5

U.S. Mortality Data: Data Quality, Methodology, and Recommendations

Monitoring of trends and disparities in U.S. working-age mortality requires accurate vital statistics data on deaths, including causes of death, as well as accurate estimates of the size of the population at risk of death, to calculate consistent and accurate mortality rates. This chapter reviews the sources used to generate these estimates, including the quality and limitations of this information and how these data were used to produce the analyses presented in this report. The chapter covers issues related to the methodologies used to collect death certificate data and link them to survey data, the advantages and limitations of these types of mortality data, and the analytical methodology used by the committee in conducting its analyses. The chapter also includes the committee’s recommendations for improving data quality to expand the capacity for future research on trends and disparities in U.S. working-age mortality.

THE U.S. NATIONAL VITAL STATISTICS SYSTEM (NVSS) AND THE CONSTRUCTION OF MORTALITY RATES

Death certificate records are an important component of the U.S. system of vital records. Within the United States, the responsibility for collecting death records is delegated to individual U.S. states and territories. They report this information to the federal government, which serves as the national repository of these records. This national repository of vital records, the NVSS, is maintained by the National Center for Health Statistics (NCHS) of the Centers for Disease Control and Prevention (CDC). NCHS death record files are considered 100 percent complete, although

there may be a small number of deaths for which registration is delayed, such as when an American dies outside of the United States or in the case of “missing persons” for whom the courts have not (yet) assigned certification of death. Vital statistics death certificate data include a limited amount of information about each decedent, including age at death, sex, race and ethnicity, educational attainment, place of residence, location of death, and cause(s) of death, as well as a few other items.

In addition to serving as the final repository for vital statistics records, NCHS assists the states in maintaining the best-quality vital records possible and provides resources and guidance for the structure and collection of vital records data, including recommendations and guidelines regarding the coding of such demographic information as race and ethnicity and educational attainment to improve the uniformity of coding across U.S. states. NCHS also cooperates with the World Health Organization (WHO), helping to improve comparability with international vital statistics data, particularly with respect to cause-of-death coding.

NCHS vital statistics data are released annually in data files that are structured to provide information about all deaths that occurred during a given year. Most studies of cause-specific mortality rely on the “underlying cause of death” coded in the files, which is defined as “the disease or injury which initiated the train of morbid events leading directly to death, or the circumstances of the accident or violence which produced the fatal injury” (World Health Organization [WHO], 2011, p. 31). In practice, the underlying cause is selected from the conditions entered by the medical certifier on the death certificate, thus making the training and qualifications of medical certifiers critical for accurate documentation of causes of death (Kochanek et al., 2019). When more than one cause or condition is entered, “the underlying cause is determined by the sequence of conditions on the certificate, provisions of ICD [International Classification of Diseases], and associated selection rules and modifications” (Kochanek et al., 2019, p. 62). Because most death certificates list more than one cause of death, more medical information is reported on death certificates than is directly reflected in the underlying cause of death. This additional information is available in the NCHS multiple cause-of-death data files.

Vital statistics data files do not include information about the size of the population that was at risk of death, which, as noted above, is necessary to calculate mortality rates. Therefore, this information must be drawn from a separate source—the U.S. Census. The U.S. Census Bureau conducts a census of the U.S. population every 10 years. During the intercensal period, the Census Bureau generates annual midyear population estimates for the country as a whole, as well as for demographic subgroups (e.g., age, sex, race and ethnicity), states, and local areas. Together, the demographic information included in the NCHS death certificate data files and the U.S.

Census Bureau’s population estimates are used to calculate mortality rates by age, sex, and race and Hispanic ethnicity, both for the nation as a whole and for smaller geographic areas. Thus, the calculation of subgroup-specific mortality rates requires subgroup-specific population estimates that are based on comparable definitions of these subgroups.

Because denominator data for the calculation of U.S. mortality rates come from decennial U.S. Census data or the U.S. Census Bureau’s population estimates for intercensal years, it is critically important that the decennial U.S. Census count be as accurate as possible. Similarly, the accuracy of the annual estimates of the size of the population by age, sex, and race and ethnicity, as well as geographic area, are vital for generating estimates of U.S. mortality patterns and trends.

