cases that were only decedents by virtue of VA utilization files alone rather than BIRLS, although Dominitz, et. al., identified only 2.7% of deaths this way [11]. We did not compare VA files or death certificates to the National Death Index (NDI), as have previous studies that have used the NDI as a gold standard. The NDI is a central data repository of state vital statistics that is often used as a gold standard in US mortality studies [28]. We report findings for only one state, but given similar findings in Washington State, it would be helpful to determine if this is a national issue or there are particular state issues related to BIRLS death file agreement, or concerns related to veteran morality ascertainment with California death certificates.

Researchers, managers, and policy makers should understand the limitations of sources of mortality ascertainment. The relationship of missing data to bias is related somewhat to how "missingness" is distributed by the outcome of interest. Our findings suggest these concerns may be relatively more important for studies involving veterans and racial-ethnic disparities, co-morbidity, certain disease comparisons, or settings of care. Additional study is needed to compare BIRLS, death certificates, and the NDI for mortality ascertainment in veterans. If our findings are confirmed, the VA may need to consider improving its system for mortality ascertainment through routine linkages to national mortality data. Studies of end-of-life care using decedent cohorts need to pay particular attention to the incompleteness of VA data as the sole source of mortality information.