As of this writing, the U.S. Census Bureau is weighing strategies for implementing a statistical practice known as “differential privacy” to protect the privacy of respondents to the decennial Census and other U.S. Census Bureau survey products by reducing the risk of disclosing information that, when combined with other publicly or privately available data sources, such as social media accounts, could allow respondents to be personally identified. In essence, the application of differential privacy would infuse statistical “noise” into the data, potentially affecting the accuracy of population counts for important subgroups of the population that are used to calculate mortality rates. The effects of introducing noise into population estimates are certain to be variable across subgroups and potentially could be large for small geographic areas and racial/ethnic groups with small populations. The magnitude of the impact of this change on subgroup population counts, and therefore its effect on the accuracy of mortality rates, will depend on how differential privacy is implemented by the U.S. Census Bureau. Evaluating the impact of these changes on mortality estimates will be of crucial importance for future mortality researchers.

Limitations and Quality of Mortality Data

Although they serve as a complete record of all U.S. deaths, death certificate data have important limitations that restrict the types of analysis and questions that researchers can address. The process by which death certificate data are collected can also result in issues of data quality and accuracy that affect the quality of mortality estimates and the comparability of these estimates over time. This section outlines the limitations of death certificate data and the steps that data providers have taken to address these issues so as to improve the utility of these data and expand the types of research questions they can be used to address. It then reviews known issues with the quality and accuracy of death record data that affect the quality of mortality estimates.

Limitations of Death Certificate Data and Use of Linked Mortality Data

The relatively limited information about the decedent on U.S. death certificates noted earlier (age, sex, race and ethnicity, educational attainment, place of residence, location of death, and cause(s) of death, as well as a few other items) is useful for examining disparities in mortality by these factors, but it limits the characteristics that can be examined. For example, educational attainment is only one dimension of socioeconomic status; other socioeconomic factors, such as income and wealth, may be important for understanding trends and disparities in working-age mortality, but this information is not available on death certificates. In addition to restricting the types of mortality disparities that can be examined, the relatively modest set of characteristics available on death certificates restricts researchers’ ability to examine the factors that might explain mortality trends and disparities.

To examine a wider range of characteristics that might be related to mortality, researchers tend to rely on death certificates that are statistically linked to other data sources, most notably large social and health surveys. In some countries, death records are routinely linked to population registries (e.g., census records) to provide additional demographic and contextual information for death data. Such linked datasets have become especially important because they provide the research community with rich survey data on individuals, who are then followed statistically across multiple years to document who lives and who dies. But linking U.S. survey data to death certificates poses its own difficulties and limitations. To link death records to individual-level surveys, one must be able to identify individuals within both the death record data and the survey data, putting such linkages beyond the reach of most researchers.

To address this difficulty, several nationally representative governmental surveys, such as the National Health Interview Survey and the National Health and Nutrition Examination Survey, merge death record information from the annual National Death Index (NDI) with the survey data so that users can track mortality among those who appear in the surveys.¹ Because these surveys collect detailed demographic, income, behavioral, and health data and are representative of the U.S. population, this information can be used to calculate mortality risks by individual-level characteristics. However, one limitation of these linked datasets is that deaths occur quite infrequently during the follow-up period unless a survey includes a very large sample size, has a high proportion of elderly respondents, and/or has a long follow-up period. This limitation can lead to imprecise estimates of

___________________

¹ See https://www.cdc.gov/nchs/data-linkage/mortality-public.htm.

mortality and hamper the ability to compare mortality risks across groups; states and local areas; or social, behavioral, and health characteristics.

Some surveys with large sample sizes (e.g., the National Health Interview Survey) or long follow-up periods (e.g., the Panel Study of Income Dynamics and the Health and Retirement Survey that covers individuals ages 50 and over) may include enough deaths to enable the calculation of stable mortality rates. While these surveys greatly improve upon the quality and quantity of socioeconomic, behavioral, and health information available for studying relationships between individual-level factors and mortality risk, they also suffer from their own limitations, many of which may lead to underestimation of mortality rates. First, because these surveys are generally limited to the noninstitutionalized U.S. population, the individuals they include are healthier than the U.S. population as a whole, and underestimation of mortality rates may result (Keyes et al., 2018). Second, deaths are assessed by linking death records to the survey data, often using algorithms based on a set of respondent characteristics. When an individual is not linked, that individual is usually assumed to be alive; because of imperfect linkages between the death record data and the survey data, however, not all deaths may be recorded for the survey respondents. This is particularly the case, for example, when an individual in one of these surveys moves out of the United States and dies in a different country. Third, such datasets tend to be exceptionally useful for nationally based mortality estimates but typically do not include enough deaths to enable estimation of state- or local-level mortality patterns and trends. Finally, the small sample size and number of linked deaths among those at the oldest ages often lead to improbably low mortality rates for the oldest old within these surveys, among whom the proportion institutionalized grows with age.

Each decennial Census contains demographic information for every U.S. resident, but linkages between decennial Censuses and death certificates are not routinely conducted by the U.S. Census Bureau. However, a subset of records from the 1980 U.S. Census are linked to death certificate data as part of the National Longitudinal Mortality Study (NLMS). The NLMS is an important example of a linked dataset that was created to enable the study of mortality. It uses data from the Current Population Survey Annual Social and Economic Supplements and the 1980 U.S. Census linked to death certificate information for many years after individuals were included in the survey or U.S. Census. The dataset contains information from 3.8 million individuals and more than 550,000 death certificates.² The large sample size and number of deaths, combined with detailed information on socioeconomic status (SES), make the NLMS an important resource for studying the relationship between SES and mortality. In addition to the

___________________

² See https://www.census.gov/did/www/nlms.

above linked data sources, the linkage of the 2008 American Community Survey (ACS) to the NDI provides another large data source for studying mortality by individual-level characteristics.³ However, another limitation of these linked datasets, other than the Panel Study of Income Dynamics and the Health and Retirement Study, is that survey information is collected at a single point in time; this point in time may precede death by a decade or more and therefore may not reflect the individual’s SES at the time of death.

Despite these limitations, the above linked mortality datasets provide a wealth of information that would otherwise not be available using vital statistics data alone. The ability to link the NDI data to existing survey datasets provides an invaluable resource for researchers, public health officials, and policy makers. Thus, the committee concluded that no reduction or changes in the content of the information collected for these datasets is warranted, and that the research community would be well served if the datasets were made as widely and easily available to researchers as possible. Currently, many survey datasets that are linked to death records do not make these linked data publicly accessible or provide only limited mortality information in public-use data files; more detailed information on cause of death and geography are available only through restricted-use data files. Making available the linking algorithm and weighting scheme used to produce linkages between the NDI and individual-level survey data would also help researchers determine any potential biases in the linkages and enable them to better assess how such biases might influence their estimates.

Quality and Accuracy of Death Certificate Data

Several well-known data collection and coding issues affect the quality and accuracy of the data on U.S. death certificates, related to the coding of cause-of-death information by medical examiners and the assignment of other demographic information. While the accuracy and quality of these records is constantly evolving, there have also been ongoing efforts to progressively improve their accuracy and utility for public health purposes.

A number of factors can influence the quality and accuracy of the cause-of-death information on death certificates, such as the era (period) in which the death occurred, changes in the ICD, place of death, available local resources, training of the medical certifiers, and the complexity of the chain of medical diagnoses that led to death. The era in which the death record was generated matters for several reasons, but in particular because advances in diagnostic and forensic technology can change how causes of death are identified and coded.

The United States transitioned from the ICD-9 to the ICD-10 coding system for cause-of-death data in 1999. This was the first major change in

___________________

³ See https://www.census.gov/mdac.

the cause-of-death coding system in the United States since the implementation of the ICD-9 system in 1979. One of the features of the ICD-10 system is a standardized schedule for introducing updates to the codes to ensure that the system is flexible and remains consistent with current medical practice and knowledge. The advent of new versions or modifications of the current ICD and related taxonomic rubrics, such as environmental causes of death, can spur training in new coding systems and improve the accuracy of data coding as knowledge evolves, but can also lead to inconsistencies with coding from previous eras that impede comparisons over time. WHO will introduce ICD-11 in 2022, and the National Committee on Vital and Health Statistics (2019) advised the U.S. Department of Health and Human Services to take a proactive approach to preparing for its release.⁴

The quality of death record reporting is also affected by the training and resources available to local medical certifiers. The availability of modern local resources and quality control programs in a region or jurisdiction affects the provision of the services of medical examiners and coroners with appropriate forensic, clinical pathology, and toxicology expertise. Consistent and accurate coding of causes of death can be difficult given the frequent complexity and uncertainty of clinical diagnoses, and a lack of training can lead to inconsistent coding of more complex causes of death. In general, diseases that are clinically easier to diagnose and/or have longer clinical courses are the most likely causes of death to be identified accurately. In contrast, diseases with short and more diverse clinical presentations and courses are more likely to be misclassified (Mieno et al., 2016). NCHS has developed an automated system to standardize and improve the coding of underlying cause of death. This system, the Automated Classification of Medical Entities (ACME), reads in the multiple cause-of-death data reported on the death certificate and applies decision rules developed by WHO to assign the underlying cause of death.⁵ Misclassification may also depend on the specific level of the condition at hand. For example, there is evidence that liver disease in general is underrepresented in death records (Durante et al., 2008), whereas primary liver cancer may be overrepresented (Polednak, 2013) because of misclassification of metastatic disease. In addition, chronic conditions are often missing from death certificates or assigned the status of a contributing cause rather than the underlying cause of death, even though they are relatively easy to diagnose (Gao et al., 2018). The training of certifying professionals is an important factor in the accuracy and completeness of these cause-of-death reports. Because the quality and training of medical certifiers are of paramount importance for

___________________

⁴ See https://ncvhs.hhs.gov/wp-content/uploads/2019/12/Recommendation-Letter-Preparingfor-Adoption-of-ICD-11-as-a-Mandated-US-Health-Data-Standard-final.pdf.

⁵ See https://www.cdc.gov/nchs/nvss/mmds/about_mmds.htm.

high-quality cause-of-death data, states and local health agencies need to ensure that certifiers are well trained in cause-of-death recording, including the coding of both underlying and multiple causes of death.

Differences in resources and training across jurisdictions can lead to local, state, and regional variations in standard practices for cause-of-death coding (Cheng et al., 2012; Cheng, Lu, and Kawachi, 2012). Although NCHS produces decision tables in the ACME to improve the consistency of cause-of-death reporting and identification of the underlying cause of death in death certificate data, these decision tables rely on not only the conditions reported as cause of death but also the causal sequencing of these reports. Reporting errors that can lead to discrepancies in the identification of the underlying cause of death remain common (Lu, Anderson, and Kawachi, 2010) and show considerable variation by cause of death (Falci et al., 2018), particularly when multiple comorbidities are present and may contribute to death (Lu, Anderson, and Kawachi, 2010). The level of reporting errors varies substantially across states, which affects the comparability of cause-specific mortality rates (Cheng, Lu, and Kawachi, 2012).

Data quality issues are a particular concern in examining deaths from acute poisoning and drug overdose. The term “drug overdose” is often used synonymously with acute poisoning, but it has at least three different coding definitions (Slavova, Bunn, and Talbert, 2014) and is often difficult to define toxicologically and pharmacologically. Moreover, many decedents are found to have multiple substances in their blood or other issues that create uncertainty in identifying the specific cause of death (Ruhm, 2018a, p. 1339). In some cases, for example, drugs associated with the treatment of drug addiction, such as methadone or buprenorphine, are present alongside other substances, further complicating the assignment of a specific cause of death and raising the question of whether these treatment medications should be assigned a dedicated cause-of-death code (Darke et al., 2019).

U.S. states vary substantially with respect to laws regarding the types of deaths that require a formal autopsy, as well as the resources provided to coroners and medical examiners for conducting the autopsies and biochemical determinations necessary to record cause of death accurately. Taken together, these variations lead in turn to local, regional, and national variation in the quality and accuracy of cause-of-death recording, making trend analyses and geographic comparisons of causes of death challenging. In light of the state-by-state variation and increasing geographic inequalities in U.S. mortality, studies to evaluate state-level variation in coding practices are warranted (Dwyer-Lindgren et al., 2016).

Given uncertainties in the coding of underlying causes of death, useful and detailed information may be contained in the contributing causes of death, depending on the goal of the analysis (Remund, Camarda, and Riffe, 2018). Yet most analyses, including those presented in this report and the

standard reports on mortality published by NCHS, document mortality patterns and trends according to underlying cause of death while ignoring associated causes. This is the case largely because of the complexity involved in coding and use of the full range of causes on death certificates. However, methods that account for multiple causes of death can reduce the impact of coding errors in the underlying cause of death that arise from misreporting of the causal ordering of conditions on the death certificate. In light of developments in data analysis (e.g., machine learning, weighting of multiple causes of death) and computing power, analysts could consider ways in which the full range of causes available on the death certificate can be used to document the complexity of causes from which U.S. adults are dying (Dwyer-Lindgren et al., 2016; Eberstein, Nam, and Heyman, 2008; Piffaretti et al., 2016).

Accurate coding of demographic information included on death certificates depends on a different set of factors. On most surveys, including the U.S. Census, demographic information is either self-reported or reported by a knowledgeable proxy. However, decedents obviously are not able to self-report demographic information. Therefore, this information is often provided by surviving relatives or friends or can sometimes be drawn from other sources, such as medical or official records. In some cases, however, the information may be left to the medical examiner or coroner to report based on physical examination or ancillary sources, and these reports may not be consistent with how this information would have been reported by the decedent, particularly when the reports refer to such characteristics as educational attainment, race, and ethnicity. This issue is compounded when death counts based on death certificate data are divided by population estimates from the U.S. Census Bureau to calculate mortality rates. The most common U.S. Census-based sources for population estimates provide data on such characteristics as race, ethnicity, and educational attainment that are collected, reported, and coded differently from similar data appearing on death records. These inconsistencies can compound the effects of recording errors in the death certificate data, leading to biased estimates of mortality and mortality disparities.

The documentation of race and ethnicity on U.S. death certificates, while improving, continues to be far less than 100 percent accurate. This is especially the case for American Indian and Alaska Native (AI/AN) populations, among whom official mortality rates have been deemed far too low (Arias, Heron, and Hakes, 2016). The committee found that death certificates continue to misclassify AI/AN individuals to such an extent that official mortality estimates for these populations are not valid. This in turn limits the ability of public health officials to track and highlight mortality within these populations, despite indications from other data sources that they experience much higher mortality relative to most other racial/ethnic

groups, with the possible exception of Blacks (Espey et al., 2014; Hummer and Gutin, 2018). Together, researchers from NCHS, a state vital statistics agency, and the Tribal Epidemiology Center have developed a set of promising ideas for improving the estimation of AI/AN mortality (Anderson, Copeland, and Hayes, 2014); below, the committee recommends further work in this area to make such improvements.

At the same time, according to the 2016 update to an earlier NCHS evaluation of the quality of race and ethnicity data on death certificates, the classification of race and ethnicity for Asians and Pacific Islanders and Hispanics had improved relative to earlier reports, and the accuracy of reporting for these populations was almost as good as that for Whites and Blacks, which previously was of higher quality. Although improvements over time in the quality of death certificate reports for Asians limited the committee’s ability to interpret their mortality trends over the period covered by this report, reporting of Asian race on death certificates has improved substantially as well and is now of sufficient quality that researchers, policy makers, and public health officials can have confidence in mortality estimates for this population. Given that underascertainment of mortality among Asians is now estimated to be about 3 percent, similar to that for Hispanics, it is time for Asians to be included in regular NCHS reports on life expectancy. Doing so is particularly important given the rapidly increasing size of the U.S. Asian population.

An additional complication in the coding of race and ethnicity on death certificates is the change in racial/ethnic reporting categories over time. Prior to changes recommended by the Office of Management and Budget in 2003, death certificates recorded only a single race and included less-detailed race and ethnicity categories. Subsequent changes in reporting were adopted inconsistently across states, potentially leading to difficulties in creating comparable race and ethnicity groups across time. Moreover, these measurement differences can reduce the comparability of race and ethnicity reports on death records and in population data, in turn reducing the accuracy of mortality rates. NCHS has developed bridged-race estimates that use an empirically derived algorithm to reassign multiracial individuals to one of the single-race categories in use prior to 2003; however, NCHS also used these bridged-race estimates to evaluate the quality of race reporting on death certificates. To the committee’s knowledge, no research to date has evaluated the concordance between multiracial race reports on death certificates and self-reports of multiracial identity in survey data. Racial/ethnic groups that make up a smaller percentage of the population and have a higher percentage of members who identify as multiracial, such as AI/ANs and Native Hawaiians and Pacific Islanders, are more likely to be misidentified on death certificates. Finally, despite the expanded race and ethnicity options introduced in 2003, patterns and trends in working-age

mortality of “other” racial/ethnic groups, such as persons who (wish to) identify as Middle Eastern, are unknown because neither death certificate numerator data nor U.S. Census-based denominator data include options for reporting such identities.

Similar reporting issues affect the quality of educational attainment reports on death certificates. Vital statistics and census data are the only large-enough data sources that allow for detailed examination of educational disparities in mortality by specific place of residence. However, the documentation of educational attainment on U.S. death certificates is often inaccurate (Rostron, Boies, and Arias, 2010) because this information is reported by proxy sources. NCHS reported that when educational attainment on death records was compared with corresponding information from the Current Population Survey, substantial differences were found between the two sources (Rostron, Boies, and Arias, 2010). For example, when educational attainment data on death certificates are inaccurate, high school completion tends to be overreported, leading to overstatement of deaths among those with a high school degree and understatement of deaths among those with less than a high school degree (Sorlie and Johnson, 1996).

Often the denominator information for the size of educational attainment groups is drawn from population data, such as the decennial U.S. Census, or federal data, such as population projections or the ACS. These sources are more likely than death records to contain self-reported data or data reported by a knowledgeable respondent. In calculating mortality rates, data on educational attainment from death records are combined with population estimates from the decennial U.S. Census to calculate mortality rates. When census population data are combined with death counts that underreport the number of deaths among those with less education because of misclassification, lower mortality rates among this group result. Similarly, when deaths for those with higher levels of education are overcounted and the resulting numbers are combined with accurate population size estimates, the resulting mortality rates overstate mortality in this population. This inaccuracy affects not only the estimates of mortality within each educational attainment group but also estimates of the disparities among those groups.

These accuracy issues raise important concerns about the validity of using death records to study educational disparities in mortality. States and local health agencies may need to initiate additional training and guidance for those who provide this information on death certificates. Given not only the documented inaccuracy of educational attainment data on death certificates but also possible state-by-state and local-area variation in the recording of this information, researchers would be well advised to use studies based on self-reported survey data linked to death certificate

mortality data to assess the accuracy of data on educational attainment and their geographic variation on death certificates.

Place of birth is also recorded on U.S. death certificates, but data on place of birth are not made available to researchers in public-use vital statistics mortality files. This is an important concern given the growth of the foreign-born population in recent decades, along with its heterogeneity. Moreover, variation in mortality by nativity has become critical for understanding U.S. mortality trends and differentials (Dupre, Gu, and Vaupel, 2012; Lariscy, Hummer, and Hayward, 2015; Lauderdale and Kestenbaum, 2002; Turra and Elo, 2008). In addition, the size of the foreign-born population in the U.S. Census denominator data is subject to underestimation, especially if large segments of the foreign-born population are missed because of their documentation status. Given the current lack of information on nativity (i.e., foreign-born vs. U.S.-born) in publicly available vital statistics data, researchers need to tackle both racial/ethnic and nativity disparities in U.S. working-age mortality using survey-based data linked to the NDI or Social Security Administration data on older Americans, or make a special request for restricted death record files.

ESTIMATION OF MORTALITY RATES IN THIS REPORT

The data used to produce the mortality rates presented in this report were drawn from death certificate data provided by the CDC in the NVSS restricted death certificate files for 1990–2017 (National Center for Health Statistics [NCHS], 2018). These files include information about the decedent’s date and underlying cause of death; place of residence; and a small number of demographic characteristics, including sex, age, and race and ethnicity. In 2003, during the period covered by the committee’s analyses, NVSS adopted new recommendations for the coding of racial/ethnic data on death certificates that allowed the reporting of more than one race; however, the timing of the adoption of this change varied across states. For this reason, the population counts by age, sex, and race and ethnicity that the committee combined with these mortality data to calculate mortality rates were based on the U.S. Census bridged-race estimates of the U.S. resident population on July 1 of each year, produced by the U.S. Census Bureau (U.S. Department of Health and Human Services [HHS], 2018). These bridged-race estimates reclassify multiracial individuals into one of the single-race categories that were in use prior to 2003. State-level mortality rates were drawn from the CDC WONDER database (Centers for Disease Control and Prevention [CDC], 2020b).

The analyses presented herein are based on trends for multiyear age groups. Previous researchers have noted that because mortality rates increase with age in adulthood, differences in the age composition of two

populations can affect the comparability of their mortality rates when broad age categories are used (Gelman and Auerbach, 2016). To ensure comparability over time and across subpopulations, rates were age-adjusted by single year of age and standardized to reflect the age distribution of the U.S. population in 2000, unless otherwise noted in the text. Throughout, mortality rates are presented as the number of deaths per 100,000 population. Deaths were pooled across 3-year periods from 1990–2017, with the exception of the first period (1990–1993), which includes 4 years. The data were pooled across years to smooth fluctuations in mortality trends that sometimes occur for smaller populations with relatively low death counts. Many of the analyses presented in this report rely heavily on mortality rates from four periods corresponding roughly to the beginning and end of the 1990s, the 2000s, and the 2010s: 1990–1993, 2000–2002, 2009–2011, and 2015–2017.

This report’s summary of research on mortality trends and differentials by educational attainment draws on previously published studies that use both vital statistics data and survey-based data linked with the NDI. Given the issues of the quality of educational attainment data raised above, especially in vital statistics mortality data, the committee used only the highest-quality studies in this area in which such data quality issues are best taken into account. Provided below are the committee’s recommendations for studying and improving data quality in this important area of mortality study.

In preliminary analysis, the committee also examined mortality patterns and trends for Asians and Pacific Islanders and for AI/ANs. Because of concerns about data quality—specifically, that the errors in racial classification on death certificates for these populations changed over time as discussed above (Arias, Heron, and Hakes, 2016)—those trends are not presented in this report. To the extent that such information is available, summaries of the existing literature on mortality trends for these groups are provided to ensure that their experiences are represented. Within this literature, however, researchers may have used different classifications for cause of death from those used for the original analyses presented in this chapter, and therefore these analyses may not be directly comparable.

Metropolitan status was based on a modified version of the geographic identifiers developed by the Economic Research Service (ERS) of the U.S. Department of Agriculture and NCHS.⁶ The committee classified U.S. counties into four types of metropolitan areas: large central metropolitan areas (counties in metropolitan statistical areas [MSAs] of more than 1 million population, including counties that contain all or part of the area’s inner cities), large fringe metropolitan areas (surrounding counties of the

___________________

⁶ See https://www.cdc.gov/nchs/data_access/urban_rural.htm.

large central metros, including suburbs), medium/small metropolitan areas (counties in MSAs of 50,000–999,999 population), and nonmetropolitan areas (counties outside of metropolitan areas). To maintain consistency over time, the counties’ metropolitan categories were assigned based on the 2013 ERS classification scheme (sensitivity analyses showed only minor differences using earlier classification schemes).

For the analyses in this report, cause of death was assigned to one of 20 broad, nonoverlapping categories based on the underlying cause of death identified on the decedent’s death certificate (Table 5-1). The committee identified these categories after reviewing trends across more detailed categorizations. After that detailed review, causes with low death counts and similar temporal trends were collapsed into larger categories. Causes of death were determined based on ICD-9 for 1990–1998 and ICD-10 for 1999–2017. This change in coding systems could have led to discontinuities in cause-specific mortality trends at the points at which the new ICD-10 coding system was adopted even within categories whose definitions had not changed. The committee was cognizant of this possibility and examined cause-specific mortality trends that covered the full 1990–2017 period in order to assess evidence of a discontinuity in the cause-specific trends. These results are presented in Appendix A.

Because of a small change in ICD codes used over the study period, readers should exercise caution when interpreting changes in two of the committee’s broad cause-of-death categories: alcohol-induced diseases and diseases of the digestive system. ICD-10 code K85 (acute pancreatitis) was discontinued in 2006 and replaced with several K subcodes. One of those subcodes (K85.2, alcohol-induced acute pancreatitis) is included in the committee’s “alcohol-induced” category for 2006–2017 but could not be broken out prior to 2006. From 2006 to 2017 (the years during which K85.2 was used), there were only 3,279 deaths in that category among all working-age (ages 25–64) adults (both sexes, all racial/ethnic groups). Any bias this coding change may have introduced into the committee’s temporal comparison would be observed in a jump in alcohol-related deaths and a decline in diseases of the digestive system between 2005 and 2006.

Chapter 4 presents trends over time in the percentages of deaths due to mental and behavioral disorders that were drug- or alcohol-related. The assignment of these deaths to these two categories was based on the ICD-9 and ICD-10 codes listed in Table 5-2.

RECOMMENDATIONS

Based on the issues discussed above, the committee offers the following recommendations for improving data quality and availability to support analyses of patterns and trends in U.S. working-age mortality.

TABLE 5-1 Assignment to 20 Cause-of-Death Categories

SOURCES: Data from https://wonder.cdc.gov/controller/datarequest/D16 (ICD-9) and https://wonder.cdc.gov/controller/datarequest/D76 (ICD-10).

TABLE 5-2 ICD-9 and ICD-10 Codes for Drug- and Alcohol-Related Deaths due to Mental and Behavioral Disorders

SOURCES: Data from https://wonder.cdc.gov/controller/datarequest/D16 (ICD-9) and https://wonder.cdc.gov/controller/datarequest/D76 (ICD-10).

RECOMMENDATION 5-1: The National Center for Health Statistics (NCHS), state vital statistics offices, and local-area health agencies should work together to develop a plan and set of activities for improving the accuracy of reporting on U.S. death certificates of educational attainment, American Indian and Alaska Native identity, and multiple causes of death. NCHS should also continue to conduct or facilitate studies on the accuracy of reporting on U.S. death certificates of educational attainment (particularly as such reports may vary across states and local areas) and American Indian and Alaska Native identity (particularly as such reports may vary across states, tribal affiliations, and local areas).

RECOMMENDATION 5-2: The National Center for Health Statistics and the National Institutes of Health should undertake and/or fund studies to evaluate state- and local-level variation in cause-of-death coding practices, explore how such variation may contribute to observed mortality trends, and make recommendations for reducing such variation.

RECOMMENDATION 5-3: The National Center for Health Statistics should include Asians in its regular reports on life expectancy estimates and trends in the United States and make an item on place of birth available to researchers in the public-use files, even if such information is at first categorical (e.g., foreign-born vs. U.S.-born) rather than granular.

RECOMMENDATION 5-4: To enable robust research on rural–urban trends in health and mortality, the National Institutes of Health and other research agencies and funders should support the oversampling of rural populations on national health and social surveys, including both existing (e.g., Health and Retirement Study, Behavioral Risk Factor Surveillance System, National Longitudinal Study of Adolescent to Adult Health [Add Health], National Survey on Drug Use and Health, National Health Interview Survey, National Health and Nutrition Examination Survey) and new surveys